CheXpert Plus: Augmenting a Large Chest X-ray Dataset with Text Radiology Reports, Patient Demographics and Additional Image Formats

Largest public radiology text dataset released: CheXpert Plus contains 36 million text tokens (including 13 million impression tokens), making it the largest text dataset publicly available in radiology and enabling new research on vision-language models that combine images with written reports.

Major de-identification and privacy effort: The dataset represents "the largest text de-identification effort in radiology," with almost 1 million instances of protected health information (PHI) anonymized, addressing critical privacy concerns for responsible AI development in healthcare.

First large-scale cross-institution training opportunity: As only the second large-scale paired image-text dataset in radiology, CheXpert Plus enables researchers to train AI models across multiple institutions for the first time, potentially improving model robustness and generalization to diverse patient populations.

Supports AI fairness research through demographic data: The dataset includes patient demographics and covers "various clinical and socio-economic groups," directly addressing growing demands from AI fairness researchers to develop models that perform equitably across different populations rather than perpetuating healthcare disparities.

Comprehensive metadata for responsible model development: Beyond images and reports, the dataset includes numerous clinical metadata, pathology labels, and RadGraph annotations, providing researchers with the detailed information needed to audit models for bias, understand failure modes, and build more trustworthy AI systems for clinical use.