Research map / Architectures and efficiency / Mixture of experts
Anatomy-Activated Mixture-of-Experts for 3D Medical Vision-Language Pre-training
Szymon Płotka, Gizem Mert, Pedro R. A. S. Bassi, Wenxuan Li, Zongwei Zhou, Anna M Marcinkiewicz, Wiktoria Romańczyk, Jarosław B Ćwikła, Łukasz Struski, Ewa Szczurek, Jacek Tabor, Arkadiusz Sitek
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an anatomy-activated mixture-of-experts architecture for 3D medical vision-language pre-training, routing volumetric patches with organ-level anatomical context rather than isolated local features.
Problem
Medical 3D vision-language pre-training uses anatomy-agnostic encoders that apply uniform dense parameterization to all volumetric patches, and existing MoE routers gate on isolated local features without anatomical context.
Approach
Anatomy-activated MoE routing for 3D medical VLP so that expert selection is conditioned on organ/anatomical context instead of only local patch features; the abstract is truncated so exact mechanism is not fully stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add anatomy-activated MoE routing (ablation) | SegMamba · AMOS-CT | DSC (%): 85.5 -> 87.5 | Improved |
| add anatomy-activated MoE routing (ablation) | SegMamba · AMOS-CT | HD95 (mm): 17.3 -> 7.6 | Improved |
| add OQMSP organ-query alignment after A2MoE (ablation) | A2MoE SegMamba · AMOS-MRI | DSC (%): 78.5 -> 82.3 | Improved |
| add OQMSP organ-query alignment after A2MoE (ablation) | A2MoE SegMamba · AMOS-CT | HD95 (mm): 7.6 -> 6.2 | Improved |
| add OQMSP organ-query alignment after A2MoE (ablation) | A2MoE SegMamba · GEPAR3D | HD95 (mm): 7.9 -> 8.1 | Hurt |
| add histogram embedding to router (ablation) | MoE router · AMOS-CT | DSC (%): 86.0 -> 87.0 | Improved |
| add class-to-expert affinity prior to router (ablation) | MoE router with histogram · AMOS-CT | HD95 (mm): 12.0 -> 7.6 | Improved |
| add affinity prior and histogram (ablation) | MoE router · AMOS-MRI | DSC (%): 77.4 -> 78.5 | Improved |
| use [16,16,16,16] experts instead of [8,8,8,8] (ablation) | A2MoE · PANORAMA | mDSC (%): 75.1 -> 78.2 | Improved |
| use [16,16,16,16] experts instead of [4,4,4,4] (ablation) | A2MoE · AMOS-CT | mDSC (%): 87.5 -> 86.9 | Hurt |
| use pyramid expert allocation [4,8,16,32] (ablation) | A2MoE · multiple datasets | mDSC (%): 77.4 vs 78.2 (PANORAMA); 86.1 vs 86.2 (GEPAR3D) | Mixed |
Limitations
- Full text unavailable; only the truncated abstract was used
Experimental setup
Builds on: mixture-of-experts
Closest papers by meaning
- Large-scale and Fine-grained Vision-language Pre-training for Enhanced CT Image Understanding (ICLR 2025 · similarity 0.89)
- Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.88)
- 3DMedAgent: Unified Perception-to-Understanding for 3D Medical Analysis (ICML 2026 · similarity 0.87)
- Enhancing vision-language models for medical imaging: bridging the 3D gap with innovative slice selection (NeurIPS 2024 · similarity 0.87)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.87)
- AOR: Anatomical Ontology-Guided Reasoning for Medical Large Multimodal Model in Chest X-Ray Interpretation (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
