Pathology-Genomic Cross Attention network for Cancer Survival Prediction from Whole Slide Image

With the rapid development of digital imaging and high-throughput gene sequencing technologies, multimodal data integration offers unprecedented possibilities for advancing cancer survival analysis. However, existing deep learning methods often face limitations in effectively capturing complex relationships and seamlessly integrating heterogeneous multimodal information. To address these challenges, we propose a novel Pathology-Genomic Cross Attention Network (PGCANet), which constructs hierarchical pathological structure graphs from whole slide images and achieves robust multimodal feature representation through a cross-attention mechanism. Our model not only focuses on identifying crucial tumor regions but also captures long-range dependencies within and across modalities, enabling more comprehensive data integration. Extensive experiments on large-scale datasets demonstrate that PGCANet significantly outperforms state-of-the-art methods in survival prediction, achieving a notable improvement in the C-Index. These results underscore the potential of our approach in leveraging multimodal data for more accurate and interpretable cancer prognosis.