---

title: How to register CT and MRI of the same patient in Python description: Rigid CT to MRI registration in Python with SimpleITK through one call: mutual information, multi-resolution, physical-shift optimizer scales; the transform in degrees and millimetres, the resampled volume, a checkerboard check, and what to do when it fails.

---

How to register CT and MRI of the same patient in Python

Mean squares fails across modalities (bone is bright in CT and dark in T1); mutual information does not assume an intensity relation. The call below runs SimpleITK's rigid registration with Mattes MI, three resolution levels and optimizer scales from physical shift, which is what makes one learning rate work on any pair.

import pardesline as pl

s = pl.load("rire_brain")[0]  # CT 0.65x0.65x4 mm, MR T1 1.25x1.25x4 mm, 19 MB
moved = s.ct.align(s.mr_t1)  # fixed = CT, moving = MR; 8.8 s on a laptop
t = moved.transform["sub_transforms"][0]
print(t["rotation_deg"], t["translation_mm"])  # [0.34, -0.35, -4.38] [16.2, -33.1, -21.6]
moved.save("mr_on_ct.nii.gz")  # the MR resampled on the CT grid + sidecar

Your own volumes: pl.load("ct.nii.gz").align(pl.load("mr.nii.gz")), or a DICOM folder for either. The registration runs in physical space (origin, spacing, direction from the headers), so different spacings and fields of view need no resampling first.

Check it

import pyvista as pv

ct, mr = s.ct.to_pyvista(), moved.to_pyvista()
p = pv.Plotter()
p.add_mesh(ct.slice("z"), cmap="gray")
p.add_mesh(mr.slice("z"), cmap="magma", opacity=0.5)
p.show()

Or s.ct.run("simpleitk_registration.difference_analysis", moving=moved) for a difference map. The clinical check is a target registration error on landmarks placed in both volumes and not used by the algorithm; RIRE reports about 1 mm for rigid brain CT/MR.

When it fails

Symptom	Cause	Do
metric flat from the first iterations	the initial offset is outside the capture range (above ~30° or different patient position)	s.ct.run("simpleitk_registration.advanced_rigid", moving=s.mr_t1): exhaustive rotation search first
converges, visibly wrong	wrong centre initialisation when one field of view is much larger	initializer="MOMENTS"
noisy metric, jittery result	too few samples on a small field of view	sampling_percentage=0.05 to 0.15
same modality follow-up	MI is overkill	metric="mean_squares", faster
scanner scale or shear	not rigid	simpleitk_registration.affine; deformation: bspline, demons

print(pl.describe("simpleitk_registration.rigid")) lists the fourteen parameters, their ranges and the reason for each default.

Related: Guide: multimodal rigid registration (a hand-written SimpleITK stack next to the one-liner, metric traced per iteration) · RIRE dataset · Why our defaults work