PadChest-GR is a public chest X-ray benchmark for grounded radiology report generation: each finding sentence is tied to boxes on the image, in Spanish and English. It is a dataset plus a task. It is not a datasets listicle. It is not a PYCAD product.
If you meant TCIA / MIMIC-CXR / MURA as a list → medical image datasets. If you meant how you draw the labels → image annotation tools. If you meant CAD on a real DICOM → AI for medical diagnosis.
505 already pointed here as the named-benchmark leftover. That split stands. PatchChestCT (a later chest-CT spatial-label paper on this site) is a different job.
PYCAD builds custom web DICOM viewers and imaging models. It does not host PadChest, sell a reporting SKU, or run this benchmark as a service.
The parent set, and this slice
PadChest (Bustos et al., 2020) is a large Spanish chest X-ray collection from Hospital Sant Joan d’Alacant, hosted under BIMCV: reports plus labels, not grounded boxes as the main artifact. BIMCV’s PadChest page is the parent.
PadChest-GR (Castro, Bustos, Bannur, Hyland, et al.; NEJM AI 2025; arXiv:2411.05085) takes a frontal subset, drops pediatric and originally-suboptimal studies, splits the report into single-finding sentences, translates Spanish → English, and has radiologists put boxes on every positive finding. GPT-4 was used as a preprocessing aid (split / translate / link to existing PadChest labels / progression). Fourteen radiologists at Hospital Sant Joan d’Alacant then discarded bad studies and drew the boxes. The paper is the source for every count below. Files on the BIMCV drop are PNG, not DICOM — window/level is already baked in.
What is in the box
| PadChest-GR | |
|---|---|
| Studies | 4,555 chest X-rays (3,099 abnormal / 1,456 normal). Acquired 2014–2017. Median age 69. Roughly balanced sex |
| Sentences | 7,037 positive-finding sentences; 3,422 negative-finding sentences. Each study has the full present/absent lists in both Spanish and English |
| Grounding | Every positive sentence has up to two independent bounding-box sets (two readers). Categorical labels: finding type, location, progression |
| Priors | A prior image exists for 31.7% of studies (higher in the abnormal slice) |
| Task | Grounded radiology report generation (GRRG): write the findings and say where on the image |
A fluent English paragraph that never touches the image can still win a BLEU-style score. GRRG is the opposite test: if you said “left pleural effusion,” the box has to land on that effusion. Negative sentences (“no pneumothorax”) are first-class; they are not a leftover “normal” token.
How it is not MIMIC-CXR
MIMIC-CXR is the English-report workhorse: a large MIMIC-linked chest X-ray set with free-text reports. It taught the field to generate radiology language. It does not, as a default task, give you two-reader boxes on every positive sentence in two languages. PadChest-GR is smaller on purpose. Do not flatten it into the 505 list as “another CXR dump.”
The paper’s published baseline is MAIRA-2: logical precision / recall 0.560 / 0.514 and grounding precision / recall 0.802 / 0.766 under the RadFact metric they introduce. Those are one run, not a league table. Cite the paper; do not turn them into a product claim.
What to do with it (and not)
- Train or test GRRG. Sentence + box + bilingual pair is the point.
- Measure hallucination. A finding with no box, or a box on the wrong lobe, is a failure even if the prose is pretty.
- Do not treat it as a hospital. One site, one era, frontal only, adults. Site shift is still the boss.
- Do not invent a PYCAD PadChest model. A viewer can display boxes. That is not this benchmark as a SKU.
Paper: arXiv:2411.05085. Journal version: NEJM AI AIdbp2401120. Microsoft Research also wrote a publication page. Download (on request): bimcv-projects/padchest-gr. Read the license before you train.