Fu Bangkang, He Junjie, Zhang Xiaoli, Peng Yunsong, Zhang Zhuxu, Tang Qi, Liu Xinfeng, Cao Ying, Wang Rongpin
Medical College, Guizhou University, Guiyang, 550025, Guizhou, China; Guizhou Province International Science and Technology Cooperation Base for Precision Imaging Diagnosis and Treatment, Key Laboratory of Advanced Medical Imaging and Intelligent Computing of Guizhou Province, Department of Radiology, Guizhou Provincial People's Hospital, Guiyang, 550002, Guizhou, China.
Guizhou Province International Science and Technology Cooperation Base for Precision Imaging Diagnosis and Treatment, Key Laboratory of Advanced Medical Imaging and Intelligent Computing of Guizhou Province, Department of Radiology, Guizhou Provincial People's Hospital, Guiyang, 550002, Guizhou, China.
Med Image Anal. 2026 Jan;107(Pt A):103810. doi: 10.1016/j.media.2025.103810. Epub 2025 Sep 24.
Multimodal data play a significant role in survival analysis, with pathological images providing morphological information about tumors and genomic data offering molecular insights. Leveraging multimodal data for survival analysis has become a prominent research topic. However, the heterogeneity of data poses significant challenges to multimodal integration. While existing methods consider interactions among features from different modalities, the heterogeneity of feature spaces often hinders performance in survival analysis. In this paper, we propose a hybrid supervised framework for survival analysis (HSFSurv) based on multimodal feature decomposition. This framework utilizes a multimodal feature decomposition module to partition features into highly correlated and modality-specific components, facilitating targeted feature fusion in subsequent steps. To alleviate feature space heterogeneity, we design an individual-level uncertainty minimization (UMI) module to ensure consistency in prediction outcomes. Additionally, we develop a feature-level multimodal cohort contrastive learning (MCF) module to enforce consistency across features. Moreover, a probabilistic decay detection module with a supervisory signal is introduced to guide the contrastive learning process. These modules are jointly trained to project multimodal features into a shared latent vector space. Finally, we fine-tune the framework for survival analysis tasks to achieve prognostic predictions. Experimental results on five cancer datasets demonstrate the state-of-the-art performance of the proposed multimodal fusion framework in survival analysis.
多模态数据在生存分析中发挥着重要作用，病理图像提供有关肿瘤的形态学信息，而基因组数据提供分子层面的见解。利用多模态数据进行生存分析已成为一个突出的研究课题。然而，数据的异质性给多模态整合带来了重大挑战。虽然现有方法考虑了来自不同模态的特征之间的相互作用，但特征空间的异质性常常阻碍生存分析中的性能表现。在本文中，我们提出了一种基于多模态特征分解的用于生存分析的混合监督框架（HSFSurv）。该框架利用一个多模态特征分解模块将特征划分为高度相关和特定于模态的组件，便于在后续步骤中进行有针对性的特征融合。为了缓解特征空间的异质性，我们设计了一个个体层面的不确定性最小化（UMI）模块，以确保预测结果的一致性。此外，我们开发了一个特征层面的多模态队列对比学习（MCF）模块，以加强特征之间的一致性。而且，引入了一个带有监督信号的概率衰减检测模块来指导对比学习过程。这些模块联合训练，将多模态特征投影到一个共享的潜在向量空间中。最后，我们针对生存分析任务对该框架进行微调，以实现预后预测。在五个癌症数据集上的实验结果证明了所提出的多模态融合框架在生存分析中的最优性能。
