cmtk-apply

Fast, accurate and user-friendly Python interface for applying CMTK (Computational Morphometry Toolkit) transformations, without requiring subprocess calls to external command-line tools.

Thanks to the use of numba, we are on par for inverse transforms and ~2x faster for forward transforms compared to CMTK's streamxform tool. For smaller point sets the speed-up is even more dramatic due to eliminating overhead from process startup and file I/O. See BENCHMARK_RESULTS.md for details.

Installation

From Github

pip install git+https://github.com/schlegelp/cmtk_apply

From source with development dependencies

Clone the repository and run:

pip install -e ".[dev]"

Quick start

from cmtk_apply import load_registration

# Load a registration (directory or file path)
reg = load_registration("JFRC2_FCWB.list")

# Forward transform points
points = [[10.0, 20.0, 30.0], [40.0, 50.0, 60.0]]
forward = reg.transform_points(points)
print(forward)

# Inverse transform
inverse = reg.inverse_transform_points(forward)
print(inverse)

Optional Fastmath

Enable fastmath for Numba kernels by setting the environment variable:

export CMTK_APPLY_FASTMATH=1

Expected impact (warp forward): ~1.5x speedup with max absolute drift around 1e-13 on typical inputs. Restart the Python process after changing the variable.

API Reference

load_registration(path: str) -> Registration

Load a CMTK registration from a .list folder or registration file.

Parameters:

• path: Path to a CMTK registration directory (.list folder) or a registration file (optionally gzipped).

Returns: A Registration object.

Registration.transform_points(...) -> np.ndarray

Apply the registration transformation to points.

output = reg.transform_points(
    points,                      # N×3 array-like of 3D points
    transform="warp",            # "warp" (default) or "affine"
    allow_extrapolation=True,   # Allow evaluation outside spline grid
    fallback_to_affine=True,     # Use affine if spline fails
)

Parameters:

• points: N×3 array or list of 3D coordinates
• transform: "warp" evaluates both affine and spline; "affine" evaluates affine only
• allow_extrapolation: If False, points outside the spline domain return NaN
• fallback_to_affine: If True, points that fail spline evaluation fall back to affine

Returns: N×3 NumPy array of transformed points

Registration.inverse_transform_points(...) -> np.ndarray

Apply the inverse transformation to points.

output = reg.inverse_transform_points(
    points,                      # N×3 array-like of 3D points
    transform="warp",            # "warp" (default) or "affine"
    allow_extrapolation=True,   # Allow evaluation outside spline grid
    fallback_to_affine=True,     # Use affine if spline fails
    solver="auto",               # "auto" (default), "analytical" or "numerical"
)

Parameters:

• points: N×3 array or list of 3D coordinates
• transform: "warp" evaluates both affine and spline; "affine" evaluates affine only
• allow_extrapolation: If False, points outside the spline domain return NaN
• fallback_to_affine: If True, points that fail spline evaluation fall back to affine
• solver: Method for solving inverse warp; "auto" uses analytical for >500 points and "numerical" otherwise

Returns: N×3 NumPy array of inverse transformed points

Implementation Details

Spline Warp

The library interprets the absolute flag in the spline warp configuration:

• absolute yes: Coefficients represent absolute positions → final = spline(affine_transformed_point)
• absolute no: Coefficients represent displacements → final = affine_transformed_point + spline_displacement

Affine Composition

CMTK affine parameters are decomposed into:

• xlate: Translation (tx, ty, tz)
• rotate: Rotation angles in degrees (rx, ry, rz)
• scale: Anisotropic scaling (sx, sy, sz)
• shear: Shear components (shx, shy, shz)
• center: Center of rotation (cx, cy, cz)

The matrix composition follows CMTK's post-2.4.0 convention. Legacy (<2.4.0) behavior is auto-detected from the TypedStream version field.

B-spline Basis

The cubic B-spline basis functions for parameter t ∈ [0,1] are: $$w_0(t) = \frac{(1 - 3t + 3t^2 - t^3)}{6}$$ $$w_1(t) = \frac{(4 - 6t^2 + 3t^3)}{6}$$ $$w_2(t) = \frac{(1 + 3t + 3t^2 - 3t^3)}{6}$$ $$w_3(t) = \frac{t^3}{6}$$

A 4×4×4 neighborhood of control points is evaluated for each query point using tensor products of these weights.

Testing

Run tests with pytest:

pytest tests/

To compare against CMTK's streamxform tool (if available):

pytest -v tests/test_cmtk_xf.py::test_affine_matches_streamxform
pytest -v tests/test_cmtk_xf.py::test_warp_matches_streamxform

Tests skip automatically if streamxform is not in the expected location.

Example: JFRC2 to FCWB Registration

The JFRC2_FCWB.list/ folder contains a pre-computed registration from the Drosophila reference brain JFRC2 to the FCWB template. This registration includes:

• An affine component (12 DOF)
• A 59×27×11 B-spline warp grid with deformation coefficients

File Format

CMTK's TypedStream format is a text-based nested structure:

! TYPEDSTREAM 2.4

registration {
  reference_study "..."
  floating_study "..."
  affine_xform {
    xlate 0 0 0
    rotate 0 0 0
    ...
  }
  spline_warp {
    affine_xform { ... }
    absolute yes
    dims 59 27 11
    origin -11.36 -13.24 -16.87
    domain 636.39 317.88 134.99
    coefficients ...
  }
}

Limitations

• This library focuses on the core transformation logic; it does not reformat images (use CMTK's reformatx for that).
• The active field in spline warps is parsed but not used; all control points are weighted equally (matches behavior of CMTK's streamxform).

References

• CMTK on NITRC
• nat (R package) CMTK I/O
• Drosophila brain templates (JFRC2, FCWB)

License

MIT