Reproduce “DGMIL: Distribution Guided Multiple Instance Learning for Whole Slide Image Classification” on OpenResearch

About this paper

Linhao Qu, Xiaoyuan Luo, Shaolei Liu, Manning Wang, Zhijian Song

Researchers from Fudan University developed DGMIL, a Multiple Instance Learning framework that explicitly models and refines feature distributions for Whole Slide Image classification and positive patch localization. It achieved a Slide AUC of 0.8368 and Patch AUC of 0.9045 on CAMELYON16, and a Slide AUC of 0.9702 on the TCGA Lung Cancer dataset, outperforming previous methods.

Key insights

• Explicitly modeling the underlying feature distribution of histopathological data provides an effective guide for instance classification, moving beyond purely discriminative approaches.
• An iterative refinement strategy, guided by pseudo-labeled "extreme instances," progressively enhances the separability of positive and negative features in the latent space.
• Self-supervised pre-training with a Masked Autoencoder yields robust initial feature representations, and clustering negative instances helps to better quantify the "positive-likeness" of any given patch.

Summary by alphaXiv.