Research map / Architectures and efficiency / Mixture of experts
Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation
Read the paperCheck your idea against 108,967 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 HoME, a two-level hierarchical soft mixture-of-experts layer integrated with Mamba in a U-shaped network (Mamba-HoME) for 3D medical image segmentation across CT, MRI, and ultrasound.
Problem
Efficient long-context 3D medical image segmentation across modalities, where SSMs capture global context but do not adaptively handle diverse local patterns, and global MoE routing over millions of tokens is costly.
Approach
Two-level token routing: tokens are softly assigned to slots within local groups and processed by per-group local experts, then slot outputs are routed through a global soft MoE for cross-group fusion; embedded in a Mamba block with gated spatial convolution and dynamic tanh normalization.
Key results
- On PANORAMA/in-house, Mamba-HoME (pre-trained) reaches 78.2 mDSC and 4.3 mHD95 versus 76.6 for the best prior mDSC (uC 3DU-Net, SuPreM).
- On FeTA 2022 5-fold CV, Mamba-HoME achieves 87.5 mDSC and 2.1 mHD95 (scratch) versus 86.5 for the best baseline Hermes.
- Mamba-HoME has 170.1M parameters and 11.1 GB GPU memory, with inference about 30% slower than SegMamba.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add supervised pre-training | Mamba-HoME · PANORAMA and in-house test sets | PDAC DSC (%): 54.8 -> 56.7 | Helped |
| add supervised pre-training | Mamba-HoME · AMOS-CT, CT→MRI, MRI, and CT+MRI | mDSC (%): 86.3 -> 87.3; 84.8 -> 85.0; 81.0 -> 82.3; 85.1 -> 86.4 | Helped |
| add supervised pre-training | Mamba-HoME · FeTA 2022 | mDSC (%): 87.5 -> 87.7 | Helped |
| add supervised pre-training | Mamba-HoME · MVSeg | mDSC (%): 84.8 -> 85.0 | Helped |
| increase expert counts beyond E1=[4, 8, 12, 16], E2=[8, 16, 24, 32] (ablation) | Mamba-HoME · PANORAMA | mDSC (%): 77.5 -> 76.3; 77.5 -> 75.8; 77.5 -> 77.5 | Mixed |
| use group sizes [1024, 512, 256, 128] instead of [2048, 1024, 512, 256] (ablation) | Mamba-HoME · PANORAMA | mDSC (%): 77.5 -> 77.2 | Hurt |
| set first- and second-level expert counts equal (ablation) | Mamba-HoME · PANORAMA | mDSC (%): 77.5 -> 76.8 | Hurt |
| increase group sizes to [4096, 2048, 1024, 512] (ablation) | Mamba-HoME · PANORAMA | mDSC (%): 77.5 -> 77.4 | Hurt |
| vary slots per expert from 4 to 1, 2, or 8 (ablation) | Mamba-HoME · PANORAMA | mDSC (%): 77.5 -> 76.2; 77.5 -> 77.2; 77.5 -> 76.5 | Hurt |
| replace Layer Normalization with Dynamic Tanh (ablation) | Mamba-HoME · PANO, AMOS, and FeTA | mDSC (%): 77.4 -> 77.5; 86.2 -> 86.3; 87.5 -> 87.4 | Mixed |
Surprising findings
- Dynamic Tanh gives performance on par with Layer Normalization in SSM-based segmentation while speeding training and inference by about 6%.
- Increasing the number of experts beyond the chosen configuration (e.g., 277.8M to 367.0M parameters) did not improve performance.
Limitations
- Scalability to large datasets (>10,000 scans) is unexplored.
- Behavior under large-scale self-supervised learning (>200,000 scans) is not studied.
- Inference is about 30% slower than the SegMamba baseline and the model has 170.1M parameters.
Experimental setup
Datasets: AbdomenAtlas 1.1 TotalSegmentator MRI PANORAMA AMOS FeTA 2022 MVSeg MSD Pancreas NIH Pancreas in-house CT
Benchmarks: PANORAMA AMOS-CT AMOS-MRI FeTA 2022 MVSeg in-house CT MSD Pancreas NIH Pancreas
Metrics: mDSC mHD95 parameters GPU memory inference speed sensitivity specificity
Baselines: VoCo-B Hermes Swin SMT VSmTrans uC 3DU-Net Swin UNETR SegMamba SuPreM
Builds on: Mamba Soft Mixture-of-Experts SegMamba Dynamic Tanh
Closest papers by meaning
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.86)
- Diversified and Personalized Multi-rater Medical Image Segmentation (CVPR 2024 · similarity 0.86)
- Multi-view Evidential Learning-based Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Integrating Deep Metric Learning with Coreset for Active Learning in 3D Segmentation (NeurIPS 2024 · similarity 0.85)
- E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.85)
- S³-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation (AAAI 2025 · similarity 0.85)
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.