BrainNet

Deep learning models for different tasks related to analysis of brain anatomy. Currently, two model are included:

• TREGA which estimates an affine transformation to MNI152 or MNI305 space.
• TopoFit which estimates cortical surfaces from (almost) any MRI scan.

A Note on Expected Input Orientation and Resolution

The networks expect the input scan(s) to be "conformed" meaning that it has to be (1) 1 mm isotropic resolution and (2) RAS oriented. Effectively, this means that the linear part of the affine transformation matrix should a 3x3 identity matrix (or very close to). If this is not the case, the scan can be automatically resampled to this orientation by specifying the --conform flag by doing brainnet --conform [your command]. In the examples below, we assume that the input scan is already "conformed".

Usage

TREGA

Estimate an affine registration from image space to MNI152 or MNI305.

brainnet trega image.nii.gz ./output_dir

TopoFit

Estimate cortical surfaces, uncertainty, and spherical registration.

brainnet topofit image.nii.gz ./output_dir

Default is to use the model trained on 1 mm isotropic T1w images. To specify models based on other contrasts and resolutions, use -c and -r flags, e.g.,

brainnet topofit -c synth -r random image.nii.gz ./output_dir

to use a model trained on synthetic data with random resolution.

Installation

Basic Installation

brainnet is not available on PyPI (yet). Basic installation (enough to do inference) is simple, however, for now you need to handle some dependencies manually. For example,

pip install https://github.com/simnibs/cortech/releases/download/v0.1/cortech-0.1-cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl
pip install brainsynth@git+https://github.com/simnibs/brainsynth@v0.1
pip install brainnet@git+https://github.com/simnibs/brainnet@v0.2

For an editable installation, do (with similar dependency handling as above)

git clone
cd brainnet
pip install -e .

Building the included CUDA Extension

For training we rely on a few CUDA extensions which at the moment are not build as part of the package installation and so needs to be build manually. To do this, cd into /brainnet/mesh/cuda and execute python build.py build_ext --inplace (having installed the proper CUDA libraries, torch, etc.)

# get cuda version used to compile torch
python -c "import torch; print(torch.version.cuda)"

# and use this here
conda install -c conda-forge cudatoolkit=[torch.version.cuda]

Kaolin (currently not used)

I have considered using Kaolin instead of the included CUDA extensions, however, we currently do not use Kaolin.

# torchvision is for kaolin
# torchaudio==2.6.0
pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu124
# kaolin should match torch and cuda versions
pip install kaolin==0.18.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.6.0_cu124.html