https://atlas.rsna.org/schemas/2025-11/model.json

Chest Radiography Foundation Model (as studied for bias)

https://github.com/biomedia-mira/cxr-foundation-bias

1.0

B.G.: ERC grant 757173 (Project MIRA), Innovate UK/UKRI (UKRI London Medical Imaging & AI Centre for Value Based Healthcare), Royal Academy of Engineering; employment/stock options disclosures as stated. C.J.: Microsoft Research–EPSRC iCASE PhD Scholarship. M.R.: Imperial College London President's PhD Scholarship. S.W.: Innovate UK/UKRI (UKRI London Medical Imaging & AI Centre for Value Based Healthcare).

Retrospective study using publicly available secondary data; HIPAA-compliant; exempt from ethical approval.

Proprietary chest radiography foundation model used as a fixed feature extractor (pretrained on natural images, then >800,000 chest radiographs using supervised contrastive learning). Downstream classifiers tested: single linear layer (CXR-linear) and multilayer perceptrons with 3 or 5 hidden layers (CXR-MLP-3, CXR-MLP-5). Baseline comparator: DenseNet-121 fine-tuned on CheXpert.

Access to the foundation model only via a programming interface that outputs features; network weights not publicly available.

Not established; study concludes the investigated foundation model may be unsafe for clinical applications due to subgroup performance disparities.

Research evaluation/auditing of AI models for bias and subgroup performance.

Thresholds selected per model to achieve FPR = 0.20 on the whole sample; Youden J = TPR − FPR reported at this operating point.

Feature extraction automated; downstream decision models trained by researchers; not autonomous clinical decision support.

Investigated as a feature extractor for downstream chest radiograph disease detection (labels: no finding, pleural effusion, cardiomegaly, pneumothorax) on adult patients in the CheXpert dataset; not intended for clinical use in this study.

Posterior-anterior and other view chest radiographs from the CheXpert dataset; features produced by the foundation model’s backbone.

Backbone kept frozen; extract features via provided API and train a downstream classifier (linear or MLP) on CheXpert training set with validation for model selection; evaluate on resampled balanced CheXpert test sets with subgroup analyses.

Only one proprietary foundation model analyzed; backbone weights unavailable (cannot update feature extractor or apply debiasing that requires unfreezing); training data for the foundation model primarily from India and the US (>800k CXRs, >700k from India), potentially contributing to observed biases; analysis limited to selected labels; cannot determine exact origin of biases due to limited insight into pretraining data/process.

Comprehensive bias and subgroup performance analysis should be performed before any clinical use; the studied foundation model may be unsafe for clinical applications as it could amplify health disparities.

All code and instructions to recreate data splits and analyses are available under Apache 2.0 at https://github.com/biomedia-mira/cxr-foundation-bias.