Molecular subtyping of breast cancer from routine Hematoxylin & Eosin-stained histology images using deep learning

Breast cancer is a heterogeneous malignancy comprising distinct molecular subtypes, each with varying therapeutic strategies and prognoses. Conventional molecular subtyping relies on immunohistochemistry and in situ hybridization studies, which may not be universally accessible due to financial and technical limitations. This study evaluates the utility of deep learning for identifying breast cancer molecular subtypes using histopathological images.

A dataset of 1000 H&E-stained breast cancer images (× 40 magnification) was collected and labeled into four molecular subtypes, and a modified EfficientNetB0 architecture was employed. The model was trained using the Adam optimizer with categorical cross-entropy loss over 25 epochs. Model performance was evaluated on a separate test set using accuracy, AUC, and classification metrics.

The model reached an overall classification accuracy of 0.81. Luminal A and Basal-like subgroups achieved the best diagnostic performance. The highest F1-score was observed for Luminal A (0.93), followed by Basal-like (0.86), Luminal B (0.76), and HER2-enriched (0.69). In an independent external validation cohort of 50 patients from a separate hospital, the model achieved an overall accuracy of 0.78.

Our results demonstrate that deep learning-based image analysis can effectively predict breast cancer molecular subtypes from routine histopathological images, offering a valuable adjunct when IHC and molecular tests are unavailable, inconclusive, or cost-prohibitive. Furthermore, this approach could serve as a quality control mechanism by flagging discordant cases where histological prediction and molecular tests are not compatible. The findings provide proof-of-concept evidence that H&E histopathology images contain information that can be used by deep learning to predict breast cancer molecular subtypes.