| Abstract [eng] |
Background: Cancer is acknowledged as a complex and heterogeneous disease, known for significant mortality rates, consequently demanding a comprehensive and integrated approach to improve diagnostic, prognostic, and therapeutic strategies. Such an integrative approach can be based on multimodal data, including medical imaging and omics data, which provide complementary insights into the disease. Simultaneously, advancements in Artificial Intelligence, particularly in deep learning, have shown promising results in the analysis of complex datasets, including multimodal data, thus enhancing cancer-related activities. Nonetheless, many deep learning models operate as 'BLACK-BOXES,' limiting clinical adoption due to a lack of transparency and trust. Self-Explainable Artificial Intelligence (S-XAI) aims to address this limitation by developing approaches that explain the internal processes of AI models, focusing on built-in interpretability to enhance trustworthiness and adoption. Attention mechanisms, which draw inspiration from human visual perception, facilitate AI models to focus on the key features of input data; while not explicitly designed for explainability, they naturally highlight key features, thus providing insights into the reasoning processes of the model. Despite their potential advantages, significant knowledge gaps remain in this domain, particularly regarding S-XAI integrated with Attention Mechanisms for multimodal cancer data, which remain unexplored. To address these gaps, this systematic review adheres to a pre-registered protocol. Objectives: To systematically review and synthesize existing literature concerning the deployment of attention mechanisms in S-XAI for interpretable cancer analysis using either multimodal integration or individual use of medical images and omics data. Methods: This systematic review complies with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. Comprehensive searches will be conducted across electronic databases such as PubMed/MEDLINE, Scopus, Web of Science, and IEEE Xplore for data retrieval. The studies will be screened and data extracted by a primary reviewer and verified by supervisors, employing suitable tools for bias risk assessment. Results: The findings of this systematic review will be comprehensively summarized and presented. Conclusions: This systematic review is essential to provide a comprehensive overview of the current state of S-XAI with attention mechanisms, and their potential uses to improve interpretability in multimodal cancer analysis. |
