🤖 AI Summary
Weak causal interpretability and high attribution noise due to the absence of pixel-level annotations hinder diagnostic reliability in digital pathology whole-slide image (WSI) analysis. To address this, we propose CI-MIL—a novel framework that pioneers the integration of causal inference into the multi-instance learning (MIL) paradigm. CI-MIL achieves out-of-distribution generalization and confounder correction via decorrelated reweighting and cross-correlation minimization in a random Fourier feature space, all without requiring pixel-level supervision. This significantly enhances the determinacy and clinical interpretability of evidence subregion attribution. Evaluated on Camelyon16 and TCGA-NSCLC, CI-MIL achieves 92.25% accuracy / 95.28% AUC and 94.29% accuracy / 98.07% AUC, respectively. Attribution heatmaps exhibit strong spatial alignment with ground-truth annotations, empirically validating its robust causal modeling capability.
📝 Abstract
Due to the lack of fine-grained annotation guidance, current Multiple Instance Learning (MIL) struggles to establish a robust causal relationship between Whole Slide Image (WSI) diagnosis and evidence sub-images, just like fully supervised learning. So many noisy images can undermine the network's prediction. The proposed Causal Inference Multiple Instance Learning (CI-MIL), uses out-of-distribution generalization to reduce the recognition confusion of sub-images by MIL network, without requiring pixelwise annotations. Specifically, feature distillation is introduced to roughly identify the feature representation of lesion patches. Then, in the random Fourier feature space, these features are re-weighted to minimize the cross-correlation, effectively correcting the feature distribution deviation. These processes reduce the uncertainty when tracing the prediction results back to patches. Predicted diagnoses are more direct and reliable because the causal relationship between them and diagnostic evidence images is more clearly recognized by the network. Experimental results demonstrate that CI-MIL outperforms state-of-the-art methods, achieving 92.25% accuracy and 95.28% AUC on the Camelyon16 dataset (breast cancer), while 94.29% accuracy and 98.07% AUC on the TCGA-NSCLC dataset (non-small cell lung cancer). Additionally, CI-MIL exhibits superior interpretability, as its selected regions demonstrate high consistency with ground truth annotations, promising more reliable diagnostic assistance for pathologists.