Chen Genlang, Sui Sixuan, Zhang Jiajian, Liu Xuan, Cai Ping
School of Computer Science and Data Engineering, NingboTech University, China.
College of Computer Science, Zhejiang University, China.
J Biomed Inform. 2025 Aug;168:104836. doi: 10.1016/j.jbi.2025.104836. Epub 2025 Jun 6.
Cancer survival prediction plays a vital role in enhancing medical decision-making and optimizing patient management. Accurate survival estimation enables healthcare providers to develop personalized treatment plans, improve treatment outcomes, and identify high-risk patients for timely intervention. However, existing methods often rely on single-modality data or suffer from excessive computational complexity, limiting their practical application and the full potential of multimodal integration.
To address these challenges, we propose a novel multimodal survival prediction framework that integrates Whole Slide Image (WSI) and genomic data. The framework employs attention mechanisms to model intra-modal and inter-modal correlations, effectively capturing complex dependencies within and between modalities. Additionally, locality-sensitive hashing is applied to optimize the self-attention mechanism, significantly reducing computational costs while maintaining predictive performance, enabling the model to handle large-scale or high-resolution WSI datasets efficiently.
Extensive experiments on the TCGA-BLCA dataset validate the effectiveness of the proposed approach. The results demonstrate that integrating WSI and genomic data improves survival prediction accuracy compared to unimodal methods. The optimized self-attention mechanism further enhances model efficiency, allowing for practical implementation on large datasets.
The proposed framework provides a robust and efficient solution for cancer survival prediction by leveraging multimodal data integration and optimized attention mechanisms. This study highlights the importance of multimodal learning in medical applications and offers a promising direction for future advancements in AI-driven clinical decision support systems.
癌症生存预测在加强医疗决策和优化患者管理方面发挥着至关重要的作用。准确的生存估计使医疗服务提供者能够制定个性化的治疗方案，改善治疗效果，并识别高危患者以便及时干预。然而，现有方法往往依赖单模态数据或存在计算复杂度过高的问题，限制了它们的实际应用以及多模态整合的全部潜力。
为应对这些挑战，我们提出了一种新颖的多模态生存预测框架，该框架整合了全切片图像（WSI）和基因组数据。该框架采用注意力机制对模态内和模态间的相关性进行建模，有效地捕捉模态内部和之间的复杂依赖关系。此外，应用局部敏感哈希来优化自注意力机制，在保持预测性能的同时显著降低计算成本，使模型能够高效处理大规模或高分辨率的WSI数据集。
在TCGA-BLCA数据集上进行的大量实验验证了所提方法的有效性。结果表明，与单模态方法相比，整合WSI和基因组数据提高了生存预测准确性。优化后的自注意力机制进一步提高了模型效率，使其能够在大型数据集上实际应用。
所提出的框架通过利用多模态数据整合和优化的注意力机制，为癌症生存预测提供了一个强大而高效的解决方案。本研究强调了多模态学习在医学应用中的重要性，并为人工智能驱动的临床决策支持系统的未来发展提供了一个有前景的方向。
