Research map / Detection and segmentation / Open-set and few-shot detection
Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline
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 IMed-361M, a 6.4M-image, 361M-mask medical interactive segmentation dataset built by SAM-based automatic mask generation with GT correction, and an IMIS-Net baseline (SAM-style ViT-B) supporting click, box and text prompts.
Problem
Interactive medical image segmentation lacks large-scale, diverse, densely annotated datasets, which limits model generalization and consistent evaluation across methods.
Approach
Scale medical segmentation supervision by using SAM's automatic mask generation, corrected against original ground truth, to produce dense interactive masks (56 per image on average), and train a SAM-style baseline on this data with simulated multi-step interactions.
Key results
- IMIS-Net achieves 89.27% average Dice on SegThor and 79.06% on TotalSegmentator MRI, the best average among compared methods.
- On ISLES stroke lesion segmentation IMIS-Net reaches 71.78 Dice, surpassing the second-best SAM-Med2D by 3.56 points.
- Scaling decoder dimension from 256 to 768 raises Dice from 84.97% to 90.60% with bbox prompts, with about 24M more trainable parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train IMIS-Net on IMed-361M with interactive masks in addition to GT masks (ablation) | IMIS-Net · IMed-361M test set | Dice | Helped |
| increase the amount of training data (ablation) | IMIS-Net · IMed-361M test set | Dice | Helped |
| combine text prompts with point prompts (ablation) | IMIS-Net · IMed-361M test set | Dice: +11.95% | Helped |
| increase decoder dimension from 256 to 768 (ablation) | IMIS-Net · IMed-361M test set | Dice: 84.97% to 90.60% | Helped |
| increase training image resolution (ablation) | IMIS-Net · IMed-361M test set | Dice | Helped |
| increase image resolution from 256×256 to 512×512 with decoder dimension 768 (ablation) | IMIS-Net · IMed-361M test set, point prompts | Dice: 0.8214 -> 0.8673 | Helped |
| increase image resolution from 256×256 to 512×512 with decoder dimension 768 (ablation) | IMIS-Net · IMed-361M test set, bbox prompts | Dice: 0.8469 -> 0.8968 | Helped |
| use IMIS-Net trained on IMed-361M instead of SAM-Med2D | IMIS-Net · ISLES | Dice: 68.22 -> 71.78 | Helped |
Surprising findings
- Radiologists rated 49 of 550 sampled interactive masks as 'Bad', concentrated in 18 source datasets, so the automatically generated masks are not uniformly reliable.
Limitations
- Interactive masks lack semantic labels and some have little clinical relevance.
- Automatic mask quality varies; 18 source datasets produced many bad masks requiring manual removal.
Experimental setup
Datasets: IMed-361M SA-Med2D-20M SegThor TotalSegmentator MRI ISLES
Benchmarks: IMed-361M test set SegThor TotalSegmentator MRI ISLES
Metrics: Dice
Baselines: SAM SAM 2 MedSAM Huang et al. SAM-Med2D
Builds on: SAM SAM 2 MedSAM SAM-Med2D SA-Med2D-20M
Closest papers by meaning
- EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation (CVPR 2024 · similarity 0.88)
- ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement (AAAI 2025 · similarity 0.87)
- UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation (NeurIPS 2026 · similarity 0.86)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.86)
- Generative Medical Segmentation (AAAI 2025 · similarity 0.86)
- Tyche: Stochastic In-Context Learning for Medical Image Segmentation (CVPR 2024 · similarity 0.86)
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.