基于多路径网络的多病种精准视网膜血管网络分割方法的建立及应用  

Development and application of an accurate retinal vascular network segmentation method for multiple diseases based on a multi-path network

作　　者：张金泽 李嘉雄 王耿媛 袁进[1] 肖鹏[1] Zhang Jinze;Li Jiaxiong;Wang Gengyuan;Yuan Jin;Xiao Peng(State Key Laboratory of Ophthalmology,Zhongshan Ophthalmic Center,Sun Yat-sen University,Guangdong Provincial Key Laboratory of Ophthalmology and Visual Science,Guangzhou 510060,China)

机构地区：[1]中山大学中山眼科中心,眼病防治全国重点实验室,广东省眼科视觉科学重点实验室,广州510060

出　　处：《中华实验眼科杂志》2024年第12期1120-1126,共7页Chinese Journal Of Experimental Ophthalmology

基　　金：国家自然科学基金重点项目(82230033);国家自然科学基金面上项目(82271133);广东省自然科学基金面上项目(2022A1515011486);中山大学高校基本科研业务费项目(24ykqb009);眼科学国家重点实验室基础研究基金。

摘　　要：目的建立一种适用于眼底多病种的精准视网膜血管网络分割方法,探讨不同类型眼底疾病的视网膜血管形态参数变化规律。方法采用回顾性研究方法,收集2020年1月至2023年12月在中山大学中山眼科中心就诊的眼底病患者829例及健康受试者146名的彩色眼底照相数据。将多路径分割网络进行微调,输入眼底图像血管分割公开数据集中糖尿病视网膜病变(DR)、青光眼以及年龄相关性黄斑变性(AMD)患者及健康成人的彩色眼底照相数据进行训练,直至模型loss值不再下降,最终得到完成训练的多病种视网膜血管分割模型。应用本课题组之前开发的视网膜血管形态特征分析方法,对受试者以黄斑为中心的彩色眼底图像进行分析,提取视网膜血管分型维数(D_(f))、血管面积比(VAR)、平均血管直径(D_(m))和扭曲度(τ)等形态特征参数,比较不同疾病组的视网膜血管形态特征参数。结果构建的多病种彩色眼底照相血管分割模型在测试集上的准确率为0.987,受试者工作特征曲线下面积为0.995。校正年龄和性别后,不同组间D_(f)校正、VAR校正、D m校正和τ总体比较差异均有统计学意义(F=27.87、47.60、26.48、4.63,均P<0.001),其中AMD组、DR组、糖尿病性黄斑水肿(DME)组、视网膜色素变性(RP)组、视网膜分支静脉阻塞(BRVO)组和视网膜中央静脉阻塞(CRVO)组D_(f)校正较健康对照组显著下降,差异均有统计学意义(均P<0.05);除视神经炎组和中心性浆液性脉络膜视网膜病变组外其他所有疾病组VAR校正较健康对照组显著下降,差异均有统计学意义(均P<0.05);DME组、青光眼组、RP组、BRVO组和CRVO组D m校正较健康对照组显著下降,差异均有统计学意义(均P<0.05)。τ不受年龄和性别影响,无需校正。DR组和DME组τ较健康对照组显著上升,差异均有统计学意义(均P<0.05)。结论成功构建了适用于眼底多病种的视网膜血管精准分割�Objective To establish an accurate retinal vascular network segmentation method for multiple fundus diseases,and to investigate the changing patterns of retinal vascular morphological parameters in these diseases.Methods A retrospective study was conducted.Color fundus photography data of 829 patients with fundus diseases and 146 healthy adults were collected at Zhongshan Ophthalmic Center,Sun Yat-sen University from January 2020 to December 2023.The multi-path segmentation network was fine-tuned,and the color fundus photography data of diabetic retinopathy(DR),glaucoma and age-related macular degeneration(AMD)patients and healthy adults in the fundus image vessel segmentation public dataset were input for training until the loss value of the model stopped decreasing,and finally the trained multi-disease retinal vascular segmentation model was obtained.The retinal blood vessel morphological characteristics analysis method previously developed by our research group was used to analyze the subjects'color fundus images centered on the macula,the retinal blood vessel fractal dimension(D_(f)),vascular area ratio(VAR),mean diameter(D_(m)),tortuosity(τ)and other morphological characteristics parameters were extracted and compared among various disease groups.This study adhered to the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of Zhongshan Ophthalmic Center,Sun Yat-sen University(No.2023KYPJ344).Written informed consent was obtained from each subject.Results The accuracy of the multi-disease color fundus photography vessel segmentation model on the test set was 0.987,and the area under the receiver operating characteristic curve was 0.995.After adjustment for age and sex,there were statistically significant differences in adjusted D_(f),adjusted VAR,adjusted D m andτamong different groups(F=27.87,47.60,26.48,4.63;all at P<0.001).Adjusted D_(f)in AMD group,DR group,diabetic macular edema(DME)group,retinitis pigmentosa(RP)group,branch retinal vein occlusion(BRVO)group and central reti

关 键 词：视网膜疾病 视网膜血管 人工智能 彩色眼底照相 形态学参数 

分 类 号：R774.1[医药卫生—眼科]