Doctor of Philosophy
In order to improve radiological diagnosis of back pain and spine disease, two new algorithms have been developed to aid the 75% of Canadians who will suffer from back pain in a given year. With the associated medical imaging required for many of these patients, there is a potential for improvement in both patient care and healthcare economics by increasing the accuracy and efficiency of spine diagnosis. A real-time spine image fusion system and an automatic vertebra/disc labeling system have been developed to address this. Both magnetic resonance (MR) images and computed tomography (CT) images are often acquired for patients. The MR image highlights soft tissue detail while the CT image highlights bone detail. It is desirable to present both modalities on a single fused image containing the clinically relevant detail. The fusion problem was encoded in an energy functional balancing three competing goals for the fused image: 1) similarity to the MR image, 2) similarity to the CT image and 3) smoothness (containing natural transitions). Graph-Cut and convex solutions have been developed. They have similar performance to each other and outperform other fusion methods from recent literature. The convex solution has real-time performance on modern graphics processing units, allowing for interactive control of the fused image. Clinical validation has been conducted on the convex solution based on 15 patient images. The fused images have been shown to increase confidence of diagnosis compared to unregistered MR and CT images, with no change in time for diagnosis based on readings from 5 radiologists. Spinal vertebrae serve as a reference for the location of surrounding tissues, but vertebrae have a very similar appearance to each other, making it time consume for radiologist to keep track of their locations. To automate this, an axial MR labeling algorithm was developed that runs in near real-time. Probability product kernels and fast integral images combined with simple geometric rules were used to classify pixels, slices and vertebrae. Evaluation was conducted on 32 lumbar spine images and 24 cervical spine images. The algorithm demonstrated 99% and 79% accuracy on the lumbar and cervical spine respectively.
Miles, Brandon, "Image Fusion and Axial Labeling of the Spine" (2014). Electronic Thesis and Dissertation Repository. 1911.