Breast Cancer Recurrence Risk Prediction Based on Multiple Instance Learning

Predicting breast cancer recurrence risk is a critical clinical challenge. This study investigates the potential of computational pathology to stratify patients using deep learning on routine Hematoxylin and Eosin (H&E) stained whole-slide images (WSIs). We developed and compared three Multiple Instance Learning (MIL) frameworks – CLAM-SB, ABMIL, and ConvNeXt-MIL-XGBoost – on an in-house dataset of 210 patient cases. The models were trained to predict 5-year recurrence risk, categorized into three tiers (low, medium, high), with ground truth labels established by the 21-gene Recurrence Score. Features were extracted using the UNI and CONCH pre-trained models. In a 5-fold cross-validation, the modified CLAM-SB model demonstrated the strongest performance, achieving a mean Area Under the Curve (AUC) of 0.836 and a classification accuracy of 76.2%. Our findings demonstrate the feasibility of using deep learning on standard histology slides for automated, genomics-correlated risk stratification, highlighting a promising pathway toward rapid and cost-effective clinical decision support.

💡 Research Summary

This paper investigates whether routine H&E‑stained whole‑slide images (WSIs) can be used to predict five‑year breast‑cancer recurrence risk, defined in three tiers (low, medium, high) based on the clinically approved 21‑gene Oncotype DX Recurrence Score. The authors assembled an in‑house cohort of 210 patients from the China‑Japan Friendship Hospital, each with a digitized WSI and a corresponding risk label determined by board‑certified pathologists integrating the genomic score with other clinicopathologic factors. Because the dataset is modest, the study relies heavily on transfer learning: two state‑of‑the‑art foundation models—UNI (a ViT‑L/16 encoder pretrained with DINOv2 on 100 M patches) and CONCH (a vision‑language model trained on image‑caption pairs)—are used to extract high‑dimensional feature vectors from non‑overlapping 256 × 256‑pixel patches. The preprocessing pipeline converts proprietary .sdpc files to .svs, selects an optimal low‑magnification level, applies adaptive Gaussian blur, converts to HSV, thresholds the Saturation channel with a modified Otsu method, and performs morphological cleaning to generate tissue masks. Patch coordinates and masks are stored in HDF5 for efficient loading.

Three Multiple‑Instance Learning (MIL) frameworks are built on top of the extracted patch features: (1) CLAM‑SB, a modified version of the Clustering‑constrained Attention MIL (CLAM) model that adds slide‑by‑slide clustering constraints and uses a gated attention mechanism (two learnable weight matrices combined with sigmoid and tanh activations) to weight instances; (2) ABMIL, an attention‑based MIL that learns a single attention vector for each bag without clustering; and (3) ConvNeXt‑MIL‑XGBoost, which feeds ConvNeXt‑derived patch embeddings into an XGBoost meta‑classifier, thereby blending deep visual representations with a powerful gradient‑boosted decision tree.

All models are evaluated using five‑fold cross‑validation (168 training, 42 validation slides per fold). Performance metrics include macro‑averaged AUC, overall accuracy, and per‑class precision/recall. CLAM‑SB achieves the best results with a mean AUC of 0.836 and an accuracy of 76.2 %, outperforming ABMIL (AUC 0.782, accuracy 70.9 %) and ConvNeXt‑MIL‑XGBoost (AUC 0.801, accuracy 73.5 %). Attention heatmaps from CLAM‑SB highlight histopathologic regions that contribute most to the prediction, offering a degree of interpretability that could be valuable for pathologists.

The authors position their work against prior studies that either used large public WSI collections, multimodal omics data, or handcrafted nuclear features. By leveraging pre‑trained foundation models, they demonstrate that robust risk stratification is feasible even with a limited number of slides, reducing the data‑collection burden that has hampered many previous efforts. They also argue that the complementary nature of UNI (pure visual self‑supervision) and CONCH (vision‑language) captures both low‑level texture and higher‑level semantic cues, potentially explaining the strong performance despite the small cohort.

Limitations are acknowledged: the medium‑risk class contains only 21 cases, leading to class imbalance; external validation on independent cohorts is absent, so generalizability remains uncertain; and the study does not address regulatory or workflow integration issues required for clinical deployment. Future directions include expanding to multi‑institutional datasets, incorporating additional modalities (clinical variables, genomics), and applying explainable‑AI techniques to further bridge the gap between algorithmic predictions and pathologist reasoning.

In summary, this work provides a compelling proof‑of‑concept that pre‑trained pathology foundation models combined with attention‑based MIL can predict breast‑cancer recurrence risk from routine histology slides with clinically relevant accuracy. It offers a cost‑effective, scalable alternative to genomic assays for risk stratification, and lays groundwork for future AI‑assisted decision support tools in oncology.