Source-linked AI summary
SymmAdapt: Symmetrical Flow Matching for Source-Free Domain Adaptation in Medical Image Segmentation
Tal Grossman, Noa Cahan, Hayit Greenspan
TL;DR
Medical segmentation must handle modality and site shifts, but source-free adaptation is constrained to a pretrained model and unlabeled target data. SymmAdapt uses Symmetrical Flow Matching to segment targets, synthesize source-like replay images, and jointly fine-tune on target and replay data. Across abdominal, cardiac, and prostate benchmarks, it outperforms SFUDA baselines and remains competitive with non-source-free methods.
Problem
Domain shifts hinder medical segmentation deployment, while many UDA methods require source data that cannot be shared during adaptation.
Method
SymmAdapt trains a bidirectional SymmFlow on labeled source data, then uses Gaussian-origin inference, pseudo-label filtering, source-like replay, and dual-stream fine-tuning on unlabeled target data.
Results
SymmAdapt outperforms SFUDA baselines across abdominal, cardiac, and prostate benchmarks and remains competitive with non-source-free methods.
Takeaways & Limitations
Gaussian-origin inference and synthetic replay preserve anatomical and source semantic priors during adaptation without requiring source-data access.
Abstract
from arXiv · showhide
Domain shift across imaging modalities and acquisition sites remains a significant barrier to the clinical deployment of segmentation models. Source-free unsupervised domain adaptation (SFUDA) addresses this by adapting a pretrained model to an unlabeled target domain without requiring access to sensitive source data. We introduce a novel SFUDA framework built on Symmetrical Flow Matching, a unified generative model that segments an input image and synthesizes a source-like image from a mask within the same learned flow. By initializing inference from a domain-agnostic Gaussian origin, the model preserves structural consistency across domains and grounds predictions in learned anatomy rather than shifted texture statistics. Our pipeline leverages this symmetry to generate reliable pseudo-labels and corresponding source-like synthetic images from unlabeled target data, creating a generative replay buffer that anchors source knowledge during a generative self-training stage that fine-tunes on a joint set of real target and synthetic source-like images. We evaluate on abdominal multi-organ and cardiac segmentation, covering cross-modality MRI<->CT shifts, and multi-site prostate segmentation. Our approach outperforms SFUDA baselines and is competitive with conventional UDA methods.
1 Introduction
Medical segmentation models face scanner, protocol, and modality shifts, while conventional UDA often requires inaccessible source data. SymmAdapt addresses the source-free setting with a generative bidirectional flow that uses Gaussian-origin inference and source-like replay.
- Domain shift across scanners, protocols, and modalities remains a barrier to clinical deployment of medical segmentation models.
- SFUDA adapts a pretrained source model using only unlabeled target images, avoiding source-data access during adaptation.
- Existing SFUDA methods primarily use discriminative architectures and do not exploit the generative structure of segmentation.
- Conditional flow-based segmenters can diverge when initialized from target images outside the source manifold.
- SymmFlow instead models joint image-mask distributions from a fixed Gaussian origin, encouraging anatomy-based inference despite modality shifts.
- SymmAdapt synthesizes source-like replay data through its bidirectional flow and reports strong performance across abdominal, cardiac, and prostate benchmarks.
2 Method
SymmAdapt trains a bidirectional image-mask flow on labeled source data, then adapts without source access by filtering generated pseudo-label pairs and jointly fine-tuning on target and replay data. Its design uses Gaussian-origin inference and synthetic replay to preserve source semantics during adaptation.
- SFUDA starts with a source-trained SymmFlow model and adapts it to unlabeled target images without accessing the source domain.
- SymmFlow learns opposite-direction image and mask transport paths with a shared network that predicts their joint velocity.
- Segmentation integrates the mask ODE backward from Gaussian noise while conditioning on the input image.
- The pipeline uses target predictions to generate source-like images, retains cycle-consistent pairs, and fine-tunes with real target and synthetic replay streams.
- Synthetic replay rehearses the source manifold and provides anatomical anchoring against degenerate solutions under large domain shifts.
3 Experiments
SymmAdapt is evaluated across abdominal, cardiac, and prostate domain-shift benchmarks using standardized splits, comparisons with UDA/SFUDA baselines, and ablations of its adaptation design. It achieves strong segmentation results, while efficiency, pseudo-label retention, and stochastic stability analyses characterize practical behavior.
- Datasets and setup: SymmAdapt is evaluated on abdominal, cardiac, and prostate cross-modality or cross-site segmentation benchmarks.The evaluation includes abdominal MRI↔CT, cardiac MRI→CT, and prostate RUNMC→QUBIQ/BMC tasks.
- Benchmark results: 87.2% and 88.0% average Dice on abdominal MRI→CT and CT→MRI outperform the best SFUDA baseline by +2.8 and +1.7 points, respectively.The method remains competitive with non-source-free methods on these tasks.
- Benchmark results: 87.2% Dice on cardiac MRI→CT surpasses ProtoContra at 84.5% and non-source-free C3R at 86.1%.The comparison uses the MMWHS cardiac dataset.
- Benchmark results: SymmAdapt is best on both prostate RUNMC→QUBIQ and RUNMC→BMC, exceeding DFG and AdaMI on Dice and ASSD.The prostate evaluation is reported as a cross-site MRI task.
- Efficiency and stability: Round-trip filtering retains 1142 of 1356 slices (84.2%) with mean round-trip Dice 0.903, while 5 Euler steps provide 87.2 Dice and best ASSD of 2.07 mm.The filtering threshold is reported as insensitive across tested ODE-step and threshold settings.
4 Conclusion
The paper presents SymmAdapt as a Symmetrical Flow Matching framework for source-free adaptation under imaging domain shifts. Across abdominal, cardiac, and prostate benchmarks, it reports strong SFUDA performance while remaining competitive with non-source-free methods.
- Conclusion: SymmAdapt uses Gaussian-origin inference, synthetic source-like replay, and round-trip filtering to adapt on unlabeled target images without source data.The framework is powered by Symmetrical Flow Matching.
- Conclusion: Across abdominal, cardiac, and prostate benchmarks, SymmAdapt achieves strong SFUDA performance and remains competitive with non-source-free methods.The conclusion frames the approach for clinic-focused segmentation when source data cannot be shared and domain shifts hinder deployment.