Comparison of five methods for deformable, multi-modal image registration in prostate and pelvic area

Radiation therapy (RT) planning in prostate cancer typically uses both CT images and T2weighted MR images, for dose calculations and treatment volume delineation, respectively. By using MRI to generate synthetic attenuation maps based on a T1-weighted, double-echo, gradient echo sequence with Dixon reconstruction (mDixon), CT imaging can be discarded from the planning process, and dose delivery uncertainties due to e.g. patient positioning and image registration errors decrease. Multi-modal image registration may still be needed to ensure that the mDixon and T2-weighted images are aligned. The aim of this thesis was to compare five deformable methods for this registration problem; the Demons algorithm, B-spline registration, a fast elastic image registration (FEIR) method, a locally affine hierarchical and a locally rigid hierarchical method. The methods were evaluated qualitatively (visual assessment) and quantitatively for 19 cases. In the quantitative analysis anatomical landmark accuracy, normalised cross-correlation of gradient images and the Dice similarity coefficient (DSC) of the bladder, prostate and seminal vesicles were used as evaluation metrics. The results showed that in cases following the normal RT clinical work-flow, image alignment was good even without registration, but was improved with registration. When intentional patient movement impaired the initial alignment, the benefit of image registration increased. FEIR was the best method in terms of visual assessment, quantitative evaluation performance and computational speed, and seems well-suited for use in the MR-only radiotherapy work-flow for prostate cancer.