DEFORMABLE IMAGE REGISTRATION OF SPINE MAGNETIC RESONANCE IMAGES WITH ELASTIX FOR DETAILED CHARACTERIZATION OF DISC LOADING BEHAVIORS

Abstract: Purpose: This work aimed to develop an optimized strategy for registration of spine magnetic resonance (MR) images acquired with and without spinal loading, as well as evaluate the quality of the registration strategy with commonly used quality assessment methods. Method: Previously collected MR images of the lumbar spine of 37 individuals were re analyzed. Multi-echo T2-mapping had been obtained from the subjects in both unloaded and axial-loaded condition. The signal intensity-based registration software Elastix were selected for the registrations between the two image data sets. To facilitate automatic and advanced image analysis, the MICE Toolkit was used as a graphical user interface for Elastix. The registration was done in two steps using the short echo-time raw data images: (1) rigid registration and (2) deformable registration (i.e. non-rigid) with the first registration as input. For comparison, registration was also performed using the long echo-time raw data images. To reduce the calculation time and possibly improve the quality of the image registration, the image was limited by a binary image mask that was applied over the spinal column. To optimize the registration, the parameters FinalGridSpacingInVoxels and BSplineInterpolationOrder were varied. The registration quality was evaluated with Dice similarity coefficient (DSC) and Jaccard coefficient, where a semi-automatic software based on region-growing was used for the delineation of the intervertebral discs (IVDs). The Jacobian determinant was also calculated to ensure that the deformation was realistic with non-zero positive values. The intra- and inter-observer reliability between readers were also determined. Result: The DSC (0.845 ± 0.059) and the Jaccard coefficient (0.735 ± 0.084) were high for all individuals. The mean of the Jacobian determinant was close to one (1.035 ± 0.043). Analysis based on the whole image stack, also showed high DSC (0.845 ± 0.057) and the Jaccard coefficient (0.736 ± 0.081). Significant difference was neither observed between registration with short and long TE images (p = 0.205), nor between registration with or without mask (p = 0.247). A slightly higher accuracy was obtained with less final B-spline grid spacing. The inter- and intra-observer agreement were excellent (ICC = 0.99) Conclusion: Deformable image registration of spine MR images with the optimized registration strategy seems feasible and useful for detailed characterization of disc loading behaviours. The registration using the strategy showed high quality evaluated with DSC and Jaccard coefficient. The evaluation of the Jacobian determinant indicated a preserved topology of the deformed image.