scalesufer

scalesurfer is a repository for fast FreeSurfer inference. The volumetric model uses UNet with a Transformer bottleneck. The stats models predict FreeSurfer measures, including cortical thickness, surface area, curvature, and folding index.

Installation

pip install scalesurfer

Usage

See the inference notebook for additional settings for faster processing.

from scalesurfer import ScaleSurfer

# Anatomical images
adni_dir = "/home/rph/scalesurfer/data/adni_bids"
subjects = ["sub-002S0559", "sub-002S0619"]

anat_files = [
    f"{adni_dir}/sub-002S0559/ses-20060627/anat/sub-002S0559_ses-20060627_T1w.nii.gz",
    f"{adni_dir}/sub-002S0619/ses-20060601/anat/sub-002S0619_ses-20060601_T1w.nii.gz"
]

# Predict aparc+aseg and stats tables
surfer = ScaleSurfer(anat_files, subjects, "/tmp/scalesurfer_subjects", device="cuda")
surfer.prepare_images()
surfer.predict_volumes()
surfer.plot_volume(subjects[0])
df_stats = surfer.predict_stats()

GPU

All models are implemented with pytorch and inference time depends on GPU. CPU-based inference will be much slower. These models were developed on an NVIDIA card with 32 GB of VRAM and 64 RAM. With this hardware, the settings in the inference notebook worked well. Please open an issue if inference fails on your hardware.