Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
One-Prompt to Segment All Medical Images
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.
Introduces One-Prompt Segmentation, a single-forward-pass medical segmentation model that takes one clinician-prompted template image to segment unseen tasks, trained on 64 medical datasets and tested on 14 held-out datasets.
Problem
Medical image segmentation foundation models either need user prompts for every image (interactive models like SAM) or labeled support samples (few/one-shot models), both of which are costly for clinical use on unseen tasks.
Approach
A One-Prompt Former decoder with a Prompt-Parser (adaptive attentive mask with Gaussian masking) fuses a prompted template image with a query image across feature scales, so a single prompted example suffices for a new task in one forward pass; four prompt types (Click, BBox, Doodle, SegLab) are supported.
Key results
- On 14 held-out tasks with one prompted template, One-Prompt outperforms few/one-shot baselines (e.g., second-best averages 11.2% lower with SegLab prompts).
- In zero-shot 'segment everything' setting on 11 unseen datasets, One-Prompt achieves 64.0% average Dice, surpassing the second-highest (MSA-style SAM methods) by 10.7%.
- Efficiency: average user cost 2.28 s per image prompt versus 27.47 s for annotating; One-Prompt reaches 73.98 Dice vs. 77.21 for task-specific TransUNet upper bound.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use One-Prompt instead of the strongest listed baseline, MedSAM, under segment-everything evaluation | One-Prompt · 11 unseen tasks average | Dice (%): 53.9 -> 64.0 | Helped |
| Use One-Prompt instead of the supervised TransUNet upper-bound | One-Prompt · 14 held-out test tasks | Dice: 77.21 -> 73.98 | Hurt |
| Use Stack MLP prompting with Gaussian Masking instead of the alternative Prompt-Parser combinations (ablation) | One-Prompt Model · held-out test set | Dice score | Helped |
| Increase prompt quality from Low toward Human | One-Prompt Model · REFUGE and WBC | segmentation performance | Helped |
| Use fine-grained SegLab prompts rather than other prompt types | One-Prompt Model · REFUGE and WBC | segmentation performance | Helped |
Surprising findings
- Interactive-style use of One-Prompt (prompting the query image itself) outperforms dedicated interactive SAM-style models without being trained for that setting.
Limitations
- Performance depends on prompt quality; lower-quality prompts degrade results notably.
- Variance across random template choices is up to about 13%, with larger variance on diverse targets like optic cup and pancreas.
- Few/one-shot baselines were given segmentation labels as prompts, and SAM-based comparisons used simulated oracle prompts, so comparisons are not fully matched.
Experimental setup
Datasets: KiTS23 ATLAS23 TDSC WBC SegRap CrossM23 REFUGE Pendal Pancreas-CT LNQ23 CAS23 CadVidSet
Benchmarks: KiTS23 ATLAS23 TDSC WBC SegRap CrossM23 REFUGE Pendal Pancreas-CT LNQ23 CAS23 CadVidSet
Metrics: Dice score
Baselines: PANet ALPNet SENet UniverSeg DAT ProbOne HyperSegNas LT-Net SAM SAM-U VMN iSegFormer
Builds on: Segment Anything Model UniverSeg TransUNet
Closest papers by meaning
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.89)
- AoP-SAM: Automation of Prompts for Efficient Segmentation (AAAI 2025 · similarity 0.86)
- Rethinking Interactive Image Segmentation with Low Latency High Quality and Diverse Prompts (CVPR 2024 · similarity 0.86)
- Prompting Diffusion Models for Zero-Shot Instance Segmentation (NeurIPS 2026 · similarity 0.86)
- CAT: Coordinating Anatomical-Textual Prompts for Multi-Organ and Tumor Segmentation (NeurIPS 2024 · similarity 0.86)
- Personalize Segment Anything Model with One Shot (ICLR 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.