Computational Pathology for Accurate Prediction of Breast Cancer Recurrence: Development and Validation of a Deep...

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate breast cancer recurrence risk stratification is vital for personalized treatment.
• Current genomic assays are limited by cost and accessibility, especially for underserved populations.
• Hematoxylin and eosin (H&E) stained whole slide images (WSIs) are routinely available in pathology.

Purpose of the Study:

• To develop and validate a deep learning computational pathology approach, Deep-Breast-Cancer-Recurrence (BCR)-Auto, for predicting breast cancer recurrence risk.
• To assess the performance of Deep-BCR-Auto on independent datasets.
• To evaluate the potential of computational pathology as a cost-effective tool for breast cancer prognosis.

Main Methods:

• A deep learning model, Deep-BCR-Auto, was trained to predict recurrence risk from H&E-stained WSIs.
• The model was validated on two independent cohorts: The Cancer Genome Atlas (TCGA) Program breast cancer dataset and an in-house dataset from The Ohio State University.
• Performance was evaluated using metrics including area under the receiver operating characteristic curve (AUC), accuracy, specificity, and sensitivity.

Main Results:

• Deep-BCR-Auto demonstrated robust performance in stratifying patients into low- and high-recurrence risk categories.
• On the TCGA dataset, the model achieved an AUC of 0.827, outperforming existing weakly supervised models (P = .041).
• On the Ohio State University dataset, Deep-BCR-Auto achieved an AUC of 0.832, with 82.0% accuracy, 85.0% specificity, and 67.7% sensitivity, indicating strong generalizability.

Conclusions:

• Deep-BCR-Auto shows significant potential as a cost-effective computational pathology tool for breast cancer recurrence risk assessment.
• This approach can broaden access to personalized treatment strategies by leveraging routine H&E-stained WSIs.
• Integrating deep learning-based computational pathology into routine practice can enhance breast cancer prognosis across diverse clinical settings.