Research map / Detection and segmentation / Open-set and few-shot detection
Clustering Propagation for Universal 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 S2VNet, a 2D clustering-based slice-to-volume propagation framework that unifies automatic and interactive medical image segmentation in one model trained in one session.
Problem
Medical image segmentation methods are tailored separately for automatic or interactive setups, requiring duplicate models, and 3D solutions are slow and memory-heavy while failing to bridge distant slices.
Approach
Initialize k-means-style cluster centers (queries) of a 2D clustering segmenter from the previous slice's centroids so knowledge propagates through the volume; interactive clicks initialize centroids directly, so no architectural change is needed, and a recurrent centroid aggregation module fuses historic centroids.
Key results
- On WORD test automatic setup, S2VNet reaches 87.36% average DSC and 4.64 HD95, surpassing SwinMM (86.18%) by 1.18 points.
- On BTCV val, S2VNet achieves 83.81% DSC automatic and 86.11% interactive, versus nnFormer's 81.62% automatic.
- Compared to SwinMM, S2VNet reports about 15x faster inference (FPS) and 48.2% memory reduction on WORD test.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use S2VNet instead of SwinMM | Automatic segmentation · WORD test | DSC: 86.18 -> 87.36 | Helped |
| Use S2VNet instead of nnFormer | Automatic segmentation · BTCV val | DSC: 81.62 -> 83.81 | Helped |
| Add clustering-based slice-to-volume propagation (ablation) | S2VNet · WORD test | DSC: 78.67 -> 86.19 | Helped |
| Add interaction-aware centroid initialization (ablation) | S2VNet · WORD test | DSC: 87.36 -> 89.70 | Helped |
| Add adaptive pixel-feature sampling to fuse interaction rounds (ablation) | S2VNet · WORD test | DSC: 89.70 -> 90.64 | Helped |
| Add recurrent centroid aggregation (ablation) | S2VNet · WORD test | DSC: 90.64 -> 91.41 | Helped |
Surprising findings
- The unified universal model slightly outperforms task-specific automatic and interactive models (e.g., 87.36 vs 87.02 DSC automatic on WORD), despite sharing one model.
Limitations
- Interactive baselines (iSegFormer, Mem3D, SwinMM) were evaluated only on a subset of classes due to workload of training one model per class.
- Clustering centers can drift due to outliers and lose remote cues, motivating recurrent aggregation rather than solving it fully.
Experimental setup
Datasets: WORD BTCV AMOS
Benchmarks: WORD BTCV AMOS
Metrics: DSC HD95 NSD
Baselines: UNETR CoTr Swin UNETR SwinMM nnFormer 3D-UX-Net iSegFormer Mem3D TransUNet TransBTS
Builds on: Mask2Former k-means cross attention DETR Swin Transformer
Closest papers by meaning
- UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation (NeurIPS 2026 · similarity 0.88)
- E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.86)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.86)
- Integrating Deep Metric Learning with Coreset for Active Learning in 3D Segmentation (NeurIPS 2024 · similarity 0.85)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.84)
- SegVol: Universal and Interactive Volumetric Medical Image Segmentation (NeurIPS 2024 · similarity 0.84)
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.