ReXGradient-160K: A Large-Scale Publicly Available Dataset of Chest Radiographs with Free-text Reports

Josh Miller; Juli\'an N. Acosta; Ouwen Huang; Pranav Rajpurkar; Xiaoman Zhang

REVIEW 4 major objections 5 minor 5 cited by

ReXGradient-160K: A Large-Scale Publicly Available Dataset of Chest Radiographs with Free-text Reports

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper presents ReXGradient-160K, a public collection of 160,000 chest X-ray studies paired with free-text radiology reports from 109,487 patients across 79 medical sites, and claims it is the largest public chest X-ray dataset by…

desk verdict A genuinely large multi-site chest X-ray dataset with paired reports, worth using, but the paper needs to document patient identity handling before the 'largest' claim can be checked. read the letter →

arxiv 2505.00228 v2 pith:NBE7FLHW submitted 2025-05-01 eess.IV cs.CV

Xiaoman Zhang , Julián N. Acosta , Josh Miller , Ouwen Huang , Pranav Rajpurkar This is my paper · ORCID

classification eess.IVcs.CV

keywords chestX-rayradiologyreportgenerationpublicdatasetmulti-institutionalde-identificationbenchmarksplitmedicalimagingAIReXGradient-160K

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ReXGradient-160K is a new public dataset of 160,000 chest X-ray studies paired with free-text radiology reports, drawn from 109,487 patients at 79 medical sites in three U.S. health systems. The paper's central claim is that, counting unique patients, this makes it the largest publicly available chest X-ray dataset to date. The authors see it as a resource for training and evaluating AI systems that generate radiology reports, because its size, multi-site origin, and consistent four-section report format support both large-scale training and tests of generalization across institutions. The dataset is released with fixed training, validation, and public test splits, as well as a private test set used for an external benchmark.

What carries the argument

The central object is the dataset corpus itself: 273,004 unique chest X-ray images grouped into 160,000 studies, with each study linked to a four-section radiology report. The load-bearing design choices are the de-identification pipeline (HIPAA-compliant text and pixel de-identification, pseudonymized names, IDs, and date shifting that preserves within-patient time ordering), the standardized image preprocessing (DICOM-to-PNG conversion and downsampling to 25% of original dimensions), and the GPT-4o-based extraction of reports into a uniform section structure. These choices make the data usable for large-scale training and for benchmark evaluation, which is the work the central claim needs them to do.

What would settle it

Download the released metadata and check every anonymized patient identifier across the training, validation, and test splits; if any patient ID occurs in more than one split, then the splits are not patient-disjoint and the benchmark claims would need to be recomputed under exclusive splits.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is a resource: a corpus of 273,004 de-identified chest radiographs organized into 160,000 studies, each associated with a report whose sections—indication, comparison, findings, impression—have been standardized. The paper reports that these studies come from 109,487 unique patients across 79 medical sites in three U.S. health systems, and it argues that this patient count makes ReXGradient-160K the largest public chest X-ray collection with paired reports. The dataset is intended to provide a more diverse and larger training and evaluation base for report-generation models, and to support studying how well such models generalize across institutions.

Load-bearing premise

The claim depends on the all-important assumption that no patient appears in more than one of the training, validation, and test splits, but the paper never explicitly states that the splits are patient-disjoint; if a patient spans splits, the evaluation numbers become optimistic.

Editorial extensions

If this is right

If the dataset is as large and as patient-diverse as reported, training corpora for chest X-ray report generation grow to more than one hundred thousand patients, with the same report vocabulary applied across sites.
The public test split gives a fixed ground for comparing report-generation models, and the private test set provides a held-out evaluation that avoids tuning directly to public test answers.
The multi-site composition makes it possible to measure how model performance shifts across health systems, a limitation of most earlier single-institution collections.
The uniform report sections allow downstream systems to use indication, findings, and impression without extra parsing, which should lower the barrier to reproducible experiments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The 'largest public chest X-ray dataset' claim depends on the counting metric: if measured by number of studies rather than patients, an earlier collection contains more studies but fewer patients, so future comparisons should state the metric explicitly.
The report-section extraction relies on an unmeasured GPT-4o step; if extraction quality is uneven, models trained on the standardized reports could inherit errors that the paper does not quantify.
A natural next step is to test whether the private benchmark set and public splits are truly site-diverse; the paper does not report per-site image counts, so site-level leakage or imbalance remains unverified.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: ReXGradient-160K is a genuinely large, multi-site chest X-ray dataset with paired reports, and the scale is new. But the unique-patient count — the basis for the 'largest' claim — is not yet demonstrated, because the paper never says how patients were matched across the three health systems. That's a documentation gap, not proof of an inflated number, but it needs to be closed before the headline is taken at face value.

What's good: 160,000 studies, 273,004 images, 109,487 patients across 3 health systems and 79 sites; public train/val/test splits plus a private test set for ReXrank; demographics look balanced across splits; internal counts are consistent (95,716 + 6,964 + 6,807 = 109,487). The release plan on HuggingFace is clear. This is a real new artifact, not a re-coordinatization of existing data, and no circular reasoning is involved. The authors cite their own ReXrank and MAIDA work, but the dataset stands independent of those references.

Soft spots, in order of weight. First, cross-institutional patient de-duplication: the de-identification section covers pseudonymization but says nothing about linking records for the same person across sites. If a patient appears in more than one health system, they'd be counted multiple times. Even a few percent duplication would erode the margin over MIMIC-CXR's 65,379 and CheXpert Plus's 64,725 patients. This must be disclosed. Second, the paper doesn't explicitly state that the train/val/test splits are patient-disjoint. The counts sum exactly, which suggests they are, but that should be one line. Third, GPT-4o section extraction is described with a post-processing step but no validation—no human annotation or spot-check—so we don't know how clean the report sections are. Fourth, study-selection criteria are absent; how were the 160,000 studies chosen? These are all addressable documentation issues, not evidence that the data itself is bad.

Who should read it: researchers working on report generation, chest X-ray representation learning, and multi-site generalization. This dataset is likely to become a useful resource even if the paper needs revision. I'd cite it once the dedup question is answered. A serious editor should send it to peer review; the issues are verification and documentation, not conceptual soundness. The authors should be asked to add a clear 'patient identity handling across sites' paragraph, a statement about disjoint splits, and a small validation study of the section extractor.

Referee Report

4 major / 5 minor

Summary. The manuscript introduces ReXGradient-160K, a multi-institutional chest X-ray dataset described as containing 160,000 studies with paired radiology reports from 109,487 unique patients across 3 U.S. health systems (79 medical sites). The dataset is split into training (140,000 studies), validation (10,000 studies), and public test (10,000 studies) sets, with an additional private test set of 10,000 studies reserved for the ReXrank benchmark. Images are de-identified and converted from DICOM to PNG, and reports are processed with GPT-4o into four sections: Indication, Comparison, Findings, and Impression. The central claim is that this is the largest publicly available chest X-ray dataset in terms of the number of patients.

Significance. If the unique-patient count and split construction are verified, this dataset is a valuable community resource: it would exceed MIMIC-CXR (65,379 patients) and CheXpert Plus (64,725 patients) in patient count, add multi-site diversity across 79 sites, and provide paired free-text reports at scale with open release on Hugging Face. The paper's internal arithmetic is consistent (273,004 images; patient counts summing to 109,487), and the demographic table suggests broadly balanced splits. I found no circularity: the dataset's size and composition are not derived from the authors' own benchmarks, and the ReXrank citations are contextual rather than load-bearing. The main risks are transparency about patient identity across institutions and about split disjointness, both of which are fixable with additional disclosure and analysis.

major comments (4)

[Section 2.1 (Data De-identification)] The unique-patient count of 109,487 is the basis for the 'largest publicly available chest X-ray dataset' claim, but the paper does not explain how patient identity was matched across the three contributing health systems. The de-identification description states that names, IDs, and dates are pseudonymized, but it does not say whether the same individual receiving imaging at more than one system was assigned the same de-identified PatientID. If cross-system matching was not performed, a patient treated at multiple sites would be counted multiple times, and because the comparator datasets are MIMIC-CXR (65,379 patients) and CheXpert Plus (64,725 patients), even a modest duplication rate could invalidate the headline comparison. Please state explicitly whether cross-institutional patient linkage was performed before de-identification (e.g., via a master patient index), or, if not, report per-system unique-patient counts and revise the claim to describe the sum of site-specific patients rather than unique persons.
[Section 2, Table 1] The paper never states whether the training, validation, and test splits are disjoint by patient. The row-wise patient counts (95,716 + 6,964 + 6,807 = 109,487) strongly suggest that no patient appears in more than one split, but this is not declared. If a single patient appears in multiple splits, any benchmark results obtained from these splits would be inflated and the reported split statistics would be misleading. The authors should add an explicit statement of patient-level disjointness and describe the procedure used to assign patients to splits, including how longitudinal studies for the same patient were handled.
[Section 2.3 (Report Structure)] The extraction of four report sections using GPT-4o is described, but no validation of this extraction is reported. The text mentions a 'robust post-processing validation step' but gives no human evaluation, sample size, agreement metric, or failure analysis. Since downstream users will rely on the structured reports for training and evaluating report-generation models, the manuscript should report extraction accuracy on a held-out sample, describe how missing sections were handled, and quantify the frequency of extraction errors. Without this, the reliability of the report content as a dataset feature is unverified.
[Section 2 (Dataset Composition)] The paper omits the study-selection criteria used to obtain the 160,000 studies from the three health systems. It is not stated whether this represents a complete consecutive sample, a random sample, or a filtered subset of available studies. This matters for assessing sampling bias and for reproducing the dataset construction, and it is directly relevant to the claim of multi-site representativeness. Please specify the inclusion and exclusion criteria, the date range of acquisition, and the number of contributing sites per health system.

minor comments (5)

[Section 2.3 (Report Structure)] The bullet for 'Comparison' is identical in wording to the bullet for 'Indication' ('Provides relevant patient background and reason for examination'). This appears to be a copy-paste error and should be corrected to describe the comparison with prior studies.
[Section 2.2 (Image Characteristics)] The paper states that images were downsampled to 25% of their original dimensions but does not report the resulting typical resolution in pixels. Reporting the final image dimensions, or at least the range of dimensions, would help users plan model training.
[Section 3 (Data Records)] The metadata list includes PatientWeight, StudyDate, and EthnicGroup, but the units for weight and the date format after the 365-day shift are not specified. These should be documented in the data record description.
[Section 3 (Data Records)] The paper says access requires signing the license on Hugging Face but does not state the license type. Please name the license (e.g., CC BY-NC 4.0 or a custom research license).
[Section 2.1 (Data De-identification)] The statement that dates are 'shifted to within 365 days of the actual date' is ambiguous; HIPAA-compliant date shifting typically uses a random shift of up to 364 days applied consistently per patient. Clarify the exact procedure and how leap years are handled.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the dataset's size and patient counts are empirical artifacts, and the self-citations are contextual rather than load-bearing.

full rationale

ReXGradient-160K is a data descriptor rather than a derivation: it makes no model predictions, fits no parameters, and proves no theorem from assumptions. The central claim that 160,000 studies and 109,487 patients constitute the largest public chest X-ray cohort is an empirical counting claim, not a quantity obtained by construction from any input. The self-citations to ReXrank (Zhang et al., 2024) and MAIDA (Saenz et al., 2024) are contextual: ReXrank appears only to identify the private test set's role, and MAIDA is cited as related work on data sharing. Neither is used to justify the dataset's size, composition, or the 'largest' comparison. The text itself leaves a verification gap in Section 2.1: pseudonymization of names, IDs, and dates is described, but cross-institutional patient linkage is not stated, so the 109,487 'unique patients' count is not shown to be deduplicated across the three health systems. That is a correctness and reproducibility concern about an upper-bound count, not a circularity, because the count is not defined as a function of the claim it is used to support. No circular step can be exhibited with the required quote and reduction, so the honest finding is no significant circularity, score 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or invented entities. The axioms are practical assumptions about the dataset construction pipeline and split design that are not fully verified in the paper.

assumptions (3)

domain assumption GPT-4o accurately extracts the four report sections (Indication, Comparison, Findings, Impression) from original free-text reports.
Section 2.3 uses GPT-4o for section extraction with a post-processing step, but no human evaluation, error rate, or agreement metric is reported.
domain assumption The pseudonymization and date-shifting de-identification fully removes protected health information while preserving clinical utility.
Section 2.1 asserts HIPAA compliance but provides no formal re-identification risk assessment or audit.
domain assumption Train, validation, and test splits are patient-disjoint, preventing data leakage across splits.
The paper does not explicitly declare patient-level splitting; the patient counts in Table 1 sum to the total, which suggests disjoint patient sets, but this is not stated and is load-bearing for benchmark validity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ReXGradient-160K: A Large-Scale Publicly Available Dataset of Chest Radiographs with Free-text Reports." pith.science (2026). https://pith.science/paper/NBE7FLHW

@misc{pith2026250500228,
  author       = {Pith},
  title        = {Pith review of: ReXGradient-160K: A Large-Scale Publicly Available Dataset of Chest Radiographs with Free-text Reports},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NBE7FLHW}},
  note         = {Machine review of arXiv:2505.00228}
}

read the original abstract

We present ReXGradient-160K, representing the largest publicly available chest X-ray dataset to date in terms of the number of patients. This dataset contains 160,000 chest X-ray studies with paired radiological reports from 109,487 unique patients across 3 U.S. health systems (79 medical sites). This comprehensive dataset includes multiple images per study and detailed radiology reports, making it particularly valuable for the development and evaluation of AI systems for medical imaging and automated report generation models. The dataset is divided into training (140,000 studies), validation (10,000 studies), and public test (10,000 studies) sets, with an additional private test set (10,000 studies) reserved for model evaluation on the ReXrank benchmark. By providing this extensive dataset, we aim to accelerate research in medical imaging AI and advance the state-of-the-art in automated radiological analysis. Our dataset will be open-sourced at https://huggingface.co/datasets/rajpurkarlab/ReXGradient-160K.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

RadHarmony: Radiological Data Handling in the Era of Agentic AI
cs.AI 2026-07 conditional novelty 6.0 of 10

RadHarmony is a unified Python library and AI-agent workflow for harmonizing 24 public radiology datasets, demonstrated by a multi-dataset self-supervised chest X-ray model with no dataset-specific code.
Prospective clinical indication, post-hoc report leakage, and fusion design in multi-image chest radiograph classification: a patient-clustered evaluation
cs.CV 2026-07 accept novelty 6.0 of 10

Report-derived chest X-ray labels are almost perfectly predicted by the report text itself (AUROC 0.98), while images plus prospective indication reach only AUROC 0.78, quantifying report-label circularity.
Scaling medical imaging report generation with multimodal reinforcement learning
cs.CV 2026-01 conditional novelty 6.0 of 10

UniRG-CXR, a Qwen3-VL-8B model trained with SFT plus GRPO reinforcement learning that directly optimizes the ReXrank metric components, reports state-of-the-art 1/RadCliQ-v1 results on all four ReXrank chest X-ray dat...
ReXVQA: A Large-scale Visual Question Answering Benchmark for Generalist Chest X-ray Understanding
cs.CV 2025-06 conditional novelty 6.0 of 10

ReXVQA is a large-scale chest X-ray VQA benchmark generated by an LLM pipeline, with a small reader study claiming MedGemma surpasses radiology residents.
Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context
cs.CV 2026-07 conditional novelty 5.0 of 10

Harrison.Rad 1.5 is a radiology-specific multimodal LLM that passes simulated FRCR 2B Short Case examinations and outperforms general-purpose frontier models on plain-film radiography reporting tasks.

Reference graph

Works this paper leans on

11 extracted references · 3 canonical work pages · cited by 5 Pith papers

[1]

Maira-2: Grounded radiology report generation

Shruthi Bannur, Kenza Bouzid, Daniel C Castro, Anton Schwaighofer, Sam Bond-Taylor, Maximilian Ilse, Fernando P \'e rez-Garc \' a, Valentina Salvatelli, Harshita Sharma, Felix Meissen, et al. Maira-2: Grounded radiology report generation. arXiv preprint arXiv:2406.04449, 2024

arXiv 2024
[3]

Preparing a collection of radiology examinations for distribution and retrieval

Dina Demner-Fushman, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Clement J McDonald. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Association, 23 0 (2): 0 304--310, 2016

work page 2016
[4]

Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports

Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6 0 (1): 0 317, 2019

2019
[5]

Radialog: A large vision-language model for radiology report generation and conversational assistance

Chantal Pellegrini, Ege \"O zsoy, Benjamin Busam, Nassir Navab, and Matthias Keicher. Radialog: A large vision-language model for radiology report generation and conversational assistance. arXiv preprint arXiv:2311.18681, 2023

arXiv 2023
[6]

The maida initiative: establishing a framework for global medical-imaging data sharing

Agustina Saenz, Emma Chen, Henrik Marklund, and Pranav Rajpurkar. The maida initiative: establishing a framework for global medical-imaging data sharing. The Lancet Digital Health, 6 0 (1): 0 e6--e8, 2024

work page 2024
[7]

Interactive and explainable region-guided radiology report generation

Tim Tanida, Philip M \"u ller, Georgios Kaissis, and Daniel Rueckert. Interactive and explainable region-guided radiology report generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7433--7442, 2023

work page 2023
[8]

Rexrank: A public leaderboard for ai-powered radiology report generation

Xiaoman Zhang, Hong-Yu Zhou, Xiaoli Yang, Oishi Banerjee, Juli \'a n N Acosta, Josh Miller, Ouwen Huang, and Pranav Rajpurkar. Rexrank: A public leaderboard for ai-powered radiology report generation. arXiv preprint arXiv:2411.15122, 2024

arXiv 2024
[9]

A generalist learner for multifaceted medical image interpretation

Hong-Yu Zhou, Subathra Adithan, Juli \'a n Nicol \'a s Acosta, Eric J Topol, and Pranav Rajpurkar. A generalist learner for multifaceted medical image interpretation. arXiv preprint arXiv:2405.07988, 2024

arXiv 2024

Show all 11 references

[10]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...
[11]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...
[12]

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

2019 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Maira-2: Grounded radiology report generation

Shruthi Bannur, Kenza Bouzid, Daniel C Castro, Anton Schwaighofer, Sam Bond-Taylor, Maximilian Ilse, Fernando P \'e rez-Garc \' a, Valentina Salvatelli, Harshita Sharma, Felix Meissen, et al. Maira-2: Grounded radiology report generation. arXiv preprint arXiv:2406.04449, 2024

arXiv 2024

[2] [3]

Preparing a collection of radiology examinations for distribution and retrieval

Dina Demner-Fushman, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Clement J McDonald. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Association, 23 0 (2): 0 304--310, 2016

work page 2016

[3] [4]

Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports

Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6 0 (1): 0 317, 2019

2019

[4] [5]

Radialog: A large vision-language model for radiology report generation and conversational assistance

Chantal Pellegrini, Ege \"O zsoy, Benjamin Busam, Nassir Navab, and Matthias Keicher. Radialog: A large vision-language model for radiology report generation and conversational assistance. arXiv preprint arXiv:2311.18681, 2023

arXiv 2023

[5] [6]

The maida initiative: establishing a framework for global medical-imaging data sharing

Agustina Saenz, Emma Chen, Henrik Marklund, and Pranav Rajpurkar. The maida initiative: establishing a framework for global medical-imaging data sharing. The Lancet Digital Health, 6 0 (1): 0 e6--e8, 2024

work page 2024

[6] [7]

Interactive and explainable region-guided radiology report generation

Tim Tanida, Philip M \"u ller, Georgios Kaissis, and Daniel Rueckert. Interactive and explainable region-guided radiology report generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7433--7442, 2023

work page 2023

[7] [8]

Rexrank: A public leaderboard for ai-powered radiology report generation

Xiaoman Zhang, Hong-Yu Zhou, Xiaoli Yang, Oishi Banerjee, Juli \'a n N Acosta, Josh Miller, Ouwen Huang, and Pranav Rajpurkar. Rexrank: A public leaderboard for ai-powered radiology report generation. arXiv preprint arXiv:2411.15122, 2024

arXiv 2024

[8] [9]

A generalist learner for multifaceted medical image interpretation

Hong-Yu Zhou, Subathra Adithan, Juli \'a n Nicol \'a s Acosta, Eric J Topol, and Pranav Rajpurkar. A generalist learner for multifaceted medical image interpretation. arXiv preprint arXiv:2405.07988, 2024

arXiv 2024