Multi-Modal Cross Attention Network for Predicting Pathological Complete Response in Breast Cancer MRI

Predicting pathological complete response (pCR) to chemotherapy in breast cancer patients is a critical clinical endeavor, as it provides insights into individual patient responses and potential treatment outcomes. Despite the wealth of information available from MRI techniques such as DCE-MRI, pMRI, and SER, gleaning actionable insights from such complex data remains a challenging feat due to inherent variations in tumor biology, size, location, and blood flow patterns. Addressing this challenge, our study introduces the Multi-modal Cross Attention Network (MMCA), a novel deep learning approach designed to synergistically harness and emphasize the complementary information present across different MRI modalities. MMCA leverages an attention mechanism to adeptly weigh and integrate critical features, ensuring a more comprehensive and accurate pCR prediction. Comparative experiments against conventional deep learning architectures, including ResNet and DenseNet, showcased the prowess of MMCA. Specifically, through a 5-fold cross-validation, MMCA achieved an accuracy of 0.797 and an AUC of 0.758, outperforming other methods. This research not only demonstrates the potential of advanced deep learning algorithms combined with multimodal fusion techniques, but also sets a new benchmark for non-invasive pCR prediction using complex MRI data.