MIMIC-CXR

MIMIC-CXR provides a large, de-identified dataset of chest radiographs paired with semi-structured radiology reports for training and evaluating computer vision, natural language processing, and clinical data mining methods.

Key Features:

• Dataset composition: 227,835 imaging studies from 65,379 patients collected at Beth Israel Deaconess Medical Center Emergency Department between 2011 and 2016, comprising 377,110 de-identified images.
• Image views: Studies may include one or more images, typically frontal and lateral views.
• Radiology reports: Semi-structured free-text reports authored contemporaneously by practicing radiologists are paired with each imaging study.
• Labels: Includes labeled chest radiographs suitable for supervised learning and algorithm evaluation.
• De-identification: Images and reports are de-identified to remove protected health information while preserving diagnostic content.

Scientific Applications:

• Computer vision: Development and evaluation of automated chest radiograph interpretation algorithms and machine learning models.
• Natural language processing: Extraction of structured information from semi-structured free-text radiology reports.
• Clinical data mining: Investigation of imaging-derived and report-derived features for clinical research and phenotype extraction.

Methodology:

Training and evaluation of supervised machine learning and computer vision models on labeled radiographs; application of natural language processing to semi-structured radiology reports; de-identification of images and reports.