Prostate segmentation and multimodal registration in 3D ultrasound images

Abstract

Prostate cancer has been ranked as one of the leading causes of cancer death for men. This thesis addresses two important issues that arise in the transrectal ultrasound (TRUS) guided prostate cancer biopsy and treatment. The first is the prostate boundary detection in ultrasound images. In this work, we proposed a novel method based on the statistical shape model, to automatically segment the prostate from the real-time 3D ultrasound images. Prior knowledge of the prostate in 3D TRUS images is characterized by a few model parameters and used in the boundary search procedure as constraints. The second is the multimodal image registration that aims at bringing the suspected cancer map which is obtained in the magnetic resonance imaging and spectroscopy techniques (MRI/MRS) to the intra-operative ultrasound images. We proposed a generic framework of a surface-to-image registration technique applicable to ultrasound image, and investigate its use with organs of different rigidity and the formulation of the similarity measurements based on the acoustic appearance. To account for the prostate deformation occurred between modalities, we explored the use of parametric surface matching and the thin-plate spline transformation to establish the deformation transformation. An elastic pelvic phantom was designed for the validation.