718 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 48, NO.

6, JUNE 2001

Automatic Identification of Human Helminth Eggs on

Microscopic Fecal Specimens Using Digital Image
Processing and an Artificial Neural Network
Yoon Seok Yang, Student Member, IEEE, Duck Kun Park, Hee Chan Kim*, Member, IEEE, Min-Ho Choi, and
Jong-Yil Chai

Abstract—In order to automate routine fecal examination for aged more overseas travel and international food fads, allows
parasitic diseases, we propose in this study a computer processing parasitic diseases to migrate from one country to other distant
algorithm using digital image processing techniques and an countries. And despite high levels of sanitation, food inspec-
artificial neural network (ANN) classifier. The morphometric
characteristics of eggs of human parasites in fecal specimens tion, and vector control in developed countries, individuals who
were extracted from microscopic images through digital image may have acquired parasitic infections elsewhere require med-
processing. An ANN then identified the parasite species based on ical treatment. In addition, the role of zoonotic parasite is be-
those characteristics. We selected four morphometric features coming more important as agents of human diseases.
based on three morphological characteristics representing shape, Human parasitic infections are recognized through the iden-
shell smoothness, and size. A total of 82 microscopic images con-
taining seven common human helminth eggs were used. The first tification of parasitic organizms in the feces, urine, blood, and
stage (ANN-1) of the proposed ANN classification system isolated tissues by using the proper diagnostic methods [1]. Fecal exam-
eggs from confusing artifacts. The second stage (ANN-2) classified ination is the most common diagnostic method. Trained experts
eggs by species. The performance of ANN was evaluated by the examine the fecal specimens, searching for parasitic organizms,
tenfold cross-validation method to obviate the dependency on usually the eggs of helminths and cysts of protozoa. If the orga-
the selection of training samples. Cross-validation results showed
86.1% average correct classification ratio for ANN-1 and 90.3% nizms are present, they examine the sizes, shapes, and numbers
for ANN-2 with small variances of 46.0 and 39.0, respectively. to identify the species of parasites, the degree of infection, and
The algorithm developed will be an essential part of a completely appropriate therapeutic modalities. But diagnosticians are reluc-
automated fecal examination system. tant to perform fecal examinations; with decreasing importance
Index Terms—Artificial neural network, automatic identifica- of most parasites, less material is submitted for examination,
tion, helminth egg, pattern classification. which results in an insufficiency of training material. Thus, it
is difficult to maintain enough people with expertise in diag-
I. INTRODUCTION nostic parasitology. Worst of all, diagnostic parasitologists are
expected to maintain high standards of proficiency in the iden-

H UMAN parasites are classified taxonomically into

protozoa and helminths (nematodes, trematodes, and
cestodes). Parasites enter the human body mainly through soil,
tification of rare parasites as well as common endemic ones.
To resolve the issue of a lack of trained personnel and relieve
the pressure on them, the automation of routine fecal examina-
water, vegetables, fruits, or other host animals, then inhabit tions is necessary. The complete automation of check-up rou-
internal organs such as the small intestine, liver, and lung. tines would allow inspection of a great number of specimens
The infected human body suffers from malnutrition, irritation, with a high degree of reliability and objectivity. This would be
damaged tissue, parasitic contamination, greater likelihood of useful in developed countries as well as in underdeveloped and
microorganizm intrusion, allergic response by the body itself, developing countries.
and in the worst case, death. Several precedent studies employed microscopic image
In developed countries, parasitoses are less prevalent, with analysis to diagnose parasites [2]–[6]. These studies suc-
fewer people becoming ill from parasitic diseases. However, cessfully differentiated several parasites in pigs and cattle
the extended global transportation network, which has encour- with numerically extracted morphologies through the use of
interactive image analysis tools [2], [3], [6]. They extracted
Manuscript received January 4, 2000; revised July 2, 2001. This work was morphometric features, including length, smoothness, and
supported in part by the Year 2000 BK 21 Project for Medicine, Dentistry and sphericity with subjectively chosen parameters from digital
Pharmacy. Asterisk indicates corresponding author.
Y. S. Yang and D. K. Park are with the Interdisciplinary Program Biomedical microscopic images [2], [3]. These features were subsequently
Engineering Major, Seoul National University, Seoul 110-744, Korea. involved in the computation of classification indices; the
*H. C. Kim is with the Departments of Biomedical Engineering, Seoul parasite species were then identified using the calculated
National University College of Medicine and Institute of Biomedical Engi-
neering, Medical Research Center, Seoul National University, 28 Yongon indices. Their works were concentrated on statistical analysis
Dong, Chongno Gu, Seoul 110-744, Korea (e-mail: hckim@snu.ac.kr). of morphologic values and round-robin discrimination tests.
M.-H. Choi and J.-Y. Chai are with the Department of Parasitology, Seoul However, no truly automatic and computer-assisted algorithm
National University College of Medicine and Institute of Endemic Diseases,
Seoul National University Medical Research Center, Seoul 110-744, Korea. has been developed. Specifically, this would entail automation
Publisher Item Identifier S 0018-9294(01)04170-2. of the entire procedure, including image enhancement, extrac-
0018–9294/01$10.00 © 2001 IEEE
YANG et al.: AUTOMATIC IDENTIFICATION OF HUMAN HELMINTH EGGS ON MICROSCOPIC FECAL SPECIMENS 719

Fig. 1. Representatives of the seven helminth eggs.

tion of morphological characteristics, isolation of parasitic formalin-preserved specimens containing eggs of seven human
organizms from unwanted dreg materials, and identification helminths. There were about 12 images/parasite species. The
of parasite species. To develop a totally automated system, seven typical species of human helminth were chosen as
image segmentation, the algorithm for morphometric feature follows [1]: ascaris lumbricoides, trichuris trichiura, capillaria
extraction, and classification should be automated. In addition, philippinensis, clonorchis sinensis, paragonimus westermani,
dedicated hardware equipment is necessary to scan complete diphyllobothrium latum, and taenia. Representatives of each of
images of each specimen under microscope. the egg types are shown in Fig. 1 to help the reader understand
We propose in this paper an automated method to detect the classification problem visually.
common helminth eggs in microscopic fecal specimen images To localize and classify helminth eggs in a given image, we
and to identify their correct species using digital image pro- first found meaningful objects in the image and extracted their
cessing techniques [7] and an artificial neural network (ANN) features using digital image processing methods. Then a two-
system [8]. The ANN classifier is widely used in pattern stage ANN classified them based on their numeric features.
recognition problems owing to its good performance and easy All processing routines were written in MATLAB (Mathworks,
application [8]–[10]. Natick, MA) and overall processing steps are outlined in Fig. 2.

II. MATERIALS AND METHODS A. Digital Image Processing

Trained laboratory staff normally examines fecal specimen Digital image processing consists of four operational steps:
visually under 100-fold magnification. Although the higher median filtering, binary thresholding, segmentation, and feature
magnification views are occasionally used to examine cell extraction.
material inside the egg shell, it is sufficient for pinpointing 1) Median Filtering: Since fecal material normally includes
a parasitic egg on the specimen and checking its size and blood residue, various products of digestion, sloughed epithelial
shape. A relatively faster scanning speed is one advantage over cells, mucus, and microorganizms such as bacteria and yeast, the
the higher magnification. For digital processing, microscopic digitized images show nonhomogeneous background character-
images were digitized by using a color video camera (JVC, istics as in Fig. 3 [1]. Therefore, all images were preprocessed
Japan) embedded in the microscope (Olympus, Japan) and with a 4 4 median filter to remove small unimportant dregs
an image capturing board (MERIT, Korea) on a Pentium II without degrading overall image quality. Filter dimensions were
PC. Images were captured under 200-fold magnification to experimentally determined to remove small particles while pre-
provide comparable quality with a natural view under 100-fold serving target objects.
magnification. Each digital image is 560 by 420 pixels large 2) Binary Thresholding: A relatively simple binary thresh-
with 8-bit intensity depth, covering a slide area of 130 m olding method was applied to the median filtered image. Com-
100 m. This consequently yields a spatial resolution of pared with other edge detection methods that yield problems
4.2 pixels/ m. A total of 82 images were acquired from with cracked boundary lines, it was proven to have an acceptable
720 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 48, NO. 6, JUNE 2001

a mammillated external shell layer, the smoothness of

the shell is an important feature in separating eggs from
vegetable cells or other plant material that may have
irregularly shaped membranes.
To obtain numeric values representing each characteristic, we
calculated the polar coordinates ( ) of all detected boundary
pixels with the origin at their gravitational center. An one ob-
ject’s boundary line can then be represented by the trajectory of
its radial distance, ), as spans from zero to , as illustrated
in Fig. 5 [7]. After normalization of by its mean value, three
numeric features representing shape and shell characteristics are
calculated. The object’s size in total pixels number is included
as the fourth numeric feature. All four features are summarized
as follows.
a) , the ratio of minimum in
to maximum in .
b) , the ratio of low
frequency components to high frequency components
from the magnitude of discrete Fourier transform of
,
c) , standard deviation of .
d) , the number of pixels in an object area or pixel dimen-
sion.

Fig. 2. Overall processing steps.

B. ANN Classifier
Species identification was performed by an ANN classifica-
performance with minimal complexity in this specific applica- tion system which consists of two stages; the first stage (ANN-1)
tion. The threshold was empirically set at 35% of the highest separates parasite eggs from the artifact, while the second stage
intensity level in all cases. It was determined based on the fact (ANN-2) identifies the species of the detected eggs. This two-
that most of the histograms have bimodal distribution and their stage classification scheme is adopted to reduce the total com-
valley points coincide with this level within reasonable range. It puting time that is one of the critical parameters in realizing a
is well known that the valley point of the bimodal distribution fully automated system. The goal of ANN-1 is to examine all
ensures the optimal classification result in terms of Bayesian de- segmented objects and isolate parasitic organizms from other
cision errors [10], [11]. elements in feces with similar sizes. The following ANN-2 clas-
3) Segmentation: From the thresholded binary results, we sifies the species of the detected eggs by ANN-1 using the same
segmented only objects whose sizes were above the minimum numeric features. Both ANNs are general multilayer percep-
of egg sizes (600-pixel dimension) by the connected compo- trons (MLP) neural networks; the back propagation learning
nent operator [7]. Binary thresholding may have caused the seg- rule was used in both cases [8]. Generally, the MLP network
mented objects not to appear as one full mass. In order to make consists of set of neurons (perceptrons) in several layers and
later numerical feature extraction easier, every segmented ob- inter neuronal connections. Input neurons take input values and
ject was filled by the four-directional inward region growing al- each neuron in subsequent layers takes a sum of neuronal output
gorithm so that there were no internal holes. Then the single activities in the previous layers weighted by their interconnec-
skeleton algorithm produced a boundary line of an object with tions and gives its nonlinear output activity. It is basic and the
one-pixel thickness for truly quantitative representations of the most widely used neural network in pattern classification be-
object’s morphology [7]. cause of its capability of making almost any kind of linear or
4) Feature Extraction: As criteria for detection and identi- nonlinear decision boundaries between clustered patterns. MLP
fication of helminth eggs, three morphological characteristics network capacity is limited by the number of neurons and their
were selected as follows [1]. inter connections and by the nonlinear activation functions of
a) Size: Helminth eggs ranged in length from 25 m to 150 these neurons.
m, and in diameter from 12–14 m to 90 m. Typical Network topology was systematically determined in terms of
sizes of common helminth eggs are compared in Fig. 4. variance of the classification results. For all devisable network
b) Shape: Each parasite species produces eggs that are uni- topologies, classification results were calculated for different
form in shape. The shape of eggs, varying from spherical data sets selected by the cross-validation method. The number
to elongated, is a prominent feature for most species of of layers and neurons were chosen to provide minimum vari-
helminths. ance and reasonably high average classification results ( 85%).
c) Eggshell: Helminth eggs usually have a smooth shell. Typical results are summarized in Tables I and II for ANN-1 and
With the exception of A. lumbricoides, these eggs have ANN-2, respectively.
YANG et al.: AUTOMATIC IDENTIFICATION OF HUMAN HELMINTH EGGS ON MICROSCOPIC FECAL SPECIMENS 721

Fig. 3. One of the typical microscopic images of fecal specimen under 200-fold magnification.

Fig. 4. Comparative sizes of typical helminth eggs. Dark bars denote ranges of length.

The cross validation took up the leave-one-out method [5] axis. The smoothness may be rephrased as the trajectory of its
with the alteration of tenfold crossover to avoid excessive time- radial distance, , which has smaller high frequency compo-
consuming repetitions. It generally provides network perfor- nents than those of the artifacts. The input layer has four input
mance evaluations independent of the choice of training and test nodes to receive the morphological feature values. It has one
set. output node which indicates whether the input morphologies are
1) Artifacts Removal with ANN-1: Significant morpholog- those of eggs or artifacts. The desired output values were de-
ical differences exist between parasite eggs and artifacts, even rived from the ground truth obtained from diagnostic experts. A
though they are similar in size, as mentioned earlier. The egg standard sigmoid activation function was used at the input and
cell has a shell membrane which makes its boundary look more hidden nodes, while a linear activation function was used at the
circular and smoother than those of the artifacts. The circularity output node [8]. 187 samples (87 eggs and 100 artifacts in total)
of shell may be expressed by the ratio of its short axis and long were randomly divided into ten subsets of approximately equal
722 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 48, NO. 6, JUNE 2001

0!R(!) where F denotes the

Fig. 5. Schematic diagram of calculating angular trajectory r() from boundary line and its frequency components R(! ). r ( )
Fourier Transform.

TABLE I Among segmented objects in images, those which yielded

CROSS-VALIDATION RESULTS OF ANN-1 FOR DIFFERENT NETWORK
outputs above cutoff value at the trained ANN-1 were deter-
TOPOLOGIES
mined to be eggs and transferred to the next identification stage.
2) Identification of Helminth Species: The ANN-2 has three
input nodes and seven output nodes. The three input nodes re-
ceive the same numeric morphological feature values that are
used in the ANN-1 except for the third element. A preparatory
classifier using all 4 features including the third element ( )
produced no significant improvement over the three input clas-
sifier for all network topologies. A reasonable explanation for
this is that most of the eggs, once distinguished from various
artifacts, have s within a narrow range. The computed vari-
TABLE II
ance of the over all detected eggs was smaller than those of
CROSS-VALIDATION RESULTS OF ANN-2 FOR DIFFERENT NETWORK other three features. The features used in ANN-1 and ANN-2 are
TOPOLOGIES summarized in Table III for clarity. Contributions of individual
features to the ANN-2 classification are described by the fol-
lowing examples. , the size of the egg, is the strongest charac-
teristic in finding C. sinensis, which is remarkably smaller than
the other kinds of eggs, as shown in Fig. 4. The smoothness
of shell is well represented by ; this is important especially
in distinguishing A. lumbricoides from D. latum. Despite their
similarity in size, they are quite different in terms of since
the former, especially in unfertilized cases, has a rougher shell
than the latter. Finally, , which corresponds to the ratio of the
size ( ) for tenfold cross validation. After system- short axis to the long axis of the oval boundary helps to sepa-
atic evaluation, an ANN topology having eight neurons in the rate circular eggs of Taenia from the more elliptical eggs of T.
first hidden layer and eight neurons in the second hidden layer trichiura and C. philippinensis.
was chosen since it shows the maximum average and the min- The seven output nodes correspond to the seven helminth
imum variance of classification results. species selected for this study. ANN-2 has a competitive output
With this topology, the neural network was trained with layer in which only one node produces a valid output value for
randomly selected halves of 87 eggs and 100 artifacts as each input pattern. Such “winner-take-all” competition is suit-
training samples. Then classification results with the remaining able for this type of classification problem [8]–[10]. Each ar-
data samples were computed. Besides, since the results can be tificial neuron’s activation function and training rules were the
affected by cutoff at the output neuron, the receiver operating same that used for ANN-1. 87 samples (87 eggs in total) were di-
characteristic (ROC) was also analyzed to show the efficiency vided into ten subsets of approximately equal size ( )
in true positive (TP)/false positive (FP) tradeoff [11]. for cross validation. After systematic evaluation, the network
YANG et al.: AUTOMATIC IDENTIFICATION OF HUMAN HELMINTH EGGS ON MICROSCOPIC FECAL SPECIMENS 723

TABLE III
SUMMARY OF FEATURES USED IN ANN-1 AND ANN-2

image has one helminth egg from C. sinensis in the middle of

the top-right quadrant. A blurred image in Fig. 7(a) is the me-
dian-filtered result of the original. It contains reduced speckles
and dregs in background compared to the original. Therefore,
the thresholded result in Fig. 7(b) has a suitable form for ob-
ject segmentation. In this case, a total of 11 objects were se-
lected and segmented automatically, as explained previously.
They are marked on the image by square boxes. Other black
spots remaining unmarked were discarded because their sizes
were smaller than the minimum egg size. The segmented ob-
jects include one parasite egg and ten artifacts. In all, averaged
over all images, about 15 objects were segmented and one par-
asite egg appeared in each image.
Fig. 8 displays four of the segmented objects consisting of
three artifacts (a), (b), (c) and one parasite egg (d). It shows
each object’s appearance, obtained boundary line, normalized
radial distance trajectory, , and log amplitude of the Fourier
transform (calculated by fast Fourier transform), . The
Fig. 6. ROC analysis for ANN-1. The normalized area under the curve is 0.87. boundary line is actually the outermost pixels of filled-up
version of the object and shows no seriously damaged part as
having two hidden layers with eight neurons in both layers was expected. The radial trajectory in (d) shows a typical form of
chosen since it showed the maximum average and the minimum smooth elliptical egg group to which this egg belongs. The
variance of classification results. circular eggs showed more uniform distance trajectories with
Then, as in the ANN-1 case, the ANN-2 with the chosen smaller variations than this. The other trajectories from artifacts
topology was trained with randomly selected half of eggs (87 also reflect their respective shape characteristics. The log-am-
in total) and tested with the remaining half. plitude plots of the Fourier transform from artifacts and eggs
The neural network classifier was compared with the simple exhibit noticeable differences. The amplitude spectrums show
-nearest neighbor classifier for performance. The -nearest that the eggs contain negligible high frequency components in
neighbor classifier assigns an unknown test sample to the most contrast to the artifacts.
prevailing class among its nearest neighbors in feature space. In Fig. 9, which shows the ANN results for two different
All the neighbors are preclassified samples given to it prior to cases of eggs, numbers at the right side of each rectangle in
the test [11]. The performance of -nearest neighbor classifier Fig. 9(a) and (c) represents the numeric morphologic features,
was evaluated by the same cross-validation method to be com- , and extracted from and of the object.
pared with the ANN classifier. The values are arranged in that order.

B. ANN Classification
III. RESULTS In all, the finally trained ANN-1 showed 100% correct clas-
sification in the training phase and 84% correct classification in
A. Segmentation and Feature Extraction
the test phase. In terms of detection accuracy, it showed 83%
Typical digital processing results with the original image of TP ratio and 10% FP ratio on the test samples with the cutoff at
Fig. 3 are shown in Figs. 7 and 8. In this case, the original the half of output scale. The results are summarized in Table IV.
724 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 48, NO. 6, JUNE 2001

(a)

(b)
Fig. 7. Typical results of digital image processing: (a) median filtered image, (b) black and white image obtained from binary thresholding. A total of 11 objects
marked by rectangular boxes are detected as meaningful objects. The rectangles represent the segmented objects.

And the ROC curve in Fig. 6 shows the efficiency in TP/FP re- neighbor classifier are compared with ANN-1 and ANN-2,
lationship. The normalized area under the curve is 0.87. respectively. The number of three and eight were chosen
The finally trained ANN-2 showed 100% correct classifica- since they showed better performances than other values
tion in learning stage and 83% correct classification at test stage. in each application. The results show the superiority of
For more details, identification results for individual species are the proposed ANN classifiers over the -nearest neighbor
presented in a confusion matrix form in Table V. classifiers.
In Table VI, the comparison between performance evalua- Fig. 9 shows the typical results of final classification for the
tions of these two ANNs and -nearest neighbor classifiers C. sinensis and C. philippinensis. The white boxes in Fig. 9(a)
are summarized. The three-nearest neighbor and eight-nearest and (c) denotes the detected parasite eggs by ANN-1. Then the
YANG et al.: AUTOMATIC IDENTIFICATION OF HUMAN HELMINTH EGGS ON MICROSCOPIC FECAL SPECIMENS 725

(a)

(b)

(c)

(d)
Fig. 8. Comparison of meaningful objects, segmented from Fig. 7. The leftmost images show objects and their obtained boundary lines. The plots in the middle
column represent their normalized radial trajectories, r (). The plots in the last column show the logarithms of amplitude spectra against discrete frequencies
ranging from zero to 2 . (a), (b), and (c) Artifacts. (d) Egg cell.

TABLE IV Total computation time involving the digital image pro-

RESULTS OF THE ANN-1 CLASSIFIER TRAINED WITH HALF OF TOTAL SAMPLES cessing and ANN classifications was around 15 s/image.

IV. DISCUSSIONS

An intelligent computer program to identify helminth eggs

from fecal microscopic images has been developed and veri-
ANN-2 identified their species and displayed their names in fied through sample images. The developed program consists
Fig. 9(b) and (d), respectively. of two major steps: digital image processing and a two-stage
726 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 48, NO. 6, JUNE 2001

TABLE V
DETAILS OF ANN-2 CLASSIFICATION RESULTS IN A CONFUSION MATRIX FORM

TABLE VI
COMPARISONS OF ANN-1 AND ANN-2 WITH SIMPLE k -NEAREST NEIGHBOR CLASSIFIERS

ANN-based classifier. The results showed acceptable perfor- produced erroneous results in minor cases especially when
mances with an 84% detection ratio and an 83% identification there is any nearby material around the object of concerns.
ratio and proved the applicability of the developed algorithms More complicated and accurate segmentation techniques
to full automatic examination system. such as [12]–[17] are expected to improve the results and
The visual recognition results used as ground truths in this being investigated for possible simpler form.
study are considered very reliable since they had been derived The practical fecal examination is based not only on the exis-
from well-trained human experts and also cross checked by tence of helminth eggs but also on the number of their findings.
other diagnostic methods, for example, DNA analysis. Since the helminth normally produces a large number of eggs,
The ANN errors revealed in the detection and identifica- there must be many of them in a fecal specimen. Therefore, any
tion results are thought to be caused mainly by incomplete few positive findings, whether True or False, need further ex-
segmentation results rather than by the ANN performance aminations such as blood and tissue analysis.
itself. The errors involved in poorly segmented objects There is no general rule for selection of proper network
affect ANN at the training phase as well as the test (ap- topology and proper data sample size for the best performance.
plication) phase by providing inaccurate numeric features. We studied the dependency of the optimal network topology
Despite its advantages over the common edge detection on the available data size by comparing the results with
methods specifically for microscopic images, the binary available data of 52 and 82. The numbers of neurons (input–1st
thresholding method used in this study results in tradeoff hidden–2nd hidden–output layer) in ANN-1 and ANN-2 for
between accuracy and complexity. In majority of images, the best results were changed from 4-6-6-1 to 4-8-8-1 and from
this method showed good segregation result. However, it 3-6-6-7 to 3-8-8-7, respectively. This indicates that there exists
YANG et al.: AUTOMATIC IDENTIFICATION OF HUMAN HELMINTH EGGS ON MICROSCOPIC FECAL SPECIMENS 727

(a)

(b)
Fig. 9. Typical results of ANN-1 and ANN-2: (a) detection result of C. sinensis. (b) Identification result of (a).

a relationship between the optimal network topology and the In this study, ANN classification was performed by two
available sample size in our specific application. The adequacy sub classifiers. This separation made the training easier and
of the sample size of 82 still has not been verified. The proper reduced total processing time in real application to images.
size of data sample is highly related to the distributional Most ANNs classify given patterns according to their geo-
characteristics of the given data samples. For example, in a metric clustering in the pattern space [8]. In the case of
case where the given data samples represent the classification fecal images, however, the artifacts are too dispersed in the
problem well, relatively small number of data would provide feature space to be treated as one or few couples of classes.
satisfactory result. This would increase the complexity of ANN, that is, the
728 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 48, NO. 6, JUNE 2001

(c)

(d)
Fig. 9. (Continued.) Typical results of ANN-1 and ANN-2. (c) Detection result of C. philippinensis and (d) identification result of (c).

number of nodes and connections. Therefore, egg finding in The computed time for the overall procedure, including digital
the segmented objects was implemented separately from the image processing and ANN classifiers was approximately 15
identification of parasite species. The resultant two ANNs s/image. By discarding objects smaller than the minimum egg
have much simpler architectures than they would have in size at the segmentation, which eliminated unnecessary feature
one incorporated network. Moreover, this separation reduced extraction, the computational load was alleviated. By the way, one
the number of computations by skipping the application of specimen slide usually consists of 60 view fields under 200-fold
ANN-2 to the artifacts preremoved by ANN-1. magnified views. Therefore, if the developed algorithm is applied
YANG et al.: AUTOMATIC IDENTIFICATION OF HUMAN HELMINTH EGGS ON MICROSCOPIC FECAL SPECIMENS 729

to the entire section of slide plate, it will take about 15 min to [16] N. Shareef, D. L. Wang, and R. Yagel, “Segmentation of medical images
process. This length of time, which appear rather lengthy, can be using LEGION,” IEEE Trans. Med. Imag., vol. 18, pp. 74–91, Jan. 1999.
[17] M. D. Srinath, P. K. Rajasekaran, and R. Viswanathan, Introduction to
easilyreducedbyhalformorebyimplementingroutinesoptimally Statistical Signal Processing with Applications. Englewood Cliffs, NJ:
with professionalprogramming languages and compilation. Prentice-Hall, 1996.
The targets for the developed algorithm were seven typical
human helminth eggs. For the identification of more parasitic
species than in this study, a detailed analysis of materials in- Yoon Seok Yang (S’00) received the B.S. degree in
side the shell membrane [4] under 400-fold or higher magnifi- control and instrumentation engineering from Seoul
cation is indispensable. Much more features than employed in National University, Seoul, Korea, in 1996 and M.S.
degree in interdisciplinary course (biomedical engi-
this study may be possible and even the best features, in terms neering major) from Yonsei University, Seoul, Korea,
of maximal variances and minimal redundancy, would be able in 1998. He is currently a Ph.D. degree candidate in
to derived from them using dimensionality reduction methods the Medical Electronics Laboratory (MELab) of the
Seoul National University.
like Karhunen–Loéve transformation [17]. Future work will ac- From 1996 to 1998, he worked as a Research
commodate this procedure. Member at the Department of Biomedical En-
The ultimate goal of this study is to develop an automatic exam- gineering of Yonsei University Medical Center.
His main research topic was an early detection of breast cancer in digital
ination system. Two different types of processing are possible. In mammogram. In 1998, he joined the Medical Electronics Laboratory (MELab)
an online processing system, the program performs the parasitic of the Seoul National University, where the major research activities are
egg identification on one digital image before the next scan. In development of portable electronics and intelligent algorithms for medical and
biological applications such as artificial pancreas, biosensors, telemedicine,
this case, total processing time should be reduced significantly. and man-machine interface. His interests include signal processing. From
UsingthelargestoragecapacityofamodernPC,anoff-linesystem 1999 to 2000, he worked on the ultrasonic biologic tissue characterization
scans the whole slide first and stores all digital images. Then, the project funded by MEDISON, Korea, that develops various ultrasonic medical
equipments.
program retrieves the stored images and carries out the identifica- Mr. Yang is a student member of Korea Society of Medical and Biological
tion. The developed algorithm will be an essential part of the main Engineering (KOSOMBE).
software of a completely automated system.

Duck Kun Park received the B.S. degree in elec-

REFERENCES trical engineering from Seoul National University,
[1] L. R. Ash and T. C. Orihel, Atlas of Human Parasitology, 3rd Seoul, Korea, in 2000. He is a M.S. degree candidate
ed. Chicago, IL: ASCP, 1990. in the Medical Electronics Laboratory (MELab) of
[2] A. Joachim, N. Dulmer, and A. Daugschies, “Differentiation of two Oe- the Seoul National University.
sophagostomum spp. from pigs, O. dentatum and O. quadrispinulatum, Since 2000, he has been working in MELab. His
by computer-assisted image analysis of fourth-stage larvae,” Parasitol. interests include man-machine interface, augmented
Int.l, vol. 48, pp. 63–71, 1999. reality, neuroscience, bioMEMS, wireless portable
[3] A. Daugschies, S. Imarom, and W. Bollwahn, “Differentiation of porcine health care system. He helped the production of
Eimeria spp. by morphologic algorithms,” Veterinary Parasitol., vol. 81, EOG-based glasses type wireless man-machine
pp. 201–210, 1999. interface. He has developed a game interface using
[4] C. Sommer, “Quantitative characterization of texture used for identifi- EMG and acceleration together.
cation of eggs of bovine parasitic nematodes,” J. Helminthol., vol. 72, Mr. Park is a student member of Korea Society of Medical and Biological
no. 2, pp. 179–182, 1998. Engineering (KOSOMBE).
[5] , “Digital image analysis and identification of eggs from bovine
parasitic nematodes,” J. Helminthol., vol. 70, pp. 143–151, 1996.
[6] , “Quantitative characterization, classification and reconstruction
of oocyst shapes of Eimeria species from cattle,” Parasitology, vol. 116, Hee Chan Kim (M’95) received the Ph.D. degree in
pp. 21–28, 1998. control and instrumentation engineering (biomedical
[7] R. C. Gonzalez, Digital Image Processing. New York: Ad- engineering major) from Seoul National University,
dison-Wesley, 1993. Seoul, Korea, in 1989.
[8] J. A. Freeman and D. M. Skapura, Neural Networks: Algorithms, Appli- From 1982 to 1989, he has worked as a Research
cations, and Programming Techniques. New York: Addison-Wesley, Member at the Department of Biomedical Engi-
1992. neering of Seoul National University Hospital. From
[9] R. J. Sohalkoff, Pattern Recognition: Statistical, Structural and Neural 1989 to 1991, he was a Staff Engineer working on
Approaches. New York: Wiley, 1992. National Institute of Health (NIH) funded electro-
[10] J. T. Tou and R. C. Gonzalez, Pattern Recognition Principles. New hydraulic total artificial heart project at the Artificial
York: Addison-Wesley, 1981. Heart Research Laboratory of the University of Utah,
[11] K. Fukunaga, Introduction to Statistical Pattern Recognition. New Salt Lake City, USA. He joined the faculty of the Department of Biomedical
York: Academic, 1990. Engineering, College of Medicine, Seoul National University in 1991, where
[12] R. N. Czerwinski, D. L. Jones, and W. D. O’Brien, “Line and boundary he is currently an Associate Professor. He worked with the University of Utah
detection in Speckle images,” IEEE Trans. Image Processing, vol. 7, pp. Group again at the Department of Pharmaceutics and the Artificial Heart
1700–1714, Dec. 1998. Research Laboratory as a Visiting Professor from 1993 to 1994. He is now
[13] C. W. Chen, J. Luo, and K. J. Parker, “Image segmentation via adap- leading the Medical Electronics Laboratory (MELab) of the Seoul National
tive K-Mean clustering and knowledge-based morphological operations University, where the major research activities are development of intelligent
with biomedical applications,” IEEE Trans. Image Processing, vol. 7, algorithms and electronic instrumentations for medical and biological applica-
pp. 1673–1683, Dec. 1998. tions including artificial organs such as artificial heart and artificial pancreas,
[14] J. M. Gauch, “Image segmentation and analysis via multiscale gradient biosensors, telemedicine, and man-machine interface. In these areas, he has
watershed hierarchies,” IEEE Trans. Image Processing, vol. 8, pp. published over 38 peer-reviewed scientific papers in international journals.
69–79, Jan. 1999. Dr. Kim is a member of Korea Society of Medical and Biological Engineering
[15] M. A. Kupinski and M. L. Giger, “Automated seeded lesion segmen- (KOSOMBE), Institute of Electrical and Electronics Engineers/Engineering in
tation on digital mammograms,” IEEE Trans. Med. Imag., vol. 17, pp. Medicine and Biology Society (IEEE/EMBS), and American Society of Artifi-
510–517, Aug. 1998. cial Internal Organs (ASAIO).
730 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 48, NO. 6, JUNE 2001

Min-Ho Choi received the Ph.D. degree in parasite Jong-Yil Chai received the M.D. and Ph.D. (medical
biochemistry (medical parasitology major) from parasitology) degrees from Seoul National University
Seoul National University, Seoul, Korea, in 1996. College of Medicine, Seoul, Korea, in 1976 and 1984,
From 1989 to 1993, he has worked as a Research respectively.
Assistant at the Department of Parasitology, Seoul He has been with the Department of Parasitology,
National University College of Medicine, Seoul Seoul National University College of Medicine since
Korea. He was a Public Health Doctor of National 1976. He is now Professor and Chairperson of the De-
Institute of Health, Korea, from 1993 to 1996, where partment. For abroad studies, he worked on mucosal
he counseled HIV-infected persons in Korea. He immunity to protozoan parasites, from 1986 to 1987,
joined the faculty of the Department of Parasitology, in Animal Parasitology Institute, US Department of
Seoul National University College of Medicine in Agriculture, Beltsville, MD, as a Visiting Scientist.
1996, where he is currently an Assistant Professor. He is now a Visiting Scholar He also studied on immunity to intestinal parasites, in 1993, in the Department
at the Department of Pathology, Division of Infectious Diseases, University of of Parasitology, Institute of Medical Science, The University of Tokyo, Japan, as
California, San Diego Medical Center, San Diego, CA, and he is studying the a Visiting Professor. He has published over 300 peer-reviewed academic papers
pathogenesis of amebiasis and transgenic amebae. in the field of medical parasitology and one textbook on clinical parasitology,
Dr. Choi is a member of Korea Society for Parasitology and the American and contributed many chapters of textbooks on clinical medicine. His research
Society of Parasitology. area includes a wide range of parasites of human beings including protozoa
and helminths, but he has been mostly interested in the researches on intestinal
flukes.
Dr. Chai is currently the President of Korean Society for Parasitology, and
member of American Society of Tropical Medicine and Hygiene, American
Society of Parasitologists, Society of Protozoologists, Society for Mucosal Im-
munology, and Helminthological Society of Washington, Royal Society of Trop-
ical Medicine and Hygiene, and Japanese Society of Parasitology.