MIMIC-CXR Chest Radiographs and Radiology Report Impressions

CXR-PRO contains a subset of chest radiographs from the MIMIC-CXR database saved in HDF5 format, accompanied by a tabular compilation of radiology report impressions. The dataset includes training and test CSV files with fields such as dicom_id, study_id, and subject_id, prepared by the rajpurkarlab and last updated in November 2022.

Use Cases

• Train a multi-modal model by aligning chest radiographs from the HDF5 file with their corresponding text impressions in the CSV to generate or classify radiology findings.
• Link patient imaging studies across the healthcare system using the subject_id, study_id, and dicom_id fields for longitudinal analysis.
• Fine-tune a vision transformer on the HDF5-stored chest radiographs using the impressions text as weak supervision for pathology detection.

Strengths

• Derived from the established MIMIC-CXR database, a known benchmark in medical imaging.
• Provides structured linkage between images and report text via identifiers like dicom_id and subject_id.
• Includes separate training and test splits (mimic_train_impressions.csv, mimic_test_impressions.csv) for evaluation.

Limitations

• Core metadata like row count, column count, image dimensions, and dataset size are unspecified.
• The impressions text has had references to prior reports removed, which may limit contextual clinical understanding.
• Temporal coverage, geographic origin, and license terms are not provided.

Provenance

Source

Subset of MIMIC-CXR chest radiographs.

Collection Method

Radiographs saved in HDF5 format; report impressions compiled into CSV with prior references removed.

Time Range

null

Freshness

2022-11-14

Geography

null