Multimodal Fusion of Imaging and Genomic Data for Enhanced Breast Cancer Detection
Open Access DepositedThis thesis presents a biologically interpretable multimodal deep learning framework for binary breast cancer detection, integrating imaging data—MRI and histopathology (tissue and diagnostic slides)—with genomic profiles including RNA-Seq, copy number variation, DNA methylation, miRNA, and proteomics. Designed to distinguish cancerous from non-cancerous cases, the system learns coherent patterns across radiological structures and molecular signatures. Prototype-based clustering is refined using Supervised Contrastive Learning (SCL), enhancing feature separation and boosting interpretability. Intra-modality structure is preserved through Graph Neural Networks (GNNs) for spatial tumor architecture and Graph Attention Networks (GATs) for gene-level regulatory interactions. These biologically structured embeddings are dynamically fused via a Cross-Attention Transformer enhanced with Flash Attention, enabling efficient integration across eight heterogeneous modalities. Rigorous validation on TCGA-BRCA datasets, combined with testing on synthetically mismatched biological samples, demonstrates the framework’s robustness in rejecting implausible inputs. Unlike earlier multimodal models that relied on shallow fusion without structural modeling or interpretability, this work introduces biologically grounded graph modeling, contrastive prototype refinement, and scalable Transformer-based integration. It establishes a robust, biologically faithful benchmark for multimodal fusion in accurate breast cancer detection.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.