JID: COMM

ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

Coronary angiography image segmentation based on PSPNet

Xiliang Zhu, Zhaoyun Cheng, Sheng Wang∗, Xianjie Chen, Guoqing Lu
Department of Cardiovascular Surgery, Henan Province People’s Hospital, Fuwai Central China Cardiovascular Hospital, Henan Cardiovascular Hospital and
Zhengzhou University, Zhengzhou, China

a r t i c l e i n f o a b s t r a c t

Article history: Purpose: Coronary artery disease (CAD) is known to have high prevalence, high disability and mortality.
Received 13 September 2020 The incidence and mortality of cardiovascular disease are also gradually increasing worldwide. Therefore,
Accepted 30 November 2020
our paper proposes to use a more efficient image processing method to extract accurate vascular struc-
Available online xxx
tures from vascular images by combining computer vision and deep learning.
Keywords: Method: Our proposed segmentation of coronary angiography images based on PSPNet network was com-
Coronary angiography images pared with FCN, and analyzed and discussed the experimental results using three evaluation indicators of
blood vessel segmentation
precision, recall and Fl-score. Aiming at the complex and changeable structure of coronary angiography
deep learning
multi-scale convolutional neural network
images and over-fitting or parameter structure destruction, we implemented the parallel multi-scale con-
transfer learning volutional neural network model using PSPNet, using small sample transfer learning that limits parameter
learning method.
Results: The accuracy of our technique proposed in this paper is 0.957. The accuracy of PSPNet is 26.75%
higher than the traditional algorithm and 4.59% higher than U-Net. The average segmentation accuracy
of the PSPNet model using transfer learning on the test set increased from 0.926 to 0.936, the sensitivity
increased from 0.846 to 0.865, and the specificity increased from 0.921 to 0.949. The segmentation effect
in this paper is closest to the segmentation result of the human expert, and is smoother than that of
U-Net segmentation.
Conclusion: The PSPNet network reduces manual interaction in diagnosis, reduces dependence on med-
ical personnel, improves the efficiency of disease diagnosis, and provides auxiliary strategies for subse-
quent medical diagnosis systems based on cardiac coronary angiography.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction ability to examine the structure of coronary arteries, and is the

most commonly used technology from in the clinical diagnosis
Coronary artery disease (CAD) has a high incidence and high fa- of CAD. This diagnosis and treatment method has opened an-
tality rate, and is one of the most important causes of death world- other door for people to understand CAD. It can clearly and in-
wide [1-3]. The pathophysiological basis of CAD is atherosclerotic tuitively display the detailed images of the coronary arteries, al-
lesions in coronary arteries, which will cause vascular stenosis or lowing doctors and patients to observe that the coronary arter-
blockage. The primary way of early diagnosis of coronary heart dis- ies have vascular wall calcification and stenosis problems, which
ease is the detection and quantification of coronary artery stenosis. make the diagnosis of coronary heart disease more intuitive and
With the constant understanding of diseases and advancement scientific.
of imaging technology, these have led to the continuous emer- CAG image blood vessel segmentation is of great significance
gence of new diagnostic methods. Coronary angiography (CAG) in the assessing of the degree of vascular disease, assisting doc-
is regarded as the gold standard in clinical treatment [4]. This tors in diagnosis and treatment, and reconstruction of the three-
is a non-invasive method of examining coronary arteries. It does dimensional structure of blood vessels [6-8]. A good segmentation
not require an arterial catheter, but allows the contrast fluid to result can save the doctor’s manual segmentation time and help
flow through the heart by intravenous injection, and complete the the doctor correctly diagnose the condition. The segmented blood
heart scans [5]. X-ray angiography has a powerful fluoroscopic vessels can be used for further research, such as blood flow ve-
locity, degree of vascular stenosis, etc., to lay the foundation for
∗ subsequent vascular dynamics research and three-dimensional re-
Corresponding author.
E-mail address: wangshengxw@163.com (S. Wang).
construction.

https://doi.org/10.1016/j.cmpb.2020.105897
0169-2607/© 2020 Elsevier B.V. All rights reserved.

Please cite this article as: X. Zhu, Z. Cheng, S. Wang et al., Coronary angiography image segmentation based on PSPNet, Computer
Methods and Programs in Biomedicine, https://doi.org/10.1016/j.cmpb.2020.105897
JID: COMM
ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

X. Zhu, Z. Cheng, S. Wang et al. Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Convolutional neural network (CNN), relying on weight sharing, from the original grayscale or color value transformation [14]. The
automatic feature extraction and computer performance improve- threshold segmentation method is actually a process of assuming
ment, has achieved remarkable results [9]. The medical community a criterion function and then using it to find the optimal threshold
has also noticed the great success of deep learning methods and solution. The frequently used threshold selection methods include
hopes to apply these technologies to different tasks in medical im- the maximum inter-class variance method, the maximum entropy
age processing. automatic threshold method, and the gray histogram method.
The use of deep learning technology to segment coronary ves- At present, image threshold segmentation is also used in many
sels has three advantages. The first point is that it can automati- fields, for example, in medical applications, blood cell image seg-
cally extract features [10]. In the past, meaningful or task-related mentation, as well as image segmentation in magnetic resonance
parts were mainly identified by human experts. Based on their un- have also used threshold segmentation algorithms.
derstanding of these images, these features can be carried out by
non-experts in the field, using machine learning techniques. Now, 2.1.2. Region segmentation
deep learning has incorporated feature engineering steps into the Region-based segmentation methods comprise region growth
scope of learning steps. In other words, there is no need to manu- and split and merge. The region growing algorithm artificially di-
ally extract features. If necessary, deep learning only needs a small vides the image into n similar regions, and then gradually connects
amount of pre-processed data, and then new features can be dis- and merges the adjacent regions according to a judgment criterion.
covered in a self-learning manner. The region splitting algorithm is just the opposite. There is no pro-
The second point is that deep learning technology has good ver- cess of artificially dividing regions. First, the input image is made
satility [11]. Because of the difference in imaging environment and a complete region, then it starts to split, and finally the similar re-
the level of radiologists, even the same patient will have differ- gions are merged. These two algorithms are serial algorithms, that
ent CAG images [12]. The traditional method uses the features de- is, the process of segmentation is multiple steps in sequence, and
signed by experts, who may have a good effect on a certain kind the subsequent operations are based on the results of the previous
of image, but it is not universal. Deep learning technology uses step and then continue.
a large number of high-quality labeled samples for training. This Based on the region segmentation algorithm, it is divided into
method can be applied to various types of contrast images. seed region growth and Region growing algorithm [15]. Seed re-
The third point is that deep learning can improve efficiency and gion growth algorithm groups similar pixels to form regions. First
detection accuracy [13]. Coronary vascular images will inevitably find a seed pixel for each region that needs to be segmented as
introduce various noises during the imaging process, and the con- the starting point for growth, and then merge the pixels in the
trast agent will be unevenly distributed in the blood vessels due to neighborhood around the seed pixel that have the same or simi-
various reasons. The vascular bifurcation will cause the traditional lar properties as the seed into the seed where the pixel is located.
methods to be slow in the segmentation process. The accuracy is The Region growing algorithm inherited the idea of the seed re-
low and a lot of labor time is wasted. Deep learning technology re- gion growth algorithm. It uses the Celular Automaton (CA) as the
lies on powerful feature extraction and classification. This enables segmentation frame to transform the category information into the
deep learning technology to be more efficient and detection accu- state of the cell. According to features such as location and color,
racy than traditional methods. the location is adjacent to the color. Similar pixels change and
evolve to the same category.
2. Methodology
2.2. Image segmentation using convolution neural network
2.1. Target segmentation based on traditional algorithms
In the field of computer vision, the most basic work is image
Currently, traditional image segmentation comprises threshold- segmentation, which is to divide similar parts of the picture into
based segmentation, region-based segmentation, edge-based seg- continuous area blocks, such as the roughness, contrast, direction,
mentation, and specific theories-based segmentation. With the and compactness of the image block, to assist in image segmen-
progress of research, segmentation algorithms are divided into tation and enhance the effect. In simple terms, semantic segmen-
pixel-based classification, threshold segmentation, edge detection, tation is to classify pixels on the picture given a picture. Early re-
color image segmentation, depth image segmentation, and fuzzy searchers used decision trees and deep learning methods, which
set-based methods. are CNNs (fully connected layer type). In recent years, the network
for semantic segmentation is basically based on FCN [16]. Other
2.1.1. Threshold segmentation network structures for semantic segmentation include Dilated Con-
Threshold segmentation is a region-based image segmentation volutions, PSPNet [17], Deeplab Vl [18], Deeplab V2 [19], Deeplab
algorithm. The idea is to use the feature f(x, y) to calculate one V3 [20] and so on.
or more specific thresholds, and compare the feature value of each Next, this paper will introduce in detail the multi-scale CNN
pixel with the threshold, and then divide each pixel into its cor- proposed for the problem that the artery is a thin tubular struc-
responding appropriate category based on the comparison result, ture in the coronary vascular image, which has relatively low con-
which is shown in Equation 1. trast and artifacts, is difficult to accurately segment and effectively

1, f (x, y ) ≥ T annotate the scarcity of samples. Usually when using a deep CNN
g ( x, y ) = (1) model, the segmented object is a cropped coronary vascular image
0, f (x, y ) ≤ T
patch. Different from the past, this paper uses two different scales
The actual image target and background may not necessar- of patches for segmentation, one is obtained by cropping the orig-
ily be distributed in the two characteristic value ranges, the inal image; the other is obtained by cropping the original image
Equation (1) can be transformed into Equation 2. after downsampling, so that global features can be used to pro-
 vide local features reference to make pixel segmentation more ac-
1, T1 ≤ f ( x, y ) ≤ T2
g ( x, y ) = (2) curate. The result of segmentation is whether the patch is a proba-
0, else
bility map of a blood vessel, and then all the probability maps are
Commonly used feature values can be the grayscale or color spliced together to obtain a complete coronary blood vessel seg-
feature values of the original image, or the features obtained mentation image.

2
JID: COMM
ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

X. Zhu, Z. Cheng, S. Wang et al. Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Fig. 1. Multi-scale CNN segmentation flowchart

Fig. 2. Schematic diagram of multi-scale image

The entire coronary artery segmentation method is shown in 2.3.2. PSPNet network structure
Fig. 1. Fig. 2 is a schematic diagram of the multi-scale image that Pre-training in PSPNet uses a combination of ResNet and Dik-
can be seen. The core part is the construction of the segmentation ted Network. First, feature extraction is performed on the input
network model. In this paper, the segmentation model obtained by data, and the size of the extracted feature map is one-eighth of
multi-scale CNN and improved loss function has achieved good re- the original input image. Then input the feature map into the pyra-
sults in coronary artery segmentation. mid pooling module, and finally cascade to get the feature surface,
and then learn and adjust the parameters by calculating the loss
and back propagation algorithm. Compared with Global Pooling,
this structure is better for obtaining multi-size global information.
2.3. PSPNet network Comparing with the calculation, the FCN network will not increase
a lot. The Global Pyramid Poling model and the FCN feature extrac-
2.3.1. Introducing the PSPNet network tion model can be trained and optimized at the same time.
Using an image block around the pixel as the input of the CNN The network structure of CAG image segmentation using PSPNet
network for training and prediction is the characteristic of tradi- can be presented using Fig. 3. Firstly, the data of CAG is labeled:
tional CNN-based segmentation methods. This method has several it is divided into two steps: rough labeling and fine labeling, and
disadvantages. First, it requires a lot of storage space; second, the then classification and training test are carried out according to the
segmentation time is long, but the segmentation accuracy is not division of body positions in medicine. Then, the annotated image
high; third, the size of the pixel block often affects the size of the is processed into a grayscale image and input into the Cafe frame.
sensing area. FCN overcomes the above shortcomings by recovering After the image is combined with convolution pooling in the net-
the category to which each pixel belongs from the abstract fea- work, Feature Map is extracted, and then it is put into the pyramid
tures. pooling layer. The pyramid pooling layer will convolve the feature
Although FCN realizes the classification of each pixel, it does maps in different sizes, then cascade and integrate the output, and
not fully consider the relationship between pixels. Ignore the spa- then go through other convolutions and Dropout operations, and
tial regularization step used in the usual pixel classification-based finally use the Softmax classifier to classify the pixels, and achieve
segmentation method, which makes the network lack of spatial the final prediction, and complete the segmentation experiment.
consistency.
In the 2016 ImageNet scene segmentation challenge, Zhao et al. 2.4. U-Net
proposed Pyramid Scene Parsing Network (PSPNet) [17]. The scene
proposed in the paper is accurate. The knowledge graph relies on 2.4.1. Development of U-Net
the prior information of the scene context, and the main problem Long et al. [21] implemented fully connected neural network,
of the FCN model is the lack of appropriate strategies to use the referred to as FCN. FCN is a neural network architecture, which re-
category clues in the global scene. alizes object detection by combining low-level appearance infor-

3
JID: COMM
ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

X. Zhu, Z. Cheng, S. Wang et al. Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Fig. 3. CAG image segmentation based on PSPNet

mation with high-level semantic information. In 2017, Kim et al.

[19] suggested a multi-stage FCN segmentation method to solve
the problem of the rough segmentation boundary of skin lesions,
in which multiple FCNs learn complementary visual features. Ac-
cording to the period of skin lesions, the early FCN learns rough
appearance and positioning information, while the later FCN learns
the subtle features of the lesion boundary. They also introduced a
new parallel integration method that combines the supplementary
information from each segmentation stage to obtain the final seg-
mentation result with accurately positioned lesion boundaries.
In 2018, Nie et al. [22] presented a method for segmenting
phase brain images using multiple fully convolutional networks
(mFCNs). They used T1-weighted images, T2-weighted images, and
FA modal images instead of simply combining the original Three-
modal data feature map. The deep architecture they adopted can
effectively fuse advanced information from the three images. In
this method, the author trained a network model for each mode
to effectively use information from multiple image forms.
A relatively new CNN architecture called U-Net convolutional
network [23] is dedicated to biomedical image analysis. This net-
work was originally proposed to solve the cell segmentation task. Fig. 4. U-Net’s network structure
Essentially, U-Net is a technology that combines CNNs and data
enhancement technology to increase the amount of medical data.
On the basis of U-Net, Lien Kamp et al. [24] introduced the 3D U-
Net in 2016. Unlike ordinary images that are two-dimensional data,
many types of medical images are 3D. The adjacent image slices ages. This is also related to the characteristics of the medical image
display almost the same information, which makes it very difficult itself. First, the medical image has a single content and a relatively
and inefficient to label the sample data. And 3D U-Net directly in- fixed form. The U-Net’s U-shaped structure and layer jump connec-
puts 3D images, and uses 3D convolution, 3D pooling and other tion make it possible to better extract high-level semantic informa-
operations accordingly, which can output 3D probability maps. This tion and preserve low-level structural features. Secondly, the vol-
method only needs some labeled two-dimensional slices, making ume of data in medical images is relatively small, and the param-
its application range very wide. eters of the U-Net model are less than other deep learning mod-
els, but the effect is almost the same. Finally, medical images need
more universal models than other images. Medical image formats
2.4.2. U-Net network structure are diverse, and U-Net can input images of any size. After improve-
Fig. 4 presents the framework pertaining to U-Net. Compared ments, it can also input three-dimensional images, which has good
with other natural images, U-Net performs better on medical im- versatility.

4
JID: COMM
ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

X. Zhu, Z. Cheng, S. Wang et al. Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Fig. 5. Schematic diagram of transfer learning

2.5. Transfer learning technique angiography image data is collected. This study established a
database derived from real patient cases. We collected CAG images
Machine learning technologies including deep learning have of 109 patients from the Fuwai Central China Cardiovascular Hos-
made brilliant achievements in many fields including classification, pital.
regression and clustering. However, these methods all assume that
excellent results can be obtained only when the training and test- 3.1.2. Data preprocessing
ing data have the same feature space and the same distribution. During subsequent training, the network can only accept
As sample data distribution changes, algorithms need to add addi- grayscale images as input, so we convert the blood vessel RGB
tional training data and then start training from scratch. In many value to a grayscale value from 0 to N-1 according to the impor-
real-world application scenarios, the cost of this method is high, tance of the blood vessel, where N is the type of blood vessel in
not to mention that many tasks cannot collect the data required each position. Considering that some blood vessels have similar
for retraining. These tasks are in great need of methods that can functions, they are converted to the same gray value.
reduce recollection of data and training data. In this case, knowl-
3.1.3. Establishment of data set
edge transfer or transfer learning between application domains is
We prepared the CAG data set as follows. First, we introduced
of practical significance. Schematic diagram of transfer learning as
the source of the original data, desensitization processing, data
Fig. 5.
selection and classification, and classified it according to medical
2.6. Evaluation criteria for experimental results common sense according to different lesions and different posi-
tions; and then designed according to needs the labeling tool was
For the task of vessel segmentation in CAG images, samples are implemented, and the data was roughly labeled and fine labeled.
divided into two categories: vascular and non-vascular. By com- Part of the process of fine labeling of some angiographic images
paring the results of segmentation with the manual segmentation was introduced. Eventually, we have a series of CAG images, which
results of experts, four situations can be obtained, namely: blood are labelled using a serial number. Finely annotated images are di-
vessel pixels are predicted to be blood vessels, which is called vided into test set and experimental set with a ratio of 2:1.
true positive (TP); blood vessel pixels are predicted to be non-
3.2. Segmentation experiment results
vessels This is expressed as false negative (FN); non-vascular pixels
predicted to be non-vascular are called true negatives (TN); non-
To obtain intuitively the effectiveness of the method of using
vascular pixels predicted to be blood vessels are called false posi-
the deep learning model to segment coronary arteries proposed in
tives (FP).
this paper, we tested the 10 images of the test set and gave their
We often use three criteria to compare the performance of the
human expert results as a reference standard. Fig. 6(A) shows two
proposed method with other state-of-the-art methods: sensitivity
CAG images with different blood vessel directions selected in this
(Sen), specificity (Spe) and accuracy (Acc). The calculation method
paper (noted as CAG image serial number 1-5). The coronary artery
is shown by the following Equations (3), (4), and (5).
in the left picture is smaller and has more background noise than
TP the coronary artery in the right picture. Figure (B) is the result of
Sen = (3)
TP + FN the human expert. Figure (C) is the segmentation result of U-Net.
Figure (D) is the segmentation result of the PSPNet network. As an
TN
Spe = (4) illustration of our experimental results, we only show the segmen-
TN + FP
tation of CAG images based on serial numbers 1-5.
TP + TN Fig. 6 shows that the segmentation result based on the PSPNet
Acc = (5) network is closer to the result of the human expert segmentation.
TP + TN + FN + TN
Compared with U-Net network, the algorithm segmentation edge
3. Results of this paper is smoother.

3.1. Data set 3.3. Performance comparison

3.1.1. Data source In order to explore whether the segmentation result of the deep
We will perform segmentation tasks and experiments on neural network is better than the segmentation result of the tra-
coronary angiography images. At present, the available coronary ditional algorithm, the traditional algorithm takes the Threshold

5
JID: COMM
ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

X. Zhu, Z. Cheng, S. Wang et al. Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Fig. 6. Experimental results of segmentation results based on two image serial numbers A) Original image; B) Human experts segmentation result; C) U-Net method seg-
mentation result; D) PSPNet network segmentation result

Table 1 0.949. This shows that the transfer learning strategy in this paper
Comparison of segmentation performance of three algo-
can improve the model performance.
rithms

Algorithm Sen Spe Acc

4. Discussion
Threshold segmentation 0.758 0.830 0.802
Region growing 0.769 0.786 0.755
U-Net 0.911 0.906 0.915 At present, deep learning has achieved good diagnostic accuracy
PSPNet 0.947 0.951 0.957 comparable to clinical experts and medical doctors in the field of
medical imaging. However, in general, there are still many chal-
lenges in applying deep learning to the field of medical image pro-
segmentation and Region growing algorithm as an example, and cessing.
compares it with U-Net and PSPNet. The comparison results are as First, there is still a gap between the segmentation accuracy and
follows. the current best method. Although deep learning has proven its
Table 1 shows that the segmentation accuracy based on the great success in other fields, it has just entered the field of coro-
deep neural network is much higher than that of the traditional nary artery segmentation, and many methods still use the same
algorithm, and the CAG images based on the PSPNet network are methods when processing ordinary images. However, the tradi-
higher than U-Net in Sen, Spe, and Acc. The accuracy of PSPNet tional methods have made many improvements to the various seg-
is at least 10% higher than the traditional algorithms, and 4.59% mentation difficulties of CAG images, and there are many targeted
higher than U-Net. and effective techniques. But even so, the accuracy gap between
the two is not that big, which shows that deep learning has great
3.4. Results of transfer learning experiments potential and can be further studied and explored.
Second, there is a serious imbalance in the number of sam-
Table 1 shows that the segmentation accuracy, and based on ples in the model training process. The current mainstream clas-
the PSPNet network is the highest. On the basis of PSPNet, the ac- sification algorithms include support vector machines (SVM), neu-
curacy change when using the transfer learning model is tested. ral networks, etc., all of which assume that the samples are in a
Fig. 7 is a comparison of model performance before and af- balanced state and aim for high overall accuracy. However, there
ter using transfer learning. It can be seen that after using trans- is a serious sample imbalance in coronary vascular images, and
fer learning, the accuracy, specificity, and sensitivity of the model sometimes the ratio of positive and negative samples is very dif-
proposed in this paper have been improved in coronary artery seg- ferent. For example, the number of vascular pixels in an image is
mentation. Among them, the average segmentation accuracy on only one-fifth or less of the number of non-vascular pixels. Each
the test set increased from 0.926 to 0.936, the sensitivity increased pixel is a training sample, and the final total number of positive
from 0.846 to 0.865, and the specificity increased from 0.921 to and negative samples differs from 103 to 104 [25]. Such a severe

6
JID: COMM
ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

X. Zhu, Z. Cheng, S. Wang et al. Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Fig. 7. The improvement of model performance by transfer learning based on A) accuracy; B) sensitivity; C) specificity

situation may lead to model degradation and unable to complete and is difficult to accurately segment and effectively annotate the
normal classification tasks. scarcity of samples. A PSPNet-based multi-scale CNN model is pro-
Finally, the sample labeling quality is low and the quantity is posed. This algorithm first down-sample the pre-processed image
small. In this case, the trained model is likely to have overfitting or to obtain images of multiple scales, and sends them to the CNN.
poor robustness. Just as the computer vision field achieves break- The features of different scales are selected and fused in the fully
through improvements by training tens of millions of natural im- connected layer, and finally the features are used to segment the
age data, deep learning models also need a large, public medical CAG images.
image data set in order to find more general features, reduce train- It has been confirmed that the coronary angiography image seg-
ing time, and improve training performance [26,27]. However, the mentation technology based on PSPNet has higher segmentation
establishment of such a medical image data set requires high time accuracy than the traditional Region growing algorithm and U-Net.
and economic costs, and the workload is huge. In order to further improve the segmentation accuracy, we propose
The segmentation of CAG images is an important practice and migration learning. We use transfer learning technology to solve
innovation of computer vision in the medical field. CAG is also the problem of limited training data when using deep neural net-
widely used clinically because of its safety and accuracy, and is works to segment coronary vessels. Usually transfer learning is to
considered the gold standard for the diagnosis of coronary heart train in the source domain to obtain the model, and then train to
disease [26,28,29]. In the future, image segmentation based on CAG update the model in the target domain. However, because the tar-
can not only assist doctors in the diagnosis and treatment of clin- get domain has less data, the trained model parameters will be
ical diseases, but also become a prerequisite for the quantitative destroyed after fine-tuning, resulting in overfitting. Therefore, our
description of vascular diseases, and it can also establish a foun- paper improves the strategy of using transfer learning for small
dation for the development of three-dimensional reconstruction of samples, and makes the newly trained model parameters close to
cardiac blood vessels [30,31]. Therefore, the experiments in this pa- the source domain model by adding regular terms. Finally, after us-
per are of great significance to the accurate segmentation of CAG ing transfer learning, the accuracy, specificity and sensitivity of the
images. model proposed in this paper in coronary artery segmentation are
improved.
Note that this paper compares traditional image segmentation
5. Conclusion methods and methods using deep learning. The accuracy of deep
learning is much higher than that of traditional ones, and our pro-
Based on the coronary vascular imaging, the artery is a thin posed PSPNet-based CAG image segmentation accuracy is the high-
tubular structure, which has relatively low contrast and artifacts, est.

7
JID: COMM
ARTICLE IN PRESS [m5G;December 11, 2020;6:31]

X. Zhu, Z. Cheng, S. Wang et al. Computer Methods and Programs in Biomedicine xxx (xxxx) xxx

Declaration of Competing Interest [15] Vezhnevets V, Konouchine V. "Region growing" - Interactive Multi-Label ND
Image Segmentation by Cellular Automata [J]. Graphicon-2005, Novosibirsk
Akademgorodok, 2005:3-4.
No. [16] X. Wu, Fully Convolutional Networks for Semantic Segmentation [J], Computer
Science 8 (21) (2015) 3–6.
References [17] Liang-Chieh Chen, George Papandreou Iasonas Kokkinos, Kevin Murphy, Alan L.
Yuille. Semantic Image Segmentation with Deep Convolutional Nets and Fully
[1] R L Mcmillan, Coronary atherosclerotic heart disease [J], North Carolina medi- Connected CRFs [J]. arXiv:1412.7062, 2014:1-4.
cal journal 19 (4) (1958) 147–149. [18] Liang- Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, Alan
[2] N Sigfusson, G Sigurdsson, U Agnarsson, et al., Declining coronary heart dis- L. Yuille. Deep Lab: Semantic Image Segmentation with Deep Convolutional
ease mortality in Iceland: contribution by incidence, recurrence and case fa- Nets, Atrous Convolution, and Fully Connected CRFs [J], arXiv: 1606. 00915v2,
tality rate [J], Scandinavian Journal of Thoracic & Cardiovascular Surgery 36 (6) 2017: 2-3.
(2002) 337–341. [19] Liang-Chieh Chen, George Papandreou, Florian Schroff, Hartwig Adam Google
[3] W Yong, C. Wen-Xia, Clinical analysis of chronic cor pulmonale in elderly Inc. Rethinking Atrous Convolution for Semantic Image Segmentation, arXiv: 1
merges coronary atherosclerotic heart disease [J], Journal of Clinical Pulmonary 706.05587v117 Jun 2017:3-4.
Medicine 56 (6) (2010) 197–203. [20] J Long, E Shelhamer, T Darrell, Fully convolutional networks for semantic seg-
[4] C Jadrian, F J Novoa, F L Carlos, et al., Automatic multiscale vascular image mentation [C], in: Proceedings of the IEEE conference on computer vision and
segmentation algorithm for coronary angiography [J], Biomedical Signal Pro- pattern recognition, MA, USA, 2015, pp. 3431–3440.
cessing & Control 46 (2018) 1–9. [21] L Bi, J Kim, E Ahn, et al., Dermoscopic image segmentation via multi stage fully
[5] M T Dehkordi, S Sadri, A Doosthoseini, Retraction: A Review of Coronary Ves- convolutional networks [J], IEEE Transactions on Biomedical Engineering, IEEE
sel Segmentation Algorithms [J], Journal of Medical Signals and Sensors 9 (1) 64 (9) (2017) 2065–2074.
(2019) 76. [22] D Nie, L Wang, Y Gao, et al., Fully convolutional networks for multi –modality
[6] S Jiangping, Z Zhe, W Wei, et al., Assessment of coronary artery stenosis is intense infant brain image segmentation [C], in: 2016IEEE 13th International
by coronary angiography: a head-to-head comparison with pathological coro- Symposium on Biomedical Imaging (ISBI), IEEE, Prague, Czech Republic, 2016,
nary artery anatomy [J].Circulation: Cardiovascular Interventions, Journal of the pp. 1342–1345.
American Heart Association 6 (3) (2013) 262–268. [23] Ronnebrger O, Fischer P, BROX T. U-Net: Convolutional Networks for Biomedi-
[7] A G Blaiech, A Mansour, A Kerkeni, et al., Impact of Enhancement for Coro- cal Image Segmentation [J]. 2015, 9(20): 1-8.
nary Artery Segmentation Based on Deep Learning Neural Network[M], Pattern [24] A Abdulkadir, S S Lienkamp, et al., 3D U-Net: learning dense volumetric seg-
Recognition and Image Analysis 12 (9) (2019) 167–173. mentation from sparse annotation [C], in: International conference on medi-
[8] W Huang, L Huang, Z Lin, et al., Coronary Artery Segmentation by Deep Learn- cal image computing and computer- -assisted intervention, Springer, Athens,
ing Neural Networks on Computed Tomographic Coronary Angiographic Im- Greece, 2016, pp. 424–432.
ages[C], 2018 40th Annual International Conference of the IEEE Engineering in [25] M Li, Q Yin, M Lu, Retinal Blood Vessel Segmentation Based on Multi-Scale
Medicine and Biology Society (EMBC), IEEE, England, 2018. Deep Learning [C], in: 2018 Federated Conference on Computer Science and
[9] Huang Weimin, Lu, et al., Coronary Artery Segmentation by Deep Learning Information Systems (FedCSIS), IEEE, Poznan, Poland, 2018, pp. 1–7.
Neural Networks on Computed Tomographic Coronary Angiographic Images [26] X J Tang, S J Zhou, Q M Hu, Segmentation of Coronary CT Angiography Images
[C], Conference proceedings: Annual International Conference of the IEEE En- Based on Deformable Model with New Edge Measures [J], Applied Mechanics
gineering in Medicine and Biology Society, IEEE Engineering in Medicine and & Materials 333-335 (1) (2013) 888–896.
Biology Society, Honolulu, Hawaii, USA, 2018 Annual Conference. [27] T F Lee, C Y Lee, P J Chao, et al., Precision Segmentation Rendering for 3-D
[10] K Jo, J Kweon, H Kim Y, et al., Segmentation of the Main Vessel of the Left Coronary Angiography Medical Image [C], in: Fifth International Conference on
Anterior Descending Artery Using Selective Feature Mapping in Coronary An- Intelligent Information Hiding & Multimedia Signal Processing, IEEE Computer
giography [J], Transactions of Atmospheric Sciences 7 (2) (2019) 919–930. Society, Chicago, 2009, pp. 27–35.
[11] F Jiang, A Grigorev, S Rho, et al., Medical image semantic segmentation based [28] Au B, Shaham U, Dhruva S, et al. Automated Characterization of Stenosis in
on deep learning [J], Neural computing & applications 12 (3) (2018) 157 -132. Invasive Coronary Angiography Images with Convolutional Neural Networks [J].
[12] S Pan, W Zhang, W Zhang, et al., Diagnostic Model of Coronary Microvas- 2018, 7(27): 577-589.
cular Disease Combined With Full Convolution Deep Network With Balanced [29] A S Ashour, S Samanta, N Dey, et al., Computed Tomography Image Enhance-
Cross-Entropy Cost Function [J], IEEE Access 7 (1) (2019) 177997–178006. ment Using Cuckoo Search: A Log Transform Based Approach [J], Journal of
[13] P Moeskops, J M Wolterink, B H M V D Velden, et al., Deep Learning for Mul- Signal & Information Processing 6 (4) (2015).
ti-task Medical Image Segmentation in Multiple Modalities[C], International [30] C Adrian, F J Novoa, F L Carlos, et al., Automatic multiscale vascular image
Conference on Medical Image Computing and Computer-Assisted Intervention, segmentation algorithm for coronary angiography [J], Biomedical Signal Pro-
Springer International Publishing, Athens, Greece, 2016. cessing & Control 46 (2018) 1–9.
[14] M. Jia-Qing, Study of image segmentation based on morphology and watershed [31] H Qin, X. Huang, Coronary Angiography Image Segmentation and Skeleton Ex-
algorithms for coronary angiography [J], Modern Electronics Technique 3 (6) traction Based on Hessian Matrix [J], Journal of Data Acquisition & Processing
(2015) 115–120. 64 (5) (2016) 27–36.