Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-view Evidential Learning-based 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 a multi-view evidential learning framework that fuses SAM generalizable features with domain-specific encoder features via a Mamba-based multi-view state space model and models boundary uncertainty with a variational Dirichlet loss for medical image segmentation.
Problem
Single-view domain-specific medical segmentation models generalize poorly due to limited knowledge scope, while vision foundation models like SAM lack domain-specific knowledge without fine-tuning; boundary segmentation uncertainty also remains high.
Approach
Combine a frozen SAM encoder (generalizable features) with a trainable domain-specific encoder, fuse them with a Mamba-based multi-view state space model (MV-SSM, with pixel-window channel Mamba blocks), and reduce boundary uncertainty by training a Dirichlet-based evidential head with subjective-logic losses.
Key results
- On MoNuSeg the method reaches 81.88% Dice and 69.45% mIoU, versus 79.09%/66.68% for UCTransNet in the ablation baseline.
- On GlaS, 90.59% Dice and 83.70% mIoU; on TNBC 83.75% Dice and 72.23% mIoU; on DRIVE 84.12% Dice and 72.67% mIoU, best among compared methods.
- Ablation on MoNuSeg shows MV-SSM adds about 1.66% Dice and 2.28% mIoU, the largest gain among components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the uncertainty module (ablation) | UCTransNet · MoNuSeg | Dice: 79.09 -> 79.64 | Helped |
| add the uncertainty module (ablation) | UCTransNet · MoNuSeg | MIoU: 66.68 -> 66.33 | Hurt |
| add the uncertainty module (ablation) | UCTransNet · GlaS | Dice: 89.76 -> 89.98 | Helped |
| add the uncertainty module (ablation) | UCTransNet · GlaS | MIoU: 81.91 -> 82.43 | Helped |
| add Mamba Residual CCT (ablation) | UCTransNet + uncertainty · MoNuSeg | Dice: 79.64 -> 80.22 | Helped |
| add Mamba Residual CCT (ablation) | UCTransNet + uncertainty · MoNuSeg | MIoU: 66.33 -> 67.19 | Helped |
| add Mamba Residual CCT (ablation) | UCTransNet + uncertainty · GlaS | Dice: 89.98 -> 90.11 | Helped |
| add Mamba Residual CCT (ablation) | UCTransNet + uncertainty · GlaS | MIoU: 82.43 -> 82.45 | Helped |
| add MV-SSM (ablation) | UCTransNet + uncertainty + Mamba Residual CCT · MoNuSeg | Dice: 80.22 -> 81.30 | Helped |
| add MV-SSM (ablation) | UCTransNet + uncertainty + Mamba Residual CCT · MoNuSeg | MIoU: 67.19 -> 68.61 | Helped |
| add MV-SSM (ablation) | UCTransNet + uncertainty + Mamba Residual CCT · GlaS | Dice: 90.11 -> 90.12 | Helped |
| add MV-SSM (ablation) | UCTransNet + uncertainty + Mamba Residual CCT · GlaS | MIoU: 82.45 -> 82.94 | Helped |
Limitations
- Evaluation limited to four 2D medical segmentation datasets with a fixed 224x224 input size.
- Gains over the strongest baselines are modest (e.g., about 1.3% Dice on TNBC) and no variance or significance reported.
- Several architectural descriptions (e.g., WCA equations, dataset split details) are sparse or ambiguous.
Experimental setup
Datasets: MoNuSeg GlaS TNBC DRIVE
Benchmarks: MoNuSeg GlaS TNBC DRIVE
Metrics: Dice mIoU
Baselines: UNet UNet++ AttUNet Swin-UNet TransUNet UCTransNet VMUNet VMUNet v2 MedSAM LViT SAM TGANet
Builds on: Mamba SAM UCTransNet (CCT) Subjective logic evidential deep learning Bidirectional Mamba
Closest papers by meaning
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.86)
- Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Ophiuchus: Incentivizing Tool-augmented ''Think with Images'' for Joint Medical Segmentation, Understanding and Reasoning (ICML 2026 · similarity 0.86)
- Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation (NeurIPS 2025 · similarity 0.85)
- Training Like a Medical Resident: Context-Prior Learning Toward Universal Medical Image Segmentation (CVPR 2024 · similarity 0.85)
- Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · 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.