Research map / Detection and segmentation / Open-set and few-shot detection
Unified Medical Lesion Segmentation via Self-referring Indicator
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 SR-ICL, an in-context-learning unified medical lesion segmentation model that refines indicator vectors using self-referring cues from the query's coarse prediction, trained once across 8 lesion tasks with low memory.
Problem
In-context-learning based unified medical lesion segmentation degrades when reference sets are poorly matched to the query, because cross-fusion designs make query features depend heavily on the reference set.
Approach
Replace cross-fusion with reference indicator generation (masked average pooling of reference features into dynamic head weights) and a self-referring module that refines these weights using query features gated by the query's own coarse prediction, plus a "separate BP - average update" training strategy for constant memory across tasks.
Key results
- SR-ICL achieves the best Dice on all 8 MLS tasks in the unified comparison, e.g., 80.54 Dice on Wet AMD and 82.36 on Lung Infection using 4 reference pairs, versus 78.51 and 73.85 for Spider.
- Ablation shows the full model (79.06 avg-style Dice on the first task set) improves over a w/o SRIG variant (76.65 Dice on Wet AMD) and baseline (73.84).
- Training with separate BP - average update uses 5.7G (batch 6/2) and 8.3G (batch 12/4) GPU memory, and memory stays at 8.3G when task count doubles from 8 to 16, while the Spider-style strategy goes OOM at batch 12/4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SR-ICL instead of Spider | unified medical lesion segmentation · Wet AMD | Dice: 78.51 -> 80.54 | Helped |
| use SR-ICL instead of Spider | unified medical lesion segmentation · Brain Tumor | Dice: 71.59 -> 74.29 | Helped |
| use SR-ICL instead of SegGPT | unified medical lesion segmentation · Adenocarcinoma | Dice: 94.24 -> 94.96 | Helped |
| use SR-ICL instead of Spider | unified medical lesion segmentation · Thyroid Nodule | Dice: 86.52 -> 87.91 | Helped |
| use SR-ICL instead of Spider | unified medical lesion segmentation · Colon Polyp | Dice: 80.66 -> 83.26 | Helped |
| use SR-ICL instead of Spider | unified medical lesion segmentation · Lung Infection | Dice: 73.85 -> 82.36 | Helped |
| use SR-ICL instead of Spider | unified medical lesion segmentation · Breast Lesion | Dice: 81.16 -> 84.92 | Helped |
| use SR-ICL instead of Spider | unified medical lesion segmentation · Skin Lesion | Dice: 88.79 -> 90.85 | Helped |
| add SRIG (ablation) | SR-ICL · Lung Infection | Dice: 77.07 -> 80.07 | Helped |
| add the background stream (ablation) | SR-ICL · Lung Infection | Dice: 78.41 -> 80.07 | Helped |
| set saliency threshold τ = 0.5 instead of τ = 0.7 (ablation) | SR-ICL · Lung Infection | Dice: 80.07 -> 79.69 | Hurt |
| set saliency threshold τ = 0.6 instead of τ = 0.7 (ablation) | SR-ICL · Lung Infection | Dice: 80.07 -> 80.22 | Helped |
Surprising findings
- SR-ICL performance with only 4 reference pairs exceeds prior ICL methods that use 64 references, and is insensitive to the number of references.
Limitations
- Adenocarcinoma performance drops notably when using box or point reference annotations because lesions occupy large image portions.
- Hyperparameters tau, alpha, beta are tuned by ablation on the test-style setup and the method relies on a coarse prediction being reasonable.
Experimental setup
Datasets: AMD-SD BTD EBHI-Seg TNUI 2021 Five colon polyp datasets COVID-19 data BUSI ISIC 2018
Benchmarks: AMD-SD BTD EBHI-Seg TNUI 2021 Five colon polyp datasets COVID-19 data BUSI ISIC 2018
Metrics: Dice mIoU
Baselines: UniverSeg SegGPT Spider DSC-ICL UNet TransUNet RollingUNet MALUNet EGE-UNet
Builds on: SegGPT Spider CondInst UniverSeg
Closest papers by meaning
- LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging (CVPR 2025 · similarity 0.84)
- Show and Segment: Universal Medical Image Segmentation via In-Context Learning (CVPR 2025 · similarity 0.84)
- Minding Fuzzy Regions: A Data-driven Alternating Learning Paradigm for Stable Lesion Segmentation (CVPR 2025 · similarity 0.84)
- Unleashing the Potential of Vision-Language Pre-Training for 3D Zero-Shot Lesion Segmentation via Mask-Attribute Alignment (ICLR 2025 · similarity 0.84)
- S³-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation (AAAI 2025 · similarity 0.83)
- EasyLens: A Training-Free Plug-and-Play Subtle-Lesion Representation Amplifier for Medical Vision-Language Models (NeurIPS 2026 · similarity 0.83)
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.