Received 28 June 2024, accepted 23 July 2024, date of publication 29 July 2024, date of current version 7 August 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3435552

Combined Optic Disc and Optic Cup Segmentation

Network Based on Adversarial Learning
YONG LIU 1,2 , JIN WU1 , (Member, IEEE), YUANPEI ZHU3 , AND XUEZHI ZHOU2
1 School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, China
2 School of Medical Engineering, Xinxiang Medical University, Xinxiang 453003, China
3 School of Physics and Electronic Engineering, Xinxiang University, Xinxiang 453003, China

Corresponding author: Yong Liu ([email protected])

This work was supported in part by Henan Province Key Research and Development and Promotion Projects under Grant 232102310009,
and in part by the National Natural Science Foundation of China under Grant 82302298.

ABSTRACT Glaucoma is a group of diseases characterized by progressive optic nerve damage, ultimately
resulting in irreversible visual impairment. Early diagnosis through color fundus photography, including
measurement of the vertical cup-to-disk ratio (CDR), can help prevent vision loss. The normal range of
CDR values is usually 0.3-0.5, and if it exceeds 0.6, then there may be some problems. However, asym-
metrical thinning at the edges of the bottom-superior temporal-nasal region and large gaps in datasets pose
challenges for existing automatic segmentation methods. To address these challenges, this paper proposes a
joint segmentation method for the optic disc (OD) and optic cup (OC) based on an adversarial network,
incorporating new monitoring functions to guide the network optimization process. The effectiveness
and stability of this framework were evaluated using two public performance datasets of retinal fundus
images, namely Drishti-GS and REFUGE. On the Drishti-GS dataset, our method achieved a score of
0.850/0.964/0.086, while on the REFUGE dataset, it obtained a score of 0.887/0.975/0.061. These results
indicate the effectiveness of our approach.

INDEX TERMS Adversarial learning, deep learning, color fundus photography, glaucoma.

I. INTRODUCTION glaucoma. The CDR value is defined as the ratio of the

Glaucoma is the most common blinding disease, and can diameter of the optic disc to the diameter of the optic cup in
cause optic nerve damage, visual field damage and irre- the vertical direction, therefore, its accuracy depends on the
versible vision loss, bringing great inconvenience to patients’ accuracy of OD and OC segmentation.
live and work [1]. Owing to limited medical resources, anal- In figure.1, we can know the general position of the nerve
ysis of the optic nerve head (ONH) can help prevent the papilla in the fundus image. The yellow circle indicates the
occurrence and development of glaucoma. This can lead to optic disc area and the blue circle indicates the optic cup
differences in the diagnosis of glaucoma among different area. VCD represents the diameter of the vertical optic cup
doctors. In addition, manual diagnosis methods are time- area, and VDD represents the diameter of the optic disc
consuming, expensive, and not suitable for large-scale sample in the disposal direction. This ratio represents the vertical
screening. With the increasing number of patients, there is an cup-to-plate ratio. The difference between the cup-to-plate
urgent need for an automatic segmentation method that can ratio of glaucoma patients and normal conditions can be
help clinicians improve their work efficiency. clearly observed in (c) and (d). Under normal circumstances,
The morphology of the optic disc (OD), optic cup (OC) and the value of the cup-to-plate ratio is less than 0.6, and the
cup-to-disk ratio (CDR) are the main indices used to evaluate difference between the left and right eyes is less than 0.2.
ONH, and the CDR value could indicate the possibility of At present, many methods have been applied to OD
and OC automatic task segmentation. The most com-
The associate editor coordinating the review of this manuscript and monly used methods are edge detection, region segmenta-
approving it for publication was Mohammad Zia Ur Rahman . tion, and threshold methods (including color and contrast).
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
104898 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

Maninis et al. [13] proposed a DRIU network to segment OD

and related blood vessels, which is an improved full convolu-
tional network [14] with the advantages of VGG16 [15]. The
complexity of the calculation can be reduced using certain
sampling techniques. Zilly et al. [16] adopted entropy-based
techniques to achieve this goal. Ding et al. [17] proposed a
high-order attentional mechanism to obtain global context
information and applied this method in medical image seg-
mentation tasks. In these deep learning-based approaches,
each pixel of images must be classified. For fundus color
images, pixels can be divided into the background, OD, and
OC. This approach often requires manual labeling to monitor
network training. In supervised learning based on manual
labeling, the quality of the segmentation results is affected
by the manually labeled data. Most training schemes assume
that all the datasets have the same distribution. However, there
are significant differences between each dataset, including
FIGURE 1. Schematic diagram of the position of the optic nerve papilla in data distribution and picture quality, which results in the poor
the color fundus image. (a) is the optic nerve head topography image,
(b) is enlarged structure of the optic papilla, (c) is a color fundus image generalization ability of network models trained with a single
without glaucoma and (d) is a color image of a patient with glaucoma. datasets [18].
Several different approaches have been applied to seg-
mentation models to overcome the generalization problem
For example, Aquino et al. [2] proposed a method using between different datasets [20], [21], [22]. For exam-
morphology and edge detection techniques to obtain circular ple, in order to transmit weak class tag information,
OD boundaries by using circular transformations. If we can Hong et al. [19] used an attention network based on encoder-
capture the pixel change law of the image boundary and near decoder mechanism. Additional methods were also intro-
the boundary simultaneously, this will be very helpful for duced in the OD and OC joint segmentation tasks. Spatial
the segmentation result. Lu [3] proposed a method based adversarial networks have demonstrated a positive effect in
on circular transformation. Chakravarty and Sivaswamy [4] recent semantic segmentation. Wang et al. [18] proposed
extracted the boundaries of optic disc and cup based on the a patch-based model, which changes the discriminator in
edge detection method of conditional random field. There adversarial network into a patch discriminator. They also
are also some methods that use manual labeling of visual introduced morphological perception segmentation loss in
features to help us carry out segmentation tasks, such as the OD and OC joint segmentation tasks. This model has
stereo pair features [5], combines advanced feature extraction shown good segmentation effect on several datasets, includ-
and attention mechanisms with spatial processing [6] and ing Drishti-GS [24], RIM-ONE-r3 [25], and REFUGE [26].
based on multi-channel and spatial attention mechanisms [7]. Some researchers have found that the GAN network is unsta-
However, with the increase of model parameters, higher ble during the training process. Kadambi et al. [23] used
requirements will be put forward for training time and hard- WGAN [27] to carry out the training process and refined the
ware. Although these methods have made progress in the details in the training process. Although these methods have
automatic segmentation the optic cup from the optic disc, they made significant progress in the existing datasets, they cannot
are easily affected by training datasets, pathological changes, provide reliable support for the screening and diagnosis of
blood vessels near the optic disk and other factors, and the glaucoma because of the different data distributions in the
boundary between the OD and OC is fuzzy. application process. This prompted us to explore a network
With the development of science and technology, espe- that could provide reliable performance on different datasets
cially computers and neural networks [8], deep learning and satisfy the joint OD and OC segmentation task. The
technology has made remarkable achievements in the medical inclusion of OC in the OD region is a physiological structural
field and has achieved good results in the joint segmen- feature that is often ignored, and this relationship can be
tation task of OD and OC [9], [10], [11]. Deep learning beneficial to the segmentation results. The above methods can
networks no longer rely on manually marked features, and be summarized in Table 1.
after several rounds of iterative calculation, they can obtain In this work, a convolution network model based on adver-
the feature information of the training image, and can sarial learning is proposed for simultaneous OD and OC joint
apply them to subsequent segmentation tasks, and the seg- segmentation of fundus images. It is worth noting that OC
mentation effect is better than that of traditional methods. are always present within OD regions, and their boundaries
Wang et al. [12] proposed a method to capture OD and OC are difficult to define. The proposed method involves using
boundaries with estimated ellipses on the basis of deep learn- an extraction network to identify the region of interest, which
ing, which obtained better results under less supervision. primarily includes the region of the firm order and the region
VOLUME 12, 2024 104899
 Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

of the apparent cup. Subsequently, the extracted image is The optic disc, also known as the optic nerve papilla,
transformed using polar coordinates to increase the propor- is located at the posterior pole of the eyeball, approximately
tion of optic disc and optic cup regions in the entire image, 3 millimeters nasally and 1.5 millimeters in diameter. The
thereby enhancing the accuracy of subsequent segmentation. optic disc nerve plays an essential role in the fundus exam-
The segmentation process employs adversarial learning net- ination. Early methods of extracting optic disc boundaries
works, which are adjusted in terms of network structure and relied on the use of templates. Lowell et al. [29] utilized
loss function to improve overall performance. To evaluate image gradient changes to segment optic disc boundaries
the effectiveness of the proposed method, experiments were and incorporated the active contour method. Since both optic
conducted on two public datasets (Drishti-GS and REFUGE). discs and optic cups have ellipsoid shapes, methods based
The results demonstrate that the model achieves the desired on circular transformation technology have also been used
outcomes. [2], [3]. To enhance the robustness of the model, Fu et al
[9] incorporated local texture features in a multidimensional
TABLE 1. Advantages and disadvantages of the above methods. space. Pixel classification methods have achieved notable
results in current semantic segmentation tasks, particularly
in the field of medical image segmentation. By converting
the boundary segmentation task into a pixel classification
problem, researchers have found it to be more conducive
to solving the problem. Cheng et al. [30] used a superpixel
classifier to segment the optic disc and optic cup, and manu-
ally produced visual features to improve detection accuracy.
Abramoff et al. introduced parallax values extracted from
stereo image pairs to identify the optic disc and background
[5]. Although these methods have shown good results, they
all rely on manual annotation information, and are therefore
more susceptible to image quality and pathological changes.
OC Segmentation: An important indicator for diagnosis
glaucoma is the optic cup, situated in the center of the optic
This work’s primary contributions are as follows: disc within a brighter oval depression. The anterior movement
(1) We explored supervised adversarial networks to of the optic cup obstructs the optic disc, leading to glaucoma.
improve the adaptability of different segmentation networks Under normal circumstances, the optic cup is less than 1/3
to different datasets and to improve the generalization ability the size of the optic disc, but the proportion of the optic
of networks. cup is larger patients with glaucoma. Wong et al. proposed a
(2) A method of polar coordinate transformation is pro- level-set algorithm to automatically segment the boundary of
posed to transform the image so that the accuracy of the final OC [31]. Later, the information about blood vessel curvature
segmentation task can be improved. in retinal images has been shown to be beneficial for the seg-
(3) We evaluated our model on two public fundus image mentation of OC [32]. Due to the natural distortion of fundus
datasets, and achieved good results in OD and OC joint blood vessels near the optic disc (OC) boundary, the accu-
segmentation tasks. racy of OC segmentation based on the information of blood
The remainder of this paper is organized as follows. vessel distortion information is not satisfactory. In addition,
We review related techniques in Section II, and Section III Cheng et al. introduced the method of pixel classifier method
introduces the proposed method in detail. The evaluation and into the OC segmentation task [30]. More and more useful
results are presented in Section IV. Finally, we discuss the methods are being introduced to OC segmentation tasks [33],
results and draw conclusions in Section V and Section VI. [34]. All of the aforementioned methods depend on manually
labeled visual features, primarily focusing on the contrast
II. RELATED WORKS information between the edge of the optic nerve and the optic
Nowadays, many researchers are engaged in research on cup.
task segmentation of OD and OC, and many methods are Joint OD and OC Segmentation: Optic disc and optic cup
effective. These methods rely heavily on visual features of are closely related in physiological structure, and the optic
artificial markers for segmentation, such as image gradient cup is contained in the optic disc, which means that the pixels
information, features of stereoscopic image pairs, local tex- belonging to the optic cup also belong to the optic disc. The
ture features and superpixel based classifiers. The boundary joint segmentation of the optic disc can obtain high accuracy
between OC and OD is often difficult to distinguish, so OC in calculating the value of CDR [35]. Joshi et al. divided OD
segmentation is more difficult, and it relies more on man- and OC step by step [36]. Zheng et al. integrated the prior
ually annotated features. In recent years, it has been found graph cut framework into OD and OC segmentation [37]. The
that the joint segmentation of OD and OC can improve the above methods are based on the fact that any pixel in the
performance of segmentation networks [28]. fundus image only belongs to one part, such as background,
104900 VOLUME 12, 2024
Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

OD or OC. That is to say, they believe that OD and OC

are independent, which is in contradiction with the actual
physiological structure.
Recently, advancements have been made in domain adap-
tive technology, particularly in the field of medical image
analysis. This technique is based on adversarial networks
and explores the shared feature space between the source
FIGURE 3. Segmentation process based on adversarial learning Network.
and target domains, enabling feature correspondence between PCT stands for the image after polar coordinate transformation, both the
the two domains. The network is trained using these ideas, training part and the test part require this step. The discriminator can
reduce the difference between training set and test set prediction. The
and then applied to the target domain. These methods can entire network is monitored by two functions called Lseg and Ladv . Lseg
generate realistic images in another domain without using is calculated by using the predicted value and the training labeled
paired training sets. One such method is Cycle-GAN, but datasets (yTrain ). And Ladv is calculated using the prediction of the
unlabeled test datasets (yTest ).
it requires additional constraints to guide the unsupervised
style change process. For instance, Sevastopolsky [38] used
two segmentation networks stacked behind the cycle-GAN manipulate the original image and crop it in the center of the
to achieve enhanced shape consistency. In [16], adversarial optic disc. Specifically, the extraction network is a kind of
learning of semantic perception was used to prevent semantic U-Net network. We use the cut image block and the corre-
distortion during conversion. In [39], a generative adversarial sponding OD label for training, so that we can cut out the
network was constructed to enhance segmentation consis- sub-image we are interested in with the OD center. In this
tency. However, these methods fail to consider the spatial case, the size of our clipped image is 480 × 480.
correlations between the target domain and its neighborhood.
Therefore, we propose a network that combines the joint
B. SEGMENTATION NETWORK ARCHITECTURE
segmentation task of OC and OD to improve the accuracy
of the results. We used an improved network to get better performance for
OD and OC joint split tasks.
III. METHOD
Segmentor: We used MobileNetV2 [14] to replace xcex-
Figure. 2 and figure. 3 show the overall structure diagram of ception [13] in the original DeepLabv3+ architecture [13],
OD and OC joint segmentation proposed by us. It includes which reduces the overall network computation and improves
two parts: polar coordinate transformation for cropped fundu network performance. See figure 4. We used an original
images and adversarial learning network. Firstly, we use the convolutional network and seven inverted residual blocks of
ROI extraction network to extract the OD region in the color the MobileNetV2 in the down-sampling process. The stride
of the first four convolution layers is set as the initial value,
fundus image [40]. As the proportion of the cut image of OD
and the stride of the remaining three layers are set as 1. During
part is small, the network segmentation performance is not
the down-sampling process, a total of eight down-sampling
good after they are used as the training set. We will carry
operations are performed. In order to collect as many feature
out polar coordinate transformation on the cut image, so as
maps as possible, we use ASPP (Atrous Spatial Pyramid
to improve the proportion of OD region in the whole image.
Pooling) [13]. Finally, the probability graph of OD and OC
is generated according to the multi-label Settings in [1].

FIGURE 2. Overview of our ROI Extraction framework. ROI regions (ITrain ;

ITest ) are firstly extracted from the training section (ITrain ) and the test
section (ITest ).

The network can get more useful features from them.

Finally, the transformed images are used as the inputs of
adversarial learning network to obtain the segmentation
result.

A. ROI EXTRACTION
Increasing the proportion of optic disc and optic cup in the
FIGURE 4. The proposed segmentation network architecture. It includes a
whole image can help improve the accuracy of segmentation. down-sampling part, an up-sampling part and a skip connection part.
For this purpose, we use the ROI extraction framework to And it includes the ASPP module.

VOLUME 12, 2024 104901

 Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

Joint Morphology Loss: We used new functions to guide By constantly updating the parameters of the segmentation
the learning of the network. A dice coefficient loss LDL network and discriminator, better segmentation results can be
and a smoothness loss LSL are make up this compound loss obtained. We use formula 5 as the objective function of the
function. We can express it as: discriminator:
Lseg = λ1 LDice (pd , yd ) + λ2 LDice (pc , yc )
X
LD = − zlog(D(S(IcS ))) + (1 − z)log(1 − D(S(IcT )))
h i m,n
+ λ3 LSmooth (pd , yd ) + LSmooth (pc , yc ) (1) (5)
where pd , yd , represents OD prediction probability graph and where z = 1 if the patch prediction is from the training set,
binary true mask after polar coordinate transformation; pc and z = 0 if the patch prediction is from the testing set. In the
and yc represent the prediction probability graph of OC and process of segmentation, we use the supervision function
the binary ground true mask after polar transformation; by to ensure the reliability of the segmentation results, which
adjusting λ , the weight of each loss function can be changed. includes two parts and:
The dice coefficient loss [41] measures the overlap
between the prediction and ground truth, and is written as LS = Lseg (IcS ) + Ladv (IcT ) (6)
P X
2 i∈ pi · yi Ladv (IcT ) = − log(D(S(IcT ))) (7)
LDice (p, y) = 1 − P 2
P 2
(2) m,n
i∈ p + i∈ y
We used two public datasets, their training parts contain
where,  represents all pixels in the image after polar coor-
the original images and the corresponding manual anno-
dinate transformation, p represents the predicted probability
tation images, in the optimization process of the network,
graph, and y represents the ground truth mask after polar
we can use the joint morphological perception to segment the
coordinate transformation. The smoothness loss is written as:
X X loss Lseg .
LSmooth (p, y) = Bi,j × yi × pi − pj (3)
i∈ j∈N4 (i)
( C. POLAR TRANSFORMATION FOR FUNDUS IMAGE
1 if yi = yj In order to improve the performance of our network, we use
Bi,j = (4)
0 otherwise the method of polar coordinate transformation. The image
The smoothness of the counter image can be improved by after transformation will make OD and OC have a certain
reducing the variation between adjacent pixels. Where N4 (i) constraint relationship in space, which accords with the actual
denote the four-connected neighbors of pixel i, p denote the situation of physiological structure. If we set the center of the
prediction and y denote ground truth. disk as O(u0 , v0 ), then any pixel P(u, v) in the original color
fundus image corresponds to pixel P′ (θ , r) in the transformed
image. Figure. 6 (C), if r is used to represent the radius to
the origin p, θ is used to represent the azimuth. We can make
polar coordinates correspond to Cartesian coordinates:
( ( p
u = r cos θ r = u2 + v2
⇔ (8)
v = r sin θ θ = tan−1 v/u

Select the radius R of the circular region in the original image

to correspond to the height of the transformed image, and
FIGURE 5. Network architecture of the discriminator.
2PI to correspond to the width of the transformed image.
Discriminator: In this work, we use discriminator [38], After polar coordinate transformation, it can bring outstand-
[42] to identify the results of generator, and obtain better ing improvement to the network, which is mainly manifested
segmentation results through the process of adversarial- in:
learning. This discriminator can pay more attention to local 1)Geometric constraint: In physiological structure, OD and
features. As shown in figure 5, we implement the discrimina- OC have a certain structural relationship. They tend to
tor through a full convolutional network. The discriminator present an elliptical structure, and OC is included in OD.
consists of a six-layer convolutional network, and the num- As shown in (b). After polar coordinate transformation, OC,
ber of channels in each layer is 32; 64; 128; 256; 512; 32, OD and background three parts show a regular layer structure,
respectively. with a kernel size of 4 × 4 and a stride of 2. as shown in (d). Such a relationship is conducive to image
The final layer is activated with the sigmoid function and the segmentation [43] and [44].
remaining four layers are LeakyRelu. Each patch is classified 2) Data enhancement: We can do data enhancement in dif-
into real (1) or fake (0) through the discriminator. ferent ways. If we move the center O(u0 , v0 ) of the circle, it’s
Objective Function: The segmentation network is con- equivalent to the polar transformation of the different regions.
stantly generating images in an attempt to fool the dis- When we change the radius R, we’re going to transform it
criminator, whose goal is to find the generated fake data. with a different coefficient.

104902 VOLUME 12, 2024

Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

images, divided into 50 training images and 51 test images,

all derived from clinical data at the Madurai Aravan Eye
Hospital. Centered on OD, the field of view is 30 degrees, the
size is 2896×1944 pixels, and the format is an uncompressed
PNG image. Each image was labeled by four glaucoma spe-
cialists with three, five, nine and 20 years of experience
to capture differences in labeling between observers. The
REFUGE dataset is a public data set released in MICCAI
2018 for the Retinal fundus Glaucoma Challenge to advance
the development of automated assessment technologies for
retinal diseases such as glaucoma. The dataset contained
1200 fundus images, of which 120 were of glaucoma patients
and the rest were of non-glaucoma patients. All images are
stored in JPEG format and each color channel is 8 bits.
There are two image resolutions: 2124 × 2056 pixels and
1634×1634 pixels. All images were accompanied by detailed
disc and cup labeling information and were divided into
FIGURE 6. A schematic diagram of polar coordinate transformation,
where the image is transformed from (a) to (c). (a) to (c) represents the three subsets for easy experimental evaluation. Each subset
polar coordinate transformation of the organization image, where the contains 400 fundus images. The proportion of glaucoma
transformation area selects the circular area with radius R in the original
image, as indicated by the circle in the figure. (b) to (c) marks the and non-glaucoma cases also remained consistent, i.e., each
transformation of GT, where the white area represents the optic cup, the subset contained 10% of glaucoma cases and 90% of non-
gray area represents the optic disc, and the black area
represents the background.
glaucoma cases.
Implementation Detail: Training strategy: We used the
same method in [16] to optimize the segmentation network
3) Increase OC ratio: Due to the low proportion of OC and discriminator. The image IcS in the training set and the
region in the original color fundus image, the network error image IcT in the test set are passed to the network for each
will be large in the training process, which is easy to produce iteration. Then the parameters of the network are optimized
overfitting. Even if we use the ROI extraction technology, according to the functions LS and LD . This process is repeated
this result will be greatly improved. As shown in (b) and (d), over and over again.
if the method of polar coordinate transformation is used, the We can use a flowchart to summarize the whole process of
proportion of OC can be increased from about 4% to about network hyperparameter updating:
23%. This is conducive to the performance improvement and
reduction of overfitting in the process of model training.
The Python tool contains library functions commonly used
in neural network learning today, so we use it to build our Algorithm 1 Stochastic Gradient Descent Training of
network model. During the training phase, we introduced Segmentation Network Based on Adversarial Learning. The
Stochastic Gradient Descent (SGD) to guide the training of Number of Steps to Apply to the Discriminator, k, Is a
the network. We will start with a learning rate of 0.0001 and Hyperparameter. We Used k = 10 in Our Experiments
gradually decrease, with momentum of 0.9. In order to make for number of training iterations do
for k steps do
the transformed image size 480 × 480 pixels, we set the n
1 , I2 , · · · , IN
o
• Samples from Segmentation network IcS cS cS
transform radius R to 480 and plot the direction Angle into n o
• Samples from GT IcT 1 , I2 , · · · , IN
480 different bins. The output image is the probability graph cT cT
• Update the discriminator by:
of OD and OC. We set the threshold to 0.5 and then get the
binary mask from the probability plot. We used the same " #
scheme in the [10] and [36] to generate the final segmentation 1 XN X
i i
∇d − zlog(D(S(IcS ))) + (1 − z)log(1 − D(S(IcT )))
N i=1
result. m,n
Dataset: We verified our algorithm on three public end for n o
datasets: DRISHTI-GS1 dataset and REFUGE2 dataset. The • Samples from Segmentation network IcS 1 , I2 , · · · , IN
cS cS
DRISHTI-GS dataset is a publicly available dataset dedicated
n o
• Samples from GT IcT1 , I2 , · · · , IN
cT cT
to the study and evaluation of the optic nerve head (ONH) • Update Segmentation network:
segmentation problem. The enrolled patients ranged in age
from 40 to 80 years old, with roughly equal numbers of men 1 XN h i )+L i )
i
∇s Lseg (IcS adv (IcT
and women. The dataset consists of 101 high-quality retinal N i=1
end for
1 http://cvit.iiit.ac.in/projects/mip/drishti- gs/mip- dataset2/Home.php The gradient-based updates can use any standard gradient-based learning
2 https://refuge.grand-challenge.org rule. We used momentum in our experiments.

VOLUME 12, 2024 104903

 Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

In the formula, Ntp , Nfp , and Nfn represent true positive,

false positive and false negative respectively. VDcup and
VDdisc represent the vertical diameters of OC and OD, respec-
tively. Their ratios represent the cup-plate ratio, and CDRp
and CDRg represent the predicted and actual ratios. The error
between the predicted value and the actual value is denoted
by δ, and the smaller the value, the better prediction result.

IV. RESULTS
Experiments and Results: To verify the validity of the pro-
posed method, we conducted experiments on two publicly
available datasets, Drishti-GS and REFUGE.
According to the experimental results, our approach has
demonstrated superior performance on the two publicly avail-
able datasets. As shown in Table 2, our method outperformed
other methods in the OD and OC joint segmentation task.
FIGURE 7. The visual examples of polar coordinate transformation on
Drishti-GS and REFUGE datasets. The top three lines are from the
Specifically, on the Drishti-GS dataset, our method achieved
Drishti-GS dataset, and the bottom three lines are from REFUGE datasets, a score of 0.850/0.964/0.086, while on the REFUGE dataset,
Where a) fundus image, b) ROI region c) the polar coordinate it obtained a score of 0.887/0.975/0.061. These results indi-
transformation corresponding to the ROI region, d) the polar
transformation corresponding to the label, d) the label for the ROI region. cate the effectiveness of our approach.
Figure 8 presents the segmentation results on the
Drishti-GS and REFUGE datasets. Based on these results,
The method is verified on an Intel i5-9400, 32G RAM, we can conclude that our method’s segmentation results
Nvidia 2080Ti GPU and Windows10 platform computer, (column C) are more accurate and closer to the gold stan-
which is implemented in Python language and the back-end is dard compared to U-Net (column D). This indicates that
Tensorflow. First, we crop the original image to 480×480 size our approach can provide more reliable technical support to
with the optic disc as the center, and pass this region to the clinicians.
network as the ROI region for the next operation. According In the experiment, we also calculated the CDR value
to the previous experience, too small datasets tend to cause to measure the role of the proposed method in glaucoma
overfitting of the network. Simple data enhancement, such as screening. The results of ORIGA and REFUGE dataset are
translation, rotation and other operations, can not improve the shown in columns 4 and 8 of Table 1. From the glaucoma
network overfitting problem. screening results, we have the following observations: Most
To address this issue, polar coordinate transformation is FCN-based segmentation networks rely heavily on pixel-
utilized to enhance the performance of the network. The level labels, resulting in unsatisfactory results. However, GNS
process is depicted in figure 7. During the training phase, the network has some advantages in network structure, which
Adam optimizer iteratively optimizes the network’s parame- can collect high latitude features and finally make the results
ters. The initial learning rate is set at 1e-3, which is adjusted better. As shown in figure. 9, the ROC values on Drishti-GS
in subsequent iterations. The discriminator network is also (AUC=0.8590) and REFUGE datasets (AUC=0.8788).
continuously optimized during training, with a simulation Ablation Experiments: Ablation experiments were con-
learning rate set at 2e-5 and 1e-5, respectively. The learning ducted on the Drishti-GS datasets. The result achieved by
rate is reduced by 10 times every 4 learning periods, for a total different components of the model is shown in Table 3. For
of 400 learning periods. After obtaining the predicted mask, the sake of presentation, we take the U-net framework as
morphological operation is performed as a post-processing the baseline. When GAN network is combined with baseline,
step to refine the segmentation mask. a better dice score can be obtained in optic disc segmentation,
Evaluation Metrics: We use the following indicators to with an increase of 4.65%, but the optic cup segmentation get
evaluate our segmentation results, Dice coefficient (Dice), a lower score.
Jaccard (Jac) and vertical cup-to-disc ratio (CDR). The cri-
teria are defined as:
V. DISCUSSION
2 × Ntp Glaucoma diagnosis primarily depends on retinal images,
Dice = (9) particularly the optic disc and cup region in the center of
2 × Ntp + Nfp + Nfn
the image. However, existing methods suffer from limita-
Ntp
Jaccard = (10) tions such as differences in labels, datasets, and physiological
Ntp + Nfp + Nfn structure between the optic disc and cup. To address these
VDcup issues, we propose a generative adversarial network. Our
δ = CDRp − CDRg , CDR = (11)
VDdisc approach offers several improvements over existing methods:

104904 VOLUME 12, 2024

Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

TABLE 2. Results of joint OD and OC segmentation on different datasets.

FIGURE 9. The ROC curves with AUC scores for glaucoma screening based
on the segmentation results on Drishti-GS and REFUGE datasets.

the weights every 10 rounds. However, we found that this

led to an underestimation of CDR values for fundus images,
particularly in the glaucoma group.
3. Methodology: In contrast to existing methods, our pro-
FIGURE 8. A visual example of optic disc and cup segmentation on the posed approach uses a compound feature loss to calculate the
Drishti-GS and REFUGE datasets, where the top three rows of results are difference between the predicted image and the gold standard.
from the Drishti-GS dataset and the bottom three rows of results are from
the REFUGE dataset. From left to right, each column represents: a) This facilitates the segmentation network and discriminator
original fundus image, b) ground truth, c) prediction results of our to capture hierarchical features of long- and short-distance
method, d) U-Net prediction results.
spatial relationships between pixels. Additionally, the seg-
mentation network and discriminator share loss functions,
TABLE 3. Effect of different blocks of the frame on the Drishti-GS making the training of the segmentation stable from end to
datasets. end.
We know from figure 10 that when the choice of R value
is different, the resolution of the corresponding converted
image is also different. For example, when R=800, the res-
olution of the converted image is 800 by 800 pixels, and
the field of view is wider than R=400. However, too large
input image will affect the training speed of the network, and
1. Running Time: We reduced the size of the original does not significantly improve the resolution of the network.
images during training and used an NVIDIA 2080Ti GPU to After comparison, we chose R=480, which is very network
compute 400 iterations, which took approximately 9 hours. friendly.
Our model can generate a prediction image in approximately Although the proposed model performs well in the above
0.9 seconds, which is faster than existing methods. two data sets, the test results need to be improved when
2. Improvements to the proposed approach: During our different dataset are used, such as on the HRF dataset [48],
training, we iterated 400 times, reducing the learning rate the experimental results are shown in Table 4. There is a gap
by a factor of 10 every four learning cycles and recording between the performance of the proposed framework and the

VOLUME 12, 2024 104905

 Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

FIGURE 10. Example of polar coordinate transformation, when choosing different R-values for the same image. From left to right, R=400,
500, 600, 700, 800.

TABLE 4. Performance of the proposed model on the HRF datasets.

results in the references. In the next work we will explore the [4] A. Chakravarty and J. Sivaswamy, ‘‘Joint optic disc and cup
validation of model performance across different data sets. boundary extraction from monocular fundus images,’’ Comput.
Methods Programs Biomed., vol. 147, pp. 51–61, Aug. 2017, doi:
10.1016/j.cmpb.2017.06.004.
VI. CONCLUSION [5] M. D. Abramoff, W. L. M. Alward, E. C. Greenlee, L. Shuba, C. Y. Kim,
In this study, we utilized an enhanced generative adversarial J. H. Fingert, and Y. H. Kwon, ‘‘Automated segmentation of the optic disc
from stereo color photographs using physiologically plausible features,’’
network to segment the optic disc (OD) and optic cup (OC) Investigative Opthalmol. Vis. Sci., vol. 48, no. 4, p. 1665, Apr. 2007, doi:
in color fundus data. By transforming the joint segmentation 10.1167/iovs.06-1081.
task into a multi-label segmentation task, we successfully [6] M. Hayat, S. Aramvith, and T. Achakulvisut, ‘‘SEGSRNet for stereo-
endoscopic image super-resolution and surgical instrument segmentation,’’
achieved the separation of these two structures. Our pro- 2024, arXiv:2404.13330.
posed framework comprises two components: a segmentation [7] M. Hayat, S. Aramvith, and T. Achakulvisut, ‘‘Combined channel
network and a discriminator. The segmentation network is and spatial attention-based stereo endoscopic image super-resolution,’’
in Proc. IEEE Region 10 Conf. (TENCON), Chiang Mai, Thailand,
designed to learn the conditional distribution between the Oct. 2023, pp. 920–925, doi: 10.1109/TENCON58879.2023.
fundus image and its corresponding label, while the discrim- 10322331.
inator is responsible for distinguishing between the source of [8] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘ImageNet classification
the image-label pair. To ensure that the OD and OC infor- with deep convolutional neural networks,’’ in Proc. Adv. Neural Inf. Pro-
cess. Syst. (NIPS), 2012, pp. 1097–1105, doi: 10.1145/3065386.
mation is well balanced in the image, we first employed an [9] H. Fu, J. Cheng, Y. Xu, D. W. K. Wong, J. Liu, and X. Cao,
EOI network to isolate the region of interest, followed by ‘‘Joint optic disc and cup segmentation based on multi-label deep
a polar coordinate transformation for optimal performance. network and polar transformation,’’ IEEE Trans. Med. Imag.,
vol. 37, no. 7, pp. 1597–1605, Jul. 2018, doi: 10.1109/TMI.2018.
Our experimental results on two separate datasets indicate 2791488.
the viability of our approach. This method has the potential [10] Y. Jiang, L. Duan, J. Cheng, Z. Gu, H. Xia, H. Fu, C. Li, and J. Liu, ‘‘Join-
to aid clinicians in diagnosis and we plan to evaluate its tRCNN: A region-based convolutional neural network for optic disc and
performance on additional public datasets in future work. cup segmentation,’’ IEEE Trans. Biomed. Eng., vol. 67, no. 2, pp. 335–343,
Feb. 2020, doi: 10.1109/TBME.2019.2913211.
Our approach also holds valuable insights for other related [11] Z. Gu, J. Cheng, H. Fu, K. Zhou, H. Hao, Y. Zhao, T. Zhang, S. Gao,
fields. and J. Liu, ‘‘CE-Net: Context encoder network for 2D medical image
segmentation,’’ IEEE Trans. Med. Imag., vol. 38, no. 10, pp. 2281–2292,
Oct. 2019, doi: 10.1109/TMI.2019.2903562.
REFERENCES [12] Z. Wang, N. Dong, S. D. Rosario, M. Xu, P. Xie, and E. P. Xing, ‘‘Ellipse
[1] Y.-C. Tham, X. Li, T. Y. Wong, H. A. Quigley, T. Aung, and C.-Y. Cheng, detection of optic disc-and-cup boundary in fundus images,’’ in Proc. IEEE
‘‘Global prevalence of glaucoma and projections of glaucoma bur- 16th Int. Symp. Biomed. Imag. (ISBI), Apr. 2019, pp. 601–604.
den through 2040,’’ Ophthalmology, vol. 121, no. 11, pp. 2081–2090, [13] K. K. Maninis, J. Pont-Tuset, P. Arbeláez, and L. Van Gool, ‘‘Deep retinal
Nov. 2014, doi: 10.1016/j.ophtha.2014.05.013. image understanding,’’ in Proc. Int. Conf. Med. Image Comput. Comput.-
[2] A. Aquino, M. E. Gegúndez-Arias, and D. Marín, ‘‘Detecting the Assist. Intervent. Cham, Switzerland: Springer, 2016, pp. 140–148, doi:
optic disc boundary in digital fundus images using morphological, edge 10.1007/978-3-319-46723-8_17.
detection, and feature extraction techniques,’’ IEEE Trans. Med. Imag., [14] J. Long, E. Shelhamer, and T. Darrell, ‘‘Fully convolutional
vol. 29, no. 11, pp. 1860–1869, Nov. 2010, doi: 10.1109/TMI.2010. networks for semantic segmentation,’’ in Proc. IEEE Conf. Comput.
2053042. Vis. Pattern Recognit. (CVPR), Jun. 2015, pp. 3431–3440, doi:
[3] S. Lu, ‘‘Accurate and efficient optic disc detection and segmentation by 10.1109/CVPR.2015.7298965.
a circular transformation,’’ IEEE Trans. Med. Imag., vol. 30, no. 12, [15] K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks for
pp. 2126–2133, Dec. 2011, doi: 10.1109/TMI.2011.2164261. large-scale image recognition,’’ 2014, arXiv:1409.1556.

104906 VOLUME 12, 2024

Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

[16] J. Zilly, J. M. Buhmann, and D. Mahapatra, ‘‘Glaucoma detection using [35] Y. Xu, J. Liu, S. Lin, D. Xu, C. Y. Cheung, T. Aung, and T. Y. Wong, ‘‘Effi-
entropy sampling and ensemble learning for automatic optic cup and cient optic cup detection from intra-image learning with retinal structure
disc segmentation,’’ Computerized Med. Imag. Graph., vol. 55, pp. 28–41, priors,’’ in Proc. MICCAI, vol. 15, 2012, pp. 58–65, doi: 10.1007/978-3-
Jan. 2017, doi: 10.1016/j.compmedimag.2016.07.012. 642-33415-3_8.
[17] F. Ding, G. Yang, J. Wu, D. Ding, J. Xv, G. Cheng, and X. Li, [36] G. D. Joshi, J. Sivaswamy, and S. R. Krishnadas, ‘‘Optic disk and cup seg-
‘‘High-order attention networks for medical image segmentation,’’ in mentation from monocular color retinal images for glaucoma assessment,’’
Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent. Cham, IEEE Trans. Med. Imag., vol. 30, no. 6, pp. 1192–1205, Jun. 2011, doi:
Switzerland: Springer, 2020, pp. 253–262, doi: 10.1007/978-3-030-59710- 10.1109/TMI.2011.2106509.
8_25. [37] Y. Zheng, D. Stambolian, J. O’Brien, and J. Gee, ‘‘Optic disc and cup
[18] S. Wang, L. Yu, X. Yang, C.-W. Fu, and P.-A. Heng, ‘‘Patch-based output segmentation from color fundus photograph using graph cut with pri-
space adversarial learning for joint optic disc and cup segmentation,’’ ors,’’ in Proc. MICCAI, 2013, pp. 75–82, doi: 10.1007/978-3-642-40763-
IEEE Trans. Med. Imag., vol. 38, no. 11, pp. 2485–2495, Nov. 2019, doi: 5_10.
10.1109/TMI.2019.2899910. [38] A. Sevastopolsky, ‘‘Optic disc and cup segmentation methods for glaucoma
[19] S. Hong, J. Oh, H. Lee, and B. Han, ‘‘Learning transferrable knowledge detection with modification of U-Net convolutional neural network,’’ Pat-
for semantic segmentation with deep convolutional neural network,’’ in tern Recognit. Image Anal., vol. 27, no. 3, pp. 618–624, Jul. 2017, doi:
Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2016, 10.1134/s1054661817030269.
pp. 3204–3212. [39] O. Ronneberger, P. Fischer, and T. Brox, ‘‘U-Net: Convolutional net-
[20] J. Hoffman, D. Wang, F. Yu, and T. Darrell, ‘‘FCNs in the wild: Pixel-level works for biomedical image segmentation,’’ in Proc. 18th Int. Conf.
adversarial and constraint-based adaptation,’’ 2016, arXiv:1612.02649. Med. Image Comput. Comput.-Assist. Intervent., vol. 9351. Cham,
[21] Y.-H. Chen, W.-Y. Chen, Y.-T. Chen, B.-C. Tsai, Y.-C. F. Wang, and Switzerland: Springer, 2015, pp. 234–241, doi: 10.1007/978-3-319-24574-
M. Sun, ‘‘No more discrimination: Cross city adaptation of road scene 4_28.
segmenters,’’ 2017, arXiv:1704.08509. [40] R. O. Duda and P. E. Hart, ‘‘Use of the Hough transformation to detect
lines and curves in pictures,’’ Commun. ACM, vol. 15, no. 1, pp. 11–15,
[22] M. Wang and W. Deng, ‘‘Deep visual domain adaptation:
Jan. 1972, doi: 10.1145/361237.361242.
A survey,’’ Neurocomputing, vol. 312, pp. 135–153, Oct. 2018, doi:
[41] J. Hu, L. Shen, and G. Sun, ‘‘Squeeze-and-excitation networks,’’ in
10.1016/j.neucom.2018.05.083.
Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., Jun. 2018,
[23] S. Kadambi, Z. Wang, and E. Xing, ‘‘WGAN domain adaptation for the
pp. 7132–7141, doi: 10.1109/CVPR.2018.00745.
joint optic disc-and-cup segmentation in fundus images,’’ Int. J. Com-
[42] A. Sevastopolsky, S. Drapak, K. Kiselev, B. M. Snyder, J. D. Keenan, and
put. Assist. Radiol. Surg., vol. 15, no. 7, pp. 1205–1213, Jul. 2020, doi:
A. Georgievskaya, ‘‘Stack-U-Net: Refinement network for improved optic
10.1007/s11548-020-02144-9.
disc and cup image segmentation,’’ Proc. SPIE, vol. 10949, Mar. 2019,
[24] J. Sivaswamy, S. R. Krishnadas, G. D. Joshi, M. Jain, and A. U. S. Tabish, Art. no. 1094928, doi: 10.1117/12.2511572.
‘‘Drishti-GS: Retinal image dataset for optic nerve head(ONH) segmen- [43] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen,
tation,’’ in Proc. IEEE 11th Int. Symp. Biomed. Imag. (ISBI), Apr. 2014, ‘‘MobileNetV2: Inverted residuals and linear bottlenecks,’’ in
pp. 53–56. Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., Jun. 2018,
[25] F. Fumero, S. Alayon, J. L. Sanchez, J. Sigut, and M. Gonzalez-Hernandez, pp. 4510–4520, doi: 10.1109/CVPR.2018.00474.
‘‘RIM-ONE: An open retinal image database for optic nerve evaluation,’’ [44] M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard,
in Proc. 24th Int. Symp. Comput.-Based Med. Syst. (CBMS), Jun. 2011, and Q. V. Le, ‘‘MnasNet: Platform-aware neural architecture search for
pp. 1–6. mobile,’’ 2018, arXiv:1807.11626.
[26] J. I. Orlando et al., ‘‘REFUGE challenge: A unified framework for [45] S. Ioffe and C. Szegedy, ‘‘Batch normalization: Accelerating deep network
evaluating automated methods for glaucoma assessment from fundus pho- training by reducing internal covariate shift,’’ 2015, arXiv:1502.03167.
tographs,’’ Med. Image Anal., vol. 59, Jan. 2020, Art. no. 101570, doi: [46] S. Yu, D. Xiao, S. Frost, and Y. Kanagasingam, ‘‘Robust optic disc
10.1016/j.media.2019.101570. and cup segmentation with deep learning for glaucoma detection,’’
[27] M. Arjovsky, S. Chintala, and L. Bottou, ‘‘Wasserstein GAN,’’ 2017, Computerized Med. Imag. Graph., vol. 74, pp. 61–71, Jun. 2019, doi:
arXiv:1701.07875. 10.1016/j.compmedimag.2019.02.005.
[28] P. V. C. Hough, ‘‘Method and means for recognizing complex patterns,’’ [47] X. Ren, S. Ahmad, L. Zhang, L. Xiang, D. Nie, F. Yang, Q. Wang,
U.S. Patent 3 069 654, Dec. 3, 1962. and D. Shen, ‘‘Task decomposition and synchronization for semantic
[29] J. Lowell, A. Hunter, D. Steel, A. Basu, R. Ryder, E. Fletcher, biomedical image segmentation,’’ IEEE Trans. Image Process., vol. 29,
and L. Kennedy, ‘‘Optic nerve head segmentation,’’ IEEE Trans. Med. pp. 7497–7510, 2020, doi: 10.1109/TIP.2020.3003735.
Imag., vol. 23, no. 2, pp. 256–264, Feb. 2004, doi: 10.1109/tmi.2003. [48] M. K. Kar and M. K. Nath, ‘‘Efficient segmentation of vessels and
823261. disc simultaneously using multi-channel generative adversarial network,’’
[30] J. Cheng, J. Liu, Y. Xu, F. Yin, D. W. K. Wong, N.-M. Tan, D. Tao, Social Netw. Comput. Sci., vol. 5, no. 3, p. 288, Feb. 2024, doi:
C.-Y. Cheng, T. Aung, and T. Y. Wong, ‘‘Superpixel classification based 10.1007/s42979-024-02610-0.
optic disc and optic cup segmentation for glaucoma screening,’’ IEEE
Trans. Med. Imag., vol. 32, no. 6, pp. 1019–1032, Jun. 2013, doi:
10.1109/TMI.2013.2247770.
[31] D. W. K. Wong, J. Liu, J. H. Lim, X. Jia, F. Yin, H. Li, and T. Y. Wong,
‘‘Level-set based automatic cup-to-disc ratio determination using reti-
nal fundus images in ARGALI,’’ in Proc. 30th Annu. Int. Conf. IEEE
Eng. Med. Biol. Soc., Aug. 2008, p. 10, doi: 10.1109/IEMBS.2008.
4649648.
[32] D. W. K. Wong, J. Liu, J. H. Lim, H. Li, and T. Y. Wong, ‘‘Automated
detection of kinks from blood vessels for optic cup segmentation in
retinal images,’’ Proc. SPIE, vol. 7260, no. 6, pp. 964–970, 2009, doi:
YONG LIU was born in Henan, China, in 1986.
10.1117/12.810784.
He received the bachelor’s degree in electrical
[33] Y. Xu, D. Xu, S. Lin, J. Liu, J. Cheng, C. Y. Cheung, T. Aung,
engineering and automation from Yanshan Uni-
and T. Y. Wong, ‘‘Sliding window and regression based cup detec-
tion in digital fundus images for glaucoma diagnosis,’’ in Proc. versity, in 2009, and the master’s degree in cir-
Int. Conf. Med. Image Comput. Comput.-Assist. Intervent. Berlin, cuit and system from Henan Normal University,
Germany: Springer, 2011, pp. 1–8, doi: 10.1007/978-3-642-23626- in 2013. He is currently pursuing the Ph.D. degree
6_1. in control science and engineering with Wuhan
[34] Y. Xu, L. Duan, S. Lin, X. Chen, D. Wong, T. Wong, and J. Liu, ‘‘Optic cup University of Science and Technology.
segmentation for glaucoma detection using low-rank superpixel represen- His research interests include artificial intelli-
tation,’’ in Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., gence, medical image processing, and intelligent
2014, pp. 1–8, doi: 10.1007/978-3-319-10404-1_98. control.

VOLUME 12, 2024 104907

 Y. Liu et al.: Combined OD and OC Segmentation Network Based on Adversarial Learning

JIN WU (Member, IEEE) was born in Anhui, XUEZHI ZHOU was born in Henan, China,
China, in 1967. She received the degree in elec- in 1993. He received the B.S. degree in electronic
tronic information engineering from Huazhong information engineering from Henan Normal Uni-
University of Science and Technology, in 1988, versity, Xinxiang, China, in 2014, and the Ph.D.
the master’s degree in detection technology and degree in information and communication engi-
automatic equipment from the University of Sci- neering from Xidian University, Xi’an, China,
ence and Technology Beijing, in 1997, and the in 2020. His research interests include radiomics,
Ph.D. degree in pattern recognition and intelli- medical image analysis, and bioinformatics.
gent systems Huazhong University of Science and
Technology, in 2006.
She became the Director of the Department of Electronic and Information
Engineering, Wuhan University of Science and Technology, in 1998, and
the Associate Dean of the School of Information, Wuhan University of
Science and Technology, in 2007. Her research interests include image pro-
cessing, pattern recognition and intelligent systems, signal and information
processing, and multimedia communication.
Dr. Wu is a member of the Professional Committee of Image and Video
Processing and Communication of Chinese Society of Image and Graphics,
and a Senior Member of Chinese Institute of Electronics and Chinese Optical
Society.

YUANPEI ZHU was born in Henan, China,

in 1989. She received the bachelor’s degree in
electronic information engineering and the mas-
ter’s degree in circuits and systems from Henan
Normal University, in 2009 and 2010, respectively.
Since 2013, she has been with Xinxiang Univer-
sity, Henan. Her research interests include medical
image processing and artificial intelligence.

104908 VOLUME 12, 2024