结合深度学习的糖尿病视网膜病变血管分割和重建  

Vascular segmentation and reconstruction in diabetic retinopathy based on deep learning

作　　者：许诗怡 陈明惠[1] 邵怡[2] 秦楷博 吴玉全 尹志杰 杨政奇 XU Shiyi;CHEN Minghui;SHAO Yi;QIN Kaibo;WU Yuquan;YIN Zhijie;YANG Zhengqi(Medical Optical Engineering Center,Ministry of Education/Shanghai Engineering Research Center of Interventional Medical Device/School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Department of Urolithology,Shanghai General Hospital,Shanghai 200080,China)

机构地区：[1]上海理工大学健康科学与工程学院/上海介入医疗器械工程技术研究中心/教育部医学光学工程中心,上海200093 [2]上海市第一人民医院泌尿结石科,上海200080

出　　处：《中国医学物理学杂志》2024年第10期1256-1264,共9页Chinese Journal of Medical Physics

基　　金：电子内窥镜研发(3A-23-312-042)。

摘　　要：为了助于早期诊断糖尿病性视网膜病变,提出结合视网膜血管分割与三维重建的方法。三维重建可以避免分割后血管长度、曲度和分支角度等信息误判影响早期诊断。提出IAAnet算法进行视网膜图像分割,将传统Unet网络与Inception V3、ASPP、AttentionGates相结合,较好地减少信息损失并避免过拟合的现象,提高网络对特征的提取能力。运用投影重建法来还原血管三维信息,并支持调节亮度、对比度,使医生更好地观察血管的真实状态。本文算法在准确率、召回率、F1分数、交并比、ROC曲线下面积上的结果分别是97.68%、96.07%、97.26%、92.79%、94.00%,通过与其他网络对比,IAAnet算法具有良好的分割准确性,三维投影重建后能在三维图像上获取更丰富的血管信息为早期诊断提供帮助。A method capable of retinal vessel segmentation and three-dimensional(3D)reconstruction is proposed for the early diagnosis of diabetic retinopathy.The 3D reconstruction can avoid the misjudgments of blood vessel length,curvature and branch angle after segmentation,which will affect the early diagnosis.IAAnet algorithm for retinal image segmentation combines traditional Unet with Inception V3,atrous spatial pyramid pooling and AttentionGates to reduce information loss and avoid over-fitting,thereby improving the network's ability to extract features.The projection reconstruction method is used to restore the 3D information of blood vessels,and supports the adjustments of brightness and contrast,so that doctors can better observe the real state of blood vessels.The proposed algorithm has an accuracy,recall rate,F1 score,intersection over union and area under ROC curve of 97.68%,96.07%,97.26%,92.79%and 94.00%,respectively.Compared with other networks,IAAnet algorithm exhibits higher segmentation accuracy,and can obtain more vascular information in 3D image after 3D projection reconstruction to assist in the early diagnosis.

关 键 词：深度学习 糖尿病性视网膜病变 Inception V3 注意力门 空洞金字塔池化 三维投影重建 

分 类 号：R318[医药卫生—生物医学工程] TP391[医药卫生—基础医学]