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

Improving Fairness of Automated Chest Radiograph Diagnosis by Contrastive Learning

https://doi.org/10.1148/ryai.230342

1.0

ude.llenroc.dem@2004piy

Supported by the National Library of Medicine (4R00LM013001), National Science Foundation CAREER (2145640), Intramural Research Program of the NIH, Amazon Research Award. MIDRC funded by NIBIB of the NIH under contract 75N920202D00021 and through ARPA-H.

Retrospective study with IRB approval at each clinical center and Weill Cornell Medicine; informed consent waived due to publicly available datasets (MIDRC and NIH ChestX-ray14).

DenseNet-121 backbone (pretrained on CheXpert) with supervised contrastive learning (SCL) pretraining using a single-layer perceptron contrastive head (128-dim), followed by fine-tuning with a prediction head for binary classification.

Code available at https://github.com/bionlplab/CXRFairness

Mitigates algorithmic bias across sex, race, and age subgroups in automated chest radiograph diagnosis (COVID-19 and thorax abnormalities) while maintaining overall performance.

Clinical decision support research; method intended to improve fairness of image-based diagnosis and may be suitable for clinical practice per authors.

Automated image classification model development and inference; provides decision support outputs for diagnosis tasks.

Research method for automated diagnosis of COVID-19 and thorax abnormalities from frontal chest radiographs, with explicit focus on reducing subgroup bias (sex, race, age). Intended for use on datasets similar to MIDRC and ChestX-ray14.

Frontal chest radiographs (PA/AP) preprocessed from DICOM to JPEG; demographic subgroup labels (sex, race, age) are used during SCL pretraining to form positive/negative pairs; downstream prediction uses images.

Pretrain with SCL: replace classifier with 128-dim contrastive head; define positives as same disease label from different subgroups and negatives as different label from same subgroup; temperature=0.05; Adam lr=1e-4; 10 epochs. Then replace with prediction head and fine-tune entire network for 1 epoch using binary cross-entropy. Image preprocessing: DICOM normalization [0,255], invert if needed (air white), histogram equalization, resize to 256×256, center crop 224×224, random rotation (0–10°) and random flipping. Evaluate fairness using ΔmAUC and other subgroup metrics.

Calibration was not assessed; potential over/underconfidence not evaluated. Method addresses single sensitive attributes individually rather than multiple simultaneously; limited by small sample sizes in some intersectional groups. Datasets may not fully represent all patient populations; comorbidity history not available. Only DenseNet-121 backbone was used; other architectures not evaluated here. External validation limited; generalization to other modalities/diseases not established.

PyTorch implementation with public code; pretrained on CheXpert; explicit preprocessing and training hyperparameters provided; experiments run on Intel Core i9-9960X CPU and NVIDIA Quadro RTX 6000 GPU; 200 bootstrap samples used for CIs; official splits (ChestX-ray14) and patient-level split (MIDRC).