1. Introduction

2. Methods

2.1. Generating the Network

This paper proposes a lightweight segmentation network, which inherits the idea of Deeplabv3+ [11] in its overall structure. After the feature extraction network, a spatial pyramid pooling module (Atrous Spatial Pyramid Pooling, ASPP) with atrous convolution is connected. At the same time, to integrate more feature information and fuse the shallow features of the feature extraction network with the feature map obtained by the spatial pyramid pooling module to obtain more feature information. The difference is that the feature extraction network used by the Deeplabv3+ network is the ResNet [12] series of networks; the performance of the ResNet network in feature extraction is excellent. Still, at the same time, it is accompanied by a vast amount of computation, which goes against the goal of designing a lightweight and efficient segmentation network model. Inspired by the successful application of the MobileNet [13,14,15] series of networks on the mobile side, this paper replaces the Deeplabv3+ feature extraction network ResNet with the MobileNet network, and proposes a feature extraction method based on the MobileNetv3 network.

Figure 1. Overview of MBG-NET network architecture. (a) is the feature extraction module, (b) is the boundary branch, (c) is the mask branch.

Figure 1. Overview of MBG-NET network architecture. (a) is the feature extraction module, (b) is the boundary branch, (c) is the mask branch.

In terms of specific operations, the MBG-Net utilizes the large version of the MobileNetv3 network as its feature extraction network. However, only the first convolutional layer and 15 inverted residual blocks from the large MobileNetv3 version are utilized. Additionally, the stride of the third-to-last inverted residual block is modified from 2 to 1. Experimental comparisons have demonstrated that this alteration is advantageous for extracting local features. To better extract contextual information, multi-scale feature fusion is performed on the extracted feature maps using the spatial pyramid pooling module with atrous convolution. Subsequently, the feature map obtained from the spatial pyramid pooling module undergoes operations such as batch normalization (BN) and rectified linear unit (ReLU) with a convolution kernel size of 1 × 1 to reduce the number of feature map channels. Afterwards, the feature map is upsampled four times, and the resulting feature map is combined with the feature map obtained from the third inverted residual block of MobileNetv3.

To improve the accuracy of OD and OC segmentation, a multi-label approach is utilized. The obtained feature maps are fed into the boundary prediction branch and mask prediction branch. Specifically, for the boundary branch, the feature map undergoes three convolution operations, where the output channels of the first two convolutions are set to 256. The final convolution operation produces a feature map with a single channel, resulting in the predicted boundaries for OD and OC. As for the mask branch, the feature map is concatenated with the predicted boundary feature map and then subjected to a convolution operation. The output feature map consists of 2 channels, corresponding to the predicted OD and OC masks. The obtained mask features are subsequently upsampled by a factor of four to obtain a mask prediction map with the same size as the input image. All the convolution operations mentioned above have a stride of 1. To optimize the boundary and mask predictions, the boundary regression loss and mask prediction loss are defined by equations (1) and (2), respectively.

$$L_b={\displaystyle \frac{1}{N}}\sum _i^N{(y_i^b-p_i^b)}^2$$

(1)

$$L_m=-{\displaystyle \frac{1}{N}}\sum _i^N[y_i^m*log\left(p_i^m\right)+(1-y_i^m)log(1-p_i^m)]$$

(2)

For formula 1, N represents the number of pixels, $y_i^b$ is the natural boundary map generated by morphological closure operation and Gaussian filter, and $p_i^b$ is the boundary prediction map predicted by the segmentation network. For Eq.2, $y_i^m$ is the ground-truth mask label and $p_i^m$ is the mask prediction map predicted by the segmentation network.

The method in this paper sends the prediction mask map to the patch discriminator (PatchGAN) to deceive the patch discriminator into optimizing the parameters of the entire segmentation network so that the segmentation network can generate a more realistic prediction map. The segmentation network is optimized using the subsequent adversarial loss:

$$L_{adv}=-{\displaystyle \frac{1}{N}}\sum _i^N{L}_{bce}(p^m,1)$$

(3)

Joint Loss function: We can consider loss function that is a combination of the segmentation loss and the boundary loss as follows:

$$L_{total}=L_b+L_m+L_{adv}$$

(4)

To sum up, the loss for the segmentation network consists of the boundary prediction loss, the mask prediction loss, and the adversarial loss, namely, and $\beta$ is set to 0.01 according to experience.

3. Experiments and Results

3.1. Datasets

The experiment selected three public fundus datasets, Drishti-GS, RIM-ONE-r3, and REFUGE, as the experimental datasets. The Drishti-GS dataset consists of 101 fundus images, including 31 normal fundus and 70 glaucoma fundus images. The RIM-ONE-r3 dataset consists of 159 fundus images, including 85 normal fundus and 74 glaucoma fundus images. The REFUGE dataset is currently the largest open-source glaucoma fundus dataset. It consists of 1200 fundus images, including 400 training sets, 400 validation sets, and 400 test sets. The sample graphs and annotation graphs of the three datasets are shown in Figure 3. (a) (d) is the sample map of glaucoma fundus and the corresponding annotation map of the Drishti-GS dataset, (b) (e) is the sample map of the normal eye fundus, and the corresponding annotation map of the RIM-ONE-r3 dataset, (c) (f) is the glaucoma fundus sample map and the corresponding annotation map of the REFUGE dataset. In (d) (e) (f), green represents the optic disc, and blue represents the optic cup. The experiment uses the training set in the REFUGE dataset as the experimental dataset, of which 320 images are used as training and validation sets, and 80 images are used as test sets. The specific data of the three datasets are shown in Table 1:

Table 1. Statistics of the datasets used in evaluating our method..

Table 1. Statistics of the datasets used in evaluating our method..

Dataset	Number of samples	Image size	Cameras	Release year
Drishti-GS [16]	50+51	2047×1759	unknown	2014
RIM-ONE-r3 [17]	99+60	2144×1424	unknown	2015
REFUGE [18]	320+80	2124×2056	Zeiss Visucam 500	2018

¹ Tables may have a footer.

3.4. Experimental Results

For Drishti-GS and RIM-ONE-r3 datasets ,We compare our MBG-Net with several stste-of-the-art OD and OC segmentation methods: pOSAL [17], BEAL [19], TD-GAN [20], BGA-net [21], Sevastopolsky [22], Zilly [23]. The quantitative results are shown in Table 2, and the qualitative results are shown in Figure 4 and Figure 5.

Table 2. Performance Comparison of Attention Blocks for Optic Disc and Optic Cup Segmentation on Drishti-GS and Rim_One_r3 datasets.(Use mIoU to select the best model)

Table 2. Performance Comparison of Attention Blocks for Optic Disc and Optic Cup Segmentation on Drishti-GS and Rim_One_r3 datasets.(Use mIoU to select the best model)

Methods	Datasets
	Drishti-GS			Rim_One_r3
	${\mathit{DI}}_{\mathit{disc}}$	${\mathit{DI}}_{\mathit{cup}}$	$\mathit{\delta }$	${\mathit{DI}}_{\mathit{disc}}$	${\mathit{DI}}_{\mathit{cup}}$	$\mathit{\delta }$
U-net [16]	0.904	0.852	-	0.864	0.797	-
M_UNet [24]	0.95	0.85	-	0.95	0.82	-
M-Net [25]	0.967	0.808	-	0.952	0.802	-
CE-Net [26]	0.964	0.882	-	0.953	0.844	-
CDED-Net [27]	0.959	0.924	-	0.958	0.862	-
$pOSAL$ [17]	0.965	0.858	0.082	0.865	0.787	0.081
Gan-based [28]	0.953	0.864	-	0.953	0.825	-
BEAL	0.961	0.862	-	0.898	0.810	-
TD-GAN	0.924	0.747	0.117	0.853	0.728	0.118
PDD-UNET [29]	0.963	0.848	0.105	0.970	0.876	0.066
Ours	0.974	0.900	0.045	0.966	0.875	0.043

For the REFUGE dataset, its training set is used as experimental data. Take 320 images as the training set and 80 as the test set. Compared with the BGA-net segmentation model with better effect in Table 2, the network model proposed in this paper improves the inference time by about 16% compared with BGA-net, and the segmentation accuracy of the OC is also improved. And the index also has a specific improvement. The quantitative results are shown in Table 3, and Figure 6 shows the quantitative results of the REFUGE dataset.

Table 3. Performance Comparison of Dice coefficient of OD and OC segmentation on REFUGE [31] dataset.

Table 3. Performance Comparison of Dice coefficient of OD and OC segmentation on REFUGE [31] dataset.

Methods	${\mathit{D}}_{\mathit{disc}}$	${\mathit{D}}_{\mathit{cup}}$	$\mathit{vCDR}$	inference time
U-net [16]	0.927	0.848	-	-
$pOSAL$ [17]	0.932	0.869	0.059	-
Psi-Net [18]	0.956	0.851	-	-
BEAL [32]	0.945	0.860	-	-
BGA-NET [30]	0.951	0.866	0.040	29.1ms
ours	0.951	0.869	0.013	24.3ms

¹ Tables may have a footer.

Table 4. The necessity of ROI region extraction on REFUGE [31] dataset.

Table 4. The necessity of ROI region extraction on REFUGE [31] dataset.

Dataset	${\mathit{D}}_{\mathit{disc}}$	${\mathit{D}}_{\mathit{cup}}$
Original dataset	0.920	0.820
ROI extraction dataset	0.962	0.880

¹ Tables may have a footer.

At the same time, this paper compares the model parameters, occupied memory, calculation amount, and inference time of the network models under different feature extraction networks to prove the excellent performance and lightweight of Mobilenetv3 as a feature extraction network. The specific results are shown in Table 5:

Table 5. The necessity of ROI region extraction on REFUGE [31] dataset.

Table 5. The necessity of ROI region extraction on REFUGE [31] dataset.

Backbone	Total params	Total memory	Total MAdd	Total Flops	Total MemR+W	Segmentation Time
efficientnetv2_m	54.9M	2299.8M	158.4G	79.3G	3.1GB	192.7ms
Xception	52.1M	1524.5M	165.8G	83.0G	3.1GB	170.4ms
Resnet50	38.4M	845.3M	138.5G	69.3G	1.7GB	58.9ms
Mobilenetv2	5.5M	651.1M	52.9G	26.5G	1.2GB	29.1ms
Ours	5.3M	495.4M	48.8G	24.4G	839.0M	24.3ms

¹ Tables may have a footer.

CDR is an important clinical parameter for the diagnosis of glaucoma, and it is also an essential basis for the diagnosis of glaucoma by most professional ophthalmologists. The receiver operating characteristic curve (ROC) and its corresponding area under the curve (AUC) are used to evaluate the performance of the proposed algorithm in detecting glaucoma. Generally, the higher the AUC, the better the diagnosis of the system. The higher the accuracy, the better the performance of the algorithm. This paper uses the ROC curve to evaluate the segmentation performance of BGA-Net. Figure 7 shows the ROC curve and AUC values corresponding to Drishti-GS, RIM-ONE-r3, and REFUGE-train datasets.

3.5. Ablation Experiments

The following ablation experiments are designed to verify the effectiveness of the boundary branch and the discriminator network: (1) Baseline (feature extraction network + ASPP + mask branch). (2) Baseline + boundary branch. (3) Baseline + boundary branch + discriminator network.

Ablation experiments are performed on the RIM-ONE-r3 dataset, and the experimental results of different module combinations are shown in Table 6. Experiments show that, compared with the Baseline network, the boundary branch and the discriminator network both improve the segmentation performance of the model to varying degrees. The network structure with boundary branches added to the Baseline improves the OC segmentation accuracy by 0.3%, and decreases by 0.004. The discriminator and boundary branch components are added; that is, the network structure of MBG-Net improves the OC accuracy by 0.7% compared with the Baseline+ boundary branch, and reduces by 0.003, which proves the effectiveness of the boundary branch and the discriminator network. At the same time, the qualitative segmentation results are given in Figure 8, green represents the OD, and blue represents the OC. Among them, column (a) is the ROI area corresponding to the fundus image, column (b) is the manual annotation map, column (c) is the predicted segmentation map corresponding to the Baseline, (d) column is the predicted segmentation map corresponding to the Baseline+Boundary branch, (e) The column is the predicted segmentation map corresponding to the Baseline + boundary branch + discriminator network (MBG-Net). Compared with columns (c) and (d), the boundary of the segmentation map in column (e) is smoother and more natural, and it is closer to the manual annotation map.

Table 6. Ablation study on different components.

Table 6. Ablation study on different components.

Dataset	${\mathit{DI}}_{\mathit{disc}}$	${\mathit{DI}}_{\mathit{cup}}$	$\mathit{\delta }$
Baseline	0.964	0.865	0.050
Baseline+B	0.964	0.868	0.046
Baseline+B+G	0.966	0.875	0.043

¹ Tables may have a footer.

4. Discussion

5. Results

6. Patents

References

1. Tham, Y.C.; Li, X.; Wong, T.Y.; Quigley, H.A.; Aung, T.; Cheng, C.Y. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology 2014, 121, 2081–2090. [Google Scholar] [CrossRef]
2. Mary, V.S.; Rajsingh, E.B.; Naik, G.R. Retinal fundus image analysis for diagnosis of glaucoma: a comprehensive survey. IEEE Access 2016. [Google Scholar]
3. Zhu, Q.; Chen, X.; Meng, Q.; Song, J.; Luo, G.; Wang, M.; Shi, F.; Chen, Z.; Xiang, D.; Pan, L.; et al. GDCSeg-Net: general optic disc and cup segmentation network for multi-device fundus images. Biomedical Optics Express 2021, 12, 6529–6544. [Google Scholar] [CrossRef]
4. Chen, C.; Wang, G. IOSUDA: An unsupervised domain adaptation with input and output space alignment for joint optic disc and cup segmentation. Applied Intelligence 2021, 51, 3880–3898. [Google Scholar] [CrossRef]
5. Fu, H.; Cheng, J.; Xu, Y.; Wong, D.W.K.; Liu, J.; Cao, X. Joint optic disc and cup segmentation based on multi-label deep network and polar transformation. IEEE transactions on medical imaging 2018, 37, 1597–1605. [Google Scholar] [CrossRef]
6. Al-Bander, B.; Williams, B.M.; Al-Nuaimy, W.; Al-Taee, M.A.; Pratt, H.; Zheng, Y. Dense fully convolutional segmentation of the optic disc and cup in colour fundus for glaucoma diagnosis. Symmetry 2018, 10, 87. [Google Scholar] [CrossRef]
7. Qin, P.; Wang, L.; Lv, H. Optic disc and cup segmentation based on deep learning. In Proceedings of the 2019 IEEE 3rd Information Technology, Networking, Electronic and Automation Control Conference (ITNEC); IEEE, 2019; pp. 1835–1840. [Google Scholar]
8. Zhang, Y.; Cai, X.; Zhang, Y.; Kang, H.; Ji, X.; Yuan, X. TAU: Transferable Attention U-Net for optic disc and cup segmentation. Knowledge-Based Systems 2021, 213, 106668. [Google Scholar] [CrossRef]
9. Zhao, X.; Wang, S.; Zhao, J.; Wei, H.; Xiao, M.; Ta, N. Application of an attention u-net incorporating transfer learning for optic disc and cup segmentation. Signal, Image and Video Processing 2021, 15, 913–921. [Google Scholar] [CrossRef]
10. Liu, B.; Pan, D.; Shuai, Z.; Song, H. ECSD-Net: A joint optic disc and cup segmentation and glaucoma classification network based on unsupervised domain adaptation. Computer Methods and Programs in Biomedicine 2022, 213, 106530. [Google Scholar] [CrossRef] [PubMed]
11. Chen, L.C.; Zhu, Y.; Papandreou, G.; Schroff, F.; Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the Proceedings of the European conference on computer vision (ECCV); 2018; pp. 801–818. [Google Scholar]
12. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the Proceedings of the IEEE conference on computer vision and pattern recognition; 2016; pp. 770–778. [Google Scholar]
13. Howard, A.G.; Zhu, M.; Chen, B.; Kalenichenko, D.; Wang, W.; Weyand, T.; Andreetto, M.; Adam, H. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 2017.
14. Sandler, M.; Howard, A.; Zhu, M.; Zhmoginov, A.; Chen, L.C. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the Proceedings of the IEEE conference on computer vision and pattern recognition; 2018; pp. 4510–4520. [Google Scholar]
15. Howard, A.; Sandler, M.; Chu, G.; Chen, L.C.; Chen, B.; Tan, M.; Wang, W.; Zhu, Y.; Pang, R.; Vasudevan, V.; et al. Searching for mobilenetv3. In Proceedings of the Proceedings of the IEEE/CVF international conference on computer vision, 2019. pp. 1314–1324.
16. Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical image computing and computer-assisted intervention. Springer; 2015; pp. 234–241. [Google Scholar]
17. Wang, S.; Yu, L.; Yang, X.; Fu, C.W.; Heng, P.A. Patch-based output space adversarial learning for joint optic disc and cup segmentation. IEEE transactions on medical imaging 2019, 38, 2485–2495. [Google Scholar] [CrossRef] [PubMed]
18. Murugesan, B.; Sarveswaran, K.; Shankaranarayana, S.M.; Ram, K.; Joseph, J.; Sivaprakasam, M. Psi-Net: Shape and boundary aware joint multi-task deep network for medical image segmentation. In Proceedings of the 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC); IEEE, 2019; pp. 7223–7226. [Google Scholar]
19. Wang, S.; Yu, L.; Li, K.; Yang, X.; Fu, C.W.; Heng, P.A. Boundary and entropy-driven adversarial learning for fundus image segmentation. In Proceedings of the Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part I 22. Springer, 2019, pp. 102–110.
20. Son, J.; Park, S.J.; Jung, K.H. Towards accurate segmentation of retinal vessels and the optic disc in fundoscopic images with generative adversarial networks. Journal of digital imaging 2019, 32, 499–512. [Google Scholar] [CrossRef]
21. Luo, L.; Xue, D.; Pan, F.; Feng, X. Joint optic disc and optic cup segmentation based on boundary prior and adversarial learning. International Journal of Computer Assisted Radiology and Surgery 2021, 16, 905–914. [Google Scholar] [CrossRef] [PubMed]
22. Sevastopolsky, A. Optic disc and cup segmentation methods for glaucoma detection with modification of U-Net convolutional neural network. Pattern Recognition and Image Analysis 2017, 27, 618–624. [Google Scholar] [CrossRef]
23. Zilly, J.; Buhmann, J.M.; Mahapatra, D. Glaucoma detection using entropy sampling and ensemble learning for automatic optic cup and disc segmentation. Computerized Medical Imaging and Graphics 2017, 55, 28–41. [Google Scholar] [CrossRef] [PubMed]
24. Sevastopolsky, A. Optic disc and cup segmentation methods for glaucoma detection with modification of U-Net convolutional neural network. Pattern Recognition and Image Analysis 2017, 27, 618–624. [Google Scholar] [CrossRef]
25. Fu, H.; Cheng, J.; Xu, Y.; Wong, D.W.K.; Liu, J.; Cao, X. Joint Optic Disc and Cup Segmentation Based on Multi-Label Deep Network and Polar Transformation. IEEE Transactions on Medical Imaging 2018, 37, 1597–1605. [Google Scholar] [CrossRef] [PubMed]
26. Gu, Z.; Cheng, J.; Fu, H.; Zhou, K.; Hao, H.; Zhao, Y.; Zhang, T.; Gao, S.; Liu, J. Ce-net: Context encoder network for 2d medical image segmentation. IEEE transactions on medical imaging 2019, 38, 2281–2292. [Google Scholar] [CrossRef] [PubMed]
27. Tabassum, M.; Khan, T.M.; Arsalan, M.; Naqvi, S.S.; Ahmed, M.; Madni, H.A.; Mirza, J. CDED-Net: Joint segmentation of optic disc and optic cup for glaucoma screening. IEEE Access 2020, 8, 102733–102747. [Google Scholar] [CrossRef]
28. Son, J.; Park, S.J.; Jung, K.H. Towards accurate segmentation of retinal vessels and the optic disc in fundoscopic images with generative adversarial networks. Journal of digital imaging 2019, 32, 499–512. [Google Scholar] [CrossRef]
29. Shankaranarayana, S.M.; Ram, K.; Mitra, K.; Sivaprakasam, M. Fully convolutional networks for monocular retinal depth estimation and optic disc-cup segmentation. IEEE journal of biomedical and health informatics 2019, 23, 1417–1426. [Google Scholar] [CrossRef]
30. Luo, L.; Xue, D.; Pan, F.; Feng, X. Joint optic disc and optic cup segmentation based on boundary prior and adversarial learning. International Journal of Computer Assisted Radiology and Surgery 2021, 16, 905–914. [Google Scholar] [CrossRef]
31. Orlando, J.I.; Fu, H.; Breda, J.B.; Keer, K.V.; Bathula, D.R.; Diaz-Pinto, A.; Fang, R.; Heng, P.A.; Kim, J.; Lee, J.H. REFUGE Challenge: A Unified Framework for Evaluating Automated Methods for Glaucoma Assessment from Fundus Photographs. Medical Image Analysis 2019. [Google Scholar] [CrossRef]
32. Wang, S.; Yu, L.; Li, K.; Yang, X.; Fu, C.W.; Heng, P.A. Boundary and entropy-driven adversarial learning for fundus image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer; 2019; pp. 102–110. [Google Scholar]