zeahealth/medstack-medimage-medtech-v1

MedStack Medical Image-Gen Corpus

License-clean, caption-conditioned medical images for training an SDXL per-cluster LoRA image generator. Built 2026-07-26 · version v0.2 · 906 images.

License & provenance

Tier A (CC0 / PD, any jurisdiction incl. US/India/global government PD) and Tier B (CC BY + government-open licenses: India GODL, UK/national OGL, NLOD) only — CC BY-SA / NC / ND / GFDL / research-only are excluded because a generator emits derivatives. Each row carries its own license/license_tier; LICENSE-AUDIT.json is the machine-readable audit and ATTRIBUTIONS.csv is the CC-BY NOTICE. By tier: {'B': 480, 'A': 426}.

Sources: Wikimedia Commons (per-file license filter), PMC Open Access (oa_comm, CC BY/CC0).

Privacy & governance (HIPAA / GDPR / India DPDP Act)

• Captions are de-identified (HIPAA 18 identifiers + personal data); ages >89 generalized.
• Identifiable-face / clinical-portrait figures are dropped.
• Images are re-encoded to strip EXIF/GPS/timestamps.
• All rows are curated source images, marked ai_generated: false. Generated outputs are separately watermarked + C2PA-tagged (see the generation script).

Composition

• Clusters: {'medtech': 906}
• Modalities (incl. microscopy/pathology/cytogenetics/omics + non-radiology capture: fundus, OCT, audiometry, spirometry, EMG, polysomnography, semen-analysis, HSG, …): {'device': 906}
• Specialties (adult + pediatric, all surgical/medical subspecialties + basic sciences + emergency/precision/regenerative/nuclear medicine, clinical trials, telemedicine, patient safety, ethics, plus settings/people/objects — facilities & hospitals, medical education & simulation, labs, pharma manufacturing, workforce, apparel/scrubs, smart health devices): {'physiology': 3, 'medical-devices': 134, 'infectious disease': 27, 'pediatric nutrition / dietetics': 2, 'healthcare workforce': 12, 'smart health devices / equipment': 10, 'toxicology': 1, 'nephrology': 2, 'pulmonology': 1, 'radiology / diagnostic imaging': 3, 'neurology': 1, 'audiology / speech-language pathology': 2, 'otolaryngology': 5, 'obstetrics': 2, 'transfusion medicine': 1, 'tuberculosis': 1, 'biomedical engineering / devices': 62, 'operation theatre / surgical instruments': 20, 'rehabilitation': 32, 'anesthesiology / critical care': 3, 'laboratory science / research labs': 36, 'cardiology': 9, 'anatomy': 4, 'emergency medicine': 1, 'healthcare facilities / infrastructure': 3, 'nuclear medicine': 1, 'surgical-instruments': 47, 'forensic medicine': 1, 'neurosurgery': 2, 'dentistry': 2, 'orthopedic surgery': 8, 'disaster medicine / pandemic response': 1, 'general surgery': 1, 'implants': 29, 'plastic surgery': 13, 'pharmaceutical manufacturing': 29, 'bariatric surgery': 1, 'telemedicine / digital health': 3, 'regenerative medicine': 4, 'pathology': 1, 'hematology': 3, 'breast surgery': 1, 'prosthetics': 19, 'pediatric rehabilitation': 5, 'colorectal surgery': 2, 'pharma-process': 118, 'oncology': 4, 'clinical trials & research': 11, 'pediatric pharma-process': 5, 'public health / preventive medicine': 2, 'microbiology': 4, 'lab-equipment': 206, 'medical apparel / uniforms': 1, 'biochemistry': 1, 'pediatric infectious disease': 1, 'pediatrics': 1, 'pediatric lab-equipment': 1, 'ophthalmology': 1}
• Systems of medicine: {'allopathy / modern medicine': 906}
• Representation / visual medium (AR/VR/MR, 2D/3D model, simulation, avatar, illustration, photograph): {'unknown': 894, 'photograph': 6, 'simulation': 3, '3D model / render': 2, 'medical illustration': 1}
• Pediatric images: 19
• Views/planes/sequences: {'unknown': 904, 'dwi': 1, 'axial': 1}
• Regions: {'global': 906} (India-focused AYUSH cluster included)
• Normal/abnormal: {'unknown': 877, 'abnormal': 29} · severity: {'unknown': 894, 'severe': 5, 'mild': 6, 'moderate': 1}
• Rarity: {'common': 899, 'rare': 7} · edge-cases: 0
• Diversity (age/sex/ethnicity): {'by_age_group': {'unknown': 887, 'infant': 6, 'child': 13}, 'by_sex': {'unknown': 894, 'female': 4, 'male': 8}, 'by_ethnicity': {'unknown': 899, 'caucasian': 2, 'indian': 2, 'middle eastern': 3}}
• Languages: ['en']

Clinical metadata schema (standardized taxonomy)

See clinical_metadata_schema.json. Each row carries: modality (+ DICOM code in taxonomy), anatomy/body_system, finding, diagnosis, specialty, view (AP/PA/lateral/axial/ coronal/sagittal/T1/T2/FLAIR/DWI), normal_abnormal, severity (normal/mild/moderate/severe), edge_case, rarity, age_group/sex/ethnicity (explicit-only, never inferred), system_of_medicine (allopathy/ayurveda/siddha/unani/homeopathy/yoga/naturopathy/acupuncture-TCM), representation (AR/VR/MR · 3D model/render · simulation · avatar · 2D diagram · illustration · photograph), pediatric (bool + specialty prefix), region, license/license_tier, multi-language captions, dataset_version + validation_status.

Provenance & auditability

Every row records its source/source_url, retrieved date, dataset_version, and a validation_status (starts pending-expert-review). Generated images additionally carry a C2PA-style manifest (prompt, model + LoRA version, seed, watermark) — see the generation script. LICENSE-AUDIT.json is the machine-readable composition + license audit.

Bias / fairness

Diversity is reported across age_group, sex, ethnicity, region, modality, specialty, and severity in LICENSE-AUDIT.json; the eval harness scores quality + hallucination per group. Open-access literature is not epidemiologically representative — treat counts as coverage, not prevalence. Ethnicity is recorded only when explicitly stated; we do not infer race.

Intended use & limitations

Educational / illustrative synthetic medical imagery and research. Not a medical device; not for diagnosis. Captions derive from source metadata/figure legends and may be terse; class balance reflects the open-access literature and is not epidemiologically representative. Validate with the medical-expert review packet (expert-review/).

Models trained on this dataset inherit the SDXL base's CreativeML Open RAIL++-M use restrictions (Attachment A): they may NOT be used "to provide medical advice and medical results interpretation," nor for any other Attachment A restricted use. Synthetic image generation for education / illustration / research is permitted; clinical diagnosis, screening, and interpreting real patient results are not.

Regulatory

See docs/medstack-foundry/medical-image-gen-compliance.md for the HIPAA / GDPR / DPDP / EU AI Act (Art. 50 synthetic-media transparency) / India AI-governance mapping.

SAHI — Governance Pack (India), pilot-stage

Alignment of this image-generation curation dataset with the Strategy for AI in Healthcare for India (MoHFW):

• Ethics: Synthetic-only, license-clean training data; no real PHI; no patient likeness; not a medical device and not for diagnosis.
• Transparency: Public model + dataset cards, full data-license lineage, and a C2PA-style provenance manifest declaring AI-generation on every image (EU AI Act Art. 50 aligned).
• Accountability: Human-in-the-loop required — the medical-expert review packet must sign off before any downstream use; outputs are educational/illustrative material at most.
• Privacy: India DPDP Act 2023 aligned (alongside HIPAA / GDPR) — face-filter + de-identification pass; no PHI in the corpus.
• Equity / public-health alignment: First-class AYUSH (Ayurveda / Siddha / Unani / Homeopathy / Yoga / Naturopathy) + India-weighted modality coverage; counts treated as coverage, not epidemiological prevalence.