A chest X-ray dataset that refuses to be a Kaggle competition

What it does This repository curates chest X-ray and CT images from COVID-19, SARS, MERS, ARDS, and other pneumonia cases, packaging them with metadata and a hierarchical label taxonomy. It is intended as a public resource for developing machine learning models that assist with triage and severity prediction rather than diagnostic scoreboards. The collection includes third-party annotations such as lung bounding boxes, VAE-generated masks, Brixia severity scores, and COCO-format segmentations.

The interesting bit The maintainers are refreshingly allergic to hype, explicitly warning that no one should claim diagnostic performance without a clinical study and stressing that this is not a Kaggle competition dataset. That self-awareness is paired with an ethics committee approval and a pragmatic focus on prognostic tasks—intubation, survival, and supplemental oxygen needs—over binary classification bragging rights.

Key highlights

• 481 PA/AP and 173 AP Supine view samples with hierarchical labels spanning COVID-19, bacterial, viral, and other pneumonia types
• Multiple external annotation layers including lung bounding boxes, generated lung segmentations, Brixia scores, and COCO-format masks
• Linked to the open-source Chester AI Radiology Assistant for prototyping
• University of Montreal ethics-approved (CERSES-20-058-D) with per-image licensing
• Ongoing data collection from public sources, hospitals, and published literature

Caveats

• Total image counts are modest (under 700), limiting broad generalization without external validation
• Licenses vary per image (Apache 2.0, CC BY, CC BY-NC-SA), so commercial reuse requires careful license auditing
• The project explicitly discourages benchmark-chasing and warns against unvalidated diagnostic claims

Verdict Clinicians and medical imaging researchers building triage or severity-prediction workflows should start here; data scientists seeking a massive, competition-ready, uniformly licensed corpus should look elsewhere.