计算机视觉模型在糖尿病视网膜病变转诊中的应用

Application of Computer Vision Models for Diabetic Retinopathy Referral

目的收集糖尿病患者的眼底彩色照片,建立基于深度卷积神经网络的计算机视觉模型,探讨其在糖尿病视网膜病变(Diabetic Retinopathy,DR)转诊中的应用。方法收集苏州地区糖尿病患者眼底照片,其中训练集(n=350)和内部验证集(n=100)照片来自苏州大学理想眼科医院,测试集来自苏州市立医院(n=200)。对眼底照片进行分类,分为DR无需转诊组和需转诊组。基于测试集表现评价模型分类能力,并与两名眼科医生进行比较。此外,采用Grad-CAM技术对计算机视觉模型进行可视化解释。结果EfficientNet模型在各模型中表现最佳,且优于低年资眼科医生[准确度(Accuracy,ACC)为0.840,敏感度(Sensitivity,SE)为0.793,特异性(Specificity,SP)为0.859],但略低于高年资医生(ACC为0.925,SE为0.862,SP为0.951)。在EfficientNet模型的辅助下,低年资医生的ACC升至0.915,SE为0.828,SP为0.951。同样,高年资医生的ACC达到0.965,SE为0.931,SP为0.979。结论基于深度卷积神经网络的计算机视觉模型在DR转诊二分类任务中表现较好,优于低年资眼科医生。该模型可辅助眼科医生诊断,提高准确度。

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 of EfficientNet 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.