Research map / Label-efficient and robust learning / Domain adaptation and generalization
Learning Generalized Medical Image Segmentation from Decoupled Feature Queries
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 Decoupled Feature Query (DFQ), using channel-decoupled deep features as cross-attention queries plus a relaxed deep whitening loss on a MiT-B3 Transformer, for domain generalized medical image segmentation on fundus and prostate benchmarks.
Problem
Domain generalized medical image segmentation must generalize to unseen target domains with scanner/hospital shift, where features show channel redundancy and cross-domain intra-channel misalignment.
Approach
Use channel-wise decoupled deep features as queries in cross-attention over shallow-feature keys/values, so deep queries constrain shallow representations to be consistent across domains, combined with a relaxed deep whitening transformation (normalized input, upper-triangular off-diagonal loss) to reduce channel redundancy.
Key results
- On the Prostate benchmark, DFQ reaches average DSC 89.92% and ASD 1.07%, outperforming the runner-up by 1.31% DSC (as stated in the abstract).
- On the Fundus optic cup benchmark, DFQ achieves average DSC 90.57% and ASD 9.52, versus 88.94% DSC for RAM-DSIR (second best).
- Ablation on Fundus optic cup: feature queries add 0.94% DSC and the relaxed deep whitening transformation adds a further 1.16% DSC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DFQ instead of the strongest reported baseline | MiT-B3 segmentation model · Prostate benchmark, average | DSC: 88.61 -> 89.92 | Helped |
| Use DFQ instead of the strongest reported baseline | MiT-B3 segmentation model · Fundus optic cup, average | DSC: 88.94 -> 90.57 | Helped |
| Remove feature queries from the backbone-plus-query baseline (ablation) | DFQ on Fundus optic cup · Domain-1 | DSC: 84.92 -> 86.16 | Hurt |
| Add feature queries (ablation) | Fundus optic cup segmentation backbone · Domain-1 | DSC: 84.92 -> 86.16 | Helped |
| Add relaxed deep whitening transformation (ablation) | Fundus optic cup model with feature queries · Domain-1 | DSC: 86.16 -> 87.30 | Helped |
| Add style-decoupled queries from F2 (ablation) | DFQ on Prostate · Domain-1 | DSC: 85.87 -> 87.41 | Helped |
| Add style-decoupled queries from F3 (ablation) | DFQ on Prostate · Domain-1 | DSC: 87.41 -> 87.82 | Helped |
| Add style-decoupled queries from F4 (ablation) | DFQ on Prostate · Domain-1 | DSC: 87.82 -> 88.28 | Helped |
Surprising findings
- Conventional deep whitening transformation still leaves channel correlation, since the diagonal constraint lets off-diagonal correlation be reduced by shrinking feature magnitudes.
Limitations
- Evaluated only on two medical benchmarks (fundus and prostate).
- Gains on some individual domains are modest or absent (e.g., optic disc Domain-1 DSC lower than baselines in Table 3 and prostate Domain-1 DSC lower than DCAC).
Experimental setup
Datasets: Drishti-GS RIM-ONE-r3 REFUGE (train) REFUGE (val) DG Prostate benchmark
Benchmarks: DG Fundus benchmark (optic cup and disc) DG Prostate benchmark
Metrics: DSC ASD
Baselines: BigAug SAML FedDG DoFE RAM-DSIR DCAC DeepAll Intra-domain
Builds on: deep whitening transformation SegFormer DoFE RAM-DSIR
Closest papers by meaning
- LangDAug: Langevin Data Augmentation for Multi-Source Domain Generalization in Medical Image Segmentation (ICML 2025 · similarity 0.88)
- Domain Generalized Medical Landmark Detection via Robust Boundary-Aware Pre-Training (AAAI 2025 · similarity 0.87)
- Generative Medical Segmentation (AAAI 2025 · similarity 0.87)
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.87)
- Modality-agnostic Domain Generalizable Medical Image Segmentation by Multi-Frequency in Multi-Scale Attention (CVPR 2024 · similarity 0.86)
- A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · 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.