Three-Phase Scribble-Adaptive Curriculum Learning for autoPETV Grand Challenge
Libo Zhang developed a three-phase curriculum for training a residual-encoder U-Net to segment lesions in wholebody PET/CT scans. The method involves learning automatically, then exposing the model to various forms of ground-truth guidance (scribbles), and finally adapting by simulating error corrections. This online, error-driven process significantly improves segmentation accuracy.