BPMambaMIL: A bio-inspired prototype-guided multiple instance learning for oncotype DX risk assessment in histopathology

Breast cancer remains one of the most prevalent malignancies among women, with hormone receptor-positive (HR+)/human epidermal growth factor receptor 2-negative (HER2–) breast cancers constituting a majority, with treatment decisions often guided by genomic assays such as the 21-gene recurrence score assay, Oncotype DX. Although Oncotype DX provides critical prognostic and predictive insights, its high cost and limited accessibility create substantial barriers, especially for patients with constrained financial resources. To reduce the test cost, we aim to leverage H&E-stained whole slide images (WSIs) to predict Oncotype DX risk. Since WSIs are extremely large and contain redundant information, directly processing them is both computationally expensive and prone to errors. To address these limitations, we introduce a bio-inspired prototype-guided model (BPMambaMIL), a novel weakly supervised learning framework that integrates the Mamba mechanism with prototypical guidance to predict Oncotype DX score intervals directly from pathology images. Our model was evaluated on an in-house dataset with clinical Oncotype DX scores, where it achieved an AUC of 0.839, a 5.61 % improvement over the baseline model (MambaMIL), and demonstrated robust predictive performance, particularly in identifying high-risk score ranges (accuracy: 0.714 vs 0.419). Further assessments on two public breast cancer pathology image datasets using six state-of-the-art models underscored BPMambaMIL's generalizability on research-based ODX scores and binary tumor classification tasks. By evaluating various clinical scenarios, the proposed method not only enhances the accuracy of breast cancer recurrence risk predictions but also offers a cost-effective alternative to genomic assays, thus improving clinical outcomes.