This dataset is derived from MIMIC-CXR and requires credentialed PhysioNet access. Please provide your PhysioNet information below. Requests are reviewed manually and typically processed within a few days.

By submitting this form, you confirm that you have an approved PhysioNet account with MIMIC-CXR access, and that you agree to the PhysioNet Credentialed Health Data Use Agreement (https://physionet.org/content/lunguage/view-dua/1.0.0/).

LUNGUAGE: A Benchmark for Structured and Sequential Chest X-ray Interpretation

Access restricted: This dataset is derived from MIMIC-CXR. You must be a credentialed PhysioNet user with approved MIMIC-CXR access to use this dataset. Please provide your PhysioNet username when requesting access.

Dataset Summary

LUNGUAGE is a benchmark dataset for fine-grained and temporally aware interpretation of chest radiograph reports. It is the first benchmark to support both single-report structured evaluation and longitudinal patient-level assessment across multiple studies.

Constructed from the official test split of MIMIC-CXR, LUNGUAGE comprises:

1,473 chest X-ray reports from 230 patients, annotated with over 17,000 expert-verified entities and 23,000 relation–attribute pairs across 18 relation types
186 longitudinal reports from 30 patients with sequential annotations (2–14 studies/patient), grouped into semantically and temporally coherent groups
A schema-aligned vocabulary covering 3,827 diagnostic entities and attributes with UMLS mappings

All annotations were conducted and verified by board-certified radiologists.

Abstract

Radiology reports convey detailed clinical observations and capture diagnostic reasoning that evolves over time. However, existing evaluation methods are limited to single-report settings and rely on coarse metrics that fail to capture fine-grained clinical semantics and temporal dependencies. We introduce LUNGUAGE, a benchmark dataset for structured radiology report generation that supports both single-report evaluation and longitudinal patient-level assessment across multiple studies. It contains 1,473 annotated chest X-ray reports, each reviewed by experts, and 186 of them contain longitudinal annotations to capture disease progression and inter-study intervals, also reviewed by experts. Using this benchmark, we develop a two-stage structuring framework that transforms generated reports into fine-grained, schema-aligned structured reports, enabling longitudinal interpretation. We also propose LUNGUAGESCORE, an interpretable metric that compares structured outputs at the entity, relation, and attribute level while modeling temporal consistency across patient timelines. These contributions establish the first benchmark dataset, structuring framework, and evaluation metric for sequential radiology reporting, with empirical results demonstrating that LUNGUAGESCORE effectively supports structured report evaluation.

Dataset Files

File	Rows	Description
`Lunguage.csv`	17,949	Annotated entities with relation–attribute pairs (single + sequential)
`Lunguage_vocab.csv`	3,827	Schema-aligned vocabulary with UMLS mappings

Dataset Statistics

Single Structured Reports

Statistic	Value
Total reports	1,473
Total patients	230
Studies per patient	1–15
Annotated entities	17,949
Relation–attribute pairs	23,307
Relation types	18

Sequential Structured Reports

Statistic	Value
Total reports	186
Total patients	30 (subset of 230)
Reports per patient	2–14

Data Fields

`Lunguage.csv`

Column	Type	Description
`subject_id`	str	Unique patient identifier
`study_id`	str	Unique imaging study identifier
`section`	str	Report subsection (`hist`, `find`, `imp`)
`sent_idx`	int	Sentence index within section
`ent_idx`	int	Entity index within section
`cat`	str	Entity category (`pf`, `cf`, `cof`, `ncd`, `oth`, `patient info`)
`ent`	str	Entity text as it appears in the report
`normed_ent`	str	Normalized entity form
`dx_status`	str	`positive` / `negative`
`dx_certainty`	str	`definitive` / `tentative`
`location`	str	Anatomical location
`evidence`	str	Linked entity providing image evidence
`associate`	str	Co-occurring associated entity
`morphology`	str	Shape/structural description
`distribution`	str	Spatial spread pattern
`measurement`	str	Quantitative descriptor
`severity`	str	Degree of abnormality
`onset`	str	Temporal onset descriptor
`improved`	str	Regression/resolution descriptor
`worsened`	str	Progression descriptor
`no_change`	str	Stability descriptor
`placement`	str	Device position/action (for `oth` entities)
`past_hx`	str	Patient history reference
`other_source`	str	Non-CXR modality reference
`assessment_limitations`	str	Image quality / technical constraints
`gt_entity_group`	str	Cross-time semantic grouping label (sequential)
`gt_temporal_group`	float	Temporal episode index (sequential)
`sequence`	float	Study index in patient timeline
`StudyDateTime`	str	Imaging date and time
`time_from_first`	str	Days since patient's first study
`report`	str	Full report text
`section_report`	str	Section-level report text
`sent`	str	Target sentence

`Lunguage_vocab.csv`

Column	Description
`category`	High-level relation type (e.g., `location`, `morphology`, `severity`)
`subcategory`	Fine-grained semantic class
`target_term`	Original lexical phrase from reports
`normed_term`	Normalized standard form
`UMLS (w code)`	UMLS concept name and code

Entity Schema

Entities are assigned to one of 6 category labels:

Category	Description	Examples
`pf`	Perceptual Findings — directly visible on CXR	opacity, pleural effusion
`cf`	Contextual Findings — diagnosed from image + context	pneumonia, congestive heart failure
`cof`	Clinical Objective Findings — from labs / vitals	oxygen saturation, white cell count
`ncd`	Non-CXR Diagnosis — from other modalities	stroke, seizure disorder
`oth`	Other Objects — devices / surgical hardware	central venous catheter, PICC line
`patient info`	Patient history / symptoms	fever, history of malignancy

Each entity has two core attributes:

dx_status: positive (present) / negative (absent)
dx_certainty: definitive / tentative

Relation Types

18 relation types across 4 categories:

Category	Relations
Diagnostic Reasoning	`evidence`, `associate`
Spatial & Descriptive	`location`, `morphology`, `distribution`, `measurement`, `severity`, `comparison`
Temporal Change	`onset`, `improved`, `worsened`, `no_change`, `placement`
Contextual	`past_hx`, `other_source`, `assessment_limitations`

Usage

Load with `datasets`

from datasets import load_dataset

# Load the main annotation dataset
ds = load_dataset("SuperSupermoon/Lunguage", data_files="Lunguage.csv")

# Load the vocabulary
vocab = load_dataset("SuperSupermoon/Lunguage", data_files="Lunguage_vocab.csv")

Load with `lunguage-score` package

The lunguage-score package provides the full evaluation pipeline:

pip install lunguage-score

import pandas as pd
from lunguage_score import LunguageScorer
from lunguage_score.config import Config, MetricConfig

pred_df = pd.read_csv("path/to/pred_SR_df.csv")
gt_df   = pd.read_csv("path/to/Lunguage.csv")

config = Config(metrics=MetricConfig(output_dir="./results", mode="gold_eval"))
scorer = LunguageScorer(config)
results = scorer.calculate_lunguage_score_only(pred_df, gt_df)

print(f"Structure Score: {results['avg_structure_score']:.4f}")
print(f"Precision:       {results['avg_precision']:.4f}")
print(f"Recall:          {results['avg_recall']:.4f}")

Access & License

This dataset is derived from MIMIC-CXR and is subject to the PhysioNet Credentialed Health Data License 1.5.0.

Requirements to access:

Be a credentialed user on PhysioNet
Have approved access to MIMIC-CXR
Complete CITI Data or Specimens Only Research training
Sign the PhysioNet Credentialed Health Data Use Agreement 1.5.0

The official release is also available at: PhysioNet DOI: https://doi.org/10.13026/pk42-4v91

Citation

If you use this dataset, please cite both the dataset and the paper:

@article{moon2025lunguage,
  title={Lunguage: A Benchmark for Structured and Sequential Chest X-ray Interpretation},
  author={Moon, Jong Hak and Choi, Geon and Rabaey, Paloma and Kim, Min Gwam and Hong, Hyuk Gi and Lee, Jung Oh and Yoon, Hangyul and Doe, Eunwoo and Kim, Jiyoun and Sharma, Harshita and {Coelho de Castro}, Daniel and {Alvarez Valle}, Javier and Choi, Edward},
  journal={arXiv preprint arXiv:2505.21190},
  year={2025}
}

@article{PhysioNet-lunguage-1.0.0,
  author = {Moon, Jong Hak and Choi, Geon and Rabaey, Paloma and Kim, Min Gwam and Hong, Hyuk Gi and Lee, Jung Oh and Yoon, Hangyul and Doe, Eunwoo and Kim, Jiyoun and Sharma, Harshita and {Coelho de Castro}, Daniel and {Alvarez Valle}, Javier and Choi, Edward},
  title = {{Lunguage: A Benchmark for Structured and Sequential Chest X-ray Interpretation}},
  journal = {{PhysioNet}},
  year = {2026},
  month = jan,
  note = {Version 1.0.0},
  doi = {10.13026/pk42-4v91},
  url = {https://doi.org/10.13026/pk42-4v91}
}

Please also cite PhysioNet:

@article{goldberger2000physiobank,
  title={PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals},
  author={Goldberger, Ary L and others},
  journal={Circulation},
  volume={101},
  number={23},
  pages={e215--e220},
  year={2000}
}

Contact

For questions about the dataset: [email protected]

Downloads last month: 1

Total file size:

17.3 MB

Paper for SuperSupermoon/Lunguage

Lunguage: A Benchmark for Structured and Sequential Chest X-ray Interpretation

Paper • 2505.21190 • Published May 27, 2025 • 4

Request Access to LUNGUAGE