Fig. 4.

The AUC of different models in the primary validation dataset and two external validation datasets. (a) AUCs of the ResNet model trained with original SDOCT volumes. (b) AUCs of the ResNet model trained with denoised SDOCT volumes. (c) AUCs of the SE-RexNet model trained with original SDOCT volumes. (d) AUCs of the SE-ResNet model trained with denoised SDOCT volumes. (e) AUCs of the SE-ResNeXt model trained with original SDOCT volumes. (f) AUCs of the SE-ResNeXt model trained with denoised SDOCT volumes. In general, the AUCs of the SE-ResNeXt model outperformed both ResNet and SE-ResNet models. Particularly, the SE-ResNeXt model trained with denoised data achieved the highest AUCs in both primary and external validations and an overall better diagnostic performance with regard to other metrics, such as sensitivity, specificity, and accuracy.