Fig. 3.

Training-validation loss curves of different models along training epochs. During the training, we observed the cross-entropy loss to measure the model training effect. Theoretically, a good model shall keep the lowest loss and should have the smallest difference between the training-validation losses without significant oscillation. (a) The highly overfitted ResNet model trained with original SDOCT volumes. (b) The highly overfitted ResNet model trained with denoised SDOCT volumes. (c) The stable and well-fitted SE-ResNet model trained with original SDOCT volumes. (d) The stable and well-fitted SE-ResNet model trained with denoised SDOCT volumes. (e) The stable and better-fitted SE-ResNeXt model with denoised SDOCT volumes. (f) The stable and best-fitted SE-ResNeXt model with denoised SDOCT volumes.