Drop and Align: Fusing Pathology and Genomics via Graph Learning for Cancer Survival Prediction

Pathological image and genomic data have been widely explored in cancer survival analysis. However, due to the instability of gene sequencing and the randomness of tumor tissue sampling, it remains challenging to extract robust feature representation from each modality. Besides, the high semantic heterogeneity also adds an extra burden to the integration of tissue morphological feature and genomic expression information. In this paper, we propose the node Dropping and feature Alignment based Graph (DAGraph), a unified pathology-genomics fusion framework that utilizes graph neural network to model the spatial organization of pathological image and correlation information of genomics data. Notably, we propose to randomly drop nodes in pathological graph and genomic graph. Then we introduce an intra-modal consistency loss to guide the model to extract robust and general feature via minimizing the feature discrepancy between the original and modified graph in each modality. Additionally, an inter-modal alignment loss is also adopted to reduce the feature heterogeneity and contribute to more sufficient multi-modal fusion. Transformer-based fusion layer is utilized to integrate feature embedding of both modalities. Experiments on low-grade gliomas (LGG) dataset collected from The Cancer Genome Atlas (TCGA) demonstrate the robustness and effectiveness of the proposed method.