MICIL: Multiple-Instance Class-Incremental Learning for skin cancer whole slide images

Pablo Meseguer, Rocío del Amor, Valery Naranjo

Abstract: Highlights•An incremental learning algorithm is applied under the MIL paradigm to analyze WSI.•A public dataset featuring patch-level embeddings of multiple spindle-cell neoplasms.•A WSI embedding-level distillation to prevent shift in the aggregated latent space.•Comprehensive experiments address the plasticity-stability dilemma in MIL frameworks.