Study (first author, year) | Dataset | AI architecture | Summary |
---|---|---|---|
Son, 2020 [11] | Local clinic data (ERM, 3639 eyes) + external database | Modified ResNet + classic augmentation | The deep learning model for retinal membrane feature detection showed ROC-AUCs of 0.989 and 0.997 in two validation sets |
Casado-García, 2021 [31] | A nationwide database (Spain) + RFMiD | HR-net + CycleGAN | The final model achieved a F1-score of 86.82% to detect ERM |
Shao, 2021 [10] | Local clinic data (ERM, 83 eyes / no ERM, 61 eyes) | Inception-Resnet-v2 and Xception + classic augmentation | The AI model achieved an accuracy of 77.1%. It was comparable to manual reading (accuracy, 75.7%) |
Kim, 2021 [36] | Local clinic data (ERM, 99 eyes / control, 79 eyes) | ResNet50 + classic augmentation | The deep learning model for ERM detection showed a sensitivity of 92.5% and specificity of 98.3% |
Cen, 2021 [12] | JSIEC + LEDRS + EYEPACS | Mask R-CNN + Inception-V3, Xception, InceptionResNet-V2, and modified ResNet and ResNeXt | The final model for ERM detection showed ROC-AUCs of 0.9972 and 0.9976 in two validation sets |
Li, 2022 [13] | Local clinic data (ERM, 2947 eyes) | SeResNext50 + classic augmentation | The deep learning model for ERM detection showed ROC-AUCs of 0.968 in the internal validation and 0.938 and 0.934 in the external validation |
Son, 2023 [35] | Local clinic data (ERM, 3073 eyes) + MESSIDOR | EfficientNet-B7 + classic augmentation | The deep learning model for membrane feature detection showed ROC-AUCs of 0.997 in the internal validation and 0.954 in the external validation |
Ours | Local clinic data (ERM, 302 eyes / control, 1250 eyes) + RFMiD + JSIEC | EfficientNetB0 + StyleGAN2 | The proposed model achieved ROC-AUC of 0.926 for internal validation. ROC-AUCs of 0.951 and 0.914 were obtained for the two external validation datasets |