Reproduce “Pseudo-Bag Mixup Augmentation for Multiple Instance Learning-Based Whole Slide Image Classification” on OpenResearch

About this paper

Pei Liu, Luping Ji, Xinyu Zhang, Feng Ye

Researchers from the University of Electronic Science and Technology of China and West China Hospital developed Pseudo-bag Mixup (PseMix), an augmentation method that adapts Mixup to Whole Slide Image (WSI) classification by creating semantically consistent "pseudo-bags" and employing a random mixing strategy. This approach consistently improves classification accuracy by 0.93-1.75% and AUC by 0.93-1.30% across various MIL models and datasets, while also enhancing robustness against patch occlusion and label noise.

Key insights

• Clustering WSI instances into phenotype-based "pseudo-bags" allows for effective size and semantic alignment, enabling Mixup application to irregularly structured WSI data.
• The random mixing strategy, which generates both pseudo-bag-mixed and pseudo-bag-masked samples, is critical for enhancing data diversity and providing "vicinity samples" that smooth model learning.
• Decoupling the Mixup strategy from specific MIL model predictions and avoiding significant instance dropping enables a plug-and-play solution that preserves critical information for improved performance.

Summary by alphaXiv.