Answers are grounded in this paper and cite their sources.
Reliable and consistent classification of whole slide images (WSIs) is critical for computer-aided histopathological diagnosis. Multiple instance learning (MIL) has become a standard paradigm for WSI classification but still suffers from lesion heterogeneity and visual variability within WSIs, resulting in unreliable instance selection and learning bias caused by confounders. In this paper, we propose a novel Instance Attribution framework with Mutual Information-guided Confounder-Related Surrogate Representations for Bias Adjustment (IAIC-MIL) in WSI classification. To improve the reliability of instance selection, we propose a Sobol-based instance attribution, which quantifies the contribution of each instance through Sobol sensitivity, thereby enabling interpretable and interactionaware selection of important instances. To reduce learning bias, we propose a bias adjustment module that incorporates learned confounder-related surrogate representations through approximate backdoor adjustment, thereby suppressing spurious correlations. Additionally, we propose a mutual information-guided disentanglement module to extract bias-related features, which are then clustered to obtain surrogate representations for backdoor adjustment. Experiments on four representative pathological datasets demonstrate that IAIC-MIL not only achieves state-of-theart performance but also provides reliable and consistent diagnostic insights.
No takes yet. Share an insight, caveat, or question.
Kong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: