Impact of Volumetric Representation on Deep Learning-Based Molecular Subtype Prediction from 3D Digital Breast Tomosynthesis

Background/Objectives: Accurate molecular subtype characterization is essential for breast cancer management, but currently relies on invasive biopsy procedures with limited ability to capture tumor heterogeneity and the surrounding microenvironment. In this study, we investigated the feasibility of predicting breast cancer molecular subtypes directly from three-dimensional (3D) digital breast tomosynthesis (DBT) volumes using deep learning, while also evaluating the influence of different volumetric standardization strategies on classification performance. Methods: Using the multicenter OPTIMAM database, a 3D ResNet-18 was trained to perform five-class molecular subtype classification (Luminal A, Luminal B1, Luminal B2, HER2-enriched, and triple-negative breast cancer (TNBC)) as well as one-vs-rest binary classification. To assess the impact of volumetric representation, different standardization strategies, including interpolation-based resampling and zero-padding approaches, were evaluated. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) was used to investigate the image regions contributing to model predictions. Results: Classification performance varied according to the volumetric standardization strategy. Among the evaluated approaches, central padding provided the most consistent performance across repeated holdout evaluations, achieving a mean multiclass AUC of 64.07%, with TNBC reaching 70.39%. In the binary setting, HER2-enriched tumors achieved the highest AUC (76.02%). Furthermore, Grad-CAM analysis demonstrated that model activations frequently extended beyond the lesion itself and involved surrounding tissue regions. Conclusions: Overall, these findings demonstrate that volumetric DBT combined with 3D deep learning enables non-invasive prediction of breast cancer molecular subtypes. The results further indicate that volumetric standardization influences classification performance and suggest that both intratumoral and peritumoral imaging information contribute to subtype discrimination, although generalization to previously unseen patients remains unestablished.