[Paper Review] MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs

Alistair E. W. Johnson, Tom Pollard+8 more|arXiv (Cornell University)|Jan 21, 2019

COVID-19 diagnosis using AI256 citations

TL;DR

This paper presents MIMIC-CXR-JPG v2.0.0, a large, de-identified collection of 377,110 JPEG chest radiographs with 14 pathology labels derived from radiology reports, released to benchmark medical computer vision tasks.

ABSTRACT

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.

Motivation & Objective

• Provide a processed, JPEG-formatted subset of MIMIC-CXR to facilitate computer vision research on chest radiographs.
• Offer standardized data splits and labels to enable fair benchmarking across studies.
• Ensure de-identification and HIPAA safety while preserving clinically relevant image content.

Proposed method

• Extract DICOM images from MIMIC-CXR and convert to JPEG with 12-bit to 8-bit depth normalization.
• Apply histogram equalization and orientation normalization to improve contrast.
• De-identify PHI via automated text-annotation detection and manual review for privacy.
• Generate labels from radiology reports using two open-source labelers (NegBio and CheXpert).
• Provide training, validation, and test splits with CXR images and study counts; release validation set publicly.

Experimental results

Research questions

• RQ1 Can a large, labeled JPEG chest X-ray dataset facilitate robust benchmarking for computer vision methods in radiology?
• RQ2 Do standardized splits and labeling approaches enable reproducible evaluation across models and studies?
• RQ3 How well do automated labelers (NegBio, CheXpert) perform against manual radiologist labeling in this dataset?
• RQ4 What is the impact of de-identification and image preprocessing on dataset usability for research?

Key findings

• The dataset contains 377,110 chest X-rays from 227,827 imaging studies collected 2011–2016 at BIDMC.
• Images are de-identified and converted to JPEG with standardized preprocessing and 14 labels derived from radiology reports.
• Labels were produced using NegBio and CheXpert and include a Disagreement category for label conflicts.
• Training, validation, and test splits are provided, with the test set withheld publicly for evaluation integrity.
• A validation study on 687 manually labeled reports shows variable performance across labelers and tasks, with certain labels achieving high F1 scores (e.g., Pneumonia, Atelectasis, Pleural Effusion) and others lower.