A capable adapted model usually exists. The hard part is selecting it without target labels — and that selection gap is what keeps UDA from clinical use.
Research Overview
Deploying unsupervised domain adaptation (UDA) in clinical practice requires choosing which algorithm to use and which of its trained models to ship. However, the deployment (target) domain is unlabeled, so models cannot be evaluated directly on it, leaving it unclear which to select. We address this by evaluating the complete UDA pipeline, considering both adaptation and label-free selection together. Our study covers eleven clinically relevant cross-domain scenarios from nine medical imaging datasets, with ten UDA algorithms and 13 label-free selection methods (validators), evaluating over 80,000 trained models in total. We find that a capable adapted model usually exists, but identifying it without target labels is difficult: the validator-selected models leave a large and structural target performance gap to the best available one, with no evaluated validator consistently reliable. Towards closing it, we explore ensembling and a small target-labeling budget; both narrow this gap but do not close it entirely. Overall, deployable UDA depends on the complete pipeline; addressing the less explored selection step could bring much of current UDA closer to clinical use.
Brain MRI, chest X-ray, and retinal imaging · nine datasets · balanced accuracy on the target validation set.
What's the problem
Under Oracle selection (using target labels) adaptation is often effective. The models chosen by label-free validators, however, leave a large gap to the best available one.
Target accuracy (%, mean ± std) of the checkpoint a validator selects for an algorithm. Avg. is the mean over per-algorithm columns, and Across-Algo pools the checkpoints of all algorithms and selects across them. Oracle selects using target labels, SourceOnly is trained on source data only, and TargetOnly on labeled target data. Hover a cell to see the standard deviation.
Full results for all eleven scenarios, including median with 95% confidence intervals, are reported in the paper.
Where the gap comes from
Selection is a ranking problem, so we measure validator reliability by the Spearman rank correlation ρ between validation scores and true target accuracy. Pooled across algorithms, most validators are weak or even reversed.
Towards closing the gap
In the absence of a consistently reliable validator, both strategies recover much of the achievable performance — but neither closes the gap to the Oracle entirely.
| Scenario | Ens. | B1 | B2 | B3 | B4 | B5 | Best Val. | Oracle |
|---|---|---|---|---|---|---|---|---|
| ADNI-1 → ADNI-2 | 89.3 | 78.6 | 83.0 | 84.0 | 86.5 | 87.5 | 85.1 | 93.2 |
| ADNI-1 → ADNI-3 | 87.3 | 78.7 | 80.8 | 81.1 | 81.2 | 83.3 | 84.5 | 94.2 |
| ADNI-2 → ADNI-1 | 83.2 | 80.4 | 79.1 | 81.5 | 82.6 | 84.0 | 84.4 | 89.5 |
| ADNI-2 → ADNI-3 | 87.2 | 79.5 | 79.9 | 83.5 | 81.3 | 83.6 | 84.2 | 92.1 |
| ADNI-1+2 → AIBL | 89.4 | 81.6 | 82.9 | 82.9 | 85.1 | 86.0 | 89.0 | 93.8 |
| RSNA → Child CXR | 79.8 | 76.6 | 78.4 | 81.5 | 83.1 | 85.0 | 79.1 | 89.7 |
| Child CXR → RSNA | 74.4 | 74.3 | 75.8 | 76.3 | 76.9 | 76.9 | 75.3 | 78.3 |
| LDD → CRD | 81.3 | 80.2 | 81.6 | 82.0 | 82.2 | 82.1 | 81.9 | 83.6 |
| CRD → LDD | 56.9 | 75.9 | 78.1 | 80.2 | 79.7 | 80.4 | 73.1 | 82.7 |
| SLO → OCT | 63.1 | 58.9 | 61.2 | 60.8 | 61.3 | 62.6 | 60.1 | 65.6 |
| OCT → SLO | 66.4 | 60.0 | 60.9 | 63.5 | 64.5 | 64.4 | 62.7 | 67.4 |
| Mean | 78.0 | 75.0 | 76.5 | 77.9 | 78.6 | 79.6 | 78.1 | 84.5 |
Target accuracy (%). Ens. ensembles the checkpoints selected by all validators across all algorithms. Small Labeling Budget selects the best checkpoint on a small labeled target subset, at five levels (B1–B5; 5–25 samples for brain MRI, 1–5% otherwise). On average, ensembling matches Best Val. and the labeling budget surpasses both from around B4, though a gap to the Oracle remains for both.
Takeaways
Under Oracle selection, the across-algorithm pool exceeds SourceOnly in every scenario and approaches TargetOnly (mean 84.5% vs. 87.5%). Adaptation itself is often effective.
The best validator leaves a gap to the Oracle of up to 10.6 points, averaging 6.4. Because the best validator is itself chosen using target labels, the gap at deployment can only be larger.
No evaluated validator is consistently reliable. Which validator works best shifts by algorithm and scenario, and once checkpoints from different algorithms are pooled, most correlate only weakly or in reverse with target performance.
Ensembling across algorithms and validators matches the best validator on average, and a small target-labeling budget surpasses it from around 4% labeled. Both recover much of the achievable performance but remain below the Oracle, each with practical costs.
Citation
If you find this work useful, please cite the arXiv preprint. Final WACV 2027 proceedings metadata such as page numbers and DOI are not yet assigned.
@article{xiong2026far,
title={How far from clinical deployment? evaluating the complete unsupervised domain adaptation pipeline in medical imaging},
author={Xiong, Yiheng and Gall{\'e}e, Luisa and Wolf, Daniel Santak and Hillenhagen, Heiko and G{\"o}tz, Michael},
journal={arXiv preprint arXiv:2608.12035},
year={2026}
}Click outside or press Esc to close