ChestX-ray14, originally released as ChestX-ray8, is a large public dataset of chest radiographs introduced by Xiaosong Wang, Ronald Summers, and colleagues at the US National Institutes of Health in a CVPR 2017 paper. It gave medical-imaging researchers something they had long lacked: a hospital-scale collection big enough to train modern deep networks.
The dataset contains 108,948 frontal-view X-ray images from 32,717 unique patients. Because manually labeling that many images by experts is impractical, the team text-mined disease labels from the associated free-text radiology reports using natural language processing, initially covering eight conditions and later expanded to fourteen, which is why the dataset is commonly called ChestX-ray14. The labels were estimated to be over 90 percent accurate and intended for weakly supervised learning.
The release became a default benchmark for thoracic-disease classification and localization, and it was the training data for the well-known CheXNet model. It also became a case study in dataset limitations: NLP-mined labels carry noise, the same patients appear in many images, and naive train/test splits can leak information, all of which later work had to correct for.
For a general reader, ChestX-ray14 illustrates a recurring pattern in clinical AI: progress often depends less on a clever model than on someone assembling and labeling a large enough dataset, and the quality of those labels quietly shapes everything built on top.