Organ-aware mixture-of-experts framework for generalized pan-tumor segmentation
Biomedical Signal Processing and Control, vol.126, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 126
- Publication Date: 2026
- Doi Number: 10.1016/j.bspc.2026.110937
- Journal Name: Biomedical Signal Processing and Control
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, EMBASE
- Keywords: Attention, Medical image, Mixture-of-experts, Tumor segmentation
- Dokuz Eylül University Affiliated: Yes
Abstract
Medical image segmentation plays an important role in tumor diagnosis, treatment planning, and anatomical analysis across diverse imaging modalities. However, generalized pan-tumor segmentation remains challenging because organs exhibit substantial anatomical variability, while different imaging modalities introduce distinct appearance characteristics. To address these challenges, we propose an Organ-Aware Mixture-of-Experts framework, termed OAMoE, which incorporates anatomical prior knowledge into a two-tier expert selection mechanism. Specifically, OAMoE first organizes experts into anatomically guided expert pools and then activates the top two experts within each pool through an adaptive attention-based routing strategy. This layered design promotes expert specialization while maintaining computational efficiency and interpretability. We evaluated OAMoE on eight multimodal datasets spanning four anatomical regions: the brain (BraTS2019, BraTS2020, BraTS2021, MSD, and OASIS); the breast (US3M); the lung (NSCLC); and the ovary (MMOTU). Under a unified in-house evaluation protocol, OAMoE achieved an average Dice score of 89.1% across the seen datasets. The model achieved the best Dice score of 90.83% in the BraTS2020. To further assess cross-organ generalization, we evaluated the model on the unseen Brain datasets where it achieved an average Dice score of 86.0% and outperformed the internal generic MoE and shared encoder–decoder comparison baselines. These results indicate that OAMoE can improve robustness and transferability across organs and imaging modalities. Additionally, attention visualizations suggest that the model dynamically allocates segmentation tasks to experts that are sensitive to specific anatomical and modality-related features, supporting its potential for interpretable and adaptive clinical AI. Code will be released at https://github.com/XPLab-GM/OAMoE.