UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Real-time ultrasound bone segmentation and robust US-CT registration for surgical navigation of pelvic fractures Pandey, Prashant

Abstract

Surgically repairing pelvic fractures is complex and includes intensive use of ionizing X-rays. The surgery is prone to screw insertion errors, which can harm the patient. We aim to alleviate these issues by proposing ultrasound (US), instead of X-rays, to guide the surgery. However, US images are noisy, which makes them difficulty to use intraoperatively. This thesis presents segmentation and registration methods aimed at making an US based pelvic fracture repair technique possible. We perform a scoping review of US bone segmentation literature to better understand the current state of the field. We find a lack of consistency in validation practices, especially in quantifying segmentation accuracy. We also recommend techniques based on clinical requirements. We then develop a three-dimensional (3D) US bone segmentation technique: ’Shadow Peak’ (SP), which uses simplified analysis of shadow and peak intensity information in US. In a full-sized pelvic phantom study and on pilot in-vivo pediatric data, we demonstrate SP to be more accurate than two state-of-the-art segmentation methods: phase symmetry (PS) and confidence-weighted structured phase symmetry (CSPS). SP achieves a mean F-score of 63% compared to 54% for PS and 34% for CSPS on phantom data, and 94% on in-vivo data compared to 70% and 72% for PS and CSPS. SP is real-time with a mean runtime of 0.48s per volume, compared to 18.1s and 21.95s for PS and CSPS. Thirdly, we develop a registration pipeline for aligning tracked US and preoperative CT, using the normalized cross-correlation (NCC) similarity metric. We find NCC-based registration is more accurate and robust than Gaussian Mixture Model (GMM) and Coherent Point Drift (CPD) point-set registrations, two methods previously used for US-CT bone registration. SP segmentation with NCC registration achieves a mean target registration error of 3.22mm, compared to 3.89mm with CPD, while GMM registration typically fails. All methods are evaluated on a full-sized pelvic phantom containing soft-tissue details. Our proposed methods are fast and accurate as tested on phantom and in-vivo datasets, and have the potential to make ultrasound-based pelvic fracture guidance practically feasible. Moreover, our US bone segmentation review is useful for guiding future studies in the field.

Item Citations and Data

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International