MIMIC-CXR: Chest Radiograph Corpus Overview

• MIMIC-CXR is a public chest radiograph corpus offering paired imaging studies and free-text reports in both DICOM and JPEG formats.
• The dataset standardizes labels using tools like CheXpert and NegBio to support consistent evaluation and multimodal learning research.
• It underpins diverse applications such as weakly supervised classification, report generation, and cross-modal knowledge transfer while noting limitations like lack of precise lesion localization.

Searching arXiv for recent and foundational papers on MIMIC-CXR to ground the article. MIMIC-CXR is a public chest radiograph corpus centered on paired imaging studies and radiology reports from Beth Israel Deaconess Medical Center. In the literature, the name is used for both the original DICOM-based release with free-text reports and the derived MIMIC-CXR-JPG release, which standardizes image preprocessing, labels, and reference splits. Published descriptions use more than one count: one large-scale classification study described MIMIC-CXR as 473,064 chest radiographs, 206,574 radiology reports, and 63,478 patients, whereas MIMIC-CXR-JPG v2.0.0 described 377,110 JPEG images associated with 227,827 imaging studies; the same JPG paper’s split table sums to 227,835 studies/reports and 65,379 patients, and comparative dataset work also summarized MIMIC-CXR as 377,110 DICOM images with free-text reports (Rubin et al., 2018, Johnson et al., 2019, Nguyen et al., 2020).

1. Release structure and source population

The original MIMIC-CXR release provides chest radiographs in clinical DICOM format together with free-text radiology reports, while MIMIC-CXR-JPG is a derived, convenience-oriented release that provides de-identified JPEG files, 14 report-derived labels, and standard train/validation/test splits (Johnson et al., 2019). The JPG release was created to make the corpus easier to use in mainstream computer vision pipelines, while the DICOM release remains closer to the clinical source data because it preserves richer metadata and higher dynamic range (Johnson et al., 2019).

The JPG release was sourced from the Beth Israel Deaconess Medical Center PACS and queried from all studies for patients admitted to the emergency department over the date range 2011–2016 (Johnson et al., 2019). The image conversion pipeline extracted pixel data from DICOM using pydicom, normalized pixel values to $[0,255]$, used the DICOM field PhotometricInterpretation to determine whether inversion was needed, applied histogram equalization with OpenCV, and saved the result as JPEG with quality factor 95 (Johnson et al., 2019). The paper is explicit that, aside from de-identification and conversion to JPEG, no filtering or processing of the images was performed (Johnson et al., 2019).

The distinction between the original DICOM corpus and the JPG derivative is methodologically important. DICOM-based work exploits original grayscale depth, view metadata, and acquisition context, whereas JPG-based work gains ease of use and benchmarking consistency at the cost of lossy compression, reduced bit depth, histogram equalization, and loss of some clinically relevant metadata (Johnson et al., 2019). This suggests that “MIMIC-CXR” is best treated as a family of closely related releases rather than a single invariant tensor archive.

2. Study organization, de-identification, and standardized labels

MIMIC-CXR is organized at the study level: one study may contain multiple radiographs, and the report is linked to the study rather than to a single isolated image (Johnson et al., 2019). This structure is preserved in many later modeling papers, which explicitly treat the study as the natural prediction unit rather than the image. The JPG release introduced official patient-based splits intended to support consistent evaluation and reduce identity leakage across train and evaluation partitions (Johnson et al., 2019).

Split	Images	Studies / patients
Train	368,960	222,758 / 64,586
Validate	2,991	1,808 / 500
Test	5,159	3,269 / 293

The split table also reports view composition. In train, validation, and test respectively, frontal images number 248,020, 2,041, and 3,653; lateral images number 120,795, 949, and 1,502; and “other” views are rare at 145, 1, and 4 images (Johnson et al., 2019). The test split was constructed from all studies for patients who had at least one report manually reviewed, and the paper notes that the test set is enriched for pathology relative to train and validation (Johnson et al., 2019).

The de-identification pipeline addressed burned-in annotations in the pixel data. The authors used an ensemble of image preprocessing and OCR approaches to detect text with significant contrast and consistent pixel value, and removed suspected PHI by setting all pixel values in a bounding box covering the PHI to black (Johnson et al., 2019). Manual privacy review covered 6,900 radiographs, each reviewed by two independent annotators; 180 images underwent secondary consensus review, and none were ultimately judged to contain PHI (Johnson et al., 2019).

Standardized label extraction in MIMIC-CXR-JPG is report-based. Labels are derived from the impression section when present, otherwise from findings, otherwise from the final section of the report; among 227,835 reports, 189,561 had an impression section, 27,684 had a findings section, 10,514 had an equivalent unlabeled section, and 8 had no text for labeling (Johnson et al., 2019). The 14 labels are Atelectasis, Cardiomegaly, Consolidation, Edema, Enlarged Cardiomediastinum, Fracture, Lung Lesion, Lung Opacity, No Finding, Pleural Effusion, Pleural Other, Pneumonia, Pneumothorax, and Support Devices (Johnson et al., 2019). Two open-source rule-based tools, NegBio and CheXpert, are used; the release includes positive, negative, uncertain, disagreement, and blank or unmentioned states, with “Disagreement” explicitly marking cases where CheXpert and NegBio disagree (Johnson et al., 2019).

Validation of the report-derived labels used 687 reports manually reviewed by a board-certified radiologist with 8 years of experience (Johnson et al., 2019). The reported F1 scores show that mention extraction is strong for several findings, including Pneumothorax 0.970, Pleural Effusion 0.973, Pneumonia 0.956, and Atelectasis 0.944, but much weaker for No Finding 0.531, Enlarged Cardiomediastinum 0.462, and Pleural Other 0.592 (Johnson et al., 2019). Uncertainty detection is notably weaker than mention extraction and often weaker than negation detection; this limitation is central to later work that treats MIMIC-CXR labels as weak supervision rather than definitive image truth (Johnson et al., 2019).

3. Study-level structure, view geometry, and multi-view modeling

A defining property of MIMIC-CXR is that frontal and lateral radiographs coexist within the same exam. One influential classification study separated PA, AP, and lateral views and argued that this view distinction is not a nuisance variable but a core part of the clinical and statistical structure of the corpus (Rubin et al., 2018). In that work, labels were extracted from the Findings and Impression sections with NegBio, mapped from 46 UMLS concepts into 13 pathology findings plus No Finding, and uncertain mentions were excluded from positive assignment (Rubin et al., 2018). The split was by subject, not by image, with 20% test, 70% train, and 10% validation (Rubin et al., 2018).

The same study reported single-view counts of 70,875 train / 10,078 validation / 20,426 test for PA, 168,525 / 24,566 / 47,689 for AP, and 81,379 / 11,519 / 23,125 for lateral, for a total of 320,779 / 46,163 / 91,240 images (Rubin et al., 2018). For paired-view experiments restricted to studies containing both a frontal and a lateral image, the authors reported 74,872 train, 10,600 validation, and 21,403 test dual-pair studies (Rubin et al., 2018). Their DualNet architecture processed frontal and lateral images in parallel and improved average AUC from 0.690 to 0.721 on PA+Lateral studies and from 0.637 to 0.668 on AP+Lateral studies, in both cases a gain of 0.031 (Rubin et al., 2018).

Later multimodal work treated these projections as separate but related modalities rather than redundant copies. The MMVM study on MIMIC-CXR-JPG defined a frontal modality $\vec{x}_f$ consisting of AP and PA views and a lateral modality $\vec{x}_l$ consisting of LL and Lateral views, then formed all possible frontal–lateral pairings within each study that contained at least one frontal and one lateral image (Agostini et al., 2024). Labels came from the 14 diagnostic labels of MIMIC-CXR-JPG obtained via the CheXpert label extraction tool, but all non-positive values, including negative, non-mentioned, and uncertain, were merged into a single negative class (Agostini et al., 2024). This framing makes the study, not the image, the fundamental unit of multimodal representation learning.

Multi-view modeling on MIMIC-CXR also extends to report generation. EVOKE is built around the fact that a MIMIC-CXR study may contain varying numbers of radiographs plus one shared report and an optional indication field; the paper explicitly reports 269,239 train images paired with 150,957 train reports in its MIMIC-CXR split, with 66.4% of training reports having an indication, and it filters out empty or clinically insignificant reference reports (Miao et al., 2024). This suggests that MIMIC-CXR’s study-level heterogeneity is not incidental metadata but a primary modeling opportunity.

4. Foundation for report-centric, multimodal, and cross-modal learning

MIMIC-CXR has been used as a large-scale weakly supervised classification corpus, a vision-language pretraining corpus, a multimodal patient-context benchmark, and even as an imaging teacher for non-image modalities. In classification, the core supervision is usually report-derived rather than manually annotated image-level ground truth. One early large-scale study used NegBio on Findings and Impression to generate multi-label targets and trained modified DenseNet-121 models on frontal and lateral radiographs, with no data augmentation and 512 × 512 inputs normalized by the maximum 12-bit grayscale value $2^{12}-1$ (Rubin et al., 2018). That work established a widely reused view-aware benchmark protocol rather than a fully curated annotation standard.

In vision-language pretraining, CXR-CLIP treated MIMIC-CXR as the foundational public image-report corpus. It used 222,628 studies from the MIMIC-CXR training split for pretraining, 1,808 studies for validation, and 3,264 studies from the test split for retrieval evaluation, extracting the findings and impression sections and exploiting both multiple images per study and multiple report sections per study (You et al., 2023). The model introduced study-aware multi-view cross-modal alignment and intra-modal image and text contrastive losses, $L_{ICL}$ and $L_{TCL}$, to exploit study structure rather than generic augmentation (You et al., 2023). This position is reinforced by the paper’s empirical observation that MIMIC-CXR report supervision is essential for retrieval, while prompt-expanded label-only datasets improve zero-shot classification but slightly degrade report retrieval (You et al., 2023).

MIMIC-CXR also underpins multimodal reasoning with patient context. WoLF uses MIMIC-CXR in two ways: as part of the CXRVQA/EHRXQA benchmark constructed from MIMIC-IV and MIMIC-CXR, whose test set comprises 11,309 VQAs for CXR, and as the report-generation benchmark following the official MIMIC-CXR split (Kang et al., 2024). The framework links chest radiographs with EHR-derived question–answer content, reorganizes reports into anatomy-specific targets through Anatomy-Specific Knowledge decoupling, and reports superior performance over prior models on MIMIC-CXR report generation, including BLEU-1 0.496, BLEU-4 0.165, METEOR 0.187, and ROUGE-L 0.370 (Kang et al., 2024). In this setting, MIMIC-CXR is no longer only an image–report corpus but a hub for image, text, and broader patient context.

The dataset is also used for cross-modal knowledge transfer. CroMoTEX links MIMIC-CXR to MIMIC-IV-ECG and uses chest radiographs as the imaging teacher modality for ECG representation learning, with downstream ECG-only binary classification for cardiomegaly, pleural effusion, and edema (Punyamoorty et al., 24 Jun 2025). The labels are inherited from MIMIC-CXR structured disease labels derived by the CheXpert labeler, and cross-modal alignment uses a supervised contrastive objective with adaptive hard negative weighting (Punyamoorty et al., 24 Jun 2025). This suggests that MIMIC-CXR functions not only as a target dataset for chest imaging models, but also as a source of pathology supervision for other clinical modalities.

5. Derived benchmarks and MIMIC-CXR-based extensions

Several important task-specific resources are derived from MIMIC-CXR or MIMIC-CXR-JPG. The long-tailed CXR-LT setting uses MIMIC-CXR-JPG in an expanded 40-class protocol with 234,800 frontal-view JPG images; classes are divided into 11 common classes with more than 10,000 samples, 17 medium-frequency classes with 1,000–10,000 samples, and 12 rare classes with fewer than 1,000 samples, with the rare classes together constituting about 2% of the dataset (Madhipati et al., 25 Jul 2025). This benchmark exposes a property that is present but less explicit in the standard 14-label setup: severe long-tail imbalance across findings.

Temporal and longitudinal resources extend the study structure further. CoCa-CXR constructs CXR-4 from MIMIC-CXR images, MIMIC-CXR reports, and Chest ImaGenome scene graphs, following the official MIMIC-CXR split while excluding the MS-CXR-T test set (Chen et al., 27 Feb 2025). CXR-4 contains 224,487 clean image-report pairs, 132,320 image pairs with filtered reports containing comparison information, 259,562 image pairs with extracted comparison text, and 758,344 image pairs with abnormal organ condition, coordinates, and progression annotations (Chen et al., 27 Feb 2025). This derived corpus makes explicit the longitudinal dimension that is only implicit in the base release.

Interpretability-oriented derivatives also exist. REFLACX provides eye gaze data for more than 2,500 CXRs from MIMIC-CXR, with gaze sequences captured at 1000 Hz, and has been used to build anatomy-specific gaze supervision and heatmaps for controllable, gaze-grounded chest X-ray interpretation (Pham et al., 2023). The I-AI pipeline constructed binary regional tasks on REFLACX-derived subsets and showed that classification could be performed on heatmap-masked images while retaining competitive accuracy (Pham et al., 2023). In this case, MIMIC-CXR supplies the imaging substrate, while the derivative dataset contributes a human-attention signal absent from the base release.

Report-generation research has likewise created MIMIC-derived variants emphasizing view structure and conditioning context. EVOKE derives Multi-view CXR and Two-view CXR from studies with multiple views in MIMIC-CXR and IU X-ray, reporting 220,978 train images with 100,505 train reports for Multi-view CXR and 165,056 train images with 82,528 train reports for Two-view CXR (Miao et al., 2024). Such derived resources formalize scenarios that already exist in the underlying corpus but are not standardized in the original release.

6. Limitations, ambiguities, and position in the chest X-ray dataset ecosystem

The principal limitation of MIMIC-CXR as a supervision source is that most widely used labels are report-derived rather than manually image-annotated. The MIMIC-CXR-JPG paper shows that mention extraction is often strong, but uncertainty detection is substantially weaker, and label states include disagreement between NegBio and CheXpert rather than a single authoritative truth (Johnson et al., 2019). Later methodological papers are explicit that MIMIC-CXR labels therefore inherit omission bias, concept mapping ambiguity, and errors in negation or uncertainty detection (Rubin et al., 2018). For this reason, performance on MIMIC-CXR benchmarks is best interpreted as performance on a large, clinically realistic weak-labeling regime rather than on dense expert annotation.

A second limitation is the coexistence of multiple ontologies and preprocessing regimes under the “MIMIC-CXR” name. MIMIC-CXR-JPG standardized 14 CheXpert-style labels and a JPG-based reference split (Johnson et al., 2019), but other influential work used UMLS-driven mappings, different treatment of uncertainty, or filtered subsets such as frontal-only, paired-view, or long-tailed CXR-LT protocols (Rubin et al., 2018, Madhipati et al., 25 Jul 2025). This complicates direct comparison across papers even when the same underlying hospital corpus is being used.

A third limitation is the absence of explicit lesion localization in the standard public benchmarking setup. Comparative dataset work tabulates MIMIC-CXR as having 14 findings, 377,110 samples, image-level labels available, and no local labels, whereas VinDr-CXR was designed with 22 box-localized findings and 6 global disease labels (Nguyen et al., 2020). That comparison is central: MIMIC-CXR is optimized for scale, report availability, classification, multimodal learning, and representation learning, but not for lesion detection or precise grounding (Nguyen et al., 2020). This suggests that MIMIC-CXR and localization-rich datasets are complementary rather than interchangeable.

Finally, MIMIC-CXR is institutionally specific. The JPG release comes from Beth Israel Deaconess Medical Center and was sourced from emergency-department admissions over 2011–2016 (Johnson et al., 2019). The convenience of standardized JPEGs, official splits, and report-derived labels has made the corpus easy to adopt, but those same design choices also introduce prevalence shift, formatting assumptions, and release-dependent artifacts. The most robust use of MIMIC-CXR therefore treats it as a large, report-centric, study-level chest radiograph corpus whose strengths are scale, multimodal linkage, and methodological breadth, while its weaknesses are weak supervision, ontology variation, and lack of native lesion localization ground truth (Johnson et al., 2019, Nguyen et al., 2020).