基于改进U-Net的联合视杯视盘分割方法  

Joint optic disc and cup segmentation method based on improved U-Net

作　　者：阎世梁 王银玲 路丹丹 熊亮[2] 卜英博 徐杨 YAN Shiliang;WANG Yinling;LU Dandan;XIONG Liang;BU Yingbo;XU Yang(Engineering Technology Center,Southwest University of Science and Technology,Mianyang 621101,Sichuan,China;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621101,Sichuan,China)

机构地区：[1]西南科技大学工程技术中心,四川绵阳621010 [2]西南科技大学信息工程学院,四川绵阳621010

出　　处：《微电子学与计算机》2023年第10期90-101,共12页Microelectronics & Computer

基　　金：四川省高等教育人才培养质量和教学改革项目(JG2021-893);国家重点研发计划子项目(2019YFB130501)。

摘　　要：为了实现眼底图像视杯视盘的精准分割,减少人工分割方法带来的不确定性和耗时性,本文提出了一种新型的卷积神经网络用于联合视杯视盘的分割,称为M2DS-TransUNet.该网络采用一种多分辨率图像结合并通过压缩与激励模块进行自适应提取的输入形式,同时结合多分辨率模块、Transformer和深度监督机制的优势,使得网络可以提取更加丰富的图像信息.采用五折交叉验证的方式对网络模型进行训练,并在当前三个主流数据集REFUGE、DRISHTI-GS和RIM-ONE-r3上进行了实验验证与评估,在最能体现分割效果的杯盘比指标上分别达到了0.0284、0.0978和0.0179,其分割效果优于当前的一些经典算法.实验结果表明,本文所提出的方法可以提取更为丰富的视杯视盘信息,且具有跨数据集的泛化能力,是一种非常有竞争力的眼底图像视杯视盘联合分割方法.In order to achieve accurate segmentation of the optic cup and optic disc of fundus images and to reduce the uncertainty and time-consuming nature of manual segmentation methods,a novel convolutional neural network for joint optic cup and optic disc segmentation,called M2DS-TransUNet,is proposed in this paper.This network adopts a multi-resolution image combination and adaptive extraction of the input form through the squeeze and excitation modules.It also combines the advantages of multi-resolution module,Transformer and depth supervision,which allows the network to extract richer image information.The network model is trained using a five-fold cross-validation approach and experimentally validated and evaluated on three current mainstream datasets REFUGE,DRISHTI-GS and RIM-ONE-r3,which achieve 0.0284,0.0978 and 0.0179 respectively in the cup-to-disc ratio index that best reflects the segmentation effect,and its segmentation effect is better than some current classical algorithms.The experimental results show that the proposed method can extract richer information of the visual cup-vision disc and has the ability of generalization across data sets,which is a very competitive method for joint optic cup and optic disc segmentation.

关 键 词：视杯视盘 分割 U-Net TRANSFORMER 深度监督机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]