This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper highlights the superior performance of deep learning in integrating diverse data types to perform survival prediction, but existing multimodal methods focus on a single cancer type and overlook the difficulty of generalizing across cancer types. The researchers first revealed that multimodal prediction models, despite the need for robustness in clinical settings, generally underperform unimodal models in situations beyond cancer types. To address this, we propose a novel task, "Cross-Cancer Single Domain Generalization for Multimodal Prognosis," to assess whether models trained on a single cancer type can generalize to unseen cancer types. We identify two key challenges: degraded weak modal features and inefficient multimodal integration. To address these challenges, we introduce two plug-and-play modules: Sparse Dirac Information Rebalancer (SDIR) and Cancer-aware Distribution Entanglement (CADE). SDIR applies Bernoulli-based sparsification and Dirac-based stabilization to mitigate the dominance of strong features and enhance weak modal signals. CADE is designed to synthesize target domain distributions by integrating local morphological cues and global gene expression in latent space. It demonstrates excellent generalization performance in benchmark experiments on four cancer types, laying the foundation for practical and robust multimodal prediction across cancer types. The code is available at https://github.com/HopkinsKwong/MCCSDG .
We first identify the cross-cancer generalization problem in multimodal survival prediction models and present new challenges and methodologies to address it.
◦
Leveraging weak modal features and improving multimodal integration efficiency through SDIR and CADE modules.
◦
Verification of the proposed method's excellent generalization performance through benchmark experiments on four cancer types.
◦
Suggesting the possibility of developing a multimodal predictive model that is practical and robust across cancer types.
•
Limitations:
◦
Generalization performance was evaluated using a limited dataset of four cancer types. Further expansion to a more diverse and comprehensive dataset of cancer types is needed.
◦
The proposed method lacks a comparative analysis with other state-of-the-art multimodal learning methods. Additional comparative experiments with other methods are needed.
◦
Lack of detailed descriptions of parameter optimization in the SDIR and CADE modules. A detailed description of the modules' hyperparameter tuning strategies is needed.
