https://atlas.rsna.org/schemas/2025-11/model.json

Brain Imaging Generation with Latent Diffusion Models

https://arxiv.org/abs/2209.07162

v1

WHLP and MJC are supported by Wellcome Innovations [WT213038/Z/18/Z]. PTD is supported by the EPSRC Research Council, part of the EPSRC DTP, grant Ref: [EP/R513064/1]. JD is supported by the Intramural Research Program of the NIMH (ZIC-MH002960 and ZIC-MH002968). PFDC is supported by the European Union’s HORIZON 2020 Research and Innovation Programme under the Marie Sklodowska-Curie Grant Agreement No 814302. PN is supported by Wellcome Innovations [WT213038/Z/18/Z] and the UCLH NIHR Biomedical Research Centre. This research has been conducted using the UK Biobank Resource (Project number: 58292).

This study explores using Latent Diffusion Models to generate synthetic high-resolution 3D brain images conditioned on age, sex, and brain structure volumes. A synthetic dataset of 100,000 brain images was created and made openly available to the scientific community.

Latent Diffusion Models (LDM) combining autoencoders for compression into a lower-dimensional latent representation and diffusion models for generative modeling. The autoencoder maps brain images to a latent representation of 20 × 28 × 20. The diffusion model converts Gaussian noise into samples via an iterative denoising process over 1000 steps. A hybrid conditioning approach combines concatenation of conditioning variables with input data and cross-attention mechanisms.

The synthetic dataset of 100,000 brain images generated by this model is openly available at Academics Torrents, FigShare, and HDRUK Gateway.

Addresses the limitation of small dataset sizes in medical imaging by generating synthetic data, complementing training datasets, and enabling medical image research at a larger scale without privacy concerns.

Fully automated generation of synthetic high-resolution 3D brain MRI images.

Generation of synthetic high-resolution 3D brain MRI images for research purposes, data augmentation for machine learning model training, and exploring probabilistic distributions of brain images.

High-resolution 3D T1w MRI images of the brain. Conditioning variables include age, sex, ventricular volume, and brain volume normalized for head size.

The model can be conditioned on age, sex, ventricular volume, and brain volume normalized for head size to control data generation. Input images are pre-processed using rigid body registration to a common MNI space, with a 1 mm3 voxel size, and cropped to a volume of 160 × 224 × 160 voxels.

While LDMs demonstrated more stable training and easier convergence compared to GAN-based baselines, generative models can still face challenges in replicating essential finer details if image resolution is too low. Extrapolation of conditioning variables outside the training range can lead to abnormally huge ventricles or signs of neurodegeneration.

LDMs demonstrated more stable training and easier convergence when compared to GAN-based baselines (VAE-GAN, LSGAN) for high-resolution 3D images. The use of the DDIM sampler reduced sampling time from 142.3 ± 1.6s per sample to 7.6 ± 0.2s per sample with minimal performance loss.