Research map / Continual learning / Domain and prompt-based
Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D 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 VPTTA, which freezes a pre-trained segmentation model and trains a lightweight low-frequency visual prompt per test image (initialized from a memory bank, with BN-statistics warm-up) for continual test-time adaptation on 2D medical segmentation.
Problem
Continual test-time adaptation for 2D medical image segmentation under distribution shifts across medical centres, where updating model parameters causes error accumulation and catastrophic forgetting.
Approach
Keep the source model frozen and instead learn a tiny low-frequency amplitude prompt for each test image, aligning BN statistics; initialize prompts from a FIFO memory bank of similar past images and use a statistics-fusion warm-up that moves from mixed to source statistics.
Key results
- On OD/OC segmentation, VPTTA reaches 71.93 average DSC versus 65.86 for Source Only and 70.03 for the best competing method (SAR).
- On polyp segmentation, VPTTA reaches 80.46 average DSC and 85.62 S-alpha, outperforming competitors that fall below Source Only (e.g., TENT-continual 58.43 DSC).
- In long-term continual adaptation over three rounds, VPTTA shows 0.42 degradation versus 0.83 for SAR and 5.12 for TENT; the prompt uses only 75 parameters on OD/OC and 27 on polyp.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VPTTA instead of the strongest listed baseline, SAR | ResUNet-34 · OD/OC segmentation, average across domains A–E | DSC: 70.03 -> 71.93 | Helped |
| Use VPTTA instead of DUA | PraNet · Polyp segmentation, average across domains A–D | DSC: 77.13 -> 80.46 | Helped |
| Use VPTTA instead of DUA | PraNet · Polyp segmentation, average across domains A–D | Emaxϕ: 86.39 -> 89.29 | Helped |
| Use VPTTA instead of DUA | PraNet · Polyp segmentation, average across domains A–D | Sα: 83.58 -> 85.62 | Helped |
| Add a low-frequency prompt to the source-only model (ablation) | ResUNet-34 · OD/OC segmentation, average across domains A–E | DSC: 65.86 -> 66.90 | Helped |
| Add a memory bank to the low-frequency-prompt method (ablation) | ResUNet-34 · OD/OC segmentation, average across domains A–E | DSC: 66.90 -> 70.24 | Helped |
| Add warm-up to the low-frequency-prompt method (ablation) | ResUNet-34 · OD/OC segmentation, average across domains A–E | DSC: 66.90 -> 69.58 | Helped |
| Add warm-up to the low-frequency-prompt-plus-memory-bank method (ablation) | ResUNet-34 · OD/OC segmentation, average across domains A–E | DSC: 70.24 -> 71.93 | Helped |
| Replace the low-frequency prompt with a low-rank prompt (r=3) (ablation) | ResUNet-34 · OD/OC segmentation | DSC: 71.93 -> 69.01 | Hurt |
| Replace the low-frequency prompt with a low-rank prompt (r=3) (ablation) | PraNet · Polyp segmentation | DSC: 80.46 -> 78.25 | Hurt |
| Replace the low-frequency prompt with a low-rank prompt (r=3) (ablation) | PraNet · Polyp segmentation | Emaxϕ: 89.94 -> 88.27 | Hurt |
| Replace the low-frequency prompt with a low-rank prompt (r=3) (ablation) | PraNet · Polyp segmentation | Sα: 86.84 -> 84.20 | Hurt |
Surprising findings
- Entropy-minimization and other model-updating CTTA methods perform worse than the Source Only baseline on polyp segmentation, as the model confidently predicts empty masks under shift.
Limitations
- Relies on BN-based statistics alignment, so applicability to architectures without BN layers is unclear.
- Polyp experiments apply prompt loss only to encoder BN layers due to PraNet decoder complexity.
Experimental setup
Datasets: RIM-ONE-r3 REFUGE ORIGA REFUGE-Validation/Test Drishti-GS BKAI-IGH-NeoPolyp CVC-ClinicDB ETIS-LaribPolypDB Kvasir-Seg
Benchmarks: OD/OC segmentation (5 domains) polyp segmentation (4 domains)
Metrics: DSC E-max phi S-alpha
Baselines: Source Only TENT-continual CoTTA DLTTA DUA SAR DomainAdaptor
Builds on: TENT CoTTA DUA DLTTA PraNet
Closest papers by meaning
- Efficient Deformable Convolutional Prompt for Continual Test-Time Adaptation in Medical Image Segmentation (AAAI 2025 · similarity 0.91)
- Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation (AAAI 2025 · similarity 0.89)
- Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation (NeurIPS 2025 · similarity 0.86)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.86)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation (CVPR 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.