Reliable Breast Cancer Diagnostic System Using Uncertainty-Guided Multiple Instance Learning for Weakly Localized Histopathology Classification

Abstract

Breast cancer diagnosis from histopathology images requires reliable and interpretable automated analysis. However, existing weakly supervised and multiple instance learning (MIL) approaches often overlook predictive uncertainty and tend to produce overconfident decisions in ambiguous cases, thereby limiting the reliability of automated diagnosis. To address this research gap, this article presents an uncertainty-aware MIL (UMIL) framework for breast cancer classification, localization, and interpretability in histopathology images. The proposed UMIL integrates a vision transformer (ViT) and lightweight Swin transformer encoder to provide coarse localization cues and feature extraction. To enhance the efficacy of this integration, thresholding and morphological filtering are applied to highlight discriminative tumor regions. Moreover, the region-level features are aggregated through an UMIL pooling strategy to create more confident regions. Besides this, an uncertainty-guided regularization term is incorporated into the training objective to explicitly couple patch-level uncertainty with bag-level classification. To enhance decision reliability, an inconclusive prediction mechanism is introduced to avoid making forced decisions when prediction confidence falls below a predefined threshold. For edge and point-of-care applications, the proposed framework has been deployed on an NVIDIA Jetson Nano developer kit. The effectiveness of the proposed framework has been evaluated on the openly accessible BreaKHis dataset in terms of qualitative and quantitative analysis. The quantitative evaluation shows that the proposed framework has an overall classification accuracy of 89.11 %, with 11.99 % of cases labeled as inconclusive and 97.11 % accurate for conclusive predictions. Besides that, the ablation and threshold analysis validate the accuracy–tradeoff corresponding to existing methods. The analysis reveals that the proposed framework classifies, localizes, and detects breast cancer simultaneously.