Public datasets and open-source tools

In short: Where public, de-identified cancer imaging lives, and the rules that come with it.

The Cancer Imaging Archive

TCIA hosts collections of de-identified medical images, mostly DICOM, often with clinical data and annotations. Each collection has a data usage policy and a citation requirement.

NCI Imaging Data Commons

IDC makes large amounts of public cancer imaging available in the cloud with searchable metadata and built-in viewers, alongside other NCI data commons.

Good practice

• Read and follow each collection’s usage policy and citation.
• Do not try to re-identify anyone.
• Report the exact collection version you used.

In short: How public genomic resources are organised and how some imaging collections link to them.

Open and controlled tiers

The NCI Genomic Data Commons holds data from programmes such as TCGA. Summary and some derived data are open; individual-level sequence data are controlled and require approval. RadioGx uses only open data.

cBioPortal

cBioPortal presents mutations, copy-number and expression across many studies in an interactive way. It is genomics and clinical data, not imaging.

Linkage through case identifiers

TCIA collections named after TCGA projects contain imaging for a subset of TCGA cases. The shared case identifier links imaging to genomic data in the GDC. Not every case has every data type.

In short: Why "this dataset has imaging and genomics" needs checking case by case.

Things that break pairing

• Only a subset of cases has both data types.
• Imaging and tissue were obtained at different times or treatment states.
• The biopsy came from a different lesion or region than the one imaged.
• Labs and scanners changed during data collection.

RadioGx’s rule

RadioGx says a dataset has paired genomic data only when its documentation says so, labels linkage-by-identifier separately, and otherwise says "not documented".

In short: Why some data need approval, why de-identification is hard for images, and why RadioGx accepts no uploads in the workspace.

Controlled access

Resources such as dbGaP, the European Genome-phenome Archive, UK Biobank and ADNI protect participants by requiring applications and data use agreements. Public tools must not route around them.

Images can identify people

DICOM headers can hold names, dates and identifiers. Pixels can contain burned-in text. Head CT and MRI can be rendered into recognisable faces, which is why public neuroimaging is often "defaced". The DICOM standard describes de-identification profiles.

RadioGx’s stance

The workspace takes no file uploads and rejects inputs that look like identifiers. The optional Imaging Workbench keeps files in browser memory only, warns about headers and never sends files anywhere.

In short: Stars measure attention, not validity. Here is what to look at instead.

A short checklist

• Licence: is there one, and what does it allow?
• Maintenance: recent releases, issue responses, documented maintainers.
• Documentation: installation, examples, limitations.
• Testing: automated tests and continuous integration.
• Citation: a CITATION file or paper to cite.
• Provenance of included models or data, with their own licences.

Community references

RadioGx labels repositories as community developer references. They are not curated medical knowledge, not clinical evidence and not validated devices.

In short: Permissive, copyleft and no licence at all, plus the separate licences that govern datasets and models.

Code licences

Permissive licences (MIT, BSD, Apache 2.0) allow reuse with attribution. Copyleft licences (GPL family) require derived works that are distributed to carry the same licence. Code with no licence is, by default, all rights reserved.

Data and model licences

Datasets and trained models carry their own terms, often Creative Commons licences or custom usage policies. A permissively licensed tool does not make its bundled data or model free to use.

In short: A map of widely used open tools: reading DICOM, processing images, training models, extracting features and viewing.

Reading and writing

pydicom (Python) and dcmjs (JavaScript) read and write DICOM. NiBabel handles neuroimaging formats such as NIfTI.

Processing

SimpleITK and ITK provide registration, filtering and segmentation; scikit-image and OpenCV provide general image processing.

Learning and features

MONAI and TorchIO support deep learning on medical images; nnU-Net configures segmentation networks automatically; PyRadiomics extracts radiomic features.

Viewing

OHIF is a web DICOM viewer; 3D Slicer is a desktop platform; Cornerstone is a rendering library many viewers build on.

In short: Public, de-identified data only; no PHI uploads; reproducible environments; and a clear line between research code and medical devices.

Rules of thumb

• Use public, de-identified data for learning and research prototypes.
• Never upload identifiable patient data to third-party tools or AI services.
• Pin software versions and record parameters so results can be reproduced.
• Treat model outputs as research outputs that need validation.
• Follow your institution’s governance and ethics approvals.