BrainST: Structural Volume–Guided Diffusion Modelling for Counterfactual Brain MRI and Longitudinal Prediction

BrainST is a diffusion-based framework for the controlled synthesis, anatomical transformation, and longitudinal prediction of T1-weighted brain MRI, trained entirely on cross-sectional data. It enables fine-grained, region-specific control by conditioning image generation on volumetric measurements of 18 brain regions of interest (ROIs), while preserving anatomical plausibility through a conditioning alignment penalty.

Features

• Cross-sectional MRI generation from manually specified or automatically predicted ROI volumes
• Localized anatomical transformations of existing images while preserving subject-specific anatomy
• Longitudinal prediction of brain changes associated with healthy aging or neurodegenerative diseases (e.g., Alzheimer’s disease)
• Automatic ROI volume prediction from demographic and cognitive variables
• Counterfactual image synthesis with precise region-specific control

BrainST is designed to support research on brain aging, neurodegeneration, and structural variability, particularly when longitudinal MRI data are limited.

Installation

# Clone the repository
git clone https://github.com/AgustinCartaya/BrainST.git
cd BrainST

# (Optional) Create a virtual environment
conda create --name BrainST python=3.11
conda deactivate
conda activate BrainST

# Install required packages
pip install torch torchvision monai tensorboard ipykernel tqdm matplotlib opencv-python pandas nibabel scikit-image scikit-learn SimpleITK ipympl