PadChest: A Chest Radiograph Dataset
- PadChest is a large-scale chest radiograph dataset that provides high-resolution, de-identified images with multi-label annotations derived from Spanish radiology reports.
- It features a multi-view structure with PA, lateral, and other projection classes, enabling controlled paired-view analyses and innovative multi-label classification.
- Its rich metadata and hierarchical labeling support diverse applications such as segmentation, report generation, robustness analysis, and long-tailed evaluation in medical imaging.
Searching arXiv for PadChest-related papers to ground the article in the current literature. PadChest is a large-scale chest radiograph dataset that links high-resolution de-identified images to multi-label annotations derived from Spanish radiology reports. The original release reports 160,868 labeled chest X-ray images from 69,882 patients grouped into 109,931 studies acquired at Hospital Universitario de San Juan, Spain, from 2009 to 2017, while later PadChest-based studies often summarize the resource as about 160,000 images from about 67,000 patients. Its combination of report text, projection metadata, high-resolution DICOM imaging, hierarchical labels, and substantial availability of paired frontal and lateral views has made it a recurring benchmark for multi-label classification, multiview learning, report generation, segmentation, shortcut analysis, and long-tailed evaluation in chest radiography (Bustos et al., 2019, Bertrand et al., 2019).
1. Corpus design and imaging characteristics
PadChest was curated at a single Spanish institution and released through BIMCV as a de-identified DICOM-based resource with an accompanying index CSV containing 33 fields. The original release enumerates six standardized projection classes: PA with 96,010 images, lateral with 51,124, AP-horizontal with 14,355, AP-vertical with 5,158, costal with 631, and pediatric protocol with 274. Images were kept at high resolution rather than resized; the reported image dimensions are rows mean and columns mean , with BitsStored and PhotometricInterpretation MONOCHROME2. The metadata include acquisition and scanner fields such as StudyDate, PatientSex, ViewPosition, Manufacturer, SpatialResolution, WindowCenter, WindowWidth, Rows, Columns, and exposure-related variables (Bustos et al., 2019).
A distinctive design choice is explicit projection handling. PadChest does not collapse all chest radiographs into a single frontal-view pool: it preserves multi-view structure, including PA, AP, and lateral acquisitions. For 20,367 images lacking DICOM projection information, the release reports that a ResNet-50–based classifier assigned projection classes. This view-aware organization later enabled controlled paired-view studies that were not possible in frontal-only public benchmarks (Bustos et al., 2019).
2. Taxonomy and report-derived annotation pipeline
PadChest organizes labels into three hierarchical, multi-axial taxonomies mapped to UMLS concepts: 174 radiographic findings, 19 differential diagnoses, and 104 anatomical locations. The dataset also includes special labels such as Normal, Suboptimal, Exclude, and Unchanged. Radiographic findings are intended to denote entities directly observable on the image, whereas differential diagnoses capture uncertainty and clinical interpretation beyond pure image evidence (Bustos et al., 2019).
The annotation pipeline combines manual physician review with automated report mining. In the original release, 27% of reports were manually annotated by physicians, and the remaining 73% were labeled automatically using a supervised method based on a recurrent neural network with attention mechanisms. The manual component covers 27,593 reports and 22,120 unique sentences; the automated component covers 82,338 reports. Spanish report preprocessing included lowercasing, accent removal, section extraction, removal of non-alphanumeric characters except periods and spaces, Spanish stopword removal while retaining negation-bearing terms such as “sin,” “no,” and “ni,” and Spanish stemming. The best text model, an RNN with label-wise attention, achieved a Micro-F1 of 0.939 and a WeightedF1 of 0.926 on an independent 500-sentence test set; the abstract summarizes the automatic labeling validation as achieving a 0.93 Micro-F1 score (Bustos et al., 2019).
This architecture has two important implications for downstream use. First, PadChest is richer than a flat pathology list because it preserves hierarchy and anatomical localization. Second, its labels are not uniformly image-level expert ground truth; they are largely report-derived, with expert review only on a subset. Later PadChest studies repeatedly treat this distinction as methodologically consequential rather than incidental (Bustos et al., 2019, Cheplygina et al., 2023).
3. Multiview structure and the role of lateral radiographs
PadChest is unusual among public chest X-ray datasets because it contains widespread paired posteroanterior and lateral studies. One line of work restricted the corpus to patients with both PA and lateral views and a single visit per patient, yielding 30,699 paired patients. In that cohort, 194 findings were mapped to top-level parents in the PadChest hierarchy, labels with at least 100 positive patients were retained, and the final task used 56 labels. Using simple DenseNet models, the study reported an overall weighted AUC of 0.79, found the PA view more informative for 26 labels, the lateral view more informative for 8 labels, and similar performance for 21 labels. The eight labels for which the lateral view increased AUC were pleural effusion, artificial heart valve, hemidiaphragm elevation, osteopenia, flattened diaphragm, costophrenic angle blunting, vertebral degenerative changes, and surgery (Bertrand et al., 2019).
A subsequent paired-view study used the same 30,699-patient cohort but retained labels present in at least 50 patients, producing 64 labels. It compared DenseNet-PA, DenseNet-L, Stacked, DualNet, HeMIS, AuxLoss, and curriculum-learning variants. The best joint model, AuxLoss-CL, improved overall AUC from for PA-only DenseNet to with both views, and the paper states that including the lateral view increased performance for 32 labels while being neutral for the others. It further argued that the gain from adding lateral views was comparable to the gain obtained by using only the PA view with twice the amount of patients in the training set (Hashir et al., 2020).
These results temper a common simplification. PadChest does not show that lateral radiographs are uniformly beneficial, nor that naive fusion is sufficient. Rather, it shows that lateral information is selectively useful, label dependent, and architecturally non-trivial to exploit (Bertrand et al., 2019, Hashir et al., 2020).
4. PadChest as a benchmark for classification, hierarchy, and representation learning
PadChest has served as a testbed for methods that attempt to exploit its ontology rather than ignore it. In hierarchical multi-label classification, one study filtered the dataset to 121,242 samples by removing reports containing only “Unchanged,” selected 30 out of 191 labels aligned to a PLCO-derived abnormality taxonomy, and trained a DenseNet-121 with conditional then unconditional hierarchical objectives. On PadChest, the proposed HLUP-finetune model improved leaf-label performance over a strong flat baseline from AUC/AP $0.825/0.104$ to $0.837/0.145$, corresponding to the reported gains of AUC and AP (Chen et al., 2020).
PadChest also participates in recent foundation-model style benchmarks. In the MVMAE study, PadChest contributed 106,677 training studies, 1,653 validation studies, 1,601 test studies, 66,610 patients, and 158,522 images, with its original 193 labels harmonized into the CheXpert 14-category label space. Under the unified downstream evaluation, MVMAE achieved AUROC 89.19 on the PadChest test split, while MVMAE-V2T achieved the highest F1 at 35.38; in a controlled two-view subset, PadChest macro-AUROC improved from 71.49 to 77.38 for MVMAE when moving from one-view to two-view evaluation (Laguna et al., 27 Nov 2025).
Other PadChest studies use the dataset less as a leaderboard target than as a probe of model behavior. A shortcut-analysis paper filtered PadChest to 93,203 images after removing lateral views, discarding null or undesirable labels, and requiring CheXmask lung masks with Dice RCA (Mean) . On effusion, a DenseNet-121 trained on full images achieved AUC 0.86 when evaluated on images with the lungs masked out, but only 0.47 on images containing only the precise lung mask, indicating that high discrimination can persist outside the clinically expected ROI (Sourget et al., 2024). In external generalization experiments, PadChest also exposed the benefit of fine-grained supervision: on a 24,536-image frontal subset used only for testing, CheXDet improved fracture AUC from 0.55 to 0.78 relative to CheXNet, with 0 (Luo et al., 2021).
5. Derivative resources and extensions built from PadChest
PadChest has become an ecosystem rather than a single release. Later resources either curate more reliable annotations on top of it or repurpose it for new evaluation regimes.
| Resource | PadChest-based content | Primary purpose |
|---|---|---|
| PadChest-GR | 4,555 studies; 7,037 positive and 3,422 negative finding sentences; up to two box sets per positive sentence | Grounded radiology report generation (Castro et al., 2024) |
| CheXmask | 96,184 PadChest frontal instances with left lung, right lung, and heart masks plus RCA indices | Anatomical segmentation (Gaggion et al., 2023) |
| NEATX PadChest subset | 1,011 images labeled for chest drain, tracheostomy, NSG, and endotracheal tubes | Shortcut annotation and tube classification (Cheplygina et al., 2023) |
| PSF-Med PadChest portion | 4,534 images, 14,750 questions, 68,759 paraphrases | Paraphrase sensitivity in medical VLMs (Sadanandan et al., 24 Feb 2026) |
| CXR-LT 2026 | 142,928 PadChest training images; PadChest-GR development and test sets | Long-tailed and zero-shot classification (Dong et al., 25 Feb 2026) |
| ShoViR PadChest-GR subset | 1,390 unique images with 1,997 verified boxes across 9 CheXpert classes | Shortcut evaluation in report generation (Ruffini et al., 29 Jun 2026) |
PadChest-GR is the most direct grounded extension. It is a manually curated bilingual subset with one frontal image per study, optional linked prior studies for 31.7% of cases, sentence-level positive and negative findings in Spanish and English, progression labels, and manual bounding boxes for 88.3% of positive findings. CheXmask, by contrast, turns PadChest into a large segmentation resource by supplying left lung, right lung, and heart masks for more than 96,000 frontal images, together with per-mask RCA quality estimates (Castro et al., 2024, Gaggion et al., 2023).
The same dataset has also been adapted for reliability and robustness studies. NEATX adds non-expert tube annotations to a PadChest subset and reports Cohen’s 1 for tracheostomy and 2 for chest drain agreement with PadChest labels. PSF-Med uses PadChest for visual question answering stability, where model flip rates on PadChest range from 9.9% to 58.2% across six medical VLMs. SHOVIR further extends PadChest-GR with per-box CheXpert labels and controlled occlusion experiments to distinguish direct from contextual shortcut behavior in report generation (Cheplygina et al., 2023, Sadanandan et al., 24 Feb 2026, Ruffini et al., 29 Jun 2026).
6. Limitations, biases, and recurrent methodological caveats
Several later analyses emphasize that PadChest’s usefulness is inseparable from its failure modes. The most basic is label provenance: the dataset is large because most annotations are mined from reports rather than directly adjudicated on images. This is not a minor bookkeeping detail. Tube-focused analyses note that, within their PadChest subset, only 9.6% of tube labels were expert report-derived and 90.4% were ML-extracted, and they argue that a small number of expert annotations can be preferable to mixing expert and automatically extracted labels (Cheplygina et al., 2023).
A second caveat is domain shift. In a six-label cross-dataset study restricted to PA images, PadChest contributed 20,817 preprocessed images with 15,643 No Finding, 2,745 Cardiomegaly, 1,229 Lung Cancer, 605 Pleural Effusion, 547 Pneumonia, and 48 Pneumothorax positives. Models trained on PadChest performed strongly internally—for example, DenseNet161 reached AUROC/AUPRC 0.944/0.822—but external transfer could collapse: average AUPRC from PadChest to CheXpert fell from about 0.789 internally to about 0.193 externally, and average F1 fell from about 0.775 to about 0.273, often with sensitivity near 1 and specificity near 0 under transferred thresholds. The same study showed that a DenseNet-161 source-classification model could identify PadChest with sensitivity 0.999, specificity 1.000, and F1 1.000, indicating strong site-specific signatures (Rafferty et al., 18 Sep 2025).
A third issue is subgroup imbalance. In that same analysis, PadChest subgroup performance for DenseNet-161 was markedly worse in the under-represented age 3 group, with AUROC 0.682 and F1 0.339, compared with AUROC 0.966 and F1 0.720 for ages 40–64. Performance was also higher for males than females, with AUROC 0.958 versus 0.913 and F1 0.784 versus 0.751 (Rafferty et al., 18 Sep 2025).
Three misconceptions therefore require correction. PadChest is not a purely frontal-view dataset; it includes six view classes and underpins paired-view analyses (Bustos et al., 2019, Bertrand et al., 2019). Its labels are not uniformly image-level expert truth; they are largely report-derived, with 27% manual physician annotation in the original release (Bustos et al., 2019). And high internal performance on PadChest does not imply portability across institutions, thresholds, or demographic strata (Rafferty et al., 18 Sep 2025).
PadChest remains influential precisely because it combines scale, view diversity, hierarchical semantics, and rich metadata in a single public resource. Its continued value lies less in any single benchmark number than in the range of questions it permits: ontology-aware modeling, paired-view fusion, grounded report generation, anatomical segmentation, shortcut diagnosis, long-tailed recognition, and robustness under linguistic and institutional shift.