Source-linked AI summary
MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs
Alistair E. W. Johnson, Tom J. Pollard, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih-ying Deng, Yifan Peng, Zhiyong Lu, Roger G. Mark, Seth J. Berkowitz, Steven Horng
TL;DR
Automated chest-radiograph analysis needs larger, accessible datasets because interpretation requires specialized expertise and radiologist availability is uneven. This paper constructs MIMIC-CXR-JPG by processing de-identified DICOM studies into JPEG images and attaching report-derived labels, with standardized splits intended to support computer-vision research.
Problem
Limited data constrains automated chest-radiograph analysis despite the modality’s clinical importance and the need for radiology expertise.
Method
The paper constructs a de-identified JPEG dataset from DICOM chest radiographs, standardizes image processing and splits, and derives labels with NegBio and CheXpert.
Results
MIMIC-CXR-JPG provides a large publicly available collection of chest radiographs with labels for common pathologies and standardized benchmark-oriented organization.
Takeaways & Limitations
The dataset provides a processed reference for image labels and data splits to support research in medical computer vision.
Abstract
from arXiv · showhide
Chest radiography is an extremely powerful imaging modality, allowing for a detailed inspection of a patient's thorax, but requiring specialized training for proper interpretation. With the advent of high performance general purpose computer vision algorithms, the accurate automated analysis of chest radiographs is becoming increasingly of interest to researchers. However, a key challenge in the development of these techniques is the lack of sufficient data. Here we describe MIMIC-CXR-JPG v2.0.0, a large dataset of 377,110 chest x-rays associated with 227,827 imaging studies sourced from the Beth Israel Deaconess Medical Center between 2011 - 2016. Images are provided with 14 labels derived from two natural language processing tools applied to the corresponding free-text radiology reports. MIMIC-CXR-JPG is derived entirely from the MIMIC-CXR database, and aims to provide a convenient processed version of MIMIC-CXR, as well as to provide a standard reference for data splits and image labels. All images have been de-identified to protect patient privacy. The dataset is made freely available to facilitate and encourage a wide range of research in medical computer vision.
1 Introduction
MIMIC-CXR-JPG addresses limited data for automated chest-radiograph analysis by providing a standardized, JPEG-based dataset with consistent labels and de-identification.
- Chest radiographs are widely used to assess thoracic and cardiopulmonary conditions, but radiologist shortages and geographic disparities limit timely interpretation.These constraints are especially severe in resource-poor areas.
- Existing chest-radiograph datasets were substantially smaller or focused on narrower annotations, underscoring the need for more data.Examples include JSRT with 247 images, Open-I with 8,121 images, and ChestX-ray8 with 112,120 radiographs.
- MIMIC-CXR-JPG converts clinically stored DICOM images into JPEG to reduce format complexity and storage requirements for computer-vision researchers.DICOM preserves extensive metadata but can be difficult for researchers outside radiology to use, whereas JPEG is widely used and more compact.
- The dataset standardizes JPEG conversion and pixel normalization while supplying consistent labels derived from corresponding free-text radiology reports.The conversion includes a design choice from 16-bit raw images to 12-bit JPEG depth, and labels are generated with open-source labelers.
- All images are de-identified under HIPAA Safe Harbor requirements, with protected health information removed and randomly generated identifiers linking reports and patients.
2 Chest radiographs
The dataset pipeline extracts chest radiographs from hospital DICOM files, de-identifies them, and exports standardized JPEG images after normalization, inversion correction, and contrast enhancement.
- Chest radiographs were queried from the hospital PACS for emergency-department patients admitted between 2011 and 2016 and linked to radiology reports.Source images were stored in DICOM format and could contain burned-in annotations conveying orientation, anatomy, or capture time.
- Pixel values were normalized to [0, 255], converted to unsigned integers, and inverted when necessary so air appears white and the external background black.
- OpenCV histogram equalization was applied to enhance image contrast during JPEG preparation.
- Apart from de-identification and JPEG conversion, no additional filtering or image processing was performed.Figure 1 compares a DICOM image with its histogram-equalized JPEG representation.
3 Labeling of the reports
Radiology-report labels were generated with NegBio and CheXpert, rule-based tools that identify findings and classify their assertion status, with disagreements retained as a category.
- Reports use semi-structured findings and impression sections, but templates can drift and reporters can vary or override the structure.
- NegBio detects whether pre-tagged medical findings are negative or uncertain within a sentence.
- CheXpert extracts label mentions, classifies them as positive, uncertain, or negative using local context, and aggregates multiple mentions.The extractor captures alternate spellings, synonyms, and abbreviations.
- The dataset uses CheXpert mention patterns in NegBio, so its label categories overlap with CheXpert but not ChestX-ray14.A separate Disagreement category marks cases where the two tools produce different labels.
4 Training, validation, and test sets
The dataset organizes studies into training, validation, and test sets to support consistent evaluation, with a public validation subset and a withheld test set enriched for pathology coverage.
- The test set contains all studies from patients with at least one report labeled during manual review and is not publicly released.
- The validation set contains all associated studies for a random sample of 500 patients and is publicly available in a separate valid folder.
- All remaining studies are assigned to the training set.
- Table 3 summarizes the images across the training, validation, and test splits.
5 Validation of labels
Labels were validated against 687 manually reviewed reports, evaluating mention extraction, uncertainty detection, and negation detection. The tables distinguish tasks and identify categories where a metric could not be calculated.
- 687 reports were manually labeled by a board-certified radiologist using the 14 CheXpert categories.
- The validation evaluated mention extraction, negation detection, and uncertainty detection as separate binary classification tasks.
- Table 4 evaluates whether findings were mentioned, regardless of whether the report expressed certainty or negation.
- Uncertainty was not evaluated for No Finding, Lung Opacity, Pleural Other, or Support Devices because no uncertain cases were labeled.
- Negation was not evaluated for No Finding because no negated cases were labeled.
6 Data availability
MIMIC-CXR-JPG is freely available through PhysioNet, subject to a data use agreement and training requirements inherited from its sensitive source dataset. The generation code is also publicly available.
- The dataset is freely available to researchers on PhysioNet after signing a data use agreement.
- The agreement prohibits data sharing and reidentification attempts and requires users to release relevant code with publications.
- MIMIC-CXR-JPG is wholly derived from MIMIC-CXR, whose access requires completing human-subjects research training.
- The code used to generate MIMIC-CXR-JPG and the paper's summaries is publicly available.
7 Conclusions
MIMIC-CXR-JPG is a large, publicly available chest-radiograph dataset from more than 220,000 BIDMC studies. It includes labels for common pathologies and is intended to support medically relevant computer-vision benchmarks.
- MIMIC-CXR-JPG contains chest radiographs from over 220,000 studies performed at BIDMC.
- The dataset provides labels for a number of common pathologies.
- The authors position MIMIC-CXR-JPG as a benchmark for medically relevant computer-vision tasks.