Predicting molecular subtype in breast cancer using deep learning on mammography images

Abstract

Objectives: This study aimed to develop and evaluate a deep learning model for predicting molecular subtypes of breast cancer using conventional mammography images, offering a potential alternative to invasive diagnostic techniques. Methods: A retrospective analysis was conducted on 390 patients with pathologically confirmed invasive breast cancer who underwent preoperative mammography. The proposed DenseNet121-CBAM model, integrating Convolutional Block Attention Modules (CBAM) with DenseNet121, was trained and validated for binary (Luminal vs. non-Luminal, HER2-positive vs. HER2-negative, triple-negative vs. non-TN) and multiclass (Luminal A, Luminal B, HER2+/HR+, HER2+/HR−, TN) classification tasks. Performance metrics included AUC, accuracy, sensitivity, specificity, and interpretability via Grad-CAM heatmaps. Results: The model achieved AUCs of 0.759 (Luminal vs. non-Luminal), 0.658 (HER2 status), and 0.668 (TN vs. non-TN) in the independent test set. For multiclass classification, the AUC was 0.649, with superior performance in distinguishing HER2+/HR− (AUC = 0.78) and triple-negative (AUC = 0.72) subtypes. Attention heatmaps highlighted peritumoral regions as critical discriminative features. Conclusion: The DenseNet121-CBAM model demonstrates promising capability in predicting breast cancer molecular subtypes from mammography, offering a non-invasive alternative to biopsy.