Breast Cancer Response Prediction — AMD ROCm

We present a hierarchical multimodal deep learning pipeline for predicting pathologic complete response (pCR) to neoadjuvant chemotherapy in breast cancer, validated on 199 patients from the I-SPY 2 clinical trial. The system integrates three complementary architectures: a frozen RadImageNet DenseNet-121 CNN extracts domain-specific DCE-MRI features across 6 contrast phases; a GATv2 Graph Neural Network encodes clinical molecular data (HR, HER2, TNBC status) using biologically-informed edges; and a hierarchical BiLSTM models the temporal treatment trajectory across 4 timepoints (pre- to post-chemotherapy). GroupWise normalization resolves feature scale disparity across CNN, radiomics, and spatial feature groups, enabling stable gradient flow and reducing seed-to-seed AUC variance from ±0.025 to ±0.011. Molecular subgroup analysis reveals performance consistent with clinical biology: TNBC achieves the highest AUC (0.723) due to pronounced contrast enhancement patterns, while HR+/HER2- presents challenges from severe class imbalance (18% pCR rate, only 14 positive cases). The final model achieves AUC 0.771 ± 0.011 and 81% sensitivity, identifying pCR patients who may avoid unnecessary surgery. The entire pipeline runs on AMD ROCm with optimizations for the RX 9070 XT (gfx1201), demonstrating clinical-grade medical AI on AMD hardware.