Deep Learning for Breast Cancer Subtype and Treatment Response Classification from Dynamic Contrast-Enhanced MRI : An Experimental Study on Temporal Encodings, Auxiliary Inputs, and Model Design for TNBC and pCR Prediction in Breast DCE-MRI

Sammanfattning

Triple-negative breast cancer (TNBC) is an aggressive subtype of breast cancer that lacks targeted treatment options. Pathological complete response (pCR) following neoadjuvant therapy is a crucial prognostic indicator; however, both TNBC subtype classification and pCR prediction remain challenging using conventional methods. This thesis investigates whether deep learning models trained on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) can accurately classify TNBC and predict pCR. A flexible 3D convolutional neural network pipeline was developed and applied to the MAMA-MIA dataset, a large multicenter collection of expert-annotated breast DCE-MRI scans. The study systematically evaluated three design factors: temporal encoding strategies (e.g., delta-based subtraction), auxiliary input channels (e.g., segmentation masks and parametric enhancement maps), and model capacity (ResNet18, ResNet50, CBAM). All models were trained and evaluated using five-fold cross-validation. Results showed that temporal encodings improved classification performance over static input, and that adding parametric enhancement maps yielded further gains. However, segmentation masks did not yield improvement when used as additional input channels. Increasing model capacity and incorporating attention mechanisms (CBAM) achieved the highest performance, with an AUC of 0.81 for TNBC and 0.76 for pCR. These findings suggest that deep learning models can extract clinically relevant information from spatiotemporal DCE-MRI patterns, with implications for non-invasive tumor characterization and early treatment response prediction.