计算机应用研究2026，Vol.43Issue(3)：940-947,8.DOI:10.19734/j.issn.1001-3695.2025.05.0224
青光眼视盘视杯分割中的特征增强与类平衡优化策略
Strategies for feature enhancement and class balance optimization in glaucoma optic disc and cup segmentation
摘要
Abstract
To enhance the segmentation accuracy of the optic disc and optic cup and assist in the early diagnosis of glaucoma,this paper proposed an improved deep separable multi-scale aggregation network based on the U-Net architecture(depthwise sep-arable multi-scale aggregation network,DSMA-Net).First,it introduced an adaptive separable multi-scale attention module to en-hance feature representation and effectively capture the complex structures of the optic disc and cup,thereby reducing boundary blurring.Second,it replaced some traditional convolutions in the network with depthwise separable convolutions to reduce compu-tational complexity and improve model efficiency.Additionally,it used an improved adaptive weighted Tversky loss function to fo-cus the network on optic disc and cup pixels,alleviating the dominance of background pixels and further improving segmentation accuracy.Experimental results show that DSMA-Net has significant advantages in optic disc and cup segmentation tasks on the DRISHTI-GS,RIM-ONE-R3,and ORIGA datasets,providing strong technical support for early glaucoma diagnosis.关键词
U-Net/视盘视杯分割/青光眼检测/深度学习/特征融合Key words
U-Net/optic disc and cup segmentation/glaucoma detection/deep learning/feature fusion分类
信息技术与安全科学引用本文复制引用
罗敏,曹路,曾军英,麦超云,黄秀清,赵菲菲,何锡权..青光眼视盘视杯分割中的特征增强与类平衡优化策略[J].计算机应用研究,2026,43(3):940-947,8.基金项目
广东普通高校重点领域专项(2025ZDZX1040,2024ZDZX1009,2022ZDZX1033) （2025ZDZX1040,2024ZDZX1009,2022ZDZX1033）
江门市省科技创新战略专项项目计划资助项目(江科[2023]72号) （江科[2023]72号）
江门市医疗卫生科技计划资助项目(2025YL03026) （2025YL03026）
江门市基础理论科研科技计划资助项目(2023JC010039) （2023JC010039）
五邑大学大学生创新创业训练计划资助项目(X202511349050) （X202511349050）
