Fig. 3: Label cleaning strategy guided by dataset quality scores.
From: Efficiency and safety of automated label cleaning on multimodal retinal images
a, b showed the strong linear correlation between dataset quality score (DQS) and label accuracy. c, d showed receiver operating characteristic curves for predicting label accuracy with DQS. The cutoff values of DQS for label accuracy >0.98 were determined to be 0.9965 and 0.9925. e, f The correlation between noise rate and cleaning iterations using DQS-guided strategy.
