AI and Tomosynthesis for Breast Cancer Molecular Subtyping: A step toward precision medicine

Breast cancer molecular subtyping is crucial for prognosis and treatment selection. Currently, it relies on invasive biopsies that provide only a limited representation of tumor heterogeneity. Digital breast tomosynthesis (BT), with its ability to capture volumetric breast tissue, offers a promising non-invasive alternative when combined with artificial intelligence (AI). This study investigates the potential of deep convolutional neural networks (CNNs) to classify aggressive breast cancer subtypes - Luminal B2, HER2-positive (HER2+) and triple-negative (TN) - using BT images.ResNet-101 and Inception-v3 architectures were trained on BT-derived tumor regions to distinguish each aggressive subtype from the remaining classes. The models achieved an AUC of 73.17% and 71.96% for HER2+, 65.22% and 62.07% for TN and 59.28% and 59.26% for Luminal B2, respectively, demonstrating the feasibility of AI-driven molecular subtyping from BT imaging.These findings highlight the potential of BT-based deep learning models to aid in the non-invasive classification of aggressive breast cancer subtypes, particularly HER2+, where the highest performance was achieved. While challenges remain in distinguishing Luminal B2, the use of BT imaging and AI demonstrates promise for refining molecular subtype assessment and improving clinical decision-making.Clinical Relevance-This study demonstrates that AI-driven analysis of digital breast tomosynthesis can capture tumor and microenvironment characteristics, providing valuable insights into disease staging and progression. By leveraging comprehensive imaging data rather than limited biopsy samples, this approach has the potential to improve non-invasive molecular subtyping, aiding personalized treatment decisions while reducing reliance on invasive procedures.