Multimodal Imbalance-Aware Learning for Cancer Survival Prediction

Cancer survival prediction is vital for crafting effective treatment strategies and improving patient prognoses. Although multimodal survival models integrating genomic data with pathology images have made significant progress, they suffer from the imbalanced learning problem where genomics modality, with stronger predictive capability, dominates the model training process and hinders the effective learning of pathology modality. In this paper, we propose a novel multimodal imbalance-aware learning (MIAL) method to tackle this issue. MIAL incorporates an imbalance factor (IF) to dynamically detect and measure learning disparities between modalities during training. In addition, MIAL introduces pathology compensation module (PCM) that provides extra supervision to pathology modality according to the IF, thereby promoting balanced learning across modalities. Extensive experiments conducted on many cancer datasets from the Cancer Genome Atlas (TCGA) indicate that MIAL successfully addresses the imbalanced learning problem and notably improves the performance of multimodal cancer survival prediction.