Instance causal multiple instance learning for whole slide image classification

Whole slide images (WSIs) are characterized by extremely high resolution and usually lack fine-grained annotations, making multiple instance learning (MIL) the dominant paradigm for weakly supervised WSI classification. Although existing MIL methods have achieved substantial progress in feature extraction and instance aggregation, most of them overlook the influence of confounding factors, thereby learning spurious associations that impair generalization performance. To address this issue, we develop an Instance Causal Multiple Instance Learning (ICMIL) framework for WSI classification, which aims to reduce the model’s reliance on potentially spurious instance-level associations. Specifically, unlike conventional likelihood-based strategies, the proposed method introduces a backdoor-adjustment-inspired prototype module during instance aggregation, where clustered prototypes are used as proxy contextual patterns to mitigate the influence of potential confounding cues. We conduct extensive experiments on two publicly available WSI datasets, Camelyon16 and TCGA-NSCLC. The results demonstrate that the proposed ICMIL framework achieves competitive performance and generally improves existing baselines, while also exhibiting a plug-and-play property.