MorphoLibJ-manual-v1 6 2

MORPHOLIBJ USER MANUAL
David Legland, Ignacio Arganda-Carreras
October 16, 2023

Contents
1. Overview 5
2. Installation instructions 6
2.1. Installation in ImageJ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2. Installation in Fiji . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3. Definitions and data types 9

3.1. Digital images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2. Coordinate system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3. Digital connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4. Morphological filtering 13
4.1. Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2. Grayscale morphological filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3. Directional filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4. Plugin Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5. Connected components operators 21

5.1. Morphological reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.2. Regional and extended extrema . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3. Attribute filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6. Watershed segmentation 30
6.1. Classic Watershed plugin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.2. Marker-controlled Watershed plugin . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.3. Interactive Marker-controlled Watershed plugin . . . . . . . . . . . . . . . . . . 35
6.4. Morphological Segmentation plugin . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.5. Distance Transform Watershed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7. Measurements 47
7.1. Region Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.2. Region Analysis 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.3. Intensity measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.4. Label Overlap measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.5. Microstructure analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
8. Binary images 69
8.1. Distance transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2
Contents
8.2. Connected component labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

8.3. Binary images processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
9. Label images 79
9.1. Visualization of label images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
9.2. Visualization of region features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
9.3. Label image processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
9.4. Region adjacencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
9.5. Label Edition plugin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
10.Library interoperability 90
10.1.Library organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
10.2.Scripting MorphoLibJ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Bibliography 93
A. List of operators 97
B. Computation of equivalent ellipsoid coefficients 99

B.1. Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
B.2. Computation of angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
B.3. Computation of ellipsoid radius . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
MorphoLibJ user manual page 3 / 104

Abstract The «MorphoLibJ» library for ImageJ/Fiji contains a variety of plugins for
processing and analyzing 2D or 3D images. It was originally designed for the study
of plant cell images, but the algorithms are generic and can be applied to any type of
image.
1. Overview
The library implements several functionalities that were missing in the ImageJ software
(Schneider et al. (2012); Schindelin et al. (2012)), and that were not or only partially cov-
ered by other plugins.
• morphological filtering for 2D/3D and binary or grayscale images: erosion & dilation,
closing & opening, morphological gradient & Laplacian, top-hat...
• morphological reconstruction, for 2D/3D and binary or grayscale images, allowing

fast detection of regional or extended extrema, removing of borders, or hole filling
• watershed segmentation + GUI, making it possible to segment 2D/3D images of cell

tissues
• 2D/3D measurements: volume, surface area, bounding boxes, equivalent ellipse/el-

lipsoid...
• several utilities for the manipulation of binary and label images
The home page of the project is located at http://ijpb.github.io/MorphoLibJ/, and the source
code can be found on GitHub: http://github.com/ijpb/MorphoLibJ. The exhaustive code
documentation includes many use-case examples, and can be found online1 .
The reference publication for MorphoLibJ is Legland et al. (2016).
1
http://ijpb.github.io/MorphoLibJ/javadoc/
5
2. Installation instructions
2.1. Installation in ImageJ
To install the MorphoLibJ toolkit in ImageJ, you only need to download the latest release (in
the form of a JAR file) into ImageJ’s plugins folder and restart ImageJ.
All released versions can be found at https://github.com/ijpb/MorphoLibJ/releases.
2.2. Installation in Fiji

If you use Fiji (Schindelin et al. (2012); Schneider et al. (2012)), the MorphoLibJ toolkit can
be easily installed by adding its update site to Fiji’s list of update sites:
1. Open Fiji and select Help > Update... from the Fiji menu to start the updater.
6
2.2 Installation in Fiji
2. When the updater is open, click on Manage update sites.
3. This brings up a dialog where you can activate additional update sites:

2.2 Installation in Fiji
4. Activate the MorphoLibJ update site and close the dialog. Now you should see an
additional jar file for download:
5. Finally, click on Apply changes and restart Fiji. All the plugins should be available after
restarting.

3. Definitions and data types
This chapter recalls some principles from the ImageJ software, summarizes the vocabulary
and definitions used within the library, and presents some generic concepts about image
processing and associated data structures.
Contents
3.1. Digital images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.1. Data types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.2. Image types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.3. 3D images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2. Coordinate system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3. Digital connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3.1. 2D connectivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3.2. 3D connectivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
9
3.1 Digital images
3.1. Digital images

Digital images may be defined by a mapping from a pair (i, j) of integer coordinates to a
value. The value may be an integer, a floating point number, a color... Image elements are
usually refered to as “pixels”, for “picture element”.
3.1.1. Data types

Images are usually represented as 2D arrays of elementary data types. Within ImageJ, the
pixels may be of different types, according to the bit depth:
• 8-bits images contains pixels with (positive) integer values between 0 and 255.
• 16-bits images contains pixels with (positive) integer values between 0 and 216 =
65536.
• 24-bits images corresponds to color images. Pixels are represented as triplet of red,
green and blue components, each of them ranging between 0 and 255.
• 32-bits images contains pixels with floating-points values, possibly negative.
Note that other software may use additionnal data type for representation of images, such
as 32-bits integers, complex values, booleans...
3.1.2. Image types

Apart from the data type / bit depth, image data may be interpreted in different ways. The
MorphoLibJ library focusses on the following image types:
grayscale images contains pixels with positive integer values. They can correspond to
either 8-bits or 16-bits images.
binary images are expected to contain only two possible values. Within ImageJ, they are
represented as 8-bits images containing only two values: 0 for the background, and
255 for the foreground (See chapter 8).
intensity images may correspond to any types that can be interpreted as a single scalar
value. They can be represented with 8-bits, 16-bits, or 32-bits images.
label images (also called label maps) contains values that can be interpreted as the label
or index of a region (See chapter 9). The label 0 usually refer to the background (i.e.,
no region). Label images can be represented using 8-bits, 16-bits, or 32-bits images.
Binary images can be seen as a special case of label images that contain only one label.
3.1.3. 3D images
Most algorithms from mathematical morphology (and more generally, many image process-
ing algorithms) extends naturally from two dimensions to three dimensions. Within ImageJ,
3D images are represented as “image stacks”, seen as stacks of 2D arrays. Elements of a 3D
image are usually refered to as “voxels”, for volumetric picture element.

3.2 Coordinate system
3.2. Coordinate system

Images usually refer to data obtained on some physical sample. It is therefore necessary to
be able to relate the (i, j) coordinates of a pixel to its physical location. Image pixel values
may be seen as sampled on a discrete grid, defined by an origin (the coordinates of the (0, 0)
pixel) and a spacing (the width and the height of the pixel).
Several software consider the pixel centers to be located on the grid vertices, i.e. pixels
have integer coordinates. Within ImageJ, the origin pixel is located at the point (0.5, 0.5),
i.e. it occupies the space of the square with corners (0, 0) and (1, 1).
This notion of coordinate system is necessary to interpret the results obtained with region
analysis algorithms (see section 7.1).
3.3. Digital connectivity

When reconstructing structures or regions from digital images, it is often necessary to define
an adjacency relationship, or connectivity, between pixels of voxels (Rosenfeld, 1970;
Serra, 1982; Kong & Rosenfeld, 1989; Soille, 2003). The notion of connectivity determines
the result of many algorithms such as connected component labeling (see Section 8.2), or
the result of image quantification based on Euler number (see Section 7.1.1.3).
3.3.1. 2D connectivities
For planar images, typical choices are the 4- and the 8-connectivities. The 4-connectivity
considers only the orthogonal neighbors of a given pixel (Figure 3.1-a). The 8-connectivity
also considers the diagonal pixels (Figure 3.1-b).
(a) 4-connectivity (b) 8-connectivity
Figure 3.1.: Digital connectivities for 2D images
The discrete nature of images results in potential problems when considering geometric
properties on reconstructed structures. A typical example is the Jordan curve theorem, which
states than any curve of the plane that do not self-intersects divides the plane into exactly
two regions: one interior and one exterior (Figure 3.2-a).
When choosing a digital connectivity and considering reconstructed curves, the theorem
is not always true, as a 4-connected curve may create several disconnected interior regions

3.3 Digital connectivity
(a) Continuous curve (b) Digital curve using the 4-

adjacency
Figure 3.2.: Jordan curve theorem in continuous plane and in digital plane. The digitization
of a 4-connected curve results in the creation of two inner regions that are not 4-connected.
(Figure 3.2-b). The common solution is to consider pairs of adjacencies, one for the fore-
ground (here the curve), and one for the background. One can check that the two interior
regions are 8-connected, making the Jordan curve theorem valid for the (4,8)-adjacency pair.
3.3.2. 3D connectivities
In 3D, two connectivities are commonly used. The 6-connectivity considers only orthogonal
neighbors of the center voxel along the three principal directions (Figure 3.3-a). The 26-
connectivity considers all the possible neighbors within a 3 × 3 × 3 cube centered around the
reference voxel (Figure 3.3-b). As for the 2D case, one often considers pairs of complemen-
tary adjacencies for the foreground and for the background (Kong & Rosenfeld, 1989; Ohser
& Schladitz, 2009).
Figure 3.3.: Digital connectivities for 3D images

4. Morphological filtering
Morphological filters are very common filters that can be combined together to provide large
variety of solutions. They are local filters, in the sense they consider the neighborhood of
each pixel/voxel.
Contents
4.1. Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2. Grayscale morphological filters . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2.1. Grayscale erosion and dilation . . . . . . . . . . . . . . . . . . . . . 14
4.2.2. Opening and closing . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2.3. Morphological gradients . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2.4. Top-hats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3. Directional filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4. Plugin Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.4.1. Planar images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.4.2. 3D images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4.3. Directional Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
13
4.1 Principles
4.1. Principles
The original idea was to define a methodology to describe shapes by using another shape
as test probe (Serra, 1982; Serra & Vincent, 1992). The test probe is usually referred to
as structuring element. Common structuring element include squares, discrete disks and
octagons. Linear structuring element of various orientations may also be used to assess local
orientation of the structures.
The most basic morphological operators are the morphological dilation and the mor-
phological erosion. The principle of morphological dilation is to test for each point of the
plane, if the structuring element centred on this point intersects the structure of interest
(Fig. 4.1). It results in a set larger than the original set.
The principle of morphological erosion is to test for each point of the plane if the struc-
turing element centred on this point is contained within the original set. It results in a set
smaller than original set.
(a) Binary set X and structuring el- (b) Dilation of X by B (c) Erosion of X by B
ement B
Figure 4.1.: Principle of morphological dilation and erosion on a binary set, using a disk-shaped
structuring element.
Morphological dilation and erosion change the size and the resulting set. It may also
change its topology: after a dilation, components may merge and holes be filled. After an
erosion, components may disappear, or components be separated into several parts.
4.2. Grayscale morphological filters

While originally defined for binary sets, morphological operators are of great interest when
applied to grayscale images. They can be easily be applied for noise reduction (section 4.2.2),
detection of boundary structures (section 4.2.3), or background removal (section 4.2.4).
4.2.1. Grayscale erosion and dilation

Morphological erosion and dilation may also be applied on grayscale images. In that case,
the morphological dilation computes for each pixel the maximum within its neighborhood

4.2 Grayscale morphological filters
(defined by the structuring element), whereas the morphological erosion considers the min-
imum value within the neighborhood (Fig 4.2).
(a) Original image (b) Morphological dilation (c) Morphological erosion
Figure 4.2.: Some examples of morphological filters on a grayscale image. Original image, re-
sult of dilation with a square structuring element, and result of erosion with the same structuring
element.
Applying a dilation or an erosion changes the size of the structures in the image: the grains
in the result of the dilated image are larger. As for binary sets, these operations may also
merge, separate, or make disappear some components of the image.
4.2.2. Opening and closing

Morphological dilation and erosion are often used in combination for removing noise within
images. For example, the result of a dilation followed by an erosion is called a morphologi-
cal closing, and removes dark structures smaller than the structuring element (Figure 4.3-b).
It can also connect bright structures that were separated by a thin dark space.
(a) Original image (b) Morphological closing (c) Morphological opening
Figure 4.3.: Grayscale morphological closing and opening. (a) Original image. (b) Result of
morphological closing. (c) Result of morphological opening.

In a symmetric way, the result of an erosion followed by a dilation is called a morpho-

logical opening, and removes bright structures smaller than the structuring element (Fig-
ure 4.3-c).
Note that even if opening and closing preserve the size of the structures in the origi-
nal image, their shape is slightly altered. For example, the morphological closing on Fig-
ure 4.3-b creates artificial connections between grains, whereas the morphological opening
(Figure 4.3-c) tends to remove sharp corners of the grains. Choosing the best size for the
structuring element is often a compromise between noise removal and preservation of shape
structure. In some cases, attribute filtering operators (section 5.3) may provide better results
than morphological filters.
Morphological opening and closing are also often used as post-processing operators to
enhance segmentation results (Figure 4.4). Segmentation operators such as thresholding
often results in many small components not corresponding to the structures of interest (Fig-
ure 4.4-b). Many holes may also be present within the structures.
(a) Original image (b) Raw segmentation (c) Morphological closing (d) Morphological opening
Figure 4.4.: Enhancing segmentation with morphological filters. (a) Original image represent-
ing a section of maize tissue. (b) Binarisation by thresholding (inverted LUT). (c) Morphological
closing removes most holes within the bundles (inverted LUT). (d) Morphological opening re-
moves small dirts outside of vascular bundles (inverted LUT).
The application of a morphological closing on the binary images allows to consolidate the
structures of interest by removing small holes within them (Figure 4.4-c). Then, a morpho-
logical opening removes the small particles that do not correspond to the vascular bundles
(Figure 4.4-d). The result is a binary image showing only the vascular bundles.
4.2.3. Morphological gradients

More complicated combinations of elementary operations can be used. The morphological
gradient (also known as the Beucher gradient) is obtained as the difference of the result
of a morphological dilation with the result of a morphological erosion, both computed with
the same structuring element. It reveals the boundaries of the structures within the image
(Figure 4.5-a). When applied to a binary image, the morphological gradient is a convenient
way to obtain the boundaries of the structures of interest within the image (Figure 4.5-b).
The morphological Laplacian is defined as half the sum of a morphological dilation and
a morphological erosion with the same structuring element, minus the original image. It
results in enhancing the edges of the image (Figure 4.5-c).

(a) Morphological gradient (b) Morphological gradient on a (c) Morphological Laplacian

binary image
Figure 4.5.: Some examples of morphological filters resulting from the combination of elemen-
tary morphological filters. (a) Morphological gradient on a grayscale image. (b) Morphologi-
cal gradient on the binary image represented in Figure 4.4-d (inverted LUT). (c) Morphological
Laplacian on a grayscale image.
4.2.4. Top-hats
The white top-hat operator first computes a morphological opening (resulting in removing
bright structures smaller than structuring elements), and removes the result from the original
image. When applied with a large structuring element, the result is an homogenization of
the background, making bright structures easier to segment (Figure 4.6-c).
(a) Original image (b) White Top-Hat (c) Profile plots along
vertical line
Figure 4.6.: Top-hat filtering. (a) Original grayscale image, showing inhomogeneous back-
ground. (b) Result of the White Top-Hat operator using square with radius 20 as structuring
element. (c) Profile plots along line region of interest for each image.
Similarly, the dark top-hat can be used to enhance dark structures observed on an non-
homogeneous background.

4.3 Directional filters
4.3. Directional filters

For images containing very thin curvilinear structures (for example blood vessels, cell wall
sections...), the application of common filters may be difficult due to the small size of the
structures. Even for small structuring elements, the application of a morphological opening
or closing let the structure disappear (Fig. 4.7-b, 4.7-c). Moreover, it may be difficult to
preserve the whole thickness of the structure.
(a) Original image (b) Gaussian filtering (c) Median filtering (d) Directional filtering
Figure 4.7.: Filtering of a thin structure. (a) Original image representing apple cells observed
with confocal microscopy (Legland & Devaux, 2009). The application of a Gaussian filter (b)
or median filter (c) results in noise reduction, but also in a loss of the signal along the cell walls.
The directional filtering (d) better preserves the thickness of the structure.
An alternative is to apply directional filtering. The principle is to consider an oriented

structuring element such as a line segment of a given length, and to perform morpholog-
ical operations for various orientations of the structuring element (Soille & Talbot, 2001;
Heneghan et al., 2002; Luengo Hendriks & van Vliet, 2003). For example, applying a me-
dian filter or a morphological opening with horizontal direction results in the enhancement
of horizontal parts of bright structures (Fig. 4.8-a). Similarly, using a vertical structuring
element results in the enhancement of the vertical portions of the structures (Fig. 4.8-b).
(a) Horizontal direction (b) Vertical direction (c) Two directions (d) Four directions
Figure 4.8.: Principle of directional filtering of a thin structure. (a) and (b): result of median
filter using an horizontal and a vertical linear structuring element. (c) and (d): combination of
the results obtained from two directions (horizontal and vertical) and four directions (by adding
diagonal directions).

4.4 Plugin Usage
The results of oriented filters for each direction can be combined by computing the max-
imum value over all orientations (Fig. 4.7-d). Figures 4.8-c and 4.8-d show the results ob-
tained when combining two or four directions. Here, 32 orientations of line with length 25
were used. This results in the enhancement of the image while preserving the thickness of
the bright structures.
Similar results may be obtained for enhancing dark curvilinear structures, by using mor-
phological closing or median filters, and combining the results by computing the minimum
over all directions.
4.4. Plugin Usage

The collection of morphological filters is available in the “Plugins . MorphoLibJ” menu. Fil-
ters are implemented for both 2D and 3D images, and work for binary, grayscale or color
(RGB) images.
4.4.1. Planar images

Morphological filters for planar images are available “Plugins . MorphoLibJ . Morphological
filters”. The dialog let the user choose the structuring element shape, radius, and eventually
preview the result. The following operations can be chosen:
erosion keeps the minimum value within the neighborhood defined by the structuring ele-
ment.
dilation keeps the maximum value within the neighborhood defined by the structuring el-
ement.
closing consists in the succession of a dilation with an erosion. Morphological closing makes
dark structures smaller than the structuring element disappear.
opening consists in the succession of an erosion with a dilation. Morphological opening

makes bright structures smaller than the structuring element disappear.
morphological gradient is defined as the difference of a morphological dilation and a mor-

phological erosion with the same structuring element, and enhances edges of the orig-
inal images.
morphological Laplacian is defined as half the sum of a morphological dilation and a

morphological erosion with the same structuring element, minus the original image,
and enhances edges of the image.
black top-hat consists in subtracting the original image from the result of a morphological
closing, and results in the enhancement of dark structures smaller than structuring
element.

4.4 Plugin Usage
white top-hat consists in subtracting the result of a morphological opening from the origi-
nal image, and results in the enhancement of bright structures smaller than structuring
element.
The following structuring elements can be used for 2D images:
• disk
• square
• octagon
• diamond
• line with angle of 0, 90, 45 or 135 degrees
4.4.2. 3D images
Morphological filters for 3D images are available under “Plugins . MorphoLibJ . Morphological
filters (3D)”. The dialog let the user choose the structuring element shape and radius. The
same list of operations as for planar images is provided. Planar structuring elements can be
used (the operation is simply repeated on each slice), as well as a cubic or spherical struc-
turing element. For most structuring elements, the size can be chosen for each direction.
4.4.3. Directional Filters

Directional filtering is available from “Plugins . MorphoLibJ . Directional Filtering”. It re-
quires a planar image.
The parameters are:
Type to specify how to combine the results for each oriented filter
Operation the operation to apply using each oriented structuring element
Line Length the approximated length of the structuring element.
Direction Number the number of oriented structuring elements to consider. To be in-
creased if the length of line is large.

5. Connected components operators
The “classical” morphological filters presented in chapter 4 transform an input image by us-
ing the values of pixels or voxels located in a close neighborhood, defined by the structuring
element. Such filters can be seen as “local”, as the result in a given position does not depend
on image values located at a sufficient distance.
Connected components operators are more general as they propagate information within
the image based on connectivity between pixels or voxels. More details can be found in
the review of Breen & Jones (1996). Connected components operators encompass powerful
operators, such as morphological reconstruction that allows to reconstruct a marker image
by constraining it to a mask (section 5.1). An extension of morphological reconstruction is
the detection of extended minima and maxima, that can be useful as marker detection for
segmentation (section 5.2). Finally, attribute opening and filtering algorithms can filter
images based on size or range properties, with better preservation of edges than classical
filtering (section 5.3).
Contents
5.1. Morphological reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 22
5.1.1. Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.1.2. Applications to binary images . . . . . . . . . . . . . . . . . . . . . 22
5.1.3. Applications to grayscale images . . . . . . . . . . . . . . . . . . . . 23
5.1.4. Application to the segmentation of chromocenters in a 3D image . 23
5.1.5. Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.2. Regional and extended extrema . . . . . . . . . . . . . . . . . . . . . . . 25
5.2.1. Regional minima and maxima . . . . . . . . . . . . . . . . . . . . . 25
5.2.2. Extended minima and maxima . . . . . . . . . . . . . . . . . . . . . 26
5.2.3. Minima or maxima imposition . . . . . . . . . . . . . . . . . . . . . 26
5.2.4. Plugins usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.3. Attribute filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.3.1. Application to binary images . . . . . . . . . . . . . . . . . . . . . . 28
5.3.2. Application to grayscale images . . . . . . . . . . . . . . . . . . . . 28
5.3.3. Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
21
5.1 Morphological reconstruction
5.1. Morphological reconstruction

The morphological reconstruction is at the basis of many useful algorithms, such as border
removing, hole filling, or detection of regional minima or maxima in grayscale images. It
may be defined for binary as well as for grayscale images.
5.1.1. Principle
The principle of morphological reconstruction canbe illustrated by applying conditional di-
lations or erosions until idempotence. Conditional dilation is the result of a dilation, com-
bined with a mask image using a logical operation. Conditional dilations are repeated until
no more modification occur (idempotence condition).
Figure 5.1.: Principle of the morphological reconstruction algorithm. Original image is repre-
sented in gray, with marker and result of conditional dilations with increasing sizes superim-
posed in black.
The figure 5.1 shows several steps of a morphological reconstruction by dilation on a pair
of binary images. The mask image is shown in gray, and the marker image is shown in black
on the first image. The reconstructed images at each step are shown in black. The markers
propagates within the mask until an additional conditional dilation will not modify the image
any more. The result is the set of regions that contains the markers, with the same shape as
in the original image.
In practice, morphological reconstruction is often implemented using more efficient algo-
rithms based on queues, by a adding pixels or voxels to the queue based on the connectivity
with processed elements (see Section 3.3).
5.1.2. Applications to binary images

By choosing the marker image, several operations may be automatized. For example, com-
puting morphological reconstruction with image of borders, and combining with original
image will remove particles or regions touching the borders (Figure 5.2).
In a similar way, computing morphological reconstruction by using the border of the com-
plement of the image makes it possible to fill holes that may appear in particles.

Figure 5.2.: Some applications of morphological reconstruction. From left to right: original
image, result of kill borders, result of fill holes.
5.1.3. Applications to grayscale images

Morphological reconstructions can be applied to grayscale images. By manually choosing bi-
nary markers such that they overlay specific structures, and after applying a Morphological
reconstruction by dilation, it is possible to obtain a grayscale image containing only the cho-
sen structures (Figure 5.3). The algorithms for morphological reconstruction on grayscale
images are discussed in Vincent (1993); Robinson & Whelan (2004).
Figure 5.3.: Some applications of morphological reconstruction on grayscale images. From left
to right: original image with markers superimposed in red, result of morphological reconstruc-
tion by dilation, result of border kill operation.
The border kill operation can also be applied on grayscale images, making possible to
rapidly remove structures touching the image borders, while keeping the intensity informa-
tion within the structures.
5.1.4. Application to the segmentation of chromocenters in a 3D image

An example of application is presented in figure 5.4. The original 3D image represents the
nucleus of a plant cell, with chromocenters appearing as bright structures (Fig. 5.4-a). In
order to discriminate the chromocenters, the nucleus is first identified by manually selecting
a region of interest (Fig. 5.4-b) used as a marker for a morphological reconstruction with

the original image used as mask. The result of the morphological reconstruction retains the
shape of the nucleus, while “clearing” the bright chromocenters (Fig. 5.4-c).
(a) (b) (c) (d) (e)
Figure 5.4.: Application of 3D morphological reconstruction for isolating chromocenters within

a nucleus. (a) Sample slice of a 3 image of nucleus with chromocenters represented as bright
blobs. (b) Manual selection of a marker region within the nucleus. (c) Morphological recon-
struction of 3D marker image using original 3D image as mask. (d) Difference of original image
with the reconstruction. (e) 3D representation of the chromocenters within the nucleus. Image
by courtesy of Kaori Sakai and Javier Arpon, INRA-Versailles.
By computing the difference of the original image with the result of the morphological
reconstruction, the chromocenters can be easily isolated and segmented (Fig. 5.4-d).
5.1.5. Usage
The reconstruction algorithm is often used within other operators. It is however provided as
a plugin to allow its inclusion in user-designed macros or plugins:
Morphological Reconstruction
Computes the reconstruction by erosion or dilation using a marker image and a mask image,
and a specified connectivity.
Interactive Morphological Reconstruction

Computes the reconstruction by erosion or dilation taking the current 2D image as mask
image, creating the marker image out of user-defined ROIs (for example with the point
selection tool) and using a specified connectivity. The plugin allows previewing the result.
Morphological Reconstruction 3D
Computes the reconstruction by erosion or dilation on a 3D image (marker and mask images
are defined by the user among the open ones).
Interactive Morphological Reconstruction 3D

Computes the reconstruction by erosion or dilation on using the current 3D image as mask
and creating the marker image from the user-defined point selections.

5.2 Regional and extended extrema
The kill borders and fill holes operations are also provided as plugins. Both work for 2D
and 3D images of 8, 16 or 32 bit-depth images.
Kill Borders
Removes the particles touching the border of a binary or grayscale image. See also the
Remove Border Labels Plugin (Section 9.3.1), that performs the same operation on label
maps but without requiring reconstruction.
Fill Holes
Removes holes inside particles in binary images, or remove dark regions surrounded by
bright crests in grayscale images.
5.2. Regional and extended extrema

Minima and maxima within images are important features because they often correspond to
relevant structures within the image. In mathematical morphology, the terms minima and
maxima correspond to regions, i.e. they are not restricted to single pixels or voxels.
5.2.1. Regional minima and maxima

Regional minima are defined as connected regions of elements (pixels or voxels) with the
same value, and whose neighboring elements all have values greater than that of the region.
Similarly, regional maxima are regions of connected pixels or voxels with the same value,
and whose neighbors all have smaller value (Figure 5.5). In both cases, the result depends
on the choice of the connectivity of the regions (see section 3.3).
(a) Original image (b) Regional maxima (c) Profile of gray levels
Figure 5.5.: Regional maxima on a grayscale image. (a) Original image with a line ROI su-
perimposed. (b) Result of regional maxima, showing many spurious maxima. (c) Gray-level
profile along the line ROI shown in (a).
MorphoLibJ uses the algorithm described in Breen & Jones (1996) for computing regional
minima and maxima.

5.2.2. Extended minima and maxima

One problem arising with regional minima or maxima is that they are very sensitive to noise.
It is often more convenient to use so-called extended extrema, that allows filtering spurious
extrema.
Extended minima and maxima are based on the H-extrema transformations, that remove
extrema based on a contrast criterium denoted by h. For example, the H-maxima transfor-
mation of a function f is obtained by performing a morphological reconstruction by dilation
of the function f − h (the marker) constrained by the original function f (the mask). Ex-
tended maxima are obtained by computing the regional maxima of the reconstruction result.
An illustration on a 1D signal is provided on Figure 5.6.
Figure 5.6.: Computation of extended maxima on a 1D function.
For images, extended maxima are defined as a connected region containing elements such
that the difference of the value of each element within the region with the maximal value
within the region is lower than the tolerance, and such that the neighbors of the regions all
have values smaller than the maximum within the region minus the tolerance. This definition
allows the identification of larger extrema, that better takes into account the noise within
the image (Fig. 5.7). The extended minima are defined in a similar way, and are efficiently
used as pre-processing step for watershed segmentation.
Both extended maxima and minima are computed using the morphological reconstruction
algorithm. More details can be found in the book of Soille (2003).
5.2.3. Minima or maxima imposition

It is sometimes useful to transform an input grayscale image by imposing minima or max-
ima specified by a binary image. Such process is at the basis of watershed-based morpho-
logical segmentation (Section 6.4).
The principle is to perform a morphological reconstruction using as a marker image a
combination of the input image and the minima or maxima, and as mask image the original

(a) Original image (b) Extended maxima (c) Profile of gray levels
Figure 5.7.: Extended maxima on a grayscale image. (a) Original image with a line ROI super-
imposed. (b) Result of extended maxima computed with a dynamic of 10. (c) Gray-level profile
along the line ROI shown in (a).
input image. The result is a grayscale image whose regional minima or maxima are the same
as the specified ones.
5.2.4. Plugins usage

The following operations are available in the “Plugins . MorphoLibJ . Minima and Maxima”
menu :
Regional Min / Max

Computes regional minima or extrema in grayscale or binary image, with specified connec-
tivity
Regional Min / Max 3D

Computes regional minima or extrema in 3D grayscale or binary image, with specified con-
nectivity
Extended Min / Max

Computes extended minima or extrema in grayscale or binary image, with specified connec-
tivity
Extended Min / Max 3D

Computes extended minima or extrema in 3D grayscale or binary image, with specified con-
nectivity
Impose Min / Max

Imposes binary minima or maxima on a grayscale image.
Impose Min / Max 3D

Imposes binary minima or maxima on a 3D grayscale image.

5.3 Attribute filtering
5.3. Attribute filtering

Attribute filters aim at removing components of an image based on a certain size criterion,
rather than on intensity. The most common and useful criterion is the number of pixels/vox-
els (i.e., the area or volume). For example, a size opening operation with a threshold value
equal to 20 will remove all blobs containing fewer than 20 voxels. The length of the di-
agonal of the bounding box can also be of interest to discriminate elongated versus round
component shapes.
5.3.1. Application to binary images

When applied to a binary image, attribute opening consists in identifying each connected
component, computing the attribute measurement of each component, and retain only the
connected components whose measurement is above a specified value (Figure 5.8). This
kind of processing is often used to clean-up segmentation results.
(a) Original image (b) Connected components (c) Size opening
Figure 5.8.: Example of area opening on a binary image. (a) Original binary image. (b)
Identification of connected components. (c) Only the connected components with a sufficient
size (defined by the area), have been retained.
5.3.2. Application to grayscale images

When applied to a grayscale image, attribute opening consists in generating a series of binary
images by thresholding at each distinct gray level in the image. The binary attribute opening
described above is then applied independently to each binary image and the grayscale output
is computed as the union of the binary results. The final output is a grayscale image whose
bright structures with the attribute below a given value have disappeared. A great advantage
of this filter is that the contours of the structures area better preserved than opening with a
structuring element (Figure 5.9).
As for classical morphological filters, grayscale attribute closing or tophat can be defined.
Grayscale attribute closing consists in removing dark connected components whose size is

5.3 Attribute filtering
(a) Original image (b) Area opening (c) Morphological opening
Figure 5.9.: Example of area opening on a grayscale image. (a) Original grayscale image of
a leaf (image courtesy of Eric Biot, INRA Versailles). (b) Grayscale size opening making bright
spots disappear. (c) Comparison with morphological closing using a square structuring element
of radius 1: bright spots are removed, but some veins also disappear.
smaller than a specified value. White [resp. Black] Attribute Top-Hat considers the difference
of the attribute opening [resp. closing] with the original image, and can help identifying
bright [resp. dark] structures with small size.
5.3.3. Usage
So far, the following attribute filtering plugins are available within MorphoLibJ (under “Plu-
gins . MorphoLibJ”):
Gray Scale Attribute Filtering opens a dialog to perform between attribute opening, clos-
ing, and black or white top-hat on a planar (2D) grayscale image. Two size criteria
can be used: the area (number of pixels), or the diameter (length of the diagonal of
the bounding box).
Gray Scale Attribute Filtering 3D opens a dialog to perform between attribute opening,
closing, and black or white top-hat on a 3D grayscale image. The size criterion is the
number of voxels.

6. Watershed segmentation
The watershed algorithm assimilates the grayscale image to a digital elevation model, and
aims at detecting the different catchment basins. In the grayscale image, the catchment
basins correspond to dark regions surrounded by bright structures (the “crests”). It is a very
popular technique specially used to segment touching objects. The MorphoLibJ suite con-
tains several implementations and applications of the algorithm, described in the following
sections.
Contents
6.1. Classic Watershed plugin . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.1.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.1.2. Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.1.3. Over-segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.2. Marker-controlled Watershed plugin . . . . . . . . . . . . . . . . . . . . 33
6.2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.2.2. Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.3. Interactive Marker-controlled Watershed plugin . . . . . . . . . . . . . . 35
6.3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6.3.2. Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6.4. Morphological Segmentation plugin . . . . . . . . . . . . . . . . . . . . . 38
6.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.4.2. Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.4.3. Macro language compatibility . . . . . . . . . . . . . . . . . . . . . 42
6.5. Distance Transform Watershed . . . . . . . . . . . . . . . . . . . . . . . . 43
6.5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.5.2. Distance Transform Watershed . . . . . . . . . . . . . . . . . . . . . 43
6.5.3. Distance Transform Watershed (3D) . . . . . . . . . . . . . . . . . . 45
30
6.1 Classic Watershed plugin
6.1. Classic Watershed plugin

6.1.1. Introduction
Classic Watershed is an ImageJ/Fiji plugin to perform watershed segmentation of grayscale
2D/3D images using flooding simulations as described by Soille & Vincent (1990); Vincent
& Soille (1991).
The basic idea consists of considering the input image as topographic surface and placing
a water source in each regional minimum of its relief. Next the entire relief is flooded from
the sources and dams are placed where the different water sources meet.
All points in the surface at a given minimum constitute the catchment basin associated
with that minimum. The watersheds are the zones dividing adjacent catchment basins.
The first image points that are reached by water are the points at the lowest grayscale
value hmin , then all image pixels are progressively reached up to the highest level hmax (see
Fig. 6.1).
h=40
h=90
h=160
Intensity
Figure 6.1.: Schematic overview of watershed flooding in 1D
6.1.2. Usage
The Classic Watershed plugin runs on any grayscale image (8, 16 and 32-bit) in 2D and 3D.
At least one image needs to be open in order for the plugin to run. If that’s the case, a dialog
like the following will pop up:

6.1 Classic Watershed plugin
Let’s have a look at the different parameters:
• Image parameters:
– Input image: grayscale image to flood, usually the gradient of an image.
– Mask image (optional): binary image of the same dimensions as the input image
which can be used to restrict the areas of application of the algorithm. Set to
"None" to run the method on the whole input image.
• Morphological parameters:
– Use diagonal connectivity: select to allow the flooding in diagonal directions
(8-connectivity in 2D and 26-connectivity in 3D, see Section 3.3).
– Min h: minimum grayscale value to start flooding from (by default, set to the
minimum value of the image type).
– Max h: maximum grayscale value to reach with flooding (by default, set to the
maximum value of the image type).
Output:
• Labeled image containing the resulting catchment basins (with integer values 1, 2,
3...) and watershed lines (with 0 value).
6.1.3. Over-segmentation
Normally, Classic Watershed will lead to an over-segmentation of the input image, especially
for noisy images with many regional minima. In that case, it is recommended to either pre-
process the image before running the plugin, or merge regions based on a similarity
criterion afterwards. Several de-noising methods are available in Fiji/ImageJ, namely:
median filtering, Gaussian blur, bilateral filtering, etc.
Example: This short macro runs the plugin twice in the blobs sample, first without pre-
processing and then after applying a Gaussian blur of radius 3:

6.2 Marker-controlled Watershed plugin
1 // l o a d t h e B l o b s sample image
2 run ( " B l o b s (25K) " ) ;
3 // i n v e r t LUT and p i x e l v a l u e s t o have dark b l o b s
4 run ( " I n v e r t LUT " ) ;
5 run ( " I n v e r t " ) ;
6 // run p l u g i n on image
7 run ( " C l a s s i c Watershed " , " i n p u t=b l o b s mask=None use min=0 max=150 " ) ;
8 // a p p l y LUT t o f a c i l i t a t e r e s u l t v i s u a l i z a t i o n
9 run ( "3−3−2 RGB" ) ;
10 // pre−p r o c e s s image with Gaussian b l u r selectWindow ( " b l o b s . g i f " ) ;
11 run ( " Gaussian B l u r . . . " , " sigma=3" ) ;
12 rename ( " b l o b s −b l u r . g i f " ) ;
13 // a p p l y p l u g i n on pre−p r o c e s s e d image
14 run ( " C l a s s i c Watershed " , " i n p u t=b l o b s −b l u r mask=None use min=0 max=150 " ) ;
15 // a p p l y LUT t o f a c i l i t a t e r e s u l t v i s u a l i z a t i o n
16 run ( "3−3−2 RGB" ) ;
(a) Gaussian-blurred blobs image (b) Watershed segmentation on (c) Watershed segmentation on
used as input (radius = 3). original image (hmin = 0, hmax = Gaussian-blurred original image
150). (radius = 3, hmin = 0, hmax = 150).
Figure 6.2.: Input and result images from the macro example.
6.2. Marker-controlled Watershed plugin

6.2.1. Introduction
Marker-controlled Watershed is an ImageJ/Fiji plugin to segment grayscale images of any
type (8, 16 and 32-bit) in 2D and 3D based on the marker-controlled watershed algorithm by
Meyer & Beucher (1990). As the previous method, this algorithm considers the input image
as a topographic surface (where higher pixel values mean higher altitude) but it simulates
its flooding from specific seed points or markers. A common choice for the markers are
the local minima of the gradient of the image, but the method works on any specific marker,
either selected manually by the user or determined automatically by another algorithm (see
Fig. 6.3).

6.2 Marker-controlled Watershed plugin
Figure 6.3.: Example of marker-controlled watershed segmentation on nucleus of Arabidopsis

thaliana (image by courtesy of Kaori Sakai and Javier Arpon, INRA-Versailles).
6.2.2. Usage
Marker-controlled Watershed needs at least two images to run. If that’s the case, a dialog
like the following will pop up:
Let’s have a look at the different parameters:

• Image parameters:
– The Input image: a 2D or 3D grayscale image to flood, usually the gradient of
an image.
– The Marker image: an image of the same dimensions as the input containing
the seed points or markers as connected regions of voxels, each of them with a
different label. They correspond usually to the local minima of the input image,
but they can be set arbitrarily.
• And it can optionally admit a third image:

– The Mask image: a binary image of the same dimensions as input and marker
which can be used to restrict the areas of application of the algorithm. Set to
"None" to run the method on the whole input image.
• Rest of parameters:
– Calculate dams: select to enable the calculation of watershed lines.
– Use diagonal connectivity: select to allow the flooding in diagonal directions
(more rounded objects are usually obtain by unchecking this option).

6.3 Interactive Marker-controlled Watershed plugin
Output:
• Labeled image containing the catchment basins and (optionally) watershed lines (dams).
6.3. Interactive Marker-controlled Watershed plugin

6.3.1. Introduction
Similar to the Marker-controlled Watershed plugin, this ImageJ/Fiji plugin segments grayscale
images of any type (8, 16 and 32-bit) in 2D and 3D using the marker-controlled watershed
algorithm by Meyer & Beucher (1990).and it floods the image from specific seed points, but
this time the points are introduced interactively by the user.
Figure 6.4.: Overview of the Interactive Marker-controlled Watershed plugin.
6.3.2. Usage
Interactive Marker-controlled Watershed runs on any open grayscale image, single 2D image
or (3D) stack. If no image is open when calling the plugin, an Open dialog will pop up.
The user can pan, zoom in and out, or scroll between slices (if the input image is a stack)
in the main canvas as if it were any other ImageJ window. On the left side of the canvas

there are three panels of parameters, one with the watershed parameters, one for the output
(result) options and one for post-processing the result. All buttons, checkboxes and panels
contain a short explanation of their functionality that is displayed when the cursor lingers
over them.
6.3.2.1. Interactive markers

In this plugin the markers are introduced interactively by the user using any of the selection
tools. By default, the point selection tool will be enabled in the main ImageJ toolbar. To
select markers on different slices, one option is to use the point selection tool and keep the
SHIFT key pressed each time you click to set a new marker. Another possibility is to use the
ROI Manager. In that case, all selected ROIs in the manager will be used as markers (see
Fig. 6.3for examples of selections used as markers).
Figure 6.5.: Example of interactive markers introduced by the user. From left to right: point
selections, rectangular selections and free-hand selections (stored in the ROI Manager).
6.3.2.2. Watershed Segmentation panel
This panel is reserved to the parameters involved in the segmentation pipeline:
• Calculate dams: un-check this option to produce segmentations without watershed

lines.

• Connectivity: voxel connectivity (4-8 in 2D, and 6-26 in 3D). Selecting non-diagonal
connectivity (4 or 6) usually provides more rounded objects.
Finally, click on “Run” to launch the segmentation. If your segmentation is taking too
long or you want to stop it for any reason, you can do so by clicking on the same button
(which should read “STOP” during that process).
6.3.2.3. Results panel
Only enabled after running the segmentation.
• Display: list of options to display the segmentation results (see Fig. 6.7).
– Overlaid basins: colored objects overlaying the input image (with or without
dams depending on the selected option in the Watershed Segmentation panel).
– Overlaid dams: overlay the watershed dams in red on top of the input image
(only works if “Calculate dams” is checked).
– Catchment basins: colored objects.
– Watershed lines: binary image showing the watershed lines in black and the
objects in white (only works if “Calculate dams” is checked).
• Show result overlay: toggle result overlay.
• Create image button: create a new image with the results displayed in the canvas.
6.3.2.4. Post-processing panel
Similarly to the Results panel, this panel only gets enabled after running the segmentation
pipeline.

6.4 Morphological Segmentation plugin
• Merge labels: merge together labels selected by either the “freehand” selection tool
(on a single slice) or the point tool (on single or multiple slices). The zero-value label
belongs to the watershed dams, therefore it will ignored in case of being selected. The
first selected label value will be assigned to the rest of selected labels, which will share
its color.
Note: to select labels on different slices, use the point selection tool and keep the
SHIFT key pressed each time you click on a new label.
• Shuffle colors: randomly re-assign colors to the labels. This is a very handy option
whenever two adjacent labels present a similar color.
6.4. Morphological Segmentation plugin

6.4.1. Introduction
Morphological Segmentation is an ImageJ/Fiji plugin that combines morphological opera-
tions, such as extended minima and morphological gradient, with watershed flooding algo-
rithms to segment grayscale images of any type (8, 16 and 32-bit) in 2D and 3D.
Figure 6.6.: Overview of the Morphological Segmentation plugin.

6.4.2. Usage
Morphological Segmentation runs on any open grayscale image, single 2D image or (3D)
stack. If no image is open when calling the plugin, an Open dialog will pop up.
The user can pan, zoom in and out, or scroll between slices (if the input image is a stack)
in the main canvas as if it were any other ImageJ window. On the left side of the canvas there
are four panels of parameters, one for the input image, one with the watershed parameters,
one for the output options and one for post-processing the resulting labels. All buttons,
checkboxes and panels contain a short explanation of their functionality that is displayed
when the cursor lingers over them.
Image pre-processing: some pre-processing is included in the plugin to facilitate the
segmentation task. However, other pre-processing may be required depending on the input
image. It is up to the user to decide what filtering may be most appropriate upstream.
6.4.2.1. Input Image panel
First, you need to indicate the nature of the input image to process. This is a key param-
eter since the watershed algorithm is expecting an image where the boundaries of objects
present high intensity values (usually as a result of a gradient or edge detection filtering).
You should select:
• Border Image: if your input image has highlighted object boundaries.
• Object Image: if the borders of the objects do not have higher intensity values than
the rest of voxels in the image.
When selecting “Object Image”, an additional set of options is enabled to choose the type of
gradient and radius (in pixels) to apply to the input image before starting the morphological
operations. Finally, a checkbox allows displaying the gradient image instead of the input
image in the main canvas of the plugin (only after running the watershed segmentation).

6.4.2.2. Watershed Segmentation panel
This panel is reserved to the parameters involved in the segmentation pipeline. By de-
fault, only the tolerance can be changed. Clicking on “Advanced options” enables the rest of
options.
• Tolerance: dynamic of intensity for the search of regional minima (in the extended-
minima transform, which is the regional minima of the H-minima transform, value of
h). Increasing the tolerance value reduces the number of segments in the final result,
while decreasing its value produces more object splits.
Note: since the tolerance is an intensity parameter, it is sensitive to the input image
type. A tolerance value of 10 is a good starting point for 8-bit images (with 0-255 in-
tensity range) but it should be drastically increased when using image types with larger
intensity ranges. For example to ~2000 when working on a 16-bit image (intensity
values between 0 and 65535).
• Calculate dams: un-check this option to produce segmentations without watershed

lines.
• Connectivity: voxel connectivity (4-8 in 2D, and 6-26 in 3D). Selecting non-diagonal
connectivity (4 or 6) usually provides more rounded objects.
Finally, click on “Run” to launch the segmentation.

If your segmentation is taking too long or you want to stop it for any reason, you can do
so by clicking on the same button (which should read “STOP” during that process).
6.4.2.3. Results panel

Only enabled after running the segmentation.
• Display: list of options to display the segmentation results.

– Overlaid basins: colored objects overlaying the input image (with or without
dams depending on the selected option in the Watershed Segmentation panel).
– Overlaid dams: overlay the watershed dams in red on top of the input image
(only works if “Calculate dams” is checked).
– Catchment basins: colored objects.
– Watershed lines: binary image showing the watershed lines in black and the
objects in white (only works if “Calculate dams” is checked).
• Show result overlay: toggle result overlay.
• Create image button: create a new image with the results displayed in the canvas.
Figure 6.7.: Examples of the 4 different display options.
6.4.2.4. Post-processing panel
Similarly to the Results panel, this panel only gets enabled after running the segmentation
pipeline.
• Merge labels: merge together labels selected by either the “freehand” selection tool
(on a single slice) or the point tool (on single or multiple slices). The zero-value label
belongs to the watershed dams, therefore it will ignored in case of being selected. The
first selected label value will be assigned to the rest of selected labels, which will share
its color.
Note: to select labels on different slices, use the point selection tool and keep the
SHIFT key pressed each time you click on a new label.

• Shuffle colors: randomly re-assign colors to the labels. This is a very handy option
whenever two adjacent labels present a similar color.
6.4.3. Macro language compatibility

Morphological Segmentation is completely compatible with the popular ImageJ macro lan-
guage1 . Each of the buttons in the GUI are macro-recordable and their commands can be
reproduced later from a simple macro file.
The complete list of commands is as follows:
• Start the plugin:
1 run ( " M o r p h o l o g i c a l Segmentation " ) ;
• Select input image:
1 // s e l e c t as o b j e c t image
2 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . se tIn put Im age Ty pe " , " o b j e c t " ) ;
3 // s e l e c t as b o r d e r image
4 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . se tIn put Im age Ty pe " , " bor der " ) ;
• Run segmentation with specific options:
1 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . segment " , " t o l e r a n c e =10 " , "

c a l c u l a t e D a m s=t r u e " , " c o n n e c t i v i t y =6" ) ;
• Toggle result overlay:

1 c a l l ( " in r a . i j p b . plugins . MorphologicalSegmentation . toggleOverlay " ) ;
• Set option to display gradient image:
1 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . setShowGradient " , " t r u e " ) ;
• Select display format:
1 // O v e r l a i d b a s i n s
2 c a l l ( " in r a . i j p b . plugins . MorphologicalSegmentation . setDisplayFormat " , " Overlaid
basins " ) ;
3 // O v e r l a i d dams
4 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . s e t D i s p l a y F o r m a t " , " O v e r l a i d dams "
);
5 // Catchment b a s i n s
6 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . s e t D i s p l a y F o r m a t " , " Catchment
basins " ) ;
1
http://imagej.net/developer/macro/macros.html

6.5 Distance Transform Watershed
7 // Watershed l i n e s
8 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . s e t D i s p l a y F o r m a t " , " Watershed
lines " ) ;
• Create new image with the current result:
1 c a l l ( " in r a . i j p b . plugins . MorphologicalSegmentation . createResultImage " ) ;
Complete macro example

The following macro opens the Blobs sample image, applies the plugin with a tolerance value
of 32 and displays the result as overlaid dams.
1 // l o a d t h e B l o b s sample image
2 run ( " B l o b s (25K) " ) ;
3 // run t h e p l u g i n
4 run ( " M o r p h o l o g i c a l Segmentation " ) ;
5 // w a i t f o r t h e p l u g i n GUI t o l o a d
6 w a i t (1000) ;
7 // s e l e c t i n p u t image as " o b j e c t "
8 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . set In put Im age Ty pe " , " o b j e c t " ) ;
9 // s e t g r a d i e n t r a d i u s as 1
10 c a l l ( " in r a . i j p b . plugins . MorphologicalSegmentation . setGradientRadius " , " 1 " ) ;
11 // run s e g m e n t a t i o n with t o l e r a n c e 32 , c a l c u l a t i n g t h e watershed dams ,
12 // 4− c o n n e c t i v i t y
13 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . segment " , " t o l e r a n c e =32 " , "
c a l c u l a t e D a m s=t r u e " , " c o n n e c t i v i t y =4" ) ;
14 // d i s p l a y t h e o v e r l a i d dams
15 c a l l ( " i n r a . i j p b . p l u g i n s . M o r p h o l o g i c a l S e g m e n t a t i o n . s e t D i s p l a y F o r m a t " , " O v e r l a i d dams " ) ;
6.5. Distance Transform Watershed

6.5.1. Introduction
A classic way of separating touching objects in binary images makes use of the distance
transform and the watershed method. The idea is to create a border as far as possible from
the center of the overlapping objects. This strategy works very well on rounded objects and
it is called Distance Transform Watershed. It consists on calculating the distance transform
of the binary image, inverting it (so the darkest parts of the image are the centers of the
objects) and then applying watershed on it using the original image as mask (see Figure
6.8). In our implementation, we include an option to use watershed with extended minima
so the user can control the number of object splits and merges.
MorphoLibJ provides two plugins under the “Plugins . MorphoLibJ . Binary Images...” menu
to apply this strategy on 2D and 3D images:
6.5.2. Distance Transform Watershed

Distance Transform Watershed needs one 2D 8-bit binary image to run. If that’s the case, a
dialog like the following will pop up:

Figure 6.8.: Basics of the Distance Transform Watershed algorithm. From left to right:
sample image of touching DAPI stained cell nuclei from a confocal laser scanning microscope,
binary mask calculated after filtering and thresholding input image, inverse of the distance
transform applied to the binary mask (Chamfer distance map using normalized Chessknight
weights and 32-bit output) and resulting labeled image after applying watershed to the inverse
distance image using the binary mask (dynamic of 1 and 4-connectivity).
The plugin parameters are divided between the distance transform and the watershed
options:
• Distance map options:
– Distances: allows selecting among a pre-defined set of weights that can be used
to compute the distance transform using Chamfer approximations of the Eu-
clidean metric (see section 8.1.1). They affect the location but specially the shape
of the border in the final result. The options are:
· Chessboard (1,1): weight equal to 1 for all neighbors.
· City-Block (1,2): weights 1 for orthogonal neighbors and 2 for diagonal
neighbors.
p
· Quasi-Euclidean (1,1.41): weights 1 for orthogonal neighbors and 2 for
diagonal neighbors.

· Borgefors (3,4): weights 3 for orthogonal neighbors and 4 for diagonal

neighbors (best approximation of Euclidean distance for 3-by-3 masks).
· Weights (2,3): weights 2 for orthogonal neighbors and 3 for diagonal neigh-
bors.
· Weights (5,7): weights 5 for orthogonal neighbors and 7 for diagonal neigh-
bors.
· Chessknight (5,7,11): weights 5 for orthogonal neighbors and 7 for diago-
nal neighbors, and 11 for chess-knight moves (best approximation for 5-by-5
masks).
– Output type: 16 or 32-bit, to calculate distance with short or float precision.
– Normalize weights: indicates whether the resulting distance map should be nor-
malized (divide distances by the first Chamfer weight).
• Watershed options:
– Dynamic: same as in the Morphological Segmentation plugin, this is the dynamic
of intensity for the search of regional minima in the inverse of the distance trans-
form image. Basically, by increasing its value there will be more object merges
and by decreasing it there will be more object splits.
– Connectivity: pixel connectivity (4 or 8). Selecting non-diagonal connectivity
(4) usually provides more rounded objects.
Finally, the result with the current plugin configuration can be visualized by clicking on the
Preview option.
Result: 2D 32-bit label image (one index value per object).
6.5.3. Distance Transform Watershed (3D)

Distance Transform Watershed 3D needs one 3D 8-bit binary image to run. If that’s the case,
a dialog like the following will pop up:

The parameters are the same as in the 2D version but some of them are adapted for 3D
images:
• Distance map options:

– Distances: Now the available options are:
· Chessboard (1,1,1): weight equal to 1 for all neighbors.
· City-Block (1,2,3): weights 1 for orthogonal neighbors, 2 for diagonal neigh-
bors and 3 for cube-diagonals.
p
· Quasi-Euclidean (1,1.41,1.73):
p weights 1 for orthogonal neighbors, 2 for
diagonal neighbors and 3 for cube-diagonals.
· Borgefors (3,4,5): weights 3 for orthogonal neighbors, 4 for diagonal neigh-
bors and 5 for cube-diagonals (best approximation of Euclidean distance for
3-by-3-by-3 masks).
– Output type: 16 or 32-bit, to calculate distance with short or float precision.
– Normalize weights: indicates whether the resulting distance map should be nor-
malized (divide distances by the first Chamfer weight).
• Watershed options:
– Dynamic: same as in the 2D version, this is the dynamic of intensity for the search
of regional minima in the inverse of the distance transform image. Basically, by
increasing its value there will be more object merges and by decreasing it there
will be more object splits.
– Connectivity: voxel connectivity (6 or 26). Selecting non-diagonal connectivity
(6) usually provides more rounded objects.
As it is usual in ImageJ, no preview is provided here since we are dealing with 3D images.
Result: 3D 32-bit label image (one index value per object).

7. Measurements
MorphoLibJ contains several tools for quantifying the size, the shape, the intensities, or the
spatial organization of biological structures observed within images 2D or 3D images.
Sections 7.1 and refer 7.2 to the analysis of 2D and 3D regions, respectively. Most region
analysis operators of the MorphoLibJ library assume the input image to be either a binary
image representing a single region, or a label image representing a collection of disjoint
regions (see section 9). The aim is to facilitate the management of label images, contrary to
the built-in “Analyze Particles...” function that operates directly on a grayscale image.
The library also provides plugins for intensity measurements (section 7.3), the comparison
of label images (section 7.4), and the quantification of spatial organization (section 9.4.1).
Contents
7.1. Region Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.1.1. Intrinsic volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.1.2. Geometric moments and equivalent ellipse . . . . . . . . . . . . . . 51
7.1.3. Feret diameter and oriented box . . . . . . . . . . . . . . . . . . . . 52
7.1.4. Geodesic measurements . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.1.5. Thickness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.1.6. Quantification by shape indices . . . . . . . . . . . . . . . . . . . . 54
7.1.7. Plugins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
7.2. Region Analysis 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2.1. Intrinsic volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2.2. Geometric moments and equivalent Ellipsoid . . . . . . . . . . . . . 63
7.2.3. 3D Shape indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.2.4. Plugins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.3. Intensity measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.4. Label Overlap measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.5. Microstructure analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
47
7.1 Region Analysis
7.1. Region Analysis

Region analysis usually refers to the quantification of features related to the size and the
shape of regions identified within images (Arganda-Carreras & Andrey, 2017).
Many region features have been proposed to describe morphology of regions. The Mor-
phoLibJ library proposes plugins to quantify some of the most common ones:
• intrinsic volumes (section 7.1.1) encompass the area, the perimeter, and the Euler
number, that quantifies the topology of the region
• the computation of moments of a region can be used to determine the centroid or

equivalent ellipse (section 7.1.2)
• dimension extent may be determined by use of Feret diameters (section 7.1.3)
• geodesic features allow to investigate the complexity of the region, in particular the
tortuosity (section 7.1.4)
• the use of the skeleton provide derived features such as the average thickness (sec-
tion 7.1.5)
• features describing the size may be combined to obtain shape indices that describe
the morphology of the regions independently of their size and location (section 7.1.6)
The following sections will present these featuers. For each family of features, we first pro-
vide mathematical definitions of the implemented features, eventually completed by the
methods used to measure them from discrete images. A final section (section 7.1.7) presents
the different plugins that allow to measure features from images.
7.1.1. Intrinsic volumes

The intrinsic volumes are a set of features with interesting mathematical properties that
are commonly used for describing individual particles as well as binary microstructures. In
2D, they correspond to the area, the perimeter and the Euler number.
7.1.1.1. Area
The area of a set X can be defined using integrals over the set:
Z
A(X ) = dx (7.1)
X
In image analysis, the measurement of area of a region simply consists in counting the
number of pixels that constitute it, and to multiply by the area of a single pixel.

7.1 Region Analysis
7.1.1.2. Perimeter
The perimeter of a region corresponds to the length of its boundary. Its formal definition
involves computing an integral over its boundary ∂ X :
Z
P(X ) = dx (7.2)
∂X
An intuitive approach for measuring the perimeter of a binary region is to count the
number of boundary pixels (Figure 7.1). However, this approach is rather inaccurate. For
example, when counting the boundary pixels of a discretized disk, one obtain the same
perimeter value as the length of the bounding rectangle, leading to a relative error of more
than 25%1 .
(a) Boundary pixels count (b) Crofton method
Figure 7.1.: Measurement of perimeter within a discrete image of a disk. (a) Counting the
number of boundary pixels. (b) Counting the number of intersection with parallel lines.
The discrete version of Crofton formula can provide a better estimate of the perimeter
than traditional boundary pixel count. The principle is to consider a set of parallel lines with
a random orientation, and to count the number of intersections with the region of interest
(Figure 7.1). The number of intersections is proportional to the line density and to the
perimeter (Serra, 1982; Legland et al., 2007; Ohser & Schladitz, 2009). By averaging over
all possible directions, an unbiased estimate can be obtained.
Perimeter can be measured using either two directions (horizontal and vertical), or four
directions (by adding the diagonals). Restricting the number of directions introduces an
estimation bias, with known theoretical bounds (Moran, 1966; Legland et al., 2007). The
error made by counting intersections is usually better than boundary pixel count (Lehmann
& Legland, 2012).
1
If d is the diameter of the circle, the expected perimeter is πd. The number of boundary pixels measure the
perimeter of the enclosing square, equal to 4d. The relative error is (4d − πd)/πd ' 27%

7.1 Region Analysis
7.1.1.3. Euler Number

The Euler number is a feature that describes the topology of a region. It corresponds to the
number of connected components, minus the number of holes (Fig. 7.2-a).
Figure 7.2.: Euler Number for 2D regions. Particle A is composed of a single component, its
Euler number is 1. Particles B and C present one and two holes respectively. Their corresponding
Euler numbers are equal to 0 = 1 − 1 and −1 = 1 − 2.
In 2D, the Euler number of a region with smooth boundary also equals the integral of the
curvature over the boundary of the set:
Z
1
χ(X ) = κ(x)d x (7.3)
2π ∂ X
The measurement of Euler number from binary images is usually performed by consider-
ing a reconstruction of the connected components that connects adjacent pixels or voxels
(Fig. 7.3). The resulting reconstruction depends on the choice of the connectivity (see Sec-
tion 3.3). For planar images, typical choices are the 4-connectivity, corresponding to the
orthogonal neighbors (Fig. 7.3-b), and the 8-connectivity, that also consider the diagonal
neighbors (Fig. 7.3-b). The connectivity is also considered for computing morphological re-
constructions (section 5.1), or for computing the connected component labeling of a binary
image (section 8.2).
(a) Binary image (b) 4-connectivity (c) 8-connectivity
Figure 7.3.: Determination of the Euler number from binary images. (a) Original binary im-
ages. (b) Euler number computed with 4-connectivity: five components are detected. (c) Euler
number computed using the 8-connectivity: three components are detected, two of them having
holes.

7.1 Region Analysis
Depening on the chosen connectivity, the resulting Euler number may differ. For exam-
ple, the reconstruction presented in Figure 7.3 has Euler number equal to five (the num-
ber of connected components), whereas the Euler number of the reconstruction using the
8-connectivity is equal to 1: three connected components are obtained, minus two holes
within them (Fig. 7.3-c).
7.1.2. Geometric moments and equivalent ellipse

A binary region X may be described mathematically by its geometric moments m pq of order
(p, q), which correspond to an integral of its indicator function I X , with various degrees along
the directions:
Z Z
m pq (X ) = I X (x, y)x p y q · d x · d y (7.4)
The moment of order (0, 0) simply corresponds to the area of X :
Z Z
m00 = I X (x, y) · d x · d y
= A(X )
The coordinates of the centroid of X can be expressed from the first-order moments:
Z Z
m10 1
xc = = I X (x, y) · x · d x · d y
m00 Ar ea(X )
Z Z
m01 1
yc = = I X (x, y) · y · d x · d y
m00 Ar ea(X )
It is often more convenient to work with centered moments, given by:
Z Z
µ pq = I X (x, y) (x − x c ) p ( y − y c)q · d x · d y (7.5)
The second order centered moments can be arranged into a matrix:

µ20 µ11
µ11 µ02
By computing the parameters of the ellipse that produces the same moments up to the
second order, one obtains the equivalent ellipse (Burger & Burge, 2008). In particular, the
orientation θ of the equivalent ellipse is given by θ = 12 atan2 (2 · µ11 , µ20 − µ02 ).

7.1 Region Analysis
Figure 7.4.: Equivalent ellipse of a binary region.
7.1.3. Feret diameter and oriented box

A popular way to assess the size of a region is to measure its largest Feret diameter. The
maximum Feret diameter of a region is simply the maximum distance computed over all the
pairs of points belonging to the region:
Fmax = max d(x , y) (7.6)

x ,y∈X
It is easy to realize that the search can be performed on the set of boundary points (Fig. 7.5).
In practice, the computation of Maximum Feret diameter can be accelerated by first comput-
ing the convex hull of the region. In MorphoLibJ, the maximal Feret diameter is computed
over the corners of the regions.
Figure 7.5.: Feret diameter and oriented box
The Feret diameter can be performed on any direction. Typical choice is to measure the
Feret diameter in the direction perpendicular to the direction of the maximum Feret diam-
eter. The ratio of the largest Feret diameter and of the Feret diameter measured in the
perpendicular direction gives an indication on the elongation of the region.
The notion of object-oriented bounding box (sometimes denoted as “OOBB”) is closely
related to that of Feret diameter. One possible definition of oriented bounding box is the
rectangle with smallest width that totally encloses the region. An example is illustrated on

7.1 Region Analysis
Figure 7.5. Another definition is the rectangle with smallest area: the two resulting rect-
angles are often similar, but may differ in some cases. Note that in general, the orientation
of the largest axis of the oriented box differs from both the direction of the largest Feret
diameter, and the direction of the equivalent ellipse.
7.1.4. Geodesic measurements

For particles with complex shapes, the geodesic diameter may be of interest. It corresponds
of the largest geodesic distance between two points within a region, the geodesic distance
being the length of the shortest path joining the two points while staying inside the region
(Lantuéjoul & Beucher, 1981). An example is given on Figure 7.6.
Figure 7.6.: Geodesic diameter. Left: illustration of the geodesic diameter on a simple particle.
Right: computation of the geodesic diameter on a segmented image from the DRIVE database
(Staal et al., 2004). Each connected component was associated to a label, then the longest
geodesic path within each connected component was computed and displayed as red overlay.
The implementation of geodesic diameter within MorphoLibJ relies on Chamfer distances

propagation (see section 8.1.3). Hence, the actual value may be somewhat over-estimated
due to the approximation in the measurements.
7.1.5. Thickness
The thickness is also a convenient measure to describe a region. It may be evaluated by
the diameter of the largest inscribed circle plugin (page 59, Fig. 7.7). This method however
often over-estimates the thickness, as it measures thickness at the largest point within the
region.
The MorphoLibJ library also proposes to measure average thickness by first computing
the skeleton of the region, and then to measure for each pixel of the skeleton the distance
to the closest background point in the original image (Fig. 7.7). This way, the thickness
measure is integrated over the whole region skeleton and is more representative.

7.1 Region Analysis
Figure 7.7.: Evaluation of the thickness using two methods. On the left, computation of the
largest inscribed circle. On the right, computation of the average thickness over the skeleton.
7.1.6. Quantification by shape indices

It is often convenient to describe and quantify the shape of regions independently of their
location, orientation, or relative scaling. Several indices are commonly used to describe the
shape of the particles, independently of their size. We present here the definitions used
within the MorphoLibJ plugins.
7.1.6.1. Circularity
The circularity (or “shape factor”, or “isperimetric deficit index”) is defined as the ratio of
area over the square of the perimeter, normalized such that the value for a disk equals one:
A
circularity = 4π (7.7)
P2
While the values of isoperimetric deficit index theoretically range within the interval [0; 1],
the measurements errors of the perimeter may produce circularity values above 1 (Lehmann
& Legland, 2012).
7.1.6.2. Convexity
The convexity (also known as “solidity”) is defined as the ratio of the area of the region over
the area of its convex hull.
A(X )
convexity =
A(ConvHull(X ))
The Convexify Plugin (Section 8.3.1) also allows to compute the convex image correspond-
ing to a binary region.
7.1.6.3. Elongation
The MorphoLibJ library provides several shape indices describing the elongation of the re-
gions. Their values is 1 for a disk, and increase for elongated regions.

7.1 Region Analysis
Figure 7.8.: Convexity of a planar region. The convex area is the union of the blue and yellow
regions.
The ellipse elongation factor, defined as the ratio of the largest over the smallest axis
lengths, can be used to quantify the shape of the region.
The oriented box elongation, defined as the ratio of the length of the oriented bounding
box over its width
The geodesic elongation, defined as the ratio of the geodesic diameter of the diameter
of the largest inscribed disk.
7.1.6.4. Tortuosity
The tortuosity is defined as the ratio of the geodesic diameter over the largest Feret diameter.
It describes the complexity of the region. Its theroretical value is 1 for convex regions, and
increases for regions with complex geometries.
7.1.7. Plugins
Most MorphoLibJ plugins consider the current image as input, that must be either binary
(only one region is considered), or label (typically the result of a connected components
labeling, see section 8.2). The output is a results table (ImageJ ResultsTable) containing one
row for each label actually present within the image. The spatial calibration of the image is
taken into account for most measurements. All plugins can be found under the “Plugins .
MorphoLibJ . Analyze” menu.
7.1.7.1. Analyze Regions

The global geometry of particles in 2D images can be characterized with the Analyze Re-
gions plugin (under “Plugins . MorphoLibJ . Analyze . Analyze Regions”). For 2D particles,
the area, the perimeter and derived features are implemented.
The options of the dialog correspond to the parameters that can be computed:
Area the area of the regions

7.1 Region Analysis
Perimeter the perimeter, measured using discrete version of Crofton formula (see sec-
tion 7.1.1)
Circularity the normalized ratio of the area by the square of the perimeter: 4π·A/p2 . Values
superior to 1 may appear due to discretization effect.
Equivalent Ellipse the position, size and orientation of the equivalent ellipse (see section 7.1.2)
Ellipse_Elong. the elongation of the equivalent ellipse (ratio of axis lengths)
Convexity the convexity, as defined in section 7.1.6.2
Max. Feret Diameter the value of the largest Feret diameter
Oriented Box the position, size and orientation of the oriented box with smallest width
Oriented Box Elong. the elongation of the bounding box (ratio of length over width)
Geodesic Diameter the length of the geodesic diameter
Tortuosity the tortuosity, as the ratio of geodesic diameter over largest Feret diameter
Max inscribed disc the position and the radius of the largest circle that can be inscribed
within the region
Geodesic Elong. the ratio of geodesic diameter over diameter of inscribed circle.
The first columns of the resulting ResultsTable contain the label of the regions, to facilitate
their identification.
7.1.7.2. Bounding Box

The Bounding Box plugin computes the minimum and maximum x and y coordinates of
each region within a binary or label image. The result of the plugin comprises the following
features:
Label the label of the region measured on the current line (equal to 255 for binary
images)
Box2D.XMin the minimum x coordinate
Box2D.XMax the maximum x coordinate
Box2D.YMin the minimum y coordinate
Box2D.YMax the maximum y coordinate
It is also possible to draw the resulting bounding box(es) onto another image (Fig. 7.9-a).

7.1 Region Analysis
(a) Bounding Box (b) Equivalent Ellipse (c) Feret Diameter
Figure 7.9.: Examples of graphical outputs of region analysis plugins. (a) Bounding box. (b)
Equivalent ellipse. (c) Max Feret diameter.
7.1.7.3. Equivalent Ellipse

The Equivalent Ellipse plugin computes the equivalent ellipse for each region within a
binary or label image. The result of the plugin comprises the following features:
images)
Ellipse.Center.X the x coordinate of ellipse center
Ellipse.Center.X the y coordinate of ellipse center
Ellipse.Radius1 the length of the largest semi-axis
Ellipse.Radius2 the length of the smallest semi-axis
Ellipse.Orientation the orientation of the ellipse, in degrees from the horizontal.
It is also possible to draw the resulting ellipse(s) onto another image (Fig. 7.9-b).
7.1.7.4. Max Feret Diameter

The Max Feret Diameter plugin computes the length of the largest Feret diameter of each
region within a binary or label image, as well as the position of the extreme points. The
result of the plugin comprises the following features:
Label the label of the region measured on the current line (equal to 255 for bionary
images)
Diameter the length of the Feret diameter
Orientation the orientation of the diameter, in degrees from the horizontal

7.1 Region Analysis
P1.X the x-coordinate of the first extremity
P1.Y the y-coordinate of the first extremity
P2.X the x-coordinate of the second extremity
P2.Y the y-coordinate of the second extremity
It is also possible to draw the resulting diameter as a line segment onto another image
(Fig. 7.9-c).
7.1.7.5. Oriented Box

The Oriented Box plugin computes the minimal width oriented bounding box of each region
within a binary or label image. The result of the plugin comprises the following features:
images)
OBox.Center.X the x coordinate of oriented box center
OBox.Center.X the y coordinate of oriented center
OBox.Length the length of the oriented box
OBox.Width the width of the oriented box
OBox.Orientation the orientation of the box, in degrees from the horizontal.
It is also possible to draw the resulting oriented boxe(s) onto another image (Fig. 7.10-a).
(a) Oriented box (b) Geodesic diameter (c) Inscribed circle
Figure 7.10.: Examples of graphical output of region analysis plugins. (a) Oriented bounding
box. (b) Largest geodesic diameter. (c) Largest inscribed circle.

7.1 Region Analysis
7.1.7.6. Geodesic Diameter

The Geodesic Diameter plugin computes several geodesic measures for each particle in a
label image. The result of the plugin comprises the following features:
images)
Geod. Diam. the value of the geodesic diameter.
Radius the radius of the largest inscribed circle, which is computed during the algo-
rithm.
Geod. Elong. the ratio of geodesic diameter over the diameter of the largest inscribed circle.
The values range from 1 for nearly round particles and increases for elongated
particles.
xi, yi coordinates of the largest inscribed circle.
x1, y1 coordinates of one of the geodesic extremities of the particle.
x2, y2 coordinates of another geodesic extremity of the particle.
It is also possible to draw the resulting geodesic paths onto another image (Fig. 7.10-b).
7.1.7.7. Largest inscribed circle

The largest inscribed circle plugin computes for each label the largest disk that can be
enclosed within the corresponding particle. The plugin opens a dialog that allows to choose
the label image to characterize, the choice of the method for computing distance, and even-
tually the image on which overlaid circles can be drawn. The output of the plugin includes
the following information:
images)
inscrCircle.Center.X the x-coordinate of the inscribed circle.
inscrCircle.Center.Y the y-coordinate of the inscribed circle.
inscrCircle.Radius the radius of the inscribed circle.
It is also possible to draw the resulting circle(s) onto another image (Fig. 7.10-c).

7.2 Region Analysis 3D
7.2. Region Analysis 3D

Region analysis may also be performed on 3D images. As for 2D images, input images may
be either binary (corresponding to a single region), or contain the labels of the different
regions to analyze.
7.2.1. Intrinsic volumes

For 3D particles, intrinsic volumes correspond to the volume, the surface area, the mean
breadth and the 3D Euler number. While the volume and the surface area are rather
common, the latter two are less intuitive. Both the mean breadth and the 3D Euler number
can be related to the curvatures that can be measured on smooth surfaces.
7.2.1.1. Volume
The volume and the surface area of a set X with smooth surface ∂ X can be defined using
integrals over the set, or over its boundary:
Z
V (X ) = dx (7.8)
X
The measurement of volume of 3D region is as straightforward as in 2D: it consists in

counting the voxels that belong to the 3D region, multiplied by the volume of an individual
voxel.
7.2.1.2. Surface area

The surface area of a set X with smooth surface ∂ X can be defined using integral over its
boundary:
Z
S(X ) = dx (7.9)
∂X
The measurement of surface area from 3D binary images follows the same principle
as the estimation of perimeter. One considers the intersections of the region with straight
lines, and averages over all possible directions (Lang et al., 2001; Legland et al., 2007). The
number of directions is typically chosen equal to 3 (the three main axes in image), or 13 (by
considering also diagonals). As for perimeter estimation, surface area estimation is usually
biased, but is usually more precise than measuring the surface area of the polygonal mesh
reconstructed from binary images (Lehmann & Legland, 2012).

7.2.1.3. Mean breadth

The mean breadth, also known as the mean width or the mean diameter, is a feature that
quantifies the size of a set as an average diameter. In the case of a convex set, the average
breadth corresponds to the average of the caliper diameter (or Feret diameter) measured
over all directions (Fig. 7.11-a).
(a) Convex set. (b) Non-Convex Set
Figure 7.11.: Mean breadth of a convex (a) and of a non convex (b) set. In the case of a
non-convex set, the concavities are taken into account for measuring the diameters.
In the case of a non-convex set, the concavities are taken into account for measuring the
diameters, resulting in “total projected diameter” (Fig. 7.11-b). For 3D sets, the mea-
surements of total projected diameter consists in measuring the (2D) Euler number of the
intersection of the set with planes orthogonal to the direction (Serra, 1982). The average of
total projected diameter over Ω, the set of all directions, results in the mean breadth b̄:
Z
b̄(X ) = Dω (X ) dω (7.10)
Ω
For a three-dimensional set X with smooth boundary ∂ X , an alternative definition of the

mean breadth is based on integral geometry (Serra, 1982; Ohser & Schladitz, 2009). For
each point x of the boundary ∂ X , two curvatures κ1 (x) and κ2 (x) can be defined. The
integral of their average over the surface, known as the integral of the mean curvature, is
proportional to the mean breadth:
κ1 (x) + κ2 (x)
Z
1
b̄(X ) = dx (7.11)
2π ∂ X 2
In practice, the mean breadth may be computed from digital binary images by considering
elementary configurations of 2×2×2 voxels, identifying the contribution of this configuration
to the total mean breadth, and combining with the histogram of binary configuration within
the image (Ohser & Schladitz, 2009; Legland et al., 2007).

7.2.1.4. Euler number

As in 2D, the 3D Euler number also quantifies the topology of a set. It corresponds to the
number of connected components, minus the number of “handles” or “tunnels“ through the
structure, plus the number of bubbles within the particles(Serra, 1982; Ohser & Schladitz,
2009), see Figure 7.12.
Figure 7.12.: Euler Number of a 3D particle. The Euler number equals -1, corresponding to the
subtraction of 1 connected components minus two handles.
For a set X with smooth boundary ∂ X , the 3D Euler number is proportionnal to the integral
of the gaussian curvature, corresponding to the product of the curvatures:
Z
1
χ(X ) = κ1 (x) · κ2 (x)d x (7.12)
4π ∂ X
As for 2D images, the measurement of Euler number from 3D images also relies on the
choice of a connectivity between adjacent voxels. In 3D, the 6-connectivity considers the
neighbors in the three main directions within the image, whereas the 26 connectivity also
considers the diagonals. Other connectivities have been proposed but are not implemented
in MorphoLibJ (Ohser & Schladitz, 2009). Note that depending on the choice of connec-
tivity, very different results may be obtained. Such differences may result from the small
irregularities arising in images after segmentation. A typical work-around is to regularize
the binary or label image, for example by applying morphological opening and / or closing.

7.2.2. Geometric moments and equivalent Ellipsoid

The mathematical definition of the geometric moments for 3D regions is similar to the 2D
case:
Z Z Z
m pqr (X ) = I X (x, y, z)x p y q z r · d x · d y · dz (7.13)
The first moment m000 (X ) corresponds to the volume of the particle. The normalization of
the first-order moments by the volume leads to the 3D centroid of the particle. The second-
order moments can be used to compute an equivalent ellipsoid, defined as the ellipsoid with
the same moments up to the second order as the region of interest (see appendix B).
Equivalent Ellipsoid plugin

Returns the parameters of the equivalent ellipsoid into a ResultsTable with the following
columns:
Ellipsoid.Center.X the x-coordinate of the center of gravity
Ellipsoid.Center.Y the y-coordinate of the center of gravity
Ellipsoid.Center.Z the z-coordinate of the center of gravity
Ellipsoid.Radius1 the length of the largest semi-axis
Ellipsoid.Radius2 the length of the medium semi-axis
Ellipsoid.Radius3 the length of the smallest semi-axis
Ellipsoid.Phi the azimut of the projection of the major axis on the XY plane, in degrees
(“yaw”)
Ellipsoid.Theta the elevation of the major axis on the XY plane, in degrees (“pitch”)
Ellipsoid.Psi the rotation of the ellipsoid around the main axis (“roll”), in degrees
The three angles correspond to a succession of three rotations (see Figure 7.13):
1. a rotation R x (ψ) about the x-axis by ψ degrees (positive when the y-axis moves to-
wards the z-axis)
2. a rotation R y (θ ) about the y-axis by θ degrees (positive when the z-axis moves to-
wards the x-axis)
3. a rotation Rz (ϕ) about the z-axis by ϕ degrees (positive when the x-axis moves to-
wards the y-axis)
The global rotation matrix is assimilated to the product Rz (ϕ) · R y (θ ) · R x (ψ).

Figure 7.13.: Definition of angles for representing the orientation of equivalent ellipsoid.
7.2.3. 3D Shape indices

In 3D, shape indices can be also defined to quantify the shape of the regions independently
of their size.
Sphericity
The sphericity index is the generalisation of the circularity to 3D. It can be defined as the
ratio of the squared volume over the cube of the surface area, normalized such that the value
for a ball equals one:
V2
sphericity = 36π (7.14)
S3
Other shape factors may be obtained by computing normalized ratios of volume with mean
breadth, or surface with mean breadth. Their interpretation is however often not obvious.
Elongation
It is also possible to compute elongation factors from ratio of the radiuses if the equivalent
ellipsoid. With three radius values, three ratios can be computed. In MorphoLibJ, the largest
radius is used as numerator, resulting in the following ratios:
r1 r r1
e12 = , e23 = 2 , e13 = , with r1 ≥ r2 ≥ r3
r2 r3 r3

7.3 Intensity measurements
7.2.4. Plugins
Analyze Regions 3D
The plugin “Analyze Regions 3D” gathers most 3D measures implemented within the Mor-
phoLibJ library. It can be found under “Plugins . MorphoLibJ . Analyze . Analyze Regions
3D”. The results are provided in an ImageJ ResultsTable, whose name contains the name of
the original image.
Interface surface area

The Interface Surface Area plugin allows to measure the surface area of the interface be-
tween two labels within a 3D label image. The different options are:
Label 1 the first label
Label 2 the second label
Method the number of directions to use for the computation (either 3 or 13).
One of the two labels can have the value 0. In this case, the interface of the other label with
the background is measured.
If the two regions do not touch anywhere, the resulting value will be 0. The region adja-
cency graph (section 9.4.1) can be used to identify neighbor regions.
7.3. Intensity measurements

Other measurements are provided for pairs of grayscale and label 2D or 3D images (“Plu-
gins . MorphoLibJ . Analyze . Intensity Measurements 2D/3D”). The label image can corre-
spond to a segmented particle, or to a more generic region of interest.
The plugin calculates the mean, standard deviation, maximum, minimum, median, mode,
skewness and kurtosis of the intensity value distribution of each labeled region in the grayscale
image as well as the same statistics of the neighboring (adjacent) regions of each of the la-
bels. If a regions has no adjacent labels, their measurements will appear as NaN (not a
number). The results are displayed in an ImageJ ResultsTable.

7.4 Label Overlap measures
7.4. Label Overlap measures

Given two label images, there are different measures that allow us to evaluate the overlap
agreement (or error) between the labels. Following Tustison & Gee (2009), and given a
source image S and a target image T , this plugin provides the following overlap measure-
ments in two different result tables (one with the total values for all labels and one with
values for individual labels):
• Target Overlap for each individual labeled region r:
|S r ∩ Tr |
T Or =
|Tr |
• Total Overlap (for all regions):

P
r |S r ∩ Tr |
TO = P
r |Tr |
• Jaccard Index or Union Overlap for each individual labeled region r:
|S r ∩ Tr |
UOr =
|S r ∪ Tr |
• Jaccard Index or Union Overlap for all regions:

P
|S r ∩ Tr |
UO = P r
r |S r ∪ Tr |
• Dice Coefficient or Mean Overlap for each individual labeled region r:
|S r ∩ Tr |
M Or = 2
|S r | + |Tr |
• Dice Coefficient or Mean Overlap for all regions:

P
|S r ∩ Tr |
MO = 2P r
r (|S r | + |Tr |)
• Volume Similarity for each individual labeled region r:
|S r | − |Tr |
V Sr = 2
|S r | + |Tr |
• Volume Similarity for all regions:
(|S r | − |Tr |)
P
V S = 2 Pr
r (|S r | + |Tr |)

7.5 Microstructure analysis
• False Negative Error for each individual labeled region r:
|Tr \ S r |
F Nr =
|Tr |
• False Negative Error for all regions:

P
r |Tr \ S r |
FN = P
r |Tr |
• False Positive Error for each individual labeled region r:
|S r \ Tr |
F Pr =
|S r |
• False Positive Error for all regions:

P
r|S r \ T r |
FP = P
r |S r |
7.5. Microstructure analysis

In some cases, the content of the image to analyze is better seen as a representative observa-
tion of a stationnary binary process. This context arise frequently in material sciences or in
the study of porous media such as bone material (Lang et al., 2001; Ohser & Schladitz, 2009;
Doube et al., 2010). An example of such image (obtained from a dairy gel) is represented
in Figure 7.14-a.
(a) Sample image (b) Perimeter density
Figure 7.14.: Analysis of binary microstrucure. (a) Original binary image. (b) Estimation of
perimeter density.
To analyze such images, the notion of particle analysis is not relevant anymore. One
possibility is to adapt the definitions of intrinsic volumes, by considering that the image
corresponds to a sampling window. This induces the two modifications:

7.5 Microstructure analysis
1. the integral over boundaries are restricted to the boundary within the sampling win-
dow (Fig. 7.14-b)
2. the result is normalized by the area (or the volume) of the sampling window.
For the planar case, one obtains the following definitions:

R
dx
AA(X ) = X ∩W (7.15)
A(W )
R
dx
PA(X ) = ∂ X ∩W (7.16)
A(W )
R
κ(x)d x
χA(X ) = ∂ X ∩W (7.17)
A(W )
The plugins are available in “Plugins . MorphoLibJ . Analyze . Microstructure Analysis”

and “Plugins . MorphoLibJ . Analyze . Microstructure 3D”.

8. Binary images
Binary and label images are a convenient way of representing the result of segmentation
algorithms. The MorphoLibJ library provides several plugins for the processing and the
management of binary and label images.
Contents
8.1. Distance transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
8.1.1. Distance transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
8.1.2. Chamfer distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
8.1.3. Geodesic distance transform . . . . . . . . . . . . . . . . . . . . . . 75
8.2. Connected component labeling . . . . . . . . . . . . . . . . . . . . . . . . 76
8.3. Binary images processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.3.1. Convexification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.3.2. Selection of binary regions . . . . . . . . . . . . . . . . . . . . . . . 77
69
8.1 Distance transforms
8.1. Distance transforms

8.1.1. Distance transform
When analyzing images, it is often necessary to compute distances to a particular structure
or position. A convenient operator for binary images is the distance transform. Its princi-
ple is to compute, for each foreground pixel, the distance to the nearest background pixel
(Fig. 8.1). The result is commonly referred to as Distance Map.
Figure 8.1.: Binary image, and result of computation of the distance transform.
8.1.2. Chamfer distances

Several methods exist for computing distance maps. The MorphoLibJ library implements
distance transforms based on chamfer masks, that are simpler to compute than exact Eu-
clidean distance (?Borgefors, 1984, 1986). The principle is to update the distance associated
to each pixel by considering the distances within a small neighborhood, and adding a weight
depending on the relative location of the neighbor (Figure 8.2).
Figure 8.2.: Computation of distance maps using chamfer masks. The distance associated to
each pixel is updated according to the distances associated to the neighbor pixels. Left: 3x3
neighborhood. Right: 5x5 neighborhood.

In practice, using integer weights may result in faster computations. This also allows to
store the resulting distance map within 16-bits grayscale images, resulting in more efficient
memory usage than floating point computation.
Several typical choices for chamfer weights are illustrated on Figure 8.3:
• The simplest one is the “Chessboard” mask, that considers distance equal to 1 for
both orthogonal and diagonal neighbors (Fig. 8.3-a). The propagation of distances
from a center pixel follow a square pattern.
• The City-block or Manhattan mask uses a weight equal to 2 for diagonal neighbord,
resulting in diamond-like propagation of distance (Fig. 8.3-b).
• The “Borgefors” distance uses the weight 3 for orthogonal pixels and 4 for diagonal
pixels (Borgefors, 1986). It was claimed to be best integer approximation when con-
sidering the 3-by-3 neighborhood (Fig. 8.3-c).
• The “Chess-knight” distance (Das & Chatterji, 1988) also takes into account the pixels
located at a distance from (±1, ±2) in any direction (Fig. 8.3-d). It is usually the best
choice, as it considers a larger neighborhood.
• The MorphoLibJ library also provides chamfer masks with four weights, as proposed
by Verwer et al. (1989). The fourth weight corresponds to pixels located at a distance
of (±1, ±3) from the reference pixel, in any direction.
(a) Chessboard (1,1) (b) City-block (1,2) (c) Borgefors (3,4) (d) Chess-knight (5,7,11)
Figure 8.3.: Several examples of distance maps computed with different chamfer masks. For
each image, the distance to the lower-left pixel is represented both with numeric value and with
color code. Numbers in brackets correspond to the weights associated with orthogonal and
diagonal directions. For chess-knight weights, the weight 11 corresponds to a shift of (±1, ±2)
pixels in any direction.
As chamfer masks are only an approximation of the real Euclidean distance, some differ-
ences are expected compared to the actual Euclidean distance map. Using larger weights
reduces the relative error, but the largest possible distance (that depends on the bit-depth of
the output image) may be reached faster, leading to “saturation” of the distance map.

8.1.2.1. Chamfer distance computation

A great advantage of chamfer masks is that they allow fast computation of distance maps by
dividing the process into two passes. During the first pass (in “forward” direction), pixels
are iterated from left to right and from top to bottom. The distance associated to each pixel
is updated based on pixels already updated, located upper and on the left (Figure 8.4-a).
During the second pass, pixels are iterated “backward” (from right to left and from bottom
to top), and the distance map is updated based on the pixels located bottom and on the right
(Figure 8.4-b).
(a) Forward mask (b) Backward mask
Figure 8.4.: Implementation of two-passes chamfer distances. The 5x5 chamfer mask is decom-
posed into two “half-masks”, one for each pass.
8.1.2.2. 3D chamfer masks

For 3D images, chamfer distance maps can be computed using the same principle. Three
different chamfer weights can be associated to the 6 orthogonal neighbors, to the 12 neighbor
voxel sharing an edge with reference voxel, and to the 8 neighbor voxel sharing only a
corner with reference voxel. The results can be represented via the discrete ball, obtained
by applying a threshold on the distance map computed from a single voxel located in the
center of the image (Fig. 8.5).
(a) Chessboard (1,1,1) (b) City-block (1,2,3) (c) Borgefors (3,4,5)
Figure 8.5.: Discrete 3D balls resulting from different chamfer distances. The distance map is
computed to the center voxel, then the result is thresholded and represented in 3D.
Several 3D chamfer distances can be used:

• The (3D) chessboard distance uses the same weight for all the 26 neighbors. This
results in distance propagation in a cubic shape pattern (Fig. 8.5-a).
• The (3D) Manhattan distance considers weights triplet equal to (1, 2, 3), resulting in
octagon-like distance propagation (Fig. 8.5-b).
• The Borgefors weights (3,4,5) usually give a good approximation of the Euclidean
distance when considering only the 3-by-3-by-3 neighborhood (Fig. 8.5-c).
(a) Svensson (3,4,5,7) (b) Five weights (c) Six weights
Figure 8.6.: Enhancement of 3D distance maps when using larger chamfer masks. Masks are
based on (a) four, (b) five, or (c) six weights. Using All images represent the threshold of the
distance 100 from image center.
Larger neighborhood may be considered, resulting in a better approximation (Fig. 8.6):
• Svensson and Borgefors found that adding a fourth weight for voxels shifted by (±1, ±1, ±2)
in any direction results in a substantial error reduction while preserving a low com-
plexity of the algorithm (Svensson & Borgefors, 2002b,a).
• Starting from version 1.5, MorphoLibJ allows using up to six weights (corresponding
to the complete 5–by–5-by–5 neighborhood).
As for the 2D case, the choice of the chamfer distance is compromise between the accuracy
(larger neighborhood and weights are usually more accurate) and the largest distance that
can be represented.
8.1.2.3. Case of label images

The distance transform algorithm may also be applied to label images. However, as regions
corresponding to each label may be continuous, the algorithm need to be adapted.
For label image, distances are computed between each foreground (greater than zero)
pixel and the nearest pixel with a different value : either 0 (the background) or another
label. Distances are propagated as for binary images, using a pair of forward and backward
passes. An example of result is provided on Figure 8.7.

Figure 8.7.: An example of a label image containing contiguous regions, and result of the dis-
tance transform applied on the label image.
8.1.2.4. Plugin usage

The MorphoLibJ library provides two plugins for computing chamfer distance maps for 2D
or 3D binary images:
Chamfer Distance Map

computes a chamfer distance map from a binary or label image, between each foreground
pixel to the nearest pixel with a different value (either background or another label).
Chamfer Distance Map 3D

computes a chamfer distance map from a 3D binary image between each foreground voxel
to the nearest voxel with a different value (either background or another label). Input image
is assumed to have cubic voxels.
For each plugin, a dialog opens and allows to select the following parameters:
Distances the weights of the chamfer mask used for computing the distance map
Output Type specify if the result should be stored in 16-bits image (uses less memory), or
32-bits image (larger distances can be computed).
Normalize specify if the resulting map should be divided by the weight associated with
orthogonal pixels or voxels. This may be useful for masks with large weights to
have resulting distances comparable to Euclidean ones.
Preview (2D only) enable the preview of the result

8.1.3. Geodesic distance transform

In some cases it may be useful to restrict the propagation of distances to a specific region or
mask. For example, one may be interested in the distance between two points in a vascula-
ture network, while staying within the network. The geodesic distance transform consists
in computing the distance from a given binary marker, while constraining the propagation of
the distance within a binary mask. An illustration is given in Figure 8.8, and a 3D example
is shown in Figure 8.9.
(a) Original image (b) Geodesic distance map
Figure 8.8.: Computation of the geodesic distance map on a binary image from the DRIVE
database (Staal et al., 2004). (a) Original image with marker superimposed in red. (b) Color
representation of the geodesic distance map: hot colors correspond to large distances, cold colors
correspond to small distances.
Geodesic distance maps are computed by propagating chamfer weights, as for the compu-
tation of distance maps. As chamfer weights are only an approximation of the real Euclidean
distance, some differences are expected compared to the actual geodesic distance.
Figure 8.9.: Computation of the geodesic distance map on a 3D binary image.

8.2 Connected component labeling
Geodesic Distance Map

Computes the geodesic distance between each foreground pixel of a binary mask image to
the closest pixel of a marker image, while staying within the particle represented by the mask
image (See Figure 8.8).
Interactive Geodesic Distance Map

Computes the geodesic distance between each foreground pixel of a the currently selected
image (considered the mask image) to the closest pixel of a marker image defined by the
user ROIs, while staying within the particle represented by the mask image.
Geodesic Distance Map 3D

Computes the 3D geodesic distance between each foreground pixel of a binary mask image
to the closest pixel of a marker image, while staying within the particle represented by the
mask image (See Figure 8.9). Binary image is assumed to have cubic voxels.
8.2. Connected component labeling

The different structures within a binary images can be individualised by using a connected
component labeling algorithm. The result is an image containing integer label, each label
corresponding to a set of connected pixels or voxels (Figure 8.10). Label images can be
represented with colored LUT to help distinguish adjacent labels (see section 9.1).
Figure 8.10.: Binary image, and result of connected components labeling.
The result of a connected component labeling depends on the chosen connectivity, that
corresponds to the convention used to decide wether two pixels or voxels are connected or
not (Figure 8.11). See also Section 3.3.
For planar images, a typical choice is the 4-connectivity, that considers only the orthog-
onal neighbors of a given pixel (Figure 8.11-b). When two components touch only via a
corner, they are not considered as the same one. The 8-connectivity is an alternative that
also considers the diagonal pixels as neighbors (Figure 8.11-c).
For 3D images, the 6-connectivity considers only orthogonal neighbors in the three main
directions, whereas the 26-connectivity considers all the neighbors of a given voxel in its
3-by-3-by-3 surrounding. The connectivity is also considered for computing morphological

8.3 Binary images processing
(a) Binary image (b) 4-connectivity (c) 8-connectivity
Figure 8.11.: Impact of the connectivity on the result of connected component labeling. (a)
Original binary images. (b) Connected component labeling using the 4-connectivity. (c) Con-
nected component labeling using the 8-connectivity.
reconstructions (section 5.1), or for computing the Euler number of binary or label images
(section 7.1.1.3).
The number of labels that can be represented within a label map depends on the image
type: 255 for byte images, 65535 for short images, and around 16 millions for 32-bit floating
point images (only the 24 bits of the mantissa are used for label representation, leading to
a maximum number of labels equal to 224 ).
Connected Components Labeling

transforms the binary image into a label image by assigning a specific number (label) to each
connected component
8.3. Binary images processing

8.3.1. Convexification
When processing a binary region, it can sometimes be useful to consider the binary region
corresponding to the convex hull of the original region (Fig. 8.12). For example, the con-
vexity measure (section 7.1.6.2) is based on the convex hull. The Convexify plugin allows
to generate the smallest convex region containing a binary region.
Convexify
for a binary image, generate the smallest convex region containing a region in the image.
8.3.2. Selection of binary regions

The MorphoLibJ library offers several tools for automatically selecting binary regions based
on size or position criteria. Some of them are illustrated on Figure 8.13. As the algorithms
usually requires connected component labeling, it is often more convenient to use corre-
sponding methods for label images (Section 9.3.1).
Keep Largest Region

Identifies the largest connected component, and remove all other regions.

8.3 Binary images processing
(a) Binary image. (b) Convexification. (c) Original image over its convex
hull.
Figure 8.12.: Convexification of a binary image containing several disjoint components.
Figure 8.13.: Utilities for binary images. From left to right: original image, keep largest region,
remove largest region, apply size opening for keeping only regions with at least 150 pixels.
Remove Largest Region

Identifies the largest connected component, and remove it from image.
Size Opening
computes the size (area in 2D, volume in 3D) of each connected component, and remove all
particles whose size is below the value specified by the user. Algorithms work for both 2D
or 3D images. Default connectivity 4 (resp. 6) is used for 2D (resp. 3D) images.

9. Label images
When several structures or components are present within an image, it may be convenient
to work with label images, or label maps. Within a label map, each pixel or voxel of a
label image corresponds to the integer index of the particle it belongs to. By convention, the
value 0 corresponds to the background. Label images are typically computed from binary
images by connected component algorithms (section 8.2). Segmentation algorithms such as
the Watershed may also return their result as a label image.
Contents
9.1. Visualization of label images . . . . . . . . . . . . . . . . . . . . . . . . . 80
9.1.1. Color maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
9.1.2. Plugins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
9.2. Visualization of region features . . . . . . . . . . . . . . . . . . . . . . . 82
9.3. Label image processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
9.3.1. Region and label selection . . . . . . . . . . . . . . . . . . . . . . . 83
9.3.2. Morphological filtering . . . . . . . . . . . . . . . . . . . . . . . . . 84
9.3.3. Edition of label maps . . . . . . . . . . . . . . . . . . . . . . . . . . 85
9.4. Region adjacencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
9.4.1. Region Adjacency Graph . . . . . . . . . . . . . . . . . . . . . . . . 86
9.4.2. Adjacent region boundaries . . . . . . . . . . . . . . . . . . . . . . . 87
9.4.3. Region adjacencies . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
9.5. Label Edition plugin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
79
9.1 Visualization of label images
9.1. Visualization of label images

Label images can be represented either using shades of gray, or using color maps to better
distinguish the different regions. It may also be convenient to overlay the region label on
the image to associate it to values within a table.
(a) Using Colormap (b) Draw Labels as Overlay (c) Overlay Label Image
Figure 9.1.: Various visualization modes of label images. (a) Display regions using appropriate
colormap. (b) The label associated to each region is overlaid on another image. (c) Label map
or binary images can be overlaid on another grayscale image using user-defined opacity.
9.1.1. Color maps

When the number of labels within a label image is large, it may be difficult to differentiate
regions whose labels are close. The Figure 9.2 presents various combinations of colormaps
and background colors.
(a) Glasbey, Black (b) GlasbeyBright, White (c) GlasbeyDark, White (d) Golden Angle, Black
Figure 9.2.: Several choices of colormap for visualization of a label image.
The “Glasbey” colormap, included within ImageJ, generates colors that are perceptually
contrasted (Figure 9.2-a). It however generates black and white colors, making them difficult
to distinguish from background (on the example from Figure 9.2-a, the region in the top-
right corner is not visible anymore). Therefore, the “Glasbey Bright” and “Glasbey Dark”
colormaps were made available, to provide good contrast both between labels and with

9.1 Visualization of label images
background (Glasbey et al., 2007; Kovesi, 2015). The choice of these colormaps was inspired
by the colorcet library1 . The Golden Angle colormap is an alternative to the Glasbey family,
that provides more saturated colors (Figure 9.2-d). The contrast between colors is less than
for Glasbey colormaps, making adjacent region potentially more difficult to distinguish.
9.1.2. Plugins
Several plugins allow to control the appearance of label images. It is possible to choose
a given color map, or to transform a label image into a color image. In both cases, the
background color can be specified, and the color order can be shuffled to facilitate the dis-
crimination of neighbor regions with similar labels.
Set Label Map

allows to choose the color map used to display a label image (Fig. 9.1-a). In particular,
shuffling the color map and/or choosing a specific color for background allows better visu-
alization than with only gray levels. Note that when the number of regions is large, regions
with small labels will be associated to the same color as background, and will therefore not
be visible!
Label To RGB
converts a label image to a true RGB image. Similar to ImageJ native conversion, but this
plugin avoids confusion between background pixels and labels with small values when the
number of labels is large.
Draw Labels As Overlay

Choose another image, and display the label associated to each region as a text overlay
(Fig. 9.1-b). The position of the overlay is the centroid of each region, and a small shift can
be specified.
Overlay Binary Or Label Image

Opens a dialog to choose a reference/background image, and an image to overlay, that can
be either a binary image or a label map (Fig. 9.1-c). User can specifiy the opacity of the
overlay, and generate a new Color image for the result.
1
https://colorcet.holoviz.org/index.html

9.2 Visualization of region features
9.2. Visualization of region features

When images contain many labels, it is may be difficult to interpret together the localisation
and the features describing the regions. The “Assign Measure To Label” and “Draw Label
Values” plugins provide some help to facilitate this exploration.
Figure 9.3.: Various possibilities to graphically visualize region feature. On the left, the table
used for display. Center, overlay of the elongation feature using the “Draw Label Values” plugin.
On the right, color representation of the elongation using a color code: between dark purple
(circular, compact) to white (very elongated).
Draw Label Values

draws the value of a feature measured on each label of an image using two coordinates
(bounding box, centroid...) also stored in a ResultsTable. An example is shown on Figure 9.3.
Assign Measure To Label

combines a label image with a results table, and creates a new 32-bits image for which
each pixel/voxel is assigned the measurement value corresponding to the label it belongs to
(Figure 9.3). Background pixels/voxels are associated to NaN (Not-A-Number) value.

9.3 Label image processing
9.3. Label image processing

9.3.1. Region and label selection
It is often convenient to select regions in a label image based on their size or their location
(e.g. removing the regions touching the border). The MorphoLibJ library offers several util-
ity tools for automatically selecting or removing specific regions, or to edit the label images.
Some of them are illustrated on Figure 9.4. The Label Edition Plugin (Section 9.5) provides
some of these operators integrated into a graphical user interface.
Figure 9.4.: Utilities for label images. From left to right: original label image, remove border
labels, remove largest region, apply size opening for keeping only regions with at least 150 pixels.
Remove Border Labels

Similar to the “kill borders” function, but operates faster as no morphological reconstruction
is required.
Select Label(s)
Enters a set of labels, and creates a new label image containing only the selected labels.
Keep / Remove Largest Label

Identifies the largest label, and keep it or remove it
Label Size Filtering

Filters the labels according to their size (area in 2D, volume in 3D), by specifying a compar-
ison operator and a threshold value.

9.3.2. Morphological filtering

Several morpological filters have been adapted to process label maps.
Label Morphological Filtering

Applies morphological filtering operators on the label map, by choosing the type of operation
(dilation, erosion, opening or closing), and the radius of the operation. The principle is the
same as morphological filtering on grayscale images, but using label values instead of gray
levels, and avoiding “collisions” between labels:
• when applying a dilation, only background pixels or voxels are updated, according
to the closest label within the neighborhood (Figure 9.5-b). When two regions are
adjacent without background pixel between them, the application of dilation do not
transform nor move their boundary.
• when applying an erosion, label pixels or voxels are removed depending on the labels
within the neighborhood and on the “any Label” option:
– if true, current label is removed if any other label (background or another region)
is contained within the neighborhood
– if false, current label is removed if and only if a background pixel or voxel is
contained within the neighborhood (Figure 9.5-c).
When the “Any Label” option is set to false, performing a morphological closing results in a
label map with more regular regions, with less gaps between regions, and preserved bound-
aries between adjacent regions (Figure 9.5-d).
The computation of the neighborhood, and of the closest label, is based on chamfer masks
(see Section 8.1.1).
(a) Label image (b) Dilation (c) Erosion (d) Closing
Figure 9.5.: Morphological filtering of label images. Morphological dilation (b) and erosion
(c) take care of adjacent regions.
Fill Label Holes

Fills the holes within regions. Only the holes that are adjacent to a single region are filled,
using a fill corresponding to the label of the adjacent region.

9.3.3. Edition of label maps

Several plugins allow for basic edition of filtering of label images. They result in another
label image.
Crop Label
creates a new binary image containing only the label specified by the user. The size of the
new image is fitted to the region.
Replace Value
replaces the value of a region by another value. Can be used to “clear” a label, by replacing
its value by 0, or to merge to adjacent regions, by replacing the value of a label by the value
of its neighbor.
Remap Labels
re-computes the label of each region such that the largest label is equal to the number of
labels (Figure 9.6). This tool can be useful to remove “empty labels” that could occur after
removal of one or several labels (merge labels, remove labels...).
(a) Label image (b) Original histogram (c) Remapped histogram
Figure 9.6.: Illustration of the “Remap Labels” plugin. (a) Original label image. (b) Histogram
of original label image, showing missing labels. (c) Histogram of the remapped label image,
without missing labels.
The Figure 9.6 presents a use case of the Remap Labels plugin. The left image is a label
image in which small labels as well as labels touching image borders have been removed
(Fig. 9.6-a). The histogram of the image presents some “holes”, corresponding to the labels
that have been removed (Fig. 9.6-b). The application of the plugin results in a similar images,
without missing labels (Fig. 9.6-c). The number of labels in the image corresponds to the
maximal value in the image (in this case, 71).

9.4 Region adjacencies
9.4. Region adjacencies

It may sometimes be useful to consider the adjacency relationships between regions. This
can be useful for post-processing of segmentation results (using e.g. graph processing algo-
rithms), or for exploring collections of cells within cellular tissues (Florindo et al., 2016).
Several plugins are available within the library. The definition of adjacency is not totally
consistent between the plugins, and results may greatly differ when applied on the same
input label image.
9.4.1. Region Adjacency Graph

The region adjacency graph plugin gives access to the neighborhood relationship between
adjacent regions (Figure 9.7).
Region Adjacency Graph

The plugin works for both 2D and 3D images, and requires a label image as input. The
output of the plugin is a results table with as many rows as the number of pairs of adjacent
regions, containing the labels of the two adjacent regions.
Figure 9.7.: Computation of the Region Adjacency Graph (RAG) on a microscopy image of plant
tissue. Left: original image. Middle: result of watershed segmentation. Right: overlay of edges
representing adjacent regions.
In the current implementation, MorphoLibJ considers that two regions are adjacent if they
are separated by 0 or 1 pixel (or voxel) in one of the two (or three) main directions.

9.4.2. Adjacent region boundaries

It may sometimes be useful to visualize the boundaries between adjacent regions (Fig. 9.8).
Figure 9.8.: Example use of the plugin “Label Region Boundaries” plugin. Left: a label image
containing regions with various adjacencies. Right: a new label map was created, where each
label corresponds to the boundary between two or more regions within the original image.
The “Label Boundaries” plugin generates a binary image of the boundaries between re-
gions. The “Region Boundaries Labeling” considers similar boundaries, but in addition it
associates a label to each boundary.
Label Boundaries
Creates a new binary image containing value 255 for pixels/voxels having a neighbor with a
different value. The background is taken into account for computing boundaries. Only the
bottom-left neighborhood is checked during image iteration, to generate a thinner boundary
image.
Region Boundaries Labeling

Computes a new label map, where each label corresponds to the boundary between two
(or more) regions. The background is also considered as a region. The list of identified
boundaries together with their label and bounding regions is displayed in the log window.
Most boundaries are adjacent to two regions, but the junctions between boundaries may be
adjacent to three or more regions. Label map of boundaries is represented using floating
point numbers, leading to a maximum number of 223 (around 8.3 millions) boundaries. The
number of boundaries is usually much larger than the number of regions, and the maximum
number of label may be reached quickly.
See also the Interface Surface Area plugin (section 7.2.4), that measures the surface area
between adjacent 3D regions.

9.4.3. Region adjacencies

Another usage of adjacencies is to select regions around a given one.
Select Region Neighbors

Extracts the neighbor regions within a label image from a region label, or a list of region
labels (See Fig. 9.9).
Figure 9.9.: An example of a label image containing contiguous regions, and result of the selec-
tion of the neighbors of two regions (keeping the original regions).

9.5 Label Edition plugin
9.5. Label Edition plugin

To ease the processing of label images, MorphoLibJ provides the Label Edition plugin (avail-
able under “Plugins . MorphoLibJ . Labels . Label Edition”). This plugin contains a graphical
user interface (GUI) where the users can perform the following set of editing tasks (Fig-
ure 9.10):
• Manually merge labels after their selection using the point selection tool (in 2D and
3D).
• Apply morphological erosion, dilation, opening and closing with a square/cube of ra-
dius 1 as structuring element.
• Remove labels by area or volume (size opening operation), largest size, manual selec-
tion or border location.
All operations are performed “in place”, i.e., the input image gets directly modified. However,
the initial status of the label image can be recovered by clicking on “Reset”.
Figure 9.10.: Label Edition plugin. From left to right and from top to bottom: overview of the
plugin GUI on a 2D label image, result of label merging by manual selection, removal of selected
labels, label erosion and removal of largest label (in this case, the largest label corresponds to
the background).

10. Library interoperability
A key design concept was the modularity of the implementation to facilitate its reusability.
Three layers with different programming abstraction can be identified:
• For final users, plugins provide graphical display and intuitive tuning of parameters.
Such plugins can easily be incorporated into a macro:
1 // C a l l s t h e R e g i o n a l Min/Max p l u g i n on c u r r e n t ImagePlus i n s t a n c e
2 run ( " R e g i o n a l Min & Max " , " o p e r a t i o n =[ R e g i o n a l Maxima ] c o n n e c t i v i t y =4" ) ;
• For plugins developers, operators are available through collections of static methods,
making it possible to apply most operations with a single line of code. Example:
1 // Computes r e g i o n a l maxima u s i n g t h e 4− c o n n e c t i v i t y
2 I m a g e P r o c e s s o r maxima = MinimaAndMaxima . regionalMaxima ( image , 4) ;
• For core developers, algorithms are implementations of abstract interfaces, making

it possible to choose or develop the most appropriate one, and to monitor execution
events.
1 // choose and s e t u p t h e a p p r o p r i a t e a l g o r i t h m
2 RegionalExtremaAlgo a l g o = new RegionalExtremaByFlooding ( ) ;
3 a l g o . setExtremaType ( ExtremaType . MAXIMA) ;
4 algo . s e t C o n n e c t i v i t y (4) ;
5 // add a l g o r i t h m m o n i t o r i n g
6 a l g o . a d d A l g o L i s t e n e r (new D e f a u l t A l g o L i s t e n e r ( ) ) ;
7 // compute r e s u l t on a g i v e n I m a g e P r o c e s s o r
8 I m a g e P r o c e s s o r r e s u l t = a l g o . applyTo ( image ) ;
In total, the library provides nearly two hundred classes and interfaces.
10.1. Library organization

The library follows a logic structure of folders divided by topics aiming at their re-usability
from other plugins or scripts, among others:
inra.ijpb.data contains generic data structures for manipulating 2D or 3D images

inra.ijpb.binary contains the set of utilities for working on binary images (connected com-
ponent labeling, distance transform, geodesic distance transform...)
inra.ijpb.geometry contains utility functions for geometric computing, and several classes
for representing geometric primitives (ellipse, point pair...)
90
10.2 Scripting MorphoLibJ
inra.ijpb.label contains the utilities for label images (cropping, size opening, remove bor-
der labels, etc)
inra.ijpb.measure contains the tools for geometric and gray level characterization of 2D
or 3D images
inra.ijpb.morphology contains the collection of mathematical morphology operators

inra.ijpb.watershed contains the classes implementing the different versions of the water-
shed algorithm
inra.ijpb.plugins contains the set of plugins that is accessible from ImageJ/Fiji Plugins
menu
All major methods have a general class with static methods that allow calling the methods on
2D and 3D images in a transparent way. This modularity permitted to develop other plugins
devoted to the analysis of nucleus images (Poulet et al., 2015), gray level granulometries
(Devaux & Legland, 2014) or the description of binary microstructures (Silva et al., 2015).
10.2. Scripting MorphoLibJ

One advantage of this organization of the library and the use of public static methods is that
it allows very easy and fast prototyping of morphological algorithms and pipelines.
10.2.1. Segmentation pipeline prototype

Let’s see an example in a complete Beanshell script that takes the active 2D or 3D image
and finds a reasonable segmentation combining a set of morphological operations (gradient,
extended minima and watershed):
1 // @ImagePlus imp
2
3 // ImageJ i m p o r t s
4 import i j . I J ;
5 import i j . ImagePlus ;
6 import i j . ImageStack ;
7
8 // MorphoLibJ i m p o r t s
9 import i n r a . i j p b . b i n a r y . BinaryImages ;
10 import i n r a . i j p b . morphology . MinimaAndMaxima3D ;
11 import i n r a . i j p b . morphology . Morphology ;
12 import i n r a . i j p b . morphology . S t r e l 3 D ;
13 import i n r a . i j p b . watershed . Watershed ;
14
15 // c r e a t e s t r u c t u r i n g element ( cube o f r a d i u s 1)
16 s t r e l = S t r e l 3 D . Shape . CUBE . fromRadius ( 1 ) ;
17 // a p p l y m o r p h o l o g i c a l g r a d i e n t t o i n p u t image
18 image = Morphology . g r a d i e n t ( imp . g e t Ima geS tack ( ) , s t r e l ) ;
19 // f i n d r e g i o n a l minima on g r a d i e n t image with dynamic v a l u e o f 30 and 6− c o n n e c t i v i t y
20 regionalMinima = MinimaAndMaxima3D . extendedMinima ( image , 30 , 6 ) ;
21 // impose minima on g r a d i e n t image
22 imposedMinima = MinimaAndMaxima3D . imposeMinima ( image , regionalMinima , 6 ) ;
23 // l a b e l minima u s i n g connected components (32− b i t ou t pu t )

24 labeledMinima = BinaryImages . componentsLabeling ( regionalMinima , 6 , 32 ) ;

25 // a p p l y marker−based watershed u s i n g t h e l a b e l e d minima on t h e minima−imposed
26 // g r a d i e n t image ( t h e " t r u e " v a l u e i n d i c a t e s t h e use o f dams i n t h e o ut pu t )
27 r e s u l t S t a c k = Watershed . computeWatershed ( imposedMinima , labeledMinima , 6 , t r u e ) ;
28
29 // c r e a t e image with watershed r e s u l t
30 r e s u l t I m a g e = new ImagePlus ( " watershed " , r e s u l t S t a c k ) ;
31 // a s s i g n r i g h t c a l i b r a t i o n
32 r e s u l t I m a g e . s e t C a l i b r a t i o n ( imp . g e t C a l i b r a t i o n ( ) ) ;
33 // and show i t
34 r e s u l t I m a g e . show ( ) ;
10.2.2. Label visualization in 3D viewer

Making use of MorphoLibJ’s label methods and the ImageJ 3D Viewer’s visualization tools
it is quite simple to create a script to display each label of an image as 3D surfaces of the
corresponding colors provided by the image look-up table:
1 // @ImagePlus imp
2 import i n r a . i j p b . l a b e l . LabelImages ;
3 import i j 3 d . Image3DUniverse ;
4 import i j 3 d . C o n t e n t C o n s t a n t s ;
5 import org . s c i j a v a . vecmath . C o l o r 3 f ;
6 import i j . I J ;
7 import i s o s u r f a c e . SmoothControl ;
8
9 // s e t t o t r u e t o d i s p l a y messages i n l o g window
10 verbose = f a l s e ;
11
12 // s e t d i s p l a y range t o 0−255 so t h e d i s p l a y e d c o l o r s
13 // c o r r e s p o n d t o t h e LUT v a l u e s
14 imp . s e t D i s p l a y R a n g e ( 0 , 255 ) ; imp . updateAndDraw ( ) ;
15
16 // c a l c u l a t e a r r a y o f a l l l a b e l s i n image
17 l a b e l s = LabelImages . f i n d A l l L a b e l s ( imp ) ;
18 // c r e a t e 3d u n i v e r s e
19 univ = new Image3DUniverse ( ) ;
20 univ . show ( ) ;
21 // read LUT from i n p u t image
22 l u t = imp . g e t L u t s ( ) [ 0 ] ;
23 // add a l l l a b e l s d i f f e r e n t from z e r o ( background )
24 // t o 3d u n i v e r s e
25 f o r ( i =0; i <l a b e l s . l e n g t h ; i++ )
26 {
27 if ( labels [ i ] > 0 )
28 {
29 l a b e l T o K e e p = new i n t [ 1 ] ;
30 labelToKeep [ 0 ] = l a b e l s [ i ] ;
31 i f ( verbose )
32 IJ . log ( " Reconstructing l a b e l " + l a b e l s [ i ] + " . . . " ) ;
33
34 // c r e a t e new image c o n t a i n i n g o n l y t h a t l a b e l
35 l a b e l I m p = LabelImages . k e e p L a b e l s ( imp , l a b e l T o K e e p ) ;
36 // c o n v e r t image t o 8− b i t
37 I J . run ( labelImp , "8− b i t " , " " ) ;
38 // use LUT l a b e l c o l o r
39 c o l o r = new C o l o r 3 f ( new j a v a . awt . C o l o r ( l u t . getRed ( l a b e l s [ i ] ) ,
40 l u t . getGreen ( l a b e l s [ i ] ) ,
41 l u t . getBlue ( l a b e l s [ i ] ) ) ) ;

42 i f ( verbose )
43 I J . l o g ( "RGB( " + l u t . getRed ( l a b e l s [ i ] ) +" , "
44 + l u t . getGreen ( l a b e l s [ i ] )
45 + " , " + l u t . getBlue ( l a b e l s [ i ] ) + " ) " ) ;
46 c h a n n e l s = new boolean [ 3 ] ;
47 channels [ 0 ] = f a l s e ;
50 // add l a b e l image with c o r r e s p o n d i n g c o l o r as an i s o s u r f a c e
51 univ . addContent ( labelImp , c o l o r , " l a b e l −"+l a b e l s [ i ] , 0 , channels , 2 , C o n t e n t C o n s t a n t s
. SURFACE) ;
52 }
53 }
54
55 // launch smooth c o n t r o l
56 s c = new SmoothControl ( univ ) ;
At the end of the script a dialog is shown to smooth the surfaces at will. Each label is
added to the 3D scene independently with the name "label-X" where X is its label value.
Figure 10.1.: From left to right: input label image, script output, smoothed label surfaces and
example of individually translated surfaces in the 3D viewer.

Bibliography
Arganda-Carreras, I. & Andrey, P. (2017). Designing Image Analysis Pipelines in Light Mi-
croscopy: A Rational Approach, (pp. 185–207). Springer New York: New York, NY.
Borgefors, G. (1984). Distance transformation in arbitrary dimensions. Comp. Vis. Graph.

Im. Proc., 27, 321–145.
Borgefors, G. (1986). Distance transformations in digital images. Comp. Vis. Graph. Im.
Proc., 34, 344–371.
Breen, E. J. & Jones, R. (1996). Attribute opening, thinnings, and granulometries. Computer
Vision and Image Understanding, 64(3), 377–389.
Burger, W. & Burge, M. J. (2008). Digital Image Processing, An algorithmic introduction using
Java. Springer.
Das, P. P. & Chatterji, B. N. (1988). Knight’s distance in digital geometry. Pattern Recognition
Letters, 7(4), 215–226.
Devaux, M.-F. & Legland, D. (2014). Microscopy: advances in scientific research and education,
chapter Grey level granulometry for histological image analysis of plant tissues, (pp. 681–
688). Formatex Research Center.
Doube, M., Kłosowski, M. M., Arganda-Carreras, I., Cordelières, F. P., Dougherty, R. P., Jack-
son, J. S., Schmid, B., Hutchinson, J. R., & Shefelbinea, S. J. (2010). BoneJ: free and
extensible bone image analysis in ImageJ. Bone, 47(6), 1076–1079.
Florindo, J. B., Bruno, O. M., Rossatto, D. R., Kolb, R. M., Gómez, M. C., & Landini, G.
(2016). Identifying plant species using architectural features in leaf microscopy images.
Botany, 94(1), 15–21.
Glasbey, C., van der Heijden, G., Toh, V. F. K., & et al. (2007). Colour displays for categorical
images. Color Research and Application.
Heneghan, C., Flynn, J., Keefe, M. O., & Cahill, M. (2002). Characterization of changes
in blood vessel width and tortuosity in retinopathy of prematurity using image analysis.
Medical Image Analysis, 6(4), 407 – 429.
Kong, T. Y. & Rosenfeld, A. (1989). Digital topology: Introduction and survey. Computer
Vision, Graphics, and Image Processing, 48(3), 357–393.
Kovesi, P. (2015). Good colour maps: how to design them.
94
Bibliography
Lang, C., Ohser, J., & Hilfer, R. (2001). On the analysis of spatial binary images. Journal of
Microscopy, 203(3), 303–313.
Lantuéjoul, C. & Beucher, S. (1981). On the use of geodesic metric in image analysis. Journal
of Microscopy, 121(1), 39–40.
Legland, D., Arganda-Carreras, I., & Andrey, P. (2016). MorphoLibJ: integrated library and
plugins for mathematical morphology with ImageJ. Bioinformatics, 32(22), 3532–3534.
Legland, D. & Devaux, M.-F. (2009). Détection semi-automatique de cellules de fruits charnus
observés par microscopie confocale 2d et 3d. Cahiers Techniques de l’INRA, numéro spécial
imagerie, 7–16.
Legland, D., Kiêu, K., & Devaux, M.-F. (2007). Computation of Minkowski measures on 2D
and 3D binary images. Image Analysis and Stereology, 26(6), 83–92.
Lehmann, G. & Legland, D. (2012). Efficient N-Dimensional surface estimation using Crofton
formula and run-length encoding. Technical report.
Luengo Hendriks, C. L. & van Vliet, L. J. (2003). Discrete morphology with line structuring
elements. In M. W. N. Petkov (Ed.), Computer Analysis of Images and Patterns - CAIP
2003 (Proc. 10th Int. Conf., Groningen, NL, Aug.25-27), volume 2756 of Lecture Notes in
Computer Science (pp. 722–729).: Springer Verlag, Berlin.
Meyer, F. & Beucher, S. (1990). Morphological segmentation. Journal of Visual Communica-

tion and Image Representation, 1(1), 21–46.
Moran, P. A. P. (1966). Measuring the length of a curve. Biometrika, 53(3), 359–364.
Ohser, J. & Schladitz, K. (2009). 3D Images of Materials Structures: processing and analysis.
WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim.
Poulet, A., Arganda-Carreras, I., Legland, D., Probst, A. V., Andrey, P., & Tatout, C. (2015).
NucleusJ: an ImageJ plugin for quantifying 3D images of interphase nuclei. Bioinformatics,
31(7), 1144–1146.
Robinson, K. & Whelan, P. F. (2004). Efficient morphological reconstruction: a downhill

filter. Pattern Recognition Letters, 25(15), 1759 – 1767.
Rosenfeld, A. (1970). Connectivity in digital pictures. J. ACM, 17(1), 146–160.
Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T., Preibisch,
S., Rueden, C., Saalfeld, S., Schmid, B., et al. (2012). Fiji: an open-source platform for
biological-image analysis. Nature methods, 9(7), 676–682.
Schneider, C. A., Rasband, W. S., Eliceiri, K. W., Schindelin, J., Arganda-Carreras, I., Frise,
E., Kaynig, V., Longair, M., Pietzsch, T., Preibisch, S., et al. (2012). NIH Image to ImageJ:
25 years of image analysis. Nature methods, 9(7).

Bibliography
Serra, J. (1982). Image Analysis and Mathematical Morphology. Volume 1. London: Academic
Press.
Serra, J. & Vincent, L. (1992). An overview of morphological filtering. Circuits, Systems and
Signal Processing, 11(1), 47–108.
Shoemake, K. (1994). Euler Angle Conversion. In P. S. Heckbert (Ed.), Graphics Gems (pp.
222 – 229). Academic Press.
Silva, J. V., Legland, D., Cauty, C., Kolotuev, I., & Floury, J. (2015). Characterization of
the microstructure of dairy systems using automated image analysis. Food Hydrocolloids,
44(0), 360 – 371.
Slabaugh, G. (1999). Computing Euler Angles from a Rotation Matrix.
Soille, P. (2003). Morphological Image Analysis. Springer, 2nd edition.
Soille, P. & Talbot, H. (2001). Directional morphological filtering. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 23(11), 1313–1329.
Soille, P. & Vincent, L. M. (1990). Determining watersheds in digital pictures via flooding
simulations. In Lausanne-DL tentative (pp. 240–250).: International Society for Optics and
Photonics.
Staal, J., Abramoff, M., Niemeijer, M., Viergever, M., & van Ginneken, B. (2004). Ridge based
vessel segmentation in color images of the retina. IEEE Transactions on Medical Imaging,
23, 501–509.
Svensson, S. & Borgefors, G. (2002a). Digital distance transforms in 3d images using infor-
mation from neighbourhoods up to 5 x 5 x 5. Computer Vision and Image Understanding,
88(1), 24 – 53.
Svensson, S. & Borgefors, G. (2002b). Distance transforms in 3D using four different weights.
Pattern Recognition Letters, 23, 1407–1418.
Tustison, N. & Gee, J. (2009). Introducing dice, jaccard, and other label overlap measures
to itk. Insight J, (pp. 1–4).
Verwer, B. J. H., Verbeek, P. W., & Dekker, S. T. (1989). An efficient uniform cost algorithm
applied to distance transforms. IEEE Trans. Pattern Anal. Mach. Intell., 11, 425–429.
Vincent, L. (1993). Morphological Grayscale Reconstruction in Image analysis: Applications

and Efficient Algorithm. IEEE Transaction on Image Processing, 2(2), 176–201.
Vincent, L. & Soille, P. (1991). Watersheds in digital spaces : An efficient algorithm based
on immersion simulation. IEEE Transactions on Pattern Analysis and Machine Intelligence,
13(6), 583–598.

A. List of operators
Plugin Name bin. 8 16 32 RGB 2D 3D GUI Page

Image filtering
Morphological filters * * * * * * * 19
Morphological filters (3D) * * * * * * * 20
Directional filters * * * * * * 20
Grayscale Attribute Filtering * * * * * * 28
Grayscale Attribute Filtering 3D * * * * * * 28
Morphological Reconstruction * * * * * * 22
Morphological Reconstruction (3D) * * * * * * 22
Interactive Morphological Reconstruction * * * * * * 24
Interactive Morphological Reconstruction 3D * * * * * * 24
Kill Borders * * * * * * 23
Fill Holes (Binary / Gray) * * * * * * 23
Segmentation
Classic Watershed * * * * * * 31
Marker-controlled Watershed * * * * * * 33
Interactive Marker-controlled Watershed * * * * * * 35
Morphological Segmentation * * * * * * 38
Minima and Maxima

Regional Min & Max * * * * * * 25
Regional Min & Max 3D * * * * * * 25
Extended Min & Max * * * * * * 26
Extended Min & Max 3D * * * * * * 26
Impose Min & Max * * * * * * 26
Impose Min & Max 3D * * * * * * 26
Utilities
Extend Borders * * * * * * * *
Binary Overlay * * * * n.a. * * *
Binary/Labels Overlay * * * * n.a. * * * 81
Draw Label Values * * * * n.a. * 81
Table A.1.: Image processing operators for grayscale and intensity images.
97
Plugin Name bin. 8 16 32 RGB 2D 3D GUI Page
Binary Images
Connected Components Labeling * n.a. n.a. n.a. n.a. * * * 76
Chamfer Distance Map * * * * n.a. * * 74
Chamfer Distance Map 3D * * * * n.a. * * 74
Geodesic Distance Map * * * * n.a. * * 76
Interactive Geodesic Distance Map * n.a. n.a. n.a. n.a. * * 76
Geodesic Distance Map 3D * * * * n.a. * * 76
Distance Transform Watershed * n.a. n.a. n.a. n.a. * * 43
Distance Transform Watershed 3D * n.a. n.a. n.a. n.a. * * 45
Convexify * n.a. n.a. n.a. n.a. * 77
Remove Largest Region * n.a. n.a. n.a. n.a. * * 77
Keep Largest Region * n.a. n.a. n.a. n.a. * * 77
Area Opening * n.a. n.a. n.a. n.a. * * 78
Size Opening 2D/3D * n.a. n.a. n.a. n.a. * * * 78
Label Images
Set Label Map n.a. * * * n.a. * * * 81
Labels To RGB n.a. * * * n.a. * * * 81
Draw Labels As Overlay n.a. * * * n.a. * 80
Assign Measure To Label * * * * n.a. * * * 81
Label Edition n.a. * * * n.a. * * * 89
Label Morphological Filters * * * * n.a. * * * 84
Remap Labels * * * * n.a. * * 85
Fill Label Holes * * * * n.a. * * 85
Expand Labels n.a. * * * n.a. * * * 85
Remove Border Labels n.a. * * * n.a. * * * 83
Replace / Remove Label(s) n.a. * * * n.a. * * * 83
Merge Labels n.a. * * * n.a. * * *
Select Label(s) n.a. * * * n.a. * * * 83
Crop Label n.a. * * * n.a. * * * 85
Remove Largest Label * * * * n.a. * * 85
Keep Largest Label * * * * n.a. * * 85
Label Size Filtering * * * * n.a. * * * 83
Region Adjacency Graph n.a. * * * n.a. * * 86
Region Boundaries Labeling n.a. * * * n.a. * * 86
Label Boundaries n.a. * * * n.a. * * 86
Select Neighbor Labels n.a. * * * n.a. * * * 88
Table A.2.: Image processing operators for binary and label images.

B. Computation of equivalent ellipsoid
coefficients
It is often convenient to represent the results of second order moments (section 7.2) as an
equivalent ellipsoid that has the same orientation and the same eigen values as the structure
of interest. The equivalent ellipsoid can be completely characterised by nine coefficients:
• the three coordinates of the centroid
• the three radius r1 , r2 , r3 , with r1 > r2 > r3 , corresponding to the lengths of each
semi-axis
• a 3D orientation, that can be represented by three Euler angles.
B.1. Notations
We consider euclidean spaces of dimension 3. Let X be the a 3D set representing the structure
of interest.
B.1.1. Moments
In 3D, the moments m pqr of order (p, q, r) are defined as:
Z Z Z
m pqr = I X (x, y, z)x p y q z r · d x · d y · dz (B.1)
where I X (x, y, z) is the indicator function of the set X , taking value 1 if the specified point
is within the set X , and 0 otherwise. The moment m000 corresponds to the volume of the
structure. The centered moments are expressed as:
Z Z Z
µ pqr = I X (x, y, z) (x − x c ) p ( y − yc )q (z − zc ) r · d x · d y · dz (B.2)
m m m
where (x c , yc , zc ) = ( m000
100 010
, m000 , m001
000
) is the centroid of X . The matrix of second order mo-
ments is the symmetric 3 × 3 matrix that combines all the centered second order moments:
 
µ200 µ110 µ101
M =  µ110 µ020 µ011 
µ101 µ011 µ002
99
B.2 Computation of angles
B.1.2. Eigen values decomposition

The matrix M can be factorized using eigen value decomposition:
M = QΛQ t
where
• Q is a 3 × 3 orthogonal matrix whose i-th column is the eigenvector qi of M
• Λ is a diagonal 3 × 3 matrix whose diagonal elements are the corresponding eigen

values, Λii = λi .
Eigen values are ordered in decreasing order. The matrix Q can be used to extract the ori-
entation of the main axes of the structure (section B.2), whereas the eigen values λi can be
related to the particle dimensions along these axes (section B.3).
B.2. Computation of angles

This section largely follows the document of Greg Slabaugh (Slabaugh, 1999) for obtaining
Euler angles from a rotation matrix. We mostly follow his notations. Another useful reference
may be Shoemake (1994).
B.2.1. Rotation matrices

A rotation of ψ radians about the x-axis is noted by:
 
1 0 0
R x (ψ) =  0 cos ψ − sin ψ 
0 sin ψ cos ψ
Similarily, a rotation of θ radians about the y-axis is noted by:

 
cos θ 0 sin θ
R y (θ ) =  0 1 0 
− sin θ 0 cos θ
Finally, a rotation of ϕ radians about the z-axis is noted by:

 
cos ϕ − sin ϕ 0
Rz (ϕ) =  sin ϕ cos ϕ 0 
0 0 1
A general rotation matrix may have the following form:

 
R11 R12 R13
R =  R21 R22 R23 
R31 R32 R33

B.2 Computation of angles
Such a matrix can be represented by a sequence of three successive rotations around the
main axes. As matrix multiplication does not commute, the order of the axes one rotates
about will affect the result. In MorphoLibJ, we follow the choice of G. Slabaugh and consider
rotation first about the x-axis, then about the y-axis, and finally about the z-axis. Then, the
angles ψ, θ and ϕ correspond to Euler angles. Angle ψ corresponds to the “roll”, angle θ
to the “pitch”, and angle ϕ to the “yaw”.
The global rotation matrix can be written as follow:
R = Rz (ϕ) · R y (θ ) · R x (ψ)
 
cos θ cos ϕ sin ψ sin θ cos ϕ − cos ψ sin ϕ cos ψ sin θ cos ϕ + sin ψ sin ϕ
=  cos θ sin ϕ sin ψ sin θ sin ϕ + cos ψ cos ϕ cos ψ sin θ sin ϕ − sin ψ cos ϕ 
− sin θ sin ψ cos θ cos ψ cos θ
The problem is now to identify the three Euler angles ψ, θ and ϕ from the matrix coeffi-
cients. This results in nine equations.
B.2.2. Elevation
We start considering the elevation θ , or “pitch”. From element R31 of the matrix, one finds
R31 = − sin θ
One identifies θ with the following
θ = − sin−1 (R31 )
by keeping the value of θ ∈ [− π2 ; π2 ] to remove ambiguity. The case R31 = ±1 will be

considered later. In practice, we replace the sin−1 function by the atan2 function for better
numerical stability.
Compared to the document of Slabaugh, keeping θ ∈ [− π2 ; π2 ] removes the sign ambiguity
of cos θ , and therefore simplifies the remaining computations.
B.2.3. Roll
The possible values of the roll ψ around the x-axis can be found by
R32
= tan ψ
R33
This lead to the value of ψ as
ψ = atan2(R32 , R33 )
where atan2( y, x) is the arc tangent function of the two variables y and x that extends the
atan function to the all four quadrants. The function atan2 is available in most programming
languages.

B.3 Computation of ellipsoid radius
B.2.4. Azimut
The azimut ϕ, or “yaw”, can be obtained from
R21
= tan ϕ
R11
ϕ = atan2(R21 , R11 )
B.2.5. Special case of cos θ = 0

When the matrix element R31 = ±1, we have θ = ±π/2: the main axis of the ellipsoid is
aligned with the z-axis. Developping the terms R12 , R13 , R22 , and R23 shows that the values
of ψ and ϕ are linked, and an infinite number of solutions exists (Slabaugh, 1999). This
phenomenon is called Gimbal lock. A convenient way to obtain a single valid solution is to
set the value for the azimut ϕ arbitrarily to 0, and compute ψ from R12 and R13 .
If θ = π/2 then
R12 = sin ψ cos ϕ − cos ψ sin ϕ = sin(ψ − ϕ)

R13 = cos ψ cos ϕ − sin ψ sin ϕ = cos(ψ − ϕ)
R22 = sin ψ sin ϕ − cos ψ cos ϕ = cos(ψ − ϕ) = R13
R23 = cos ψ sin ϕ − sin ψ cos ϕ = − sin(ψ − ϕ) = −R12
leading to
ψ = ϕ + atan2(R12 , R13 )
If θ = −π/2, similar development leads to:
ψ = −ϕ + atan2(−R12 , −R13 )
B.3. Computation of ellipsoid radius

Let us suppose that the ellipsoid is centered and aligned with the main axes. From centering,
we have m pqr = µ pqr . For integration, it is more convenient to use spherical coordinates
(ρ, θ , ϕ), corresponding to the distance to the origin, to the inclination with the vertical,
and to the azimuth (these notations are different from the ones used for Euler angles). This
corresponds to:
x = ρ · r1 · cos ϕ · sin θ
y = ρ · r2 · sin ϕ · sin θ
z = ρ · r3 · cos θ
with ρ ∈ [0; +∞[, ϕ ∈ [0; 2π] and θ ∈ [0; π].

The Jacobian matrix of the transformation is as follow (see example 3 in “Jacobian matrix
and determinant” on Wikipedia1 ):
∂x ∂x ∂x
   
∂ρ ∂θ ∂ϕ r1 cos ϕ sin θ ρr1 cos ϕ cos θ −ρr1 sin ϕ sin θ
 ∂y ∂y ∂y 
JΦ (x, y, z) =  ∂ ρ ∂ θ ∂ ϕ  =  r2 sin ϕ sin θ ρr2 sin ϕ cos θ ρr2 cos ϕ sin θ 
∂z ∂z ∂z r2 cos θ −ρr3 sin θ 0
∂ρ ∂θ ∂ϕ
The determinant equals ρ 2 sin θ if r1 = r2 = r3 = 1. Introducing a, b and c gives a

determinant equal to r1 r2 r3 ρ 2 sin θ . Then the moment integral is given by:
Z 2π Z π Z 1
µ pqr = r1 r2 r3 (ρ cos ϕ sin θ ) p (ρ sin ϕ sin θ )q (ρ cos θ ) r · ρ 2 sin θ · dρ · dθ · dϕ
0 0 0
(B.3)
B.3.1. Volume moment

It is easy to check that the moment m000 corresponds to the volume of the ellipsoid (equal
to 4π3 r1 r2 r3 ):
Z 2π Z π Z 1
m000 = r1 r2 r3 ρ 2 sin θ · dρ · dθ · dϕ
0 0 0
Z 2π Z π Z 2π
r1 r2 r3 a bc
= sin θ · dθ · dϕ = (− cos π + cos 0) · dϕ
3 0 0 3 0
Z 2π
2 4π
= r1 r2 r3 dϕ = r1 r2 r3
3 0 3
B.3.2. Moment m200

When considering centered ellipsoid aligned with principal axes, the eigen values λi simply
correspond to the moments µ200 , µ020 and µ002 . Let us focus on the moment µ200 :
Z 2π Z π Z 1
µ200 = r1 r2 r3 (r1 ρ cos ϕ sin θ )2 ρ 2 sin θ · dρ · dθ · dϕ
0 0 0
Z 2π Z π Z 1
= r13 r2 r3 ρ 4 dρ · cos2 ϕ sin3 θ · dθ · dϕ
0 0 0
2π Z π
r13 r2 r3
Z
= sin3 θ · dθ · cos2 ϕ · dϕ
5 0 0
1
https://en.wikipedia.org/wiki/Jacobian_matrix_and_determinant#Example_3:
_spherical-Cartesian_transformation

3
The integral over θ can be developed using the linearisation of sin3 θ = 4 sin θ − 14 sin (3θ ):
Z π Z π
3 1

3
sin θ dθ = sin θ − sin (3θ ) dθ
0 0 4 4
3 −1
= − (cos π − cos 0) − (cos 3π − cos 0)
4 12
3 1 4
= − =
2 6 3
Coming back to moment integral:
2π 2π
r13 r2 r3 4 r13 r2 r3 4
Z Z
2 1 1
µ200 = cos ϕ · dϕ == + cos (2ϕ) dϕ
5 3 0 5 3 0 2 2
r13 r2 r3 41 4π r13 r2 r3
= (2π − 0) =
5 32 3 5
r12
= m000
5
Then, the central moments are expressed as:
r12
λ1 = µ200 = m000
5
r22
λ2 = µ020 = m000
5
r32
λ3 = µ002 = m000
5
ri2
We therefore have λi = 5 m000 , leading to:
v
t5 · λ
i
ri = (B.4)
m000

MorphoLibJ-manual-v1 6 2

Uploaded by

Copyright:

Available Formats

MorphoLibJ-manual-v1 6 2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MorphoLibJ-manual-v1 6 2

Uploaded by

Copyright:

Available Formats

MORPHOLIBJ USER MANUAL

David Legland, Ignacio Arganda-Carreras

October 16, 2023

3. Definitions and data types 9

5. Connected components operators 21

8.2. Connected component labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

B. Computation of equivalent ellipsoid coefficients 99

MorphoLibJ user manual page 3 / 104

• morphological reconstruction, for 2D/3D and binary or grayscale images, allowing

• watershed segmentation + GUI, making it possible to segment 2D/3D images of cell

• 2D/3D measurements: volume, surface area, bounding boxes, equivalent ellipse/el-

• several utilities for the manipulation of binary and label images

2.2. Installation in Fiji

2. When the updater is open, click on Manage update sites.

MorphoLibJ user manual page 7 / 104

MorphoLibJ user manual page 8 / 104

3.1. Digital images

3.1.1. Data types

3.1.2. Image types

MorphoLibJ user manual page 10 / 104

3.2. Coordinate system

3.3. Digital connectivity

(a) 4-connectivity (b) 8-connectivity

Figure 3.1.: Digital connectivities for 2D images

MorphoLibJ user manual page 11 / 104

(a) Continuous curve (b) Digital curve using the 4-

Figure 3.3.: Digital connectivities for 3D images

MorphoLibJ user manual page 12 / 104

4.2. Grayscale morphological filters

4.2.1. Grayscale erosion and dilation

MorphoLibJ user manual page 14 / 104

(a) Original image (b) Morphological dilation (c) Morphological erosion

4.2.2. Opening and closing

(a) Original image (b) Morphological closing (c) Morphological opening

MorphoLibJ user manual page 15 / 104

In a symmetric way, the result of an erosion followed by a dilation is called a morpho-

4.2.3. Morphological gradients

MorphoLibJ user manual page 16 / 104

(a) Morphological gradient (b) Morphological gradient on a (c) Morphological Laplacian

MorphoLibJ user manual page 17 / 104

4.3. Directional filters

An alternative is to apply directional filtering. The principle is to consider an oriented

MorphoLibJ user manual page 18 / 104

4.4. Plugin Usage

4.4.1. Planar images

opening consists in the succession of an erosion with a dilation. Morphological opening

morphological gradient is defined as the difference of a morphological dilation and a mor-

morphological Laplacian is defined as half the sum of a morphological dilation and a

MorphoLibJ user manual page 19 / 104

The following structuring elements can be used for 2D images:

• line with angle of 0, 90, 45 or 135 degrees

4.4.3. Directional Filters

MorphoLibJ user manual page 20 / 104

5.1. Morphological reconstruction

5.1.2. Applications to binary images

MorphoLibJ user manual page 22 / 104

5.1.3. Applications to grayscale images

5.1.4. Application to the segmentation of chromocenters in a 3D image

MorphoLibJ user manual page 23 / 104