Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/dataset.json

Name

MIMIC-CXR anatomy-specific progression labels for chest radiographs (Radiol Artif Intell 2024)

Link

https://dx.doi.org/10.1148/ryai.230277

Indexing

Keywords: disease progression, weakly supervised learning, RadGraph, anatomy-specific labels, MIMIC-CXR, chest radiograph, improved, unchanged, worsened, new finding, localization, saliency maps

Content: CH, RS

RadLex: RID35976, RID29041, RID49837, RID11514, RID39271, RID39105

SNOMED: 278516003, 46621007, 19242006, 36118008, 60046008

Author(s)

Ke Yu

Shantanu Ghosh

Zhexiong Liu

Christopher Deible

Clare B. Poynton

Kayhan Batmanghelich

Organization(s)

University of Pittsburgh

Boston University

University of Pittsburgh Medical Center

Chobanian & Avedisian School of Medicine, Boston University

Funding

Pennsylvania Department of Health (4100087331); National Institutes of Health (R01HL141813); National Science Foundation (1839332; OAC-1928147 for Bridges-2).

Ethical review

Institutional review board approval and patient informed consent were not required (public de-identified dataset).

Comments

Anatomy-specific disease progression labels (improved, unchanged, worsened, new) were automatically extracted from MIMIC-CXR radiology reports using RadGraph relations and mapped with RadLex descriptors.

Date

Published: 2024-07-24

References

[1] Yu K, Ghosh S, Liu Z, Deible C, Poynton CB, Batmanghelich K. "Anatomy-specific Progression Classification in Chest Radiographs via Weakly Supervised Learning". Radiology: Artificial Intelligence. 2024-07-24. doi:10.1148/ryai.230277. PMID: 39046325. PMCID: PMC11427915.

[2] Johnson AEW, Pollard TJ, Berkowitz SJ, et al.. "MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports". Scientific Data. 2019. PMID: 31831740. PMCID: PMC6908718. Available from: https://physionet.org/content/mimic-cxr/

Dataset

Motivation

Enable classification of interval changes (improved, unchanged, worsened, new) for specific observations at specific anatomic landmarks on chest radiographs using weak labels from reports.

Sampling

Restricted to frontal radiographs; included six observations; excluded image pairs with interval time >1 year.

Partitioning scheme

Random 70:10:20 split by patient for pretraining; same split used to select consecutive pairs for fine-tuning; pipelines replicated across five different random seeds.

Missing information

No public release link for the derived progression labels was provided; split-wise counts not reported.

Relationships between instances

Progression labels are defined for pairs of consecutive studies within one year for the same patient; a single image may have multiple triplets at different anatomic landmarks.

Noise

Weak labels derived from NLP may contain errors; manual spot-checking of 100 triplets per observation showed high precision (92%–100%) depending on observation.

External data

Bounding box annotations from Chest ImaGenome were used as reference for localization evaluation; RadGraph was used to extract entities and relations from reports; RadLex descriptors were used to standardize progression terms.

Confidentiality

Derived from MIMIC-CXR, a publicly available de-identified dataset.

Re-identification

Data are de-identified in MIMIC-CXR.

Sensitive data

No PHI; de-identified free-text reports used for weak labeling.