Reproduce “Bi-directional Weakly Supervised Knowledge Distillation for Whole Slide Image Classification” on OpenResearch

About this paper

Linhao Qu, Xiaoyuan Luo, Manning Wang, Zhijian Song

A bi-directional weakly supervised knowledge distillation framework, WENO, improves Whole Slide Image classification by mutually enhancing bag and instance classifiers using only slide-level labels. This approach achieves an instance classification AUC of 0.9377 on Camelyon16 and surpasses fully supervised methods for bag classification on the same dataset with an AUC of 0.8663.

Key insights

• Mutual reinforcement between bag-level and instance-level classifiers through bi-directional knowledge distillation can effectively overcome the limitations of weakly supervised Multiple Instance Learning.
• Leveraging normalized attention scores as 'soft pseudo-labels' provides a more reliable and nuanced signal for training instance classifiers compared to traditional hard pseudo-labeling.
• An explicit strategy to re-weight or remove 'easy' positive instances during training forces the bag classifier to generalize better by learning from 'harder' or more subtle pathological features.

Summary by alphaXiv.