DeCo-MIL: Counterfactual Reasoning for Long-Tailed WSI Analysis

A new machine learning method, DeCo-MIL, addresses the challenge of long-tailed distributions in whole slide image (WSI) analysis, a technique widely used in computational pathology. The method, detailed in a preprint on arXiv (2608.14719), tackles a 'nested dual long-tail' problem: an inter-slide class imbalance where rare diseases have few training slides, and an intra-slide imbalance where diagnostic evidence is sparse within each slide. This coupling biases models toward common classes and degrades recognition of rare conditions. DeCo-MIL employs frequency-debiased counterfactual reasoning to mitigate both tails. For the inner tail, it clusters patches into tissue-morphology anchors and replaces each with a matched normal prototype, performing a counterfactual intervention to estimate causal effects. The approach aims to improve rare-class recognition in weakly supervised WSI analysis, with potential implications for diagnostic accuracy in pathology. The preprint was announced on arXiv with a cross-type classification, indicating its relevance across domains. The method is part of ongoing research to make AI systems more robust to real-world data imbalances.