Fig. 4. Classification accuracy of RETFound finetuning before and after cleaning.

The original datasets with noisy labels and the ones after 6-iterations of label cleaning were used to fine-tune the RETFound model and subsequently test on the hold-out testing sets, respectively. The classification accuracies in both CFP (a) and OCT (b) demonstrated notable enhancements, with improvements becoming more evident as the noise rates increased. Except in the clean datasets (0% noise rate), the classification accuracy of the RETFound improved using the dataset after cleaning compared to that before cleaning.