Generate synthetic medical images with medigan is a 2024 PYCAD Team how-to: call the open-source medigan library, pick a pretrained GAN, write PNG/array samples. It is not a PYCAD synthetic-data product. It is not a dataset marketplace.
If you meant what imaging ML actually returns → machine learning for medical imaging. If you meant annotation of real studies → medical image annotation. If you meant public dataset lists that is a different leftover, not this URL.
Early imaging MVPs stall on data. Public sets are the wrong anatomy, a bought set is slow, and a hospital partner is not in the room yet. Synthetic samples are one way to stand up a pipeline before the real studies exist. They are not a substitute for a site’s own labeled hold-out when you go to clinic.
What medigan is
medigan (Osuala et al., Frontiers in Oncology 2022, doi:10.3389/fonc.2022.1044496) is a Python wrapper around pretrained generative models — mostly GANs — for medical images. The 2022 paper listed 21 models, 9 GAN architectures, 11 datasets, covering mammography, endoscopy, X-ray, and MRI. The catalogue grows; check the repo for the current IDs. The library is framework-agnostic: you ask for a model_id and a count.
A GAN is two nets: a generator that fakes an image, a discriminator that tries to tell fake from real. At convergence the generator’s fakes are supposed to be hard to call. That is the method. It does not make a heart on the wrong side “creative.” If the anatomy is wrong, the sample is trash.
The call
Install is pip install medigan.
from medigan import Generators
generators = Generators()
# model_id=1 is a mammography GAN in the 2022 catalogue — confirm IDs in the repo
generators.generate(model_id=1, num_samples=6, install_dependencies=True)
install_dependencies=True pulls whatever that model needs the first time. You get arrays or files you can feed a segmentation or classification loop. Six samples is a smoke test, not a training set.
Why you would bother
- Class balance. Rare positives are the usual failure. A GAN trained on the rare class can oversample it. The discriminator still has to have seen real rares.
- A pipeline before the partner. You can wire loaders, augment, and a dummy train on fakes. You cannot claim a clinical AUC on them.
- Sharing without PHI. Weights and fakes are not the hospital’s DICOM. They are also not automatically de-identified forever — a model can memorize. Treat them as derived data, not as a waiver.
Training your own mammography GAN on a private set is a different script. One public example from the same circle: zuzaanto/mammo_gans_iwbi2022. That is their training loop, not a PYCAD SKU.
What this page is not
- Not a PYCAD synthetic-data platform, GAN service, or dataset store. The library is RichardObi’s. The note is a how-to.
- Not 673. Imaging ML is the mark on a real study. This page is how you mint extras when the real studies are thin.
- Not
/portfolio. Dropped.
If the fakes are only there to stand up a viewer or a training loop you will later swap for site data, that is the imaging piece. Case studies.
Repo: github.com/RichardObi/medigan. Paper: Osuala et al., 2022, as linked above.