Breast cancer recurrence risk prediction based on MIL

Area of Science:

• Digital pathology
• Machine learning
• Oncology

Background:

• Accurate prediction of breast cancer recurrence risk is crucial for effective patient management.
• Genomic assays like the 21-gene recurrence score are standard but can be costly and time-consuming.
• Computational pathology offers a potential alternative for risk stratification using routine histology.

Purpose of the Study:

• To evaluate the efficacy of deep learning models in predicting 5-year breast cancer recurrence risk.
• To compare the performance of different multiple instance learning (MIL) frameworks on whole-slide images (WSIs).
• To assess the feasibility of automated, genomics-correlated risk stratification using standard hematoxylin and eosin (H&E) stained slides.

Main Methods:

• Development and comparison of three MIL frameworks: CLAM-SB, ABMIL, and ConvNeXt-MIL-XGBoost.
• Training models on an in-house dataset of 210 patient cases with 5-year recurrence risk labels derived from the 21-gene recurrence score.
• Feature extraction using UNI and CONCH pre-trained models and evaluation via 5-fold cross-validation.

Main Results:

• The modified CLAM-SB model demonstrated superior performance.
• Achieved a mean area under the curve (AUC) of 0.836.
• Attained a classification accuracy of 76.2% in predicting recurrence risk tiers (low, medium, high).

Conclusions:

• Deep learning applied to standard H&E stained histology slides can effectively stratify breast cancer patients by recurrence risk.
• This computational pathology approach shows promise for rapid, cost-effective clinical decision support.
• Automated risk stratification correlates with genomic data, offering a valuable tool for personalized oncology.