Efficient Multi-Temporal 3D Learning for Breast DCE-MRI Classification with Temporal-as-Channel Encoding and Early...

Accurate breast cancer classification in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) requires effective integration of volumetric morphology and temporal contrast-enhancement kinetics. In this study, we evaluate a multi-temporal 3D framework for three-class breast classification (normal, benign, malignant) using the multi-center ODELIA dataset. The framework uses a Temporal-as-Channel (TaC) input representation that maps three DCE subtraction phases onto the native three-channel input of a pretrained R3D-18 backbone, enabling early cross-phase interaction without a dedicated temporal module. A 3D Spatial Attention Module (SAM), adapted from the spatial branch of CBAM, is inserted at the first residual stage (SAM-L1) to refine early volumetric features. Experiments identify second-phase subtraction (Sub2) as the best-performing single-volume input among those evaluated and show higher observed performance for the 3D full-volume configuration than for the 2D MIP configuration. TaC yields a larger gain in the 3D setting than in the 2D setting, while the attention ablations identify SAM-L1 as the best-performing attention configuration among those tested. The TaC+SAM-L1 configuration achieves an internal-test Average Score of 81.46%, with a micro-AUC of 91.50%, sensitivity at 90% specificity of 76.74%, and specificity at 90% sensitivity of 76.16%.