基于多模态数据的深度学习在青光眼诊断和严重程度分级中的应用  

Application of deep learning with multimodal data in glaucoma diagnosis and severity grading

作　　者：钱朝旭 周凌翔 冯雪丽 陈曦[3] 杨文艳[4] 易三莉[2] 钟华[4] Qian Chaoxu;Zhou Lingxiang;Feng Xueli;Chen Xi;Yang Wenyan;Yi Sanli;Zhong Hua(Shanghai Aier Eye Hospital,Shanghai Aier Eye Institute,Shanghai 200030,China;School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,China;Department of Ophthalmology,The First Affiliated Hospital of Nanjing Medical University,Nanjing 210029,China;Department of Ophthalmology,The First Affiliated Hospital of Kunming Medical University,Kunming 650032,China)

机构地区：[1]上海爱尔眼科医院,上海爱尔眼科研究所,上海200030 [2]昆明理工大学信息工程与自动化学院,昆明650504 [3]南京医科大学第一附属医院眼科,南京210029 [4]昆明医科大学第一附属医院眼科,昆明650032

出　　处：《中华实验眼科杂志》2024年第12期1149-1154,共6页Chinese Journal Of Experimental Ophthalmology

基　　金：国家自然科学基金(81960176);湖南省自然科学基金(2023JJ70013);上海市徐汇区医学科研项目(SHXH202317);爱尔眼科医院集团科研基金(AGK2306D03)。

摘　　要：目的基于多模态数据开发能诊断青光眼并识别严重程度的深度学习模型。方法采用诊断试验研究方法,于2023年6—12月在昆明医科大学第一附属医院眼科收集正常人群86人145眼和不同严重程度原发性开角型青光眼患者314例507眼的彩色眼底照相和视野检查结果,并根据视野的平均缺损值将青光眼分为早期154眼、中期113眼和晚期240眼。分别采用DenseNet 121、ResNet 50和VGG 19卷积神经网络(CNN)模型建立人工智能(AI)青光眼严重程度分级模型,评估单模态数据与多模态数据对于分类结果的影响,并确定适合多模态数据的CNN网络架构。结果同时具有彩色眼底照相和视野检查结果者有652眼,按照4∶1的比率采用计算机取随机数法将图片分配到训练集和测试集。不同CNN模型建立的青光眼严重程度分级模型均具有较高的准确性,DenseNet 121整体有效性指标高于ResNet 50和VGG 19。在彩色眼底照相单模态AI模型、视野单模态AI模型、彩色眼底照相联合视野的多模态AI模型中,识别早期青光眼的受试者工作特征曲线下面积分别为0.87、0.93和0.95。结论基于多模态数据能建立具有高准确性的青光眼诊断和严重程度分级工具。Objective To develop a deep learning model based on multimodal data for glaucoma diagnosis and severity assessment.Methods A diagnostic test was conducted.A total of 145 normal eyes from 86 participants and 507 eyes with primary open-angle glaucoma from 314 participants were collected at the First Affiliated Hospital of Kunming Medical University from June to December in 2023.Fundus photographs and visual field data were obtained,and glaucoma eyes were divided into three groups based on the mean deviation value of the visual field,namely mild group(154 eyes),moderate group(113 eyes),and severe group(240 eyes).Three convolutional neural network(CNN)models,including DenseNet 121,ResNet 50 and VGG 19,were used to build an artificial intelligence(AI)model.The impact of single-modal and multimodal data on the classification results was evaluated,and the most appropriate CNN network architecture for multimodal data was identified.This study adhered to the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of The First Affiliated Hospital of Kunming Medical University(No.2023L93).Written informed consent was obtained from each subject.Results A total of 652 eyes had both fundus photographs and visual field test results.Images were randomly assigned to training and test datasets in a 4∶1 ratio by using computer random number method.AI models built with different CNN models showed high accuracy,with DenseNet 121 outperforming ResNet 50 and VGG 19 on various effectiveness measures.In the single-modal algorithm using fundus photographs,single-modal algorithm using visual field tests,and multimodal algorithm combining fundus photographs and visual field data,the area under the curve for early glaucoma detection was 0.87,0.93 and 0.95,respectively.Conclusions The use of multimodal data enables the development of a highly accurate tool for the glaucoma diagnosis and severity grading.

关 键 词：青光眼 人工智能 多模态成像 彩色眼底照相 视野 

分 类 号：R775[医药卫生—眼科]