Back

Pushing a Frozen CXR Foundation Model: A LoRA Partial-Fine-Tuning Study on NIH ChestX-ray14 with a Model-Conditional Label-Flip Sensitivity Analysis

BAI, T.-C.; YEH, S.-C.

2026-08-11 radiology and imaging
10.64898/2026.08.08.26360020 medRxiv
Show abstract

Foundation models for chest X-ray interpretation make it possible to adapt specialised visual representations with relatively small trainable modules. We report a retrospective study of Low-Rank Adaptation (LoRA) of Rad-DINO Vision Transformer Base with 14x14 patches (ViT-B/14) for 14-class multi-label classification on the National Institutes of Health (NIH) ChestX-ray14 dataset. The official test labels were accessed during earlier model development and configuration comparisons; consequently, every official-test result in this manuscript is explicitly descriptive and non-confirmatory. We used a patient-disjoint 90/10 split of the official trainval pool (77,988 training and 8,536 validation images) and retained the released 25,596-image test partition. The historically selected all-linear LoRA configuration with safe augmentation and g=37 produced a descriptive test macro AUROC of 0.8462 versus the frozen baseline of 0.8295. Comparisons of target modules, patch-token grids, and a Rad-DINO-specific local query head are reported as retrospective comparisons rather than unbiased model-selection evidence. A confident-learning diagnostic flagged 17,653 of 86,524 trainval images (20.4%); this is a model-based flag rate, not a ground-truth label-error rate. A separate counterfactual relabeling sensitivity analysis, which uses the same model to identify and rescore disagreements, changed the descriptive AUROC to approximately 0.9445 after 6,509 policy-defined flips. This value is not achieved model performance and is not a radiologist-audited label-quality ceiling. We provide a validation-only threshold and artifact protocol for future locked evaluation, but a genuinely untouched holdout and new locked selection are required for a confirmatory headline. The existing Zenodo record contains the 25 publication figures only.

Matching journals

The top 8 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.