Chest X-Ray Classification

Chest X-ray classification asks models to identify one or more abnormalities, such as pneumonia, effusion, or nodules, from radiographs, often to support screening and triage. Common approaches include CNNs and vision transformers trained on labeled images, report-supervised or contrastive vision-language pretraining, and methods that use anatomical regions, disease hierarchies, or co-occurrence; weak supervision and long-tailed learning address scarce or imbalanced findings. Models are evaluated on datasets such as NIH ChestX-ray14, CheXpert, MIMIC-CXR, PadChest, and VinDr-CXR, increasingly with tests of external-hospital transfer, rare-label recognition, zero-shot performance, localization, calibration, and subgroup outcomes. Strong benchmark scores do not guarantee reliable clinical performance: results can shift with institutions, populations, label definitions, and thresholds, while shortcuts and noisy report-derived labels complicate whether predictions reflect image evidence.