DeCo-MIL debiases long-tailed whole slide analysis with counterfactual reasoning
Rare cancer subtype detection gets a boost via counterfactual intervention on tissue patches.
Whole slide image (WSI) analysis in computational pathology has long been dominated by multiple instance learning (MIL), which treats each slide as a bag of patches. But real-world datasets are painfully imbalanced: rare cancer subtypes have few training slides (inter-slide class long tail), and even those slides contain only a tiny fraction of diagnostically relevant patches buried in massive background noise (intra-slide evidence long tail). This nested dual long-tail cripples standard MIL, biasing models toward common classes and causing rare-class recognition to fail.
DeCo-MIL attacks both tails simultaneously. Internally, it clusters patches into tissue-morphology anchors, then applies a counterfactual intervention by replacing each anchor with its matched normal prototype to estimate class-frequency-corrected contributions. These contributions drive redundancy masking to preserve scarce discriminative instances. Externally, it builds anchor-stratified pseudo-bags from the reduced bags, combining tail-aware oversampling with consistency regularization to boost supervision for rare classes without distorting tissue composition. Benchmarked on three long-tailed WSI datasets, DeCo-MIL outperforms existing methods in both rare-class sensitivity and overall accuracy, marking a significant step toward clinically actionable AI pathology.
- Addresses the nested dual long-tail: class imbalance plus scarce diagnostic patches within rare-class slides
- Counterfactual intervention replaces tissue-morphology anchors with normal prototypes to estimate true class contributions
- Achieves state-of-the-art performance on three long-tailed WSI benchmarks in tail-class and overall classification
Why It Matters
Rare cancer diagnoses could become more reliable in automated pathology, reducing missed cases in imbalanced clinical datasets.