基于边缘U^(2)-Net的视盘分割方法  

Disc Segmentation Method Based on Edge U^(2)-Net

作　　者：王雪 武现阳 涂家亮 于洁茹 宁春玉[1] WANG Xue;WU Xianyang;TU Jialiang;YU Jieru;NING Chunyu(School of Life Science and Technology,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学生命科学技术学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2024年第3期93-100,共8页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技发展计划项目(20220101123JC)。

摘　　要：彩色眼底图像中的视盘分割在识别眼科疾病中起着关键作用。针对因各种因素影响的视盘边缘分割不准确及分割算法效率低问题,提出一种基于轻量级U^(2)-Net、融入边缘注意力机制的视盘自动分割方法。该方法以轻量级U^(2)-Net为主干网络,使用视盘感兴趣区域提取的预处理方式去除无关特征,同时引入边缘注意力机制增强对视盘边缘特征的提取能力。在Drishti_GS和REFUGE两个公开数据集上的F1分数分别达到97.82%和97.36%,Dice相似系数分别达到97.15%和96.64%,IOU分别达到94.47%和93.50%,与其他网络模型相比表现出优越的分割性能,具有临床应用价值。The segmentation of the optic disc in color fundus images is crucial for the identification of ophthalmic diseases.To address the issues of inaccurate disc edge segmentation due to various factors and the low efficiency of segmentation algorithms,an automatic disc segmentation is proposed which is based on the lightweight U^(2)-Net and integrated with an edge attention mechanism.This method employs the lightweight U^(2)-Net as the backbone network,utilizes a preprocessing approach of extracting the region of interest of the optic disc to remove irrelevant features,and simultaneously introduces an edge attention mechanism to enhance the extraction capability of optic disc edge features.On two public datasets,Drishti_GS and REFUGE,the F1 score reaches 97.82%and 97.36%,the Dice similarity indices reaches 97.15%and 96.64%,and the IOU reaches 94.47%and 93.50%,respectively.Compared to other network models,the proposed method demonstrates superior segmentation performance,showcasing its clinical application value.

关 键 词：彩色眼底图像 视盘分割 U^(2)-Net 感兴趣区域提取 边缘注意力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]