A Momentum-Based Acceleration of the Diffeomorphic Demons Algorithm for Registration of MRI and CT Images of the Brain

Accurate deformable registration of multiple 3D imaging modalities is vital to many areas of diagnostic and interventional radiology and surgery. Diffeomorphic Demons algorithm have been previously reported for mono-modality registration. We extend such methodology to multi-modal (MRI-CT) images of the brain using point-wise mutual information (pMI) with momentum-based acceleration of the optimization. Preprocessing via automatic histogram stretch improved robustness and accuracy of registration in studies involving CT and T1-weighted MRI of a head phantom and clinical studies of five neurosurgery patients. Performance was compared to B-spline Free-Form Deformation (FFD) and Symmetric Normalization (SyN). pMI-Demons achieved target registration error of 0.21±0.07 mm (median±iqr) in phantom and 1.57±0.52 mm in clinical studies, providing alignment comparable to the voxel size without statistically significant difference from FFD and SyN. The pMI-Demons and SyN methods yielded diffeomorphic transformations, whereas FFD yielded unrealistic deformations. pMI-Demons provided a 66% runtime reduction (10 min vs. 30 min for SyN) that facilitates application in rapid image-guided neurosurgery workflows.