Unsupervised Domain Adaptation in Medical Imaging: A Pipeline Evaluation
A recent study published on arXiv (2608.12035) investigates the entire unsupervised domain adaptation (UDA) process in medical imaging, focusing on the issue of model selection in the absence of target domain labels. The research examines eleven clinically significant cross-domain scenarios across nine medical imaging datasets, evaluating ten UDA algorithms alongside thirteen label-free selection methods (validators), leading to an analysis of over 80,000 trained models. Results indicate that while there are often capable adapted models, pinpointing them without target labels is challenging; validator-selected models show a considerable and systematic performance gap compared to the best model available, with no validator proving consistently dependable. The study proposes two strategies to mitigate this gap, although specifics are not included in the abstract. This research is crucial for the clinical application of UDA, emphasizing the need to consider adaptation and selection concurrently. The paper can be accessed on arXiv under the identifier 2608.12035.
Key facts
- Study evaluates the complete UDA pipeline in medical imaging.
- Covers 11 clinically relevant cross-domain scenarios from 9 medical imaging datasets.
- Tests 10 UDA algorithms and 13 label-free selection methods (validators).
- Evaluates over 80,000 trained models.
- Finds that capable adapted models usually exist but are hard to identify without target labels.
- Validator-selected models show a large and structural target performance gap.
- No evaluated validator is consistently reliable.
- Paper available on arXiv with ID 2608.12035.
Entities
Institutions
- arXiv