Medical diffusion models still in the waiting room

What it does The repository provides network architectures, noise schedulers, loss functions, and metrics tailored for medical imaging workflows. It wraps these into MONAI-style inferer classes and an Ignite-based trainer so researchers can train and sample from diffusion models, VQ-VAEs, and autoregressive transformers on 2D and 3D clinical data. A heavy tutorial footprint walks through conditional generation, anomaly detection, inpainting, and MRI reconstruction.

The interesting bit Most diffusion code assumes natural images; this one ships with 3D-aware perceptual losses using RadImageNet and 3DMedicalNet, plus MS-SSIM and FID metrics ready for volumetric data. The explicit “prototyping repository” label is refreshing honesty—it means the API is still fluid and the maintainers expect things to break or migrate.

Key highlights

• Diffusion, latent diffusion, and VQ-VAE + transformer inferers built in MONAI style
• Noise schedulers covering DDPM, DDIM, and PNDM
• Perceptual losses for both 2D and 3D data via LPIPS, RadImageNet, and 3DMedicalNet
• Tutorials spanning anomaly detection, super-resolution, inpainting, and classifier-free guidance
• MONAI-compatible trainer engine based on Ignite

Caveats

• The README labels this a “prototyping repository,” so expect APIs to shift before anything lands in MONAI core
• Long-term integration into MONAI core is planned but not scheduled beyond the milestone tracker

Verdict Worth a look if you are a medical-imaging researcher who wants MONAI-native diffusion or VQ-VAE primitives. Skip it if you need battle-tested, stable generative infrastructure today.