Generative Vision Atlas

core

Brain Imaging Generation with Latent Diffusion Models

Walter H. L. Pinaya, Petru-Daniel Tudosiu, Jessica Dafflon · King's College London · 2022-09 · arXiv:2209.07162

Why it matters

Applies latent diffusion to 3D brain MRI conditioned on age, sex and brain structure volumes, and releases a synthetic dataset of 100,000 brain images. An early demonstration that the latent-diffusion recipe transfers to volumetric medical data.

What this paper does

read: abstract only

Before this

Deep models for medical imaging were limited by dataset size, and diffusion had only just been shown to work in a compressed latent for natural images.

The problem

Medical imaging datasets are small relative to what deep networks need, and 3D brain volumes are too large to model at full resolution in pixel space.

The idea

Apply latent diffusion to 3D brain MRI, conditioned on the clinical variables a researcher would actually want to control, then release the synthetic data so others can use it.

How it works

Trains on 31,740 T1-weighted MRI volumes from UK Biobank and conditions generation on age, sex and brain structure volumes. The abstract does not state whether the autoencoder is trained on the medical data or borrowed from natural images, and that ambiguity is worth flagging because it is exactly the question the atlas's medical section turns on.

Evidence

Reports realistic samples with effective conditioning control, and releases a synthetic dataset of 100,000 brain images. The evidence is distributional and qualitative; no downstream clinical task gain or radiologist reader study is reported.

Limitations

The abstract discusses no limitations. It was accepted to a MICCAI workshop rather than the main conference, which is consistent with preliminary findings.

Why it matters

An early demonstration that the latent-diffusion recipe transfers to volumetric medical data at all, and the released synthetic dataset made it a common starting point for later work.

Abstract, in the authors' own words

Deep neural networks have brought remarkable breakthroughs in medical image analysis. However, due to their data-hungry nature, the modest dataset sizes in medical imaging projects might be hindering their full potential. Generating synthetic data provides a promising alternative, allowing to complement training datasets and conducting medical image research at a larger scale. Diffusion models recently have caught the attention of the computer vision community by producing photorealistic synthetic images. In this study, we explore using Latent Diffusion Models to generate synthetic images from high-resolution 3D brain images. We used T1w MRI images from the UK Biobank dataset (N=31,740) to train our models to learn about the probabilistic distribution of brain images, conditioned on covariables, such as age, sex, and brain structure volumes. We found that our models created realistic data, and we could use the conditioning variables to control the data generation effectively. Besides that, we created a synthetic dataset with 100,000 brain images and made it openly available to the scientific community.

Research line

Medical: domain-trained latentascendant

Method note — the shared flow-matching interpolation

Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:

xt=(1−t) x0+t x1,vθ(xt,t)≈x1−x0x_t = (1-t)\,x_0 + t\,x_1, \qquad v_\theta(x_t, t) \approx x_1 - x_0

Builds on

Nothing recorded yet.

Built on by

Nothing recorded yet.