Automated segmentation of the optic disc from fundus images using an asymmetric deep learning network

Abstract

Accurate segmentation of the optic disc (OD) regions from color fundus images is a critical procedure for computer-aided diagnosis of glaucoma. We present a novel deep learning network to automatically identify the OD regions. On the basis of the classical U-Net framework, we define a unique sub-network and a decoding convolutional block. The sub-network is used to preserve important textures and facilitate their detections, while the decoding block is used to improve the contrast of the regions-of-interest with their background. We integrate these two components into the classical U-Net framework to improve the accuracy and reliability of segmenting the OD regions depicted on color fundus images. We train and evaluate the developed network using three publicly available datasets (i.e., MESSIDOR, ORIGA, and REFUGE). The results on an independent testing set (n=1,970 images) show a segmentation performance with an average Dice similarity coefficient (DSC), intersection over union (IOU), and Matthew's correlation coefficient (MCC) of 0.9377, 0.8854, and 0.9383 when trained on the global field-of-view images, respectively, and 0.9735, 0.9494, and 0.9594 when trained on the local disc region images. When compared with the other three classical networks (i.e., the U-Net, M-Net, and Deeplabv3) on the same testing datasets, the developed network demonstrates a relatively higher performance.

1. Introduction

As a chronic and progressive optic neuropathy, glaucoma is one of the leading causes of irreversible vision loss in the world [1]. In its early stage, glaucoma can gradually reduce the visual field of patients without obvious symptoms and ultimately leads to a severe deterioration of vision [2]. Hence, timely screening and early detection are very important for preventing vision loss associated with glaucoma [3]. In clinical practice, ophthalmologists typically assess glaucoma by manually measuring the cup-to-disc ratio (CDR) based on color fundus images [4], where identifying the optic disc is one of the most important preprocessing steps. To improve the efficiency and consistency, especially in the screen setting, and avoid potential errors associated with this subjective assessment, it is extremely desirable to develop computerized algorithms [5-7] to extract the OD from retinal fundus images.

OD is the entry point for major blood vessels [8] and typically appears a bright yellowish region with a circular or elliptical shape on color fundus images [9,10]. By leveraging this image characteristic, a variety of unsupervised segmentation methods [11-13] have been developed to automatically identify OD and quantify its morphology. These methods generally attempted to classify each pixel as either background or disc region leveraging various hand-crafted features, such as structural textures and feature entropy [14,15]. A review of the relevant methods in this regard can be found in [16]. For instance, Welfer et al. [17] applied an adaptive mathematical morphology method to segment the OD. Septiarini et al. [18] combined an adaptive thresholding procedure and a morphological operation to locate the OD region. However, these hand-crafted features could be very sensitive to many factors [19,20], such as light exposure and the existence of diseases, which may significantly affect the appearance of the OD regions on color fundus images (Fig. 1), thereby making it challenging to accurately identify the OD regions [6,7].

Fig. 1.

Illustration of color fundus images with different qualities caused by tissue abnormality and light exposure.

In recent years, deep learning technology, especially the convolutional neural networks (CNNs), has been widely used in the area of medical image analysis [21-23] and demonstrated remarkable segmentation performance. There have been investigative efforts dedicated in this regard to segment OD [24-26]. Fu et al. [27] proposed a deep learning architecture called M-Net to jointly segment the optic disc and cup (OC) from fundus images in a one-stage segmentation procedure, where the M-Net consists of multi-scale side-input and side-output layers, and a U-shape convolutional network. Wang et al. [28] described a coarse-to-fine deep learning framework on the basis of the classical U-Net network [25] to identify the OD by combining color fundus images and their vessel density maps. Gu et al. [29] developed a context encoder network (CE-Net) based on the U-Net model by introducing a dense atrous convolution block and a residual multi-kernel pooling block to capture more high-level image features and preserve spatial information. Wang et al. [30] employed a patch-based output space adversarial learning framework (pOSAL) to jointly and robustly extract the OD and OC from multiple fundus image datasets and reduce the influence of the domain shift among different datasets. To further improve the segmentation accuracy in the disc and cup boundary regions, Wang et al. [31] proposed a boundary and entropy-driven adversarial learning (BEAL) framework. In architecture, the available networks were typically formed by an image encoder, a feature decoder, and a skip connection, and thus can be viewed as somewhat variants of the U-Net model. Additional variants can be found in [32-34]. A common limitation of these networks lies in the utilization of the consecutive pooling or strided convolutional operations, which often lead to the loss of some important features associated with object positions and boundaries [35]. Also, the extracted features are used indiscriminately in the feature decoder, making these networks insensitive to some important feature information and the morphological changes of target objects. Although the limitations can be eliminated partially by training the networks on local disc patches with a relatively large image dimension (e.g., 512×512), this strategy needs the manual location of the OD regions and has a high computational cost.

To overcome the above limitations, we propose a novel deep learning network to automatically segment the OD from fundus images. The proposed network uses several multi-scale input features to reduce the influence of the consecutive pooling operations. These multi-scale features are then incorporated into both the encoding and decoding procedures by element-wise subtraction. This motivation is to preserve key image information in the encoding procedure while highlighting the morphological and boundary changes of target objects to enable an accurate segmentation.

2. Proposed Method

2.1. Network overview

The proposed network architecture is formed by a feature detection sub-network (FDS) and a cross-connection sub-network (CCS) (Fig. 2). FDS is used to detect image features and extract desirable objects, while CCS facilitates object segmentation by providing additional vital features. FDS is derived from the classical U-Net model with different encoding and decoding convolutional blocks. The encoding block is defined by stacking two same convolutional layers, each of which includes a 3×3 convolutional kernel (Conv3×3), a batch normalization (BN), and an element-wise rectified linear unit (ReLU) activation [36]. The number of convolutional kernels or filters is initially set at 32 and gradually increases as the way shown in Fig. 2. The obtained convolutional results are processed using a 2×2 MaxPooling layer with a stride of 2 (MaxPooling2×2) to reduce redundant features and thus improve training efficiency. After the application of these encoding and MaxPooling operations, a number of high-dimensional features are extracted from the input images. These features are processed using a 2×2 UpSampling layer with a stride of 2 (UpSampling2×2), followed by a decoding block.

Fig. 2.

The architecture of the developed network for segmenting the OD regions on color fundus images.

During the processing procedure, different convolutional features from the FDS and CCS sub-networks, as shown in Fig. 2, are combined to guide the extraction of desirable objects. The last decoding features are fed into a 1×1 convolution layer (Conv1×1) with a sigmoid activation to achieve a probability map for the OD on the entire fundus images. Notably, due to the presence of the CCS sub-network, both the encoding and decoding blocks have two different versions for varying input features, and these features are combined in different manners (Section 2.2). The source codes of the developed network can be found at https://github.com/wmuLei/ODsegmentation.

2.2. Cross-connection sub-network (CCS)

Inspired by the M-Net model [27], we use the CCS sub-network to reduce the impact of multiple pooling operations and improve segmentation accuracy. CCS uses multi-scale input features for image encoding, which is similar to the side-input sub-network in the M-Net model. Specifically, CCS consists of four components by down-sampling the first four encoding features using a 4×4 MaxPooling layer with a stride of 4 (MaxPooling4×4). These components (intermediate encoding features) have relatively low convolutional hierarchies and contain a large number of structural textures associated with target objects. The structural textures are concatenated, along with image depth dimension, with the subsequent encoding features, when they have the same image width and height, as shown in Fig. 3(a). This will lead to the combination of both low- and high-level features and make additional key textures available in segmentation. These texture features are fed into the decoding blocks using the element-wise subtraction and concatenation, as illustrated in Fig. 3(b). The use of the subtraction operation can improve the sensitivity of the developed model to the OD boundaries, as compared with the U-Net and M-Net models.

Fig. 3.

The encoding (a) and decoding (b) blocks in the developed network. The operations specified by the dotted lines are activated in the presence of texture features from the CCS sub-network.

2.3. Decoding block

We use the decoding blocks to combine multiple encoding features, as displayed in Fig. 3(b), and highlight the morphological differences of target objects in these feature maps. These blocks can be fed with two input features (both from FDS) or three features (one from CCS and two from FDS). When there are three input features for these blocks, the input features from the CCS sub-network are firstly processed using a sequence of Conv3×3, BN, and ReLU operations to ensure that they have proper image dimensions for subsequent operations. The processed features are then combined with another two input features from the FDS sub-network using the element-wise subtraction and concatenation operations to enhance the objects' boundaries, as illustrated in Fig. 3(b). The utilization of the subtraction operation is motivated by the traditional edge detection method [37], where the element-wise intensity differences between a given image and its Gaussian filtered version can highlight a variety of edge information, as shown by Fig. 4. The subtraction operation allows the decoding block to capture various features associated with object boundaries and thus improve the segmentation accuracy. Also, the subtraction operation will lead to different numbers of filters for the encoding and decoding blocks, which make the developed network asymmetric and lightweight, as compared with the classical U-Net model [28].

Fig. 4.

The difference image (c) between a given image (a) and its Gaussian filtered version (b). As indicated by the arrows, subtle structural features are highlighted in the difference image.

2.4. Training the network

We implemented the proposed network using the Keras library with TensorFlow backend (https://keras.io/) and trained it using the MESSIDOR dataset [28], which is publicly available and widely used to develop algorithms for computer-assisted diagnoses of diabetic retinopathy (DR). In the MESSIDOR dataset, there are 1,200 color fundus images, including 540 normal images and 660 images diagnosed with DR. We preprocessed these images to exclude the background and resize the field-of-view regions to a uniform dimension of 256×256 pixels. Then, these images were normalized to alleviate the effects caused by different illumination conditions (e.g., over- or under-exposure), as shown in Fig. 5. The processed dataset was randomly divided into three sub-groups at a ratio of 0.75:0.15:0.1 for training (n=900), internal validation (n=180), and independent testing (n=120), respectively. The local disc regions in each sub-group were manually cropped and resized at a dimension of 256×256 pixels, resulting in an additional local disc region image dataset, which consisted of training, internal validation, and independent testing sub-groups as well. The dimensions of these local disc regions were approximately three times the radius of the OD. We used the global field-of-view images and the local disc regions to train the proposed network separately.

Fig. 5.

Illustration of three original fundus images (left column), their normalized results (middle column), and local disc regions (right column).

The Dice similarity coefficient (DSC) [38,39] was used as the loss function and given by:

$$DSC=\frac{2N(A\cap B)}{N(A)+N(B)}$$ (1)

where A and B are the segmentation results obtained by the proposed network and manual annotation, respectively. N(·) denotes the number of the pixels enclosed set. ∩ and ∪ are the intersection and union operators. The DSC was optimized using the exponential moving average variant of the adaptive moment (Adam) estimation (AMSGrad) [40]. In the AMSGrad optimization, the parameters β₁ and β₂ were set to 0.9 and 0.999, respectively, the initial learning rate was set to 1.0e-3 for the global field-of-view images and to 1.0e-4 for the local disc regions. A total of 150 epochs were used during the training, and the batch size was set at 8. The training process stopped if the DSC did not improve for consecutive 30 epochs. To improve the robustness of the training procedure, we augmented the images to improve their diversity [28,30]. Specifically, the images were randomly flipped along horizontal and vertical axes, and translated by −15 to 15 percent per axis, and rotated from −90 to 90 degrees.

2.5. Performance evaluation

We assessed the performance of the developed network using the independent testing sub-group of the MESSIDOR dataset and the publicly available ORGIA and REFUGE datasets [8]. A detailed description of the three datasets can be found in [28,36]. We also compared the performance of the developed network with three classical networks, namely the U-Net, M-Net, and Deeplabv3 with the backbone network ‘xception’ [30], under the same experiment configurations. The differences of these networks were summarized in Table 1. We used DSC, intersection over union (IOU) (Eq. (2)), Matthew's correlation coefficient (MCC) (Eq. (3)) [41], and balanced accuracy (BAC) (Eq. (4)) [28] as the performance metrics. These metrics range from 0 to 1, and a larger value means a better segmentation performance.

$$IOU=\frac{N(A\cap B)}{N(A\cup B)}$$ (2)

$$MCC=\frac{(TpTn-FpFn)}{\sqrt{(Tp+Fp)(Tp+Fn)(Tn+Fp)(Tn+Fn)}}$$ (3)

$$BAC=\frac{1}{2}(Se+Sp)$$ (4)

$$Se=\frac{Tp}{Tp+Fn}$$ (5)

$$Sp=\frac{Tn}{Tn+Fp}$$ (6)

where Tp, Tn, Fp, and Fn denote true positive, true negative, false positive and false negative, respectively. Se and Sp are the sensitivity and specificity for a segmentation method. They indicate the effectiveness in identifying the disc and background pixels, respectively. Since the optic disc only occupies a very small area in color fundus images, this makes the Sp and BAC very large in most cases. Hence, the DSC, IOU and MCC are used to indicate the overall segmentation performance. In addition, we assessed the effect of the CCS sub-network on the segmentation performance by testing the developed models with or without CCS on global field-of-view images. To statistically assess the performance differences of these segmentation methods, the paired t-test was performed. A p-value less than 0.05 was considered statistically significant.

Table 1.

The differences of the U-Net, M-Net, Deeplabv3, and the proposed network. (‘#’ denotes the number of a given variable, ‘M’ is short for Million)

Method	Layer #	up-sampling #	Parameter #
U-Net	86	5	31.4667 M
M-Net	48	4	8.5472 M
Deeplabv3	410	3	41.2530 M
The proposed	111	5	20.7984 M

3. Experimental results

3.1. Segmentation based on the global field-of-view images

We comparatively assessed the segmentation performance of the developed network with different optimization algorithms, including Adam, AMSGrad, and the stochastic gradient descent (SGD) [20], and activation functions, including ReLU and Leaky ReLU (LReLU) [42]. In the Adam optimization algorithm, the parameters β₁ and β₂ were set to 0.9 and 0.999, respectively; in the SGD, the momentum was set to 0.9, and the Nesterov momentum was enabled. The initial learning rates of these two optimizations were set to 1.0e-3. In the LReLU activation function, the scalar parameter was set to 0.00, 0.01, 0.05, 0.10, 0.15, and 0.20, respectively for the independent testing set of the MESSIDOR dataset. The segmentation results (Table 2) demonstrated that the developed CNN model had the highest accuracy in terms of DSC, IOU, and MCC when using the AMSGrad optimization algorithm and the ReLU activation function.

Table 2.

The performance of the developed network with different optimization algorithms (i.e., Adam, AMSGrad, and SGD) and activation functions (i.e., ReLU and LReLU) on the testing sub-group of the MESSIDOR dataset using the average DSC, IOU, and MCC metrics.

Optimization	Metric	ReLU	LReLU
			0.00	0.01	0.05	0.10	0.15	0.20
Adam	DSC	0.9600	0.9621	0.9611	0.9626	0.9609	0.9604	0.9601
	IOU	0.9241	0.9279	0.9262	0.9290	0.9259	0.9249	0.9243
	MCC	0.9602	0.9620	0.9613	0.9627	0.9611	0.9606	0.9600
AMSGrad	DSC	0.9646	0.9617	0.9618	0.9622	0.9641	0.9605	0.9602
	IOU	0.9326	0.9274	0.9277	0.9282	0.9315	0.9252	0.9246
	MCC	0.9646	0.9617	0.9619	0.9622	0.9640	0.9606	0.9603
SGD	DSC	0.9450	0.9385	0.9419	0.9442	0.9415	0.9408	0.9405
	IOU	0.8982	0.8876	0.8938	0.8976	0.8919	0.8911	0.8904
	MCC	0.9456	0.9393	0.9427	0.9449	0.9421	0.9414	0.9411

We summarized the performance of the developed network and three classical networks (i.e., the U-Net, M-Net, and Deeplabv3) on the testing sub-group of the MESSIDOR dataset (n=120 images) and the ORIGA (n=650 images) and REFUGE datasets (n=1,200 images) in Tables 3 and 4. The developed model achieved performance with a mean DSC, IOU, and MCC of 0.9377, 0.8854, and 0.9383, respectively for the entire test images. It significantly outperformed the U-Net (0.9316, 0.8753, and 0.9325), M-Net (0.9078, 0.8409, and 0.9105), and Deeplabv3 (0.9289, 0.8733, and 0.9299) on the same test images (Tables 3 and 4). Figs. 6 and 7 showed some segmentation examples to visually demonstrate the performance differences of the developed model and three other models.

Table 3.

The performance of the developed network and three classical networks on the testing sub-group of the MESSIDOR dataset and the whole ORIGA and REFUGE datasets using the mean and standard deviation (SD) of DSC, IOU, BAC, MCC, and Se.

Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
MESSIDOR	U-Net	0.9638±0.0271	0.9314±0.0470	0.9844±0.0137	0.9639±0.0259	0.9695±0.0278
	M-Net	0.9463±0.0449	0.9011±0.0709	0.9804±0.0292	0.9468±0.0416	0.9619±0.0587
	Deeplabv3	0.9557±0.0286	0.9165±0.0497	0.9815±0.0182	0.9558±0.0275	0.9639±0.0367
	The proposed	0.9646±0.0234	0.9326±0.0416	0.9870±0.0123	0.9646±0.0227	0.9748±0.0249
Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
ORIGA	U-Net	0.9374±0.0393	0.8845±0.0640	0.9953±0.0079	0.9386±0.0363	0.9925±0.0162
	M-Net	0.9271±0.0558	0.8680±0.0762	0.9906±0.0235	0.9286±0.0523	0.9833±0.0473
	Deeplabv3	0.9220±0.1145	0.8680±0.1232	0.9764±0.0635	0.9234±0.1100	0.9543±0.1274
	The proposed	0.9392±0.0355	0.8873±0.0589	0.9938±0.0131	0.9401±0.0333	0.9895±0.0264
Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
REFUGE	U-Net	0.9252±0.0516	0.8647±0.0812	0.9555±0.0291	0.9261±0.0494	0.9119±0.0585
	M-Net	0.8935±0.1052	0.8203±0.1375	0.9650±0.0449	0.8970±0.0982	0.9325±0.0905
	Deeplabv3	0.9300±0.0431	0.8719±0.0685	0.9621±0.0272	0.9308±0.0406	0.9251±0.0550
	The proposed	0.9342±0.0458	0.8796±0.0729	0.9627±0.0246	0.9347±0.0434	0.9262±0.0495

Table 4.

The p-values of the paired t-tests among the developed network and three classical networks on 1,970 global field-of-view images in terms of the DSC, IOU, and MCC. (‘−’ indicates invalid statistical analysis)

Method	U-Net			M-Net			Deeplabv3
	DSC	IOU	MCC	DSC	IOU	MCC	DSC	IOU	MCC
U-Net	-	-	-
M-Net	<0.0001	<0.0001	<0.0001	-	-	-
Deeplabv3	0.0918	0.2852	0.0836	<0.0001	<0.0001	<0.0001	-	-	-
The proposed	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001

Fig. 6.

Illustration of the segmentation results of the U-Net (in cyan), M-Net (in red), Deeplabv3 (in green), and the developed model (in blue) on twelve fundus images from the testing sub-group of the MESSIDOR dataset as well as their ground truth (manual delineations) (in white).

Fig. 7.

Illustration of the segmentation results of the U-Net (in cyan), M-Net (in red), Deeplabv3 (in green), and the developed model (in blue) on sixteen images from the ORGIA dataset as well as their ground truth (manual delineations) (in white).

3.2. Segmentation based on the local disc region images

Table 5 presented the segmentation performance of the developed network in different training settings (i.e., three optimization algorithms and two activation functions) on the local disc region images. Among the three optimization algorithms, Adam and AMSGrad had the initial learning rate of 1.0e-4 without changing other configurations compared to those in the global field-of-view segmentation. As demonstrated by the results, the developed model achieved reasonable accuracy when using the AMSGrad optimization algorithm and the ReLU activation function and was superior to the other training settings for the same test images in terms of DSC, IOU, and MCC.

Table 5.

The performance of the developed network with different optimization algorithms (i.e., Adam, AMSGrad, and SGD) and activation functions (i.e., ReLU and LReLU) on the testing sub-group of the MESSIDOR dataset using the average DSC, IOU, and MCC metrics.

Optimization	Metric	ReLU	LReLU
			0.00	0.01	0.05	0.10	0.15	0.20
Adam	DSC	0.9838	0.9837	0.9844	0.9840	0.9839	0.9840	0.9841
	IOU	0.9683	0.9682	0.9694	0.9686	0.9685	0.9687	0.9689
	MCC	0.9750	0.9749	0.9759	0.9753	0.9752	0.9753	0.9755
AMSGrad	DSC	0.9843	0.9840	0.9838	0.9838	0.9837	0.9840	0.9836
	IOU	0.9693	0.9687	0.9683	0.9683	0.9681	0.9686	0.9679
	MCC	0.9758	0.9753	0.9750	0.9750	0.9748	0.9753	0.9747
SGD	DSC	0.9787	0.9779	0.9785	0.9770	0.9787	0.9777	0.9776
	IOU	0.9585	0.9569	0.9581	0.9551	0.9584	0.9566	0.9564
	MCC	0.9671	0.9659	0.9669	0.9645	0.9671	0.9657	0.9654

Tables 6 and 7 show the segmentation performance of the developed network and the aforementioned three classical networks on the local disc region images. It can be seen that these networks had high segmentation accuracy comparable to manual annotations; however, their performance differences were limited. Specifically, the developed model achieved the average DSC, IOU, and MCC of 0.9735, 0.9494, and 0.9594, respectively for a total of 1,970 test regions, and was superior to the U-Net (0.9730, 0.9485, and 0.9586), M-Net (0.9594, 0.9230, and 0.9386), and Deeplabv3 (0.9571, 0.9200, and 0.9359) models. Among the four CNN models, the M-Net and Deeplabv3 had relatively low segmentation results for the same testing datasets. Some segmentation examples from the MESSIDOR and ORIGA datasets are shown in Fig. 8 for a straightforward visual comparison.

Table 6.

The performance of the developed network and three classical networks on the testing sub-group of the MESSIDOR dataset and the ORIGA and REFUGE datasets using the mean and standard deviation (SD) of DSC, IOU, BAC, MCC, and Se.

Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
MESSIDOR	U-Net	0.9839±0.0094	0.9686±0.0177	0.9874±0.0075	0.9752±0.0144	0.9829±0.0109
	M-Net	0.9796±0.0096	0.9602±0.0181	0.9837±0.0078	0.9686±0.0148	0.9773±0.0133
	Deeplabv3	0.9794±0.0105	0.9599±0.0198	0.9837±0.0083	0.9683±0.0162	0.9776±0.0117
	The proposed	0.9843±0.0095	0.9693±0.0179	0.9876±0.0076	0.9758±0.0147	0.9831±0.0109
Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
ORIGA	U-Net	0.9778±0.0075	0.9567±0.0141	0.9867±0.0056	0.9663±0.0114	0.9936±0.0072
	M-Net	0.9663±0.0121	0.9351±0.0219	0.9808±0.0073	0.9491±0.0181	0.9953±0.0068
	Deeplabv3	0.9733±0.0137	0.9483±0.0247	0.9830±0.0118	0.9595±0.0201	0.9876±0.0223
	The proposed	0.9797±0.0079	0.9604±0.0148	0.9869±0.0060	0.9692±0.0120	0.9899±0.0084
Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
REFUGE	U-Net	0.9693±0.0333	0.9421±0.0548	0.9759±0.0257	0.9527±0.0512	0.9673±0.0340
	M-Net	0.9536±0.0295	0.9127±0.0488	0.9610±0.0224	0.9298±0.0449	0.9377±0.0324
	Deeplabv3	0.9461±0.0438	0.9007±0.0721	0.9540±0.0350	0.9198±0.0649	0.9222±0.0560
	The proposed	0.9692±0.0325	0.9418±0.0533	0.9753±0.0252	0.9526±0.0499	0.9651±0.0339

Table 7.

The p-values of the paired t-tests among the developed network and three classical networks on 1,970 local disc regions in terms of the DSC, IOU, and MCC. (‘−’ indicates invalid statistical analysis)

Method	U-Net			M-Net			Deeplabv3
	DSC	IOU	MCC	DSC	IOU	MCC	DSC	IOU	MCC
U-Net	-	-	-
M-Net	<0.0001	<0.0001	<0.0001	-	-	-
Deeplabv3	<0.0001	<0.0001	<0.0001	<0.0001	0.0006	0.0003	-	-	-
The proposed	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001

Fig. 8.

Segmentation results of the local disc regions from the MESSIDOR (in the first two rows) and ORGIA (in the last two rows) datasets using four different models. The disc boundaries were obtained by the U-Net (in cyan), M-Net (in red), Deeplabv3 (in green), and the developed model (in blue) as well as manual annotations (in white), respectively.

4. Discussion

We developed and validated a novel deep learning network to accurately segment the OD regions depicted on color fundus images. Due to the introduction of a sub-network and a decoding convolutional block, this network demonstrated more sensitivity to edge information than other available models, which can capture the structural details that may be smeared out by multiple down-sampling operations. We quantitatively evaluated the performance of the developed network and compared it with three typical segmentation networks on a total of 1,970 color fundus images from the MESSIDOR, ORIGA, and REFUGE datasets. The experimental results demonstrated the unique advantage of the developed model in segmenting the OD regions.

In the developed model, the CCS sub-network and the element-wise subtraction in the decoding blocks played important roles in accurately segmenting the OD regions from fundus images. Table 8 demonstrated that the developed model achieved performance with the average DSC, IOU, and MCC of 0.9377, 0.8854, and 0.9383, respectively, in the presence of the CCS sub-network for a total of 1,970 global field-of-view images, and of 0.9268, 0.8717, and 0.9282 without the CCS sub-network for the same images. The developed model with the CCS had the potential to handle the problems caused by weak contrast or varying illumination, as shown in Fig 9. Table 9 summarized the performance of the developed model with different element-wise operations (i.e., addition, multiply, and subtraction) on both the global field-of-view images and their local disc versions from the MESSIDOR dataset. This suggests that the introduction of the CCS sub-network and the element-wise subtraction improved the performance of the developed model, which has much less convolutional parameters, as compared with the U-Net (Table 1). Their integration made our model computationally efficient and sensitive to the morphological changes of target objects in the feature maps with the same image dimension (similar to the effect described in Fig. 4).

Table 8.

The performance of the developed network with or without the CCS sub-network on a total of 1,970 global field-of-view images from the MESSIDOR, ORIGIA, and REFUGE datasets using the mean and standard deviation (SD) of DSC, IOU, BAC, MCC, and Se.

Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
MESSIDOR	with	0.9646±0.0234	0.9326±0.0416	0.9870±0.0123	0.9646±0.0227	0.9748±0.0249
	Without	0.9592±0.0320	0.9233±0.0548	0.9868±0.0143	0.9594±0.0304	0.9745±0.0289
Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
ORIGA	with	0.9392±0.0355	0.8873±0.0589	0.9938±0.0131	0.9401±0.0333	0.9895±0.0264
	Without	0.9144±0.1044	0.8530±0.1140	0.9883±0.0583	0.9169±0.0980	0.9789±0.1169
Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
REFUGE	with	0.9342±0.0458	0.8796±0.0729	0.9627±0.0246	0.9347±0.0434	0.9262±0.0495
	Without	0.9304±0.0777	0.8767±0.0981	0.9647±0.0419	0.9313±0.0748	0.9305±0.0842

Fig. 9.

Illustration of the segmentation results of the developed model with (in blue) or without (in magenta) the CCS sub-network on nine global field-of-view images from the MESSIDOR dataset as well as their ground truths (manual annotations) (in white). Only the disc regions were displayed for visual comparison purposes.

Table 9.

The performance of the developed network with different element-wise operations on the global field-of-view images and their local disc versions from the MESSIDOR dataset using the mean and standard deviation (SD) of DSC, IOU, BAC, MCC, and Se. (Note that there are different training schemes for global and local datasets.)

Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
Global	Addition	0.9626±0.0264	0.9290±0.0458	0.9871±0.0120	0.9626±0.0253	0.9749±0.0244
	Multiply	0.9622±0.0232	0.9280±0.0408	0.9858±0.0123	0.9621±0.0225	0.9724±0.0248
	Subtraction	0.9646±0.0234	0.9326±0.0416	0.9870±0.0123	0.9646±0.0227	0.9748±0.0249
Dataset	Method	DSC	IOU	BAC	MCC	Se
		Mean±SD	Mean±SD	Mean±SD	Mean±SD	Mean±SD
Local	Addition	0.9839±0.0092	0.9685±0.0173	0.9876±0.0073	0.9751±0.0142	0.9839±0.0103
	Multiply	0.9838±0.0097	0.9683±0.0184	0.9872±0.0077	0.9750±0.0150	0.9823±0.0111
	Subtraction	0.9843±0.0095	0.9693±0.0179	0.9876±0.0076	0.9758±0.0147	0.9831±0.0109

The developed model demonstrated better performance as compared with three classical segmentation models (i.e., the U-Net, M-Net, and Deeplabv3) based on both the global and local images from the MESSIDOR, ORIGA, and REFUGE datasets (Tables 3 and 6). We also summarized the segmentation performance of our developed method and other available methods in Table 10. As compared with our method, most of these methods in Table 10 were evaluated on a relatively small number of datasets formed by local disc region images. From these results, it can be seen that our developed method demonstrated a consistent and relatively higher performance for segmenting OD regions, given the size and diversity of our testing datasets as well as the small training dimension. Notably, the optic disc only occupies a very small area in a color fundus image. Consequently, the image background has a very high weight in the BAC metric, thus making this metric relatively large for different types of segmentation methods. Additionally, the U-Net had a different performance on local disc regions, in spite of the same training dimension (e.g., 256×256), because these disc regions were cropped from the global field-of-view images using different sizes, relative to the radius of the OD.

Table 10.

Performance comparison of the developed method with some available methods for the disc segmentation based on the DSC, BAC, and Se metrics. ‘Num’ and ‘Size’ indicate the number of images and their training size, respectively, ‘−’ denotes unavailable metrics in related literature.

Method	DSC	BAC	Se	Num	size	Domain	Datasets
Feng [43]	-	0.9634	0.9312	55	256×256	-	DRIONS-DB
Yin [4]	0.92	0.9184	0.8542	325	-	local	ORIGA
Lim [44]	0.9406	-	0.8742	359	157×157	local	MESSIDOR, SEED-DB DRIVE, DIARETDB1,
Abdullah [6]	0.9273	0.9926	0.8860	1,467	-	local	CHASE-DB1, MESSIDOR, DRIONS-DB,
Morales [13]	0.8774	0.9921	0.854	110	-	local	DRIONS
DRIU [24]	0.9361	0.9612	0.9252	110	565×584	local	DRIONS-DB, RIM-ONE-r3
M-Net [27]	0.9632	0.983	-	325	400×400	local	ORIGA
AG-Net [34]	0.9685	-	-	325	640×640	local	ORIGA
U-Net [28]	0.939	0.970	0.944	2,978	256×256	local	DIARETDB0, DRIE, CFI, DIARETDB1, DRIONS-DB, MESSIDOR, ORIGA
U-Net	0.9316	0.9704	0.9420	1,970	256×256	global	MESSIDOR, ORIGA, REFUGE
U-Net	0.9730	0.9802	0.9769	1,970	256×256	local	MESSIDOR, ORIGA, REFUGE
M-Net	0.9078	0.9744	0.9511	1,970	256×256	global	MESSIDOR, ORIGA, REFUGE
M-Net	0.9594	0.9689	0.9591	1,970	256×256	local	MESSIDOR, ORIGA, REFUGE
Deeplabv3	0.9289	0.9680	0.9371	1,970	256×256	global	MESSIDOR, ORIGA, REFUGE
Deeplabv3	0.9571	0.9654	0.9472	1,970	256×256	local	MESSIDOR, ORIGA, REFUGE
The proposed	0.9377	0.9744	0.9501	1,970	256×256	global	MESSIDOR, ORIGA, REFUGE
The proposed	0.9735	0.9799	0.9744	1,970	256×256	local	MESSIDOR, ORIGA, REFUGE

Finally, we are aware of some limitations in this study. First, the developed model tends to include certain background or pathological regions that had a similar appearance as the OD, as illustrated by Fig. 10. This is primarily caused by the sensitivity of the developed method to the object boundaries. Second, we conducted the segmentation experiments on images with a relatively small dimension (i.e., 256×256 pixels) due to the memory limit of the graphics process unit (GPU). This dimension may prevent the detection of certain texture information associated with the OD and its boundary regions and limit the amount of the convolutional features during the training. Third, local disc regions were manually cropped from the global field-of-view images with very small dimensions, relative to the diameter of the OD. In this way, the OD was typically located in the center of the cropped regions, and only a small amount of irrelevant background was retained and significantly magnified (Fig. 11). The background is very close to the OD and generally associated with pathological abnormalities caused by glaucoma, DR, or myopia. It may make the developed model tend to identify incorrect target boundaries. Despite these limitations, our developed method demonstrated an exciting performance in segmenting the OD regions on color fundus images.

Fig. 10.

From left to right columns showed a given fundus image from the MESSIDOR dataset, and its segmentation results obtained by the U-Net (in cyan), M-Net (in red), Deeplabv3 (in green), our developed model (in blue), and manual annotation (in white), respectively. These results were easily affected by pathological regions close to the OD.

Fig. 11.

Illustration of some local disc regions manually cropped from their global field-of-view images. In the disc regions, only a small portion of the background surrounding the OD is retained and generally associated with various pathological abnormalities caused by DR, myopia, or glaucoma.

5. Conclusion

In this study, we introduced an asymmetric segmentation network based on the U-Net model to accurately segment the OD regions from retinal fundus images. Its novelty lies in the integration of a unique cross-connection sub-network and a decoding convolutional block with the classical U-Net architecture. The integration makes the developed network able to reduce the loss of important image information and improve its sensitivity to the morphological changes of the regions-of-interest. Our segmentation experiments on 1,970 color fundus images demonstrated the relatively high performance of the developed network, as compared with other classical CNN models, including the U-Net, M-Net, and Deeplabv3. The developed CNN model has a smaller number of parameters and a better capability to alleviate the problems caused by improper light exposure and tissue abnormality close to the disc regions due to the use of the element-wise subtraction operations. The operations are very useful to locate object boundaries and exclude irrelevant optic disc pallor. We also discussed its limitations. In the future, we will test the generic characteristic of the developed CNN model by applying it to other completely different forms of medical images (e.g., computed tomography (CT) and Magnetic Resonance Imaging (MRI)).