Matija Ciganovic

Master Thesis
Supervisors: Firat Ozdemir, Dr. Christine Tanner, Prof. Orçun Göksel

Alignment of Ultrasound Images with 3D Bone Models

Orthopaedic surgery is a branch of medicine that focuses on injuries of the skeletal system. An important step in the pre-operative planning is the acquisition of a computed tomography (CT) scan, as CT is particularly well suited to visualize bone tissue and to subsequently create a 3D-model of the imaged bone. However, such a 3D model contains little to no information on soft tissue surrounding the bone of interest. While ultrasound (US) provides an affordable way to visualize such soft tissue, an optimal pre-operative planning is precluded by the fact that the surgeon then uses two different sources of information on the same body region (3D bone model from CT, soft tissue information from US) instead of having one 3D model containing both at the same time, so that the imaged soft tissue is perfectly aligned with the imaged bone it surrounds. In this work, we present a novel, feature-based approach to align a bone surface as extracted from a sequence of tracked US images with a previously generated 3D bone model. First, the bone surface is extracted from US using a factor graph-based approach, while a 3D bone model was generated a priori from an MRI volume using manual segmentation. The subsequent registration uses a multi-initialization scheme that utilizes the physics of US to help a registration algorithm find a good alignment between the two. The focus of the thesis lies on bones in the forearm, as this is a particularly likely body region to be injured. We have evaluated our registration approach on a set of simulated data, as well as on 28 in-vivo datasets. In the in-vivo study, we were able to obtain a mean distance between a manually annotated bone surface from US (not used during the registration) and the 3D bone model surface of 0.38mm (standard deviation: 0.31mm). An additional visual inspection of the registration results has confirmed a good registration quality. As a result, we conclude that our registration framework is able to return a good alignment between US and a previously generated 3D bone model in a practical setting.