Application of Computer Vision Models for Diabetic Retinopathy Referral
WANG Huixia1a, ZHANG Yuting1b, WANG Ying2, ZHU Manhui1b
1. a. Department of Ophthalmology; b. Department of Pathology, Lixiang Eye Hospital of Soochow University, Suzhou Jiangsu
215000, China; 2. Department of Ophthalmology, Suzhou Municipal Hospital, Suzhou Jiangsu 215000, China
Abstract：Objective To collect the eyeground color images of diabetic patients, and to establish a computer vision model, to explore
its application in the referral of diabetic retinopathy (DR). Methods The eyeground images of diabetic patients in Suzhou area were
collected. The training set (n=350) and the internal validation set (n=100) were from Lixiang Eye Hospital of Soochow University,
and the test set (n=200) was from Suzhou Municipal Hospital. The eyeground photos were classified into referrable DR group and
the non-referrable DR group. The performance of models was evaluated based on test set and compared with two ophthalmologists.
In addition, the Grad-CAM technology was used to visually interpret the computer vision model. Results The performance ofEfficientNet model was the best among all models and superior to specificity for ophthalmologists with low years of experience
[accuracy (ACC) was 0.840, sensitivity (SE) was 0.793, and specificity (SP) was 0.859]. However, it was slightly lower than that of
senior physicians (ACC was 0.925, SE was 0.862, SP was 0.951). With the assistance of the EfficientNet model, the ACC rose to 0.915,
SE to 0.828, and SP to 0.951. Similarly, the ACC of senior physicians reached 0.965, SE was 0.931, and SP was 0.979.
Conclusion Computer vision models based on deep convolutional neural networks show practicable performance in DR referral
binary classification, better than junior ophthalmologist. The model can assist ophthalmologists in diagnosis and improve ACC.
王慧霞1a，张玉婷1b，王莹2，朱曼辉1b. 计算机视觉模型在糖尿病视网膜
病变转诊中的应用[J]. 中国医疗设备, 2023, 38(11): 1-5.
WANG Huixia1a, ZHANG Yuting1b, WANG Ying2, ZHU Manhui1b. Application of Computer Vision Models for Diabetic Retinopathy Referral. China Medical Devices, 2023, 38(11): 1-5.
