cat-surf — Python bindings for CAT-Surface

cat-surf provides Python access to CAT-Surface, a mature C/C++ toolkit for surface-based neuroimaging analysis, focusing on the processing and analysis of cortical surface meshes.

CAT-Surface has been used internally for more than 10 years as part of CAT12, the SPM toolbox for computational anatomy. With this package, it is now distributed as an independent Python package for direct integration into external workflows to make CAT-Surface easier to use:

• in T1Prep workflows (currently relying on CAT-Surface binaries)
• in Python-based pipelines without subprocess-heavy wrappers
• in reproducible environments via published wheels

Installation

pip install cat-surf

Pre-built wheels are available for:

• macOS — arm64 (Apple Silicon), x86_64 (Intel)
• Linux — x86_64, aarch64 (manylinux)
• Python 3.9 – 3.13

Building from source

The extension modules are generated from the .pyx sources at build time; the intermediate .c files are build artifacts and are not tracked in the repository. Cython is declared in build-system.requires, so any PEP 517 build installs it for you:

# build libCAT first (see the CAT-Surface README), then:
pip install ./cat_surface_cython

For an in-place development build, Cython has to be in the ambient environment and libCAT has to be findable:

pip install cython
CAT_BUILD_DIR=/path/to/CAT-Surface python setup.py build_ext --inplace

Note that this regenerates every .c file, so a Cython version that differs from the one used elsewhere changes all of them — harmless now that they are untracked, but it is why they are not kept in git.

Basic usage

import cat_surf

# Check version
print(cat_surf.__version__)

# Load a surface file (GIFTI, FreeSurfer, BIC/MNI formats)
vertices, faces = cat_surf.read_surface("lh.central.gii")

# Per-vertex area
area, total_area = cat_surf.get_area(vertices, faces)

# Euler characteristic
chi = cat_surf.euler_characteristic(vertices, faces)

# Smooth per-vertex data (heat kernel, FWHM in mm)
smoothed = cat_surf.smooth_heatkernel(vertices, faces, area, fwhm=20.0)

Surface operations

Volume operations

Thin structures and the shrinking bias

Every isotropic regularizer — a local median, a Potts MRF, total variation — penalizes boundary area, and a thin structure has an extreme area-to-volume ratio, so deleting it is always the cheaper labelling. That is why one and the same median filter opens glued sulci in one place and closes cerebellar fissures in another, and why tuning its strength only trades one failure for the other.

vol_sheetness replaces the smoothness prior with a shape prior read off the Hessian eigenvalues, so it keeps thin sheets, ignores blobs, and shrinks nothing. Everything below is built on it, and all of it is a no-op where the sheetness is zero — the oriented operators are numerically identical to the isotropic ones they replace away from thin structures, so they are safe to switch on:

import cat_surf, nibabel as nib

t1  = nib.load("t1_corr.nii").get_fdata().astype("float32")
lab = nib.load("label.nii").get_fdata().astype("float32")
vx  = nib.load("label.nii").header.get_zooms()[:3]

# Inspect the evidence first: dark sheets are sulcal CSF, bright ones WM blades
sheet, normal = cat_surf.vol_sheetness(t1, voxelsize=vx, polarity=-1,
                                       return_normal=True)

# A median that cannot close a sulcus (orientation taken from the intensity)
clean = cat_surf.vol_oriented_median(lab, guide=t1, voxelsize=vx)

# Recover sulcal CSF the classifier discarded, then run PBT on the repaired map
lab = cat_surf.vol_sulcus_repair(t1, lab, voxelsize=vx, verbose=True)
gmt, ppm, _, _ = cat_surf.vol_thickness_pbt(lab, voxelsize=vx,
                                            oriented_filter=True)

Volume input convention

Every volume-consuming function accepts three input forms interchangeably:

import cat_surf, nibabel as nib
import numpy as np

# 1. file path
v, f = cat_surf.vol_marching_cubes("brain.nii.gz", threshold=0.5)

# 2. (ndarray, affine) tuple — supply a 4×4 RAS+ affine
v, f = cat_surf.vol_marching_cubes((array, affine), threshold=0.5)

# 3. any nibabel-image-like object (.affine + .get_fdata())
v, f = cat_surf.vol_marching_cubes(nib.load("brain.nii.gz"), threshold=0.5)

The data is auto-converted to float32 (matching the C library's expectation), and 4-D series default to the middle frame. No copy is made if the array is already float32/Fortran-order. Internally a minimal nifti_image is built from the affine — only the fields libCAT actually reads (dims, voxel sizes, sto_xyz).

Registration

Basic BBR usage

import cat_surf
import nibabel as nib

# Load surfaces (GIFTI or any format supported by cat_surf.read_surface)
lh_verts, lh_faces = cat_surf.read_surface("lh.white.surf.gii")
rh_verts, rh_faces = cat_surf.read_surface("rh.white.surf.gii")

# Full pipeline: NMI init from T1w reference, then BBR
matrix, cost = cat_surf.bbreg(
    "bold_mean.nii.gz",
    lh_surface=(lh_verts, lh_faces),
    rh_surface=(rh_verts, rh_faces),
    ref_file="T1w.nii.gz",     # NMI initialisation
    verbose=True,
)
print(f"BBR cost: {cost:.4f}")
print("EPI → T1 matrix:\n", matrix)

# Save the 4×4 transform for use with FSL / ANTs
import numpy as np
np.savetxt("epi_to_t1.txt", matrix)

Standalone volume registration

# Cross-modal (EPI ↔ T1w) — NMI
matrix, nmi = cat_surf.volume_register_nmi("T1w.nii.gz", "bold_mean.nii.gz")

# Same-modality (T1w ↔ T1w) — robust IRLS
matrix, res = cat_surf.volume_register_robust("t1_ref.nii.gz", "t1_moving.nii.gz")

Contrast auto-detection

# 0 = T1/FLAIR, 1 = T2/BOLD, -1 = undetermined
contrast = cat_surf.bbreg_detect_contrast(
    "bold_mean.nii.gz",
    lh_surface=(lh_verts, lh_faces),
    rh_surface=(rh_verts, rh_faces),
)

Conversion utilities

cat_surf.cli — drop-in replacement for the CAT binaries

The cat_surf.cli subpackage mirrors the CAT_* command-line binaries one-to-one: same names (snake_case, CAT_ prefix dropped), same positional argument order, same option semantics and defaults. Each function reads its inputs from disk, calls the in-memory numpy wrapper, and writes the outputs — ideal for porting shell scripts to Python.

from cat_surf import cli

# CAT_VolMarchingCubes brain.nii.gz brain.gii -thresh 0.5
cli.vol_marching_cubes("brain.nii.gz", "brain.gii", threshold=0.5)

# CAT_Surf2PialWhite -remove_intersect central.gii thickness.txt labels.nii \
#     pial.gii white.gii
cli.surf2pial_white("central.gii", "thickness.txt", "labels.nii",
                    "pial.gii", "white.gii", remove_intersect=True)

# CAT_SurfDistance -mean surf1.gii surf2.gii out.txt
cli.surf_distance("surf1.gii", "surf2.gii", "out.txt", mode="mean")

# CAT_VolSanlm in.nii out.nii -strength 1.0
cli.vol_sanlm("in.nii", "out.nii", strength=1.0)

The full mapping:

For composable in-memory pipelines, prefer the lower-level cat_surf API directly — the CLI shims are just thin convenience wrappers.

Typical use cases

• Cortical mesh processing (resampling, smoothing, metrics)
• Thickness and folding related computations
• Volume-to-surface projection
• Denoising and volume preprocessing for structural MRI

Citation / provenance

If you use this package in research, please cite Dahnke et al., 2013 and mention the cat-surf package version for reproducibility.

Source

• Source: https://github.com/ChristianGaser/CAT-Surface