Yan Liang

Counterintuitively, the most *unreliable* samples in medical imaging datasets—those with fluctuating confidence and frequent forgetting during training—are the *most* informative for building accurate decision boundaries.