Registration of Three-Dimensional Cranial Ultrasound Images and Segmentations using Classical and Deep Learning Approaches
While two-dimensional (2D) ultrasound (US) is currently used for monitoring neonatal cerebral lateral ventricle growth, three-dimensional (3D) US shows promise for more sensitive measurements and a better understanding of ventricle changes. To properly monitor the change in ventricle volume, image pairs, or ventricle segmentations, must be registered to each other to spatially align images. Current registration methods for this task require a manual rigid-body registration, requiring precise picking of points in each image, a technical and time-consuming procedure subject to observer variability. To this end, I developed and tested an automated method based on the iterative closest point (ICP) algorithm for registering 3D ventricle US images across time. I also evaluated two alternative methods for this registration task: statistical parametric mapping (SPM), and a deep learning method. The ICP method performed best with a target registration error (TRE) of 5.63±1.59mm and a DICE similarity coefficient of 0.656±0.104.