Enhanced segmentation of optic disc and cup using attention-based U-Net with dense dilated series convolutions

作     者：Kumar, G. Bharadwaja Kumar, Soham 

作者机构：Department of Computer Science and Engineering VIT Chennai TamilNadu Chennai600127 India 

出 版 物：《Neural Computing and Applications》 (Neural Comput. Appl.)

年 卷 期：2025年第37卷第9期

页      面：6831-6847页

核心收录：

学科分类：1201[管理学-管理科学与工程(可授管理学、工学学位)] 070207[理学-光学] 08[工学] 0837[工学-安全科学与工程] 0803[工学-光学工程] 0701[理学-数学] 0702[理学-物理学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Not applicable 

主　　题：Image segmentation 

摘      要：Delineating the boundaries of the optic disc and cup regions is a critical pre-requisite for glaucoma screening because it allows for precise measurement of key parameters, such as cup-to-disc ratio, which is a critical indicator of optic nerve head damage, a hallmark of glaucoma progression. Accurate segmentation enables early detection and monitoring of the disease, aiding in timely intervention to prevent vision loss. The main contribution of this research work is to develop an automated process to isolate and demarcate the optic disc and cup from retinal fundus images. To prevent the blood vessels from interfering with the segmentation process, a novel method is used for vessel mask generation and vessel inpainting. Most of the research works have used based encoder–decoder models like U-Net architecture or handcrafted feature extraction techniques such as hough transform, fuzzy clustering, etc. The proposed model has made significant modifications to the U-Net model. (1) Dual attention mechanism at every layer of decoder and (2) dense dilated series convolutions as skip connections to generate higher level feature map. The proposed model achieved benchmark accuracies - Dice score of 95.95% and IoU score of 92.22% for optic disc segmentation averaged over fivefold. For the task of outlining the optic cup region, it attained a Dice score of 88.7% and IoU of 79.72%. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2025.