Medical imaging has more open data than almost any other domain, and more licensing traps. Three of the nine below carry terms worth reading before you train — CheXpert is gated and non-commercial, Dental Radiography is share-alike and Augmented Alzheimer MRI is LGPL — and terms like these are usually discovered late, after a model is already trained.
These are the medical datasets in the Dataset Finder catalogue that have both a stated licence and a stated size, so you can judge fit before downloading anything. Sizes and licence wording come from each publisher, except where the original release contradicts a re-upload's label; there the entry gives the original terms.
Read the licence column first. In this category it decides more than the size does.
| Dataset | Modality | Size | Licence | Best for |
|---|---|---|---|---|
| NIH Chest X-rays | Medical imaging | 112,120 x-rays | Attribution only… | Thoracic classification |
| CheXpert | Medical imaging | 224,316 chest x-rays | Gated access | Uncertainty-aware labels |
| Chest X-ray Masks and Labels | Medical imaging | 2,403 files | CC0 | Lung segmentation |
| Knee Osteoarthritis with Severity | Medical imaging | ~9,786 x-rays | CC BY 4.0 | Ordinal grading |
| Fracture X-Ray Multi-Region | Medical imaging | 10,580 images | ODC-By / PDDL | Fracture classification |
| Pneumonia Radiography | Medical imaging | ~4,929 images | CC BY 4.0 | Binary baseline |
| Augmented Alzheimer MRI | Medical imaging | 40,400 images | LGPL | Augmentation study |
| LUNA16 | Medical imaging | 888 3D CT scans | CC BY 4.0 | 3D nodule detection |
| Dental Radiography | Medical imaging | 1,272 images | CC BY-SA 4.0 | Detector formats |
NIH Chest X-rays
- Size
- 112,120 x-rays
- Licence
- Attribution only (no usage restrictions; link to the NIH download site, cite the CVPR 2017 paper and credit the NIH Clinical Center)
112,120 frontal chest x-rays from 30,805 patients, converted from DICOM to PNG and labelled for fourteen thoracic conditions. It is the volume benchmark for chest radiography: large enough to pre-train on, and widely enough used that your results are comparable to published work.
Multi-label thoracic classification, and as a pre-training corpus before fine-tuning on a smaller in-house set.
Source: innovatiana.com/datasets
CheXpert
- Size
- 224,316 chest x-rays
- Licence
- Gated · non-commercial
Stanford's chest x-ray corpus: 224,316 chest radiographs from 65,240 patients. Its distinguishing feature is that labels carry explicit uncertainty rather than forcing a binary call, which makes it the right choice if you care about calibration rather than raw accuracy.
Uncertainty-aware classification and calibration research. Access is by request and reserved for academic, non-commercial work.
Source: innovatiana.com/datasets
Chest X-ray Masks and Labels
- Size
- 2,403 files
- Licence
- CC0 1.0
2,403 files pairing chest x-rays in PNG, from the Montgomery County and Shenzhen tuberculosis collections, with lung masks in PNG; some images have no mask. Small, but it is one of the few chest datasets that ships pixel masks rather than image-level labels, and it is labelled CC0, though the source collections ask publications to credit the National Library of Medicine and Shenzhen No.3 People's Hospital.
Lung field segmentation, and as a pre-processing step for any pipeline that crops to the lungs before classifying.
Source: innovatiana.com/datasets
Knee Osteoarthritis with Severity
- Size
- ~9,786 x-rays
- Licence
- CC BY 4.0
Roughly 9,786 knee x-rays graded on the Kellgren–Lawrence scale from 0 to 4. The grades are ordinal, not categorical, which is the interesting part: treating them as five unrelated classes throws away the ordering and usually costs you accuracy.
Ordinal regression and severity scoring. A good dataset for testing whether your loss function respects ordered classes.
Source: innovatiana.com/datasets
Fracture X-Ray Multi-Region
- Size
- 10,580 images
- Licence
- ODC-By / PDDL (both stated)
10,580 x-rays spanning multiple body regions, labelled only fractured or not fractured and already split into train, test and validation. The uploader compiled it from three earlier Kaggle sets, and the page's licence field says PDDL while its description says ODC-By. The multi-region framing matters: a model trained only on wrists tends to collapse when shown an ankle.
Fracture classification that has to generalise across anatomy rather than specialise in one joint.
Source: innovatiana.com/datasets
Pneumonia Radiography
- Size
- ~4,929 images
- Licence
- CC BY 4.0 (original; this copy labelled CC0)
Around 4,929 paediatric chest x-rays from Guangzhou Women and Children's Medical Center (Kermany et al.), sorted into pneumonia and normal, pre-split into train, test and validation folders. The original release is CC BY 4.0, so credit it whatever this copy's CC0 label says. It is the standard first exercise in medical imaging, and useful precisely because so many published baselines exist to measure yourself against.
A binary classification baseline, or a smoke test for a new training pipeline before you point it at something larger.
Source: innovatiana.com/datasets
Augmented Alzheimer MRI
- Size
- 40,400 images
- Licence
- LGPL
40,400 brain MRI images in JPEG and PNG, containing both original and augmented data. Because the augmented copies are distinguishable from the originals, you can measure what augmentation is actually buying you rather than assuming it helps.
Dementia staging, and controlled experiments on how far synthetic augmentation carries a small medical dataset.
Source: innovatiana.com/datasets
LUNA16
- Size
- 888 3D CT scans
- Licence
- CC BY 4.0 (LIDC-IDRI images CC BY 3.0)
888 3D chest CT scans, a curated subset of LIDC-IDRI, from the LUng Nodule Analysis 2016 challenge. This is volumetric data, not slices, so it needs 3D convolutions or a slice-aggregation strategy — a genuinely different problem from 2D radiography.
3D nodule detection and false-positive reduction. The data is published on Zenodo under CC BY 4.0, with the underlying LIDC-IDRI images under CC BY 3.0.
Source: innovatiana.com/datasets
Dental Radiography
- Size
- 1,272 images
- Licence
- CC BY-SA 4.0
1,272 dental x-rays with 24 metadata columns, shipped in formats adapted for CNN, YOLO and SSD pipelines. The pre-formatting is the draw: you can point a standard detector at it without writing a converter first.
Dental pathology detection, and a quick comparison of detector architectures on identical inputs.
Source: innovatiana.com/datasets
How this list was put together
Every dataset here is listed in Innovatiana's open-dataset catalogue, and each entry links to its page. Only datasets whose licence and size are explicitly published are included — nothing is estimated.
Licence strings are quoted in the condensed form shown on the chips; several are prose rather than an SPDX identifier, so follow the source link and read the original before any commercial use.