A concept-guided multimodal mixture-of-experts framework for lung cancer diagnosis and prognosis prediction
Accurate lung cancer assessment requires integrating histopathological morphology, molecular alterations, and prognostic risk. However, most computational pathology models are either task-specific or rely on high-dimensional multimodal fusion that provides limited insight into how image-derived and genomic evidence jointly supports a clinical decision. We propose CONCERTO, a concept-guided multimodal mixture-of-experts framework for multi-endpoint lung cancer analysis. The model integrates whole-slide images (WSIs) with genomics data, decomposes multimodal evidence into image-specific, omics-specific, redundant, and synergistic expert pathways, and maps expert representations to hierarchical morphology and molecular concepts. Across TCGA lung cancer tasks, CONCERTO demonstrates strong performance in subtype classification, molecular alteration prediction, and survival prediction. Ablation studies further support the contribution of concept bottlenecks, hierarchical concept organization, and interaction-aware expert specialization. These results suggest a scalable route toward multimodal lung cancer AI that is both predictive and interpretable.
Authors
- 李俞慧
- Ting Wang (ORCID: https://orcid.org/0000-0002-6569-4395)
- Fei Ye (ORCID: https://orcid.org/0009-0000-9006-884X)
- Rongshan Zheng
- Ling Weng
- Hongying Zhang
- Limin Chen
Institutions
- Fuzhou Pulmonary Hospital of Fujian (CN)
- Union Hospital (CN)
Publication Details
- Journal
- Frontiers in Oncology
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3389/fonc.2026.1957081
- Primary Topic
- AI in cancer detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00