Fig. 1. Schematic diagram of the overall label cleaning process.

a Data preparation: A total of 2263 CFP and 1316 OCT images were collected and annotated b pseudo-label strategy: Label noise was deliberately injected by randomly select 5% from one category and evenly distributed to other categories progressively to create 45 subsets. c Data cleaning: Cleanlab was applied to detect and correct label errors repeatedly. d Model comparison: The RETFound foundation model was fine-tuned on datasets before and after label cleaning and tested on a holdout testing set.