Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for 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 a multi-graph matching framework with learnable universe embeddings that encode morphological priors for test-time domain generalization in medical image segmentation, evaluated on retinal fundus and polyp segmentation.
Problem
Test-time adaptation for medical image segmentation under domain shift performs poorly because existing TTA methods ignore morphological priors of medical images, and pairwise graph matching cannot handle multiple domains consistently.
Approach
Learnable universe embeddings learned during multi-source training encode morphological priors; multi-graph matching via the universe guarantees cycle-consistency, and at test time the embeddings are frozen while the segmentation network is adapted via a matching loss on unlabeled target batches.
Key results
- Multi-source fundus segmentation: 88.46% average DSC vs 85.58% for second-best DeY-Net and 69.37% for No Adapt.
- Multi-source polyp segmentation: 87.32% average DSC, 2.52% above NC-TTT (84.80%).
- Single-source polyp generalization: 81.75% average DSC, 1.83% above NC-TTT (79.93%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TTDG multi-graph matching instead of DeY-Net | U-Net for retinal fundus segmentation · Retinal fundus segmentation, multi-source, average across Sites A–E | DSC (%): 85.58 -> 88.46 | Helped |
| Use TTDG multi-graph matching instead of NC-TTT | U-Net for polyp segmentation · Polyp segmentation, multi-source, average across Sites A–D | DSC (%): 84.80 -> 87.32 | Helped |
| Use TTDG multi-graph matching instead of NC-TTT | U-Net for polyp segmentation · Polyp segmentation, single-source, average across 12 transfers | DSC (%): 79.93 -> 81.75 | Helped |
| Add learned universe embeddings rather than using randomly initialized universe matching (ablation) | TTDG for retinal fundus segmentation · Retinal fundus segmentation, multi-source, average across Sites A–E | DSC (%): 80.47 -> 88.46 | Helped |
| Add morphological priors to universe embeddings (ablation) | TTDG for retinal fundus segmentation · Retinal fundus segmentation, multi-source, average across Sites A–E | DSC (%): 78.71 -> 88.46 | Helped |
| Replace pairwise matching with multi-matching (ablation) | TTDG for retinal fundus segmentation · Retinal fundus segmentation, Sites A–E | DSC (%): 84.63 -> 88.46 | Helped |
Surprising findings
- Using randomly initialized universe embeddings (w/o priors) performed worse than the no-universe variant only by about 1.76% DSC, showing the gain depends heavily on the pretrained universe embeddings.
Limitations
- Evaluated only on two medical segmentation benchmarks (fundus and polyp), both 2D.
- Universe size d must be chosen via the sampling step, with performance sensitive to very small or large d.
Experimental setup
Datasets: RIM-ONE REFUGE ORIGA REFUGE-Test Drishti-GS BKAI-IGH-NEOPolyp CVC-ClinicDB CVC-612 ETIS Kvasir
Benchmarks: retinal fundus OD/OC segmentation (5 sites) polyp segmentation (4 sites)
Metrics: DSC E-max phi S-alpha
Baselines: No Adapt TENT SAR TASD DLTTA DomainAdaptor VPTTA DeY-Net NC-TTT
Builds on: universe of nodes multi-graph matching VPTTA HiPPI Sinkhorn algorithm
Closest papers by meaning
- Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.82)
- Multi-view Evidential Learning-based Medical Image Segmentation (AAAI 2025 · similarity 0.81)
- Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation (AAAI 2025 · similarity 0.80)
- Generative Medical Segmentation (AAAI 2025 · similarity 0.80)
- Towards Optimizing Large-Scale Multi-Graph Matching in Bioimaging (CVPR 2025 · similarity 0.80)
- MedFlowSeg: Flow Matching for Medical Image Segmentation with Frequency-Aware Attention (NeurIPS 2026 · similarity 0.80)
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.