Evaluating Voxelmorph : a learning-based 3D non-linear registration algorithm, against the non-linear symmetric normalization technique from ANTs
Victoria Madge (Author)
"Medical image registration is the process of aligning two images of the same scene into the same image space and is a fundamental step in many image processing applications. For the registration of brain images between subjects, non-linear diffeomorphic registration is favoured since such techniques are capable of compensating for tissue deformation while maintaining brain topology. Recently, deep learning has shown success in a wide variety of medical image-analysis tasks, including image registration. VoxelMorph is a deep learning-based non-linear technique promising fast diffeomorphic registrations and claiming comparable results to Symmetric Normalization from Advanced Normalization Tools (ANTs SyN). However, the comparison between the two methods was based solely on Dice scores of automatically segmented labels. Using automatic segmentations could muddle results, and using Dice scores, an indirect evaluation measure, is an incomplete evaluation of goodness of fit. Additionally, the smoothness parameters of the ANTs SyN algorithm were altered to be more similar to those of VoxelMorph, thus restraining ANTs SyN's capacity to achieve a successful registration. This thesis presents an evaluation of VoxelMorph against the native, unaltered ANTs SyN, offering comparisons with direct and indirect evaluation metrics using data with manual gold standard segmentation labels. This evaluation was performed in experiments with three databases: a database of simulated deformations of the VoxelMorph atlas, BrainWeb20, and Neuromorphometrics. Results from the first experiment show ANTs SyN outperforms VoxelMorph in the presence of simulated deformation. Results from the second and third experiment show VoxelMorph produces inter-subject registration results comparable to those of ANTs SyN"-- Author's abstract
Thesis, Dissertation, English, 2021
McGill University Libraries, [Montreal], 2021