MobileNetV3用于移动端糖网严重程度识别的准确性评价OA

Objective To evaluate the accuracy of the lightweight deep learning model MobileNetV3 for mobile device-based diabetic retinopathy(DR)severity identification and provide technical foundation for mobile-based DR identification applications.Methods Based on the publicly available DDR(Dataset for Diabetic Retinopathy,DDR)dataset,we employed MobileNetV3-Large pre-trained models to construct DR severity identification models through transfer learning.Four identification models were developed to address DR severity requirements:presence/absence of DR,referable DR,vision-threatening DR,and proliferative DR.The training process incorporated Focal Loss to handle data imbalance issues and utilized data augmentation techniques to enhance model generalization performance.Model performance was assessed on the test set using metrics including the area under the ROC(receiver operating characteristic)curve(area under the curve,AUC),accuracy,sensitivity,and specificity.Results The MobileNetV3 models demonstrated good performance across all four DR severity identification tasks,with AUC values all above 90%.The presence/absence of DR identification model achieved an AUC of 92.8%(92.0%,93.6%),accuracy of 85.1%(83.9%,86.2%),sensitivity of 81.3%(79.4%,83.0%),and specificity of 88.9%(87.4%,90.3%).The referable DR identification model achieved an AUC of 91.0%(90.0%,91.9%),accuracy of 84.1%(82.9%,85.2%),sensitivity of 79.5%(77.5%,81.4%),and specificity of 87.8%(86.3%,89.2%).The vision-threatening DR identification model achieved an AUC of 96.5%(95.5%,97.4%),accuracy of 89.8%(88.8%,90.8%),sensitivity of 91.6%(88.2%,94.3%),and specificity of 89.6%(88.6%,90.6%).The proliferative DR identification model achieved an AUC of 98.3%(97.7%,98.8%),accuracy of 93.9%(93.1%,94.6%),sensitivity of 93.5%(89.9%,96.1%),and specificity of 93.9%(93.1%,94.7%).Conclusions MobileNetV3,as a lightweight deep learning model,demonstrates good accuracy for DR severity identification,providing a technical foundation for developing portable mobile devices based on MobileNetV3 models to achieve rapid and flexible DR identification.