Accurate quantification of local changes for carotid arteries in 3D ultrasound images using convex optimization-based deformable registration

Registration of longitudinally acquired 3D ultrasound (US) images plays an important role in monitoring and quantifying progression/regression of carotid atherosclerosis. We introduce an image-based non-rigid registration algorithm to align the baseline 3D carotid US with longitudinal images acquired over several follow-up time points. This algorithm minimizes the sum of absolute intensity differences (SAD) under a variational opticalflow perspective within a multi-scale optimization framework to capture local and global deformations. Outer wall and lumen were segmented manually on each image, and the performance of the registration algorithm was quantified by Dice similarity coefficient (DSC) and mean absolute distance (MAD) of the outer wall and lumen surfaces after registration. In this study, images for 5 subjects were registered initially by rigid registration, followed by the proposed algorithm. Mean DSC generated by the proposed algorithm was 79.3± 3.8% for lumen and 85.9± 4.0% for outer wall, compared to 73.9± 3.4% and 84.7± 3.2% generated by rigid registration. Mean MAD of 0.46±0.08mm and 0.52±0.13mm were generated for lumen and outer wall respectively by the proposed algorithm, compared to 0.55±0.08mm and 0.54±0.11mm generated by rigid registration. The mean registration time of our method per image pair was 143± 23s.