Projects per year
When metal implants are scanned during medical x-ray computed tomography (CT), the conventional model that relates the image coeﬃcients to the x-ray measurements (the sinogram) breaks down. The resulting streak and cupping artifacts aﬀect the accuracy of external beam radiotherapy (RT) treatment plans that rely on CT images for electron density and particle relative stopping power estimates. Particularly in the highly sensitive proton RT this may lead to notable errors in the estimated beam penetration depth, and consequent dosimetric errors that may exclude some patients from proton treatment that would otherwise beneﬁt from it. Metal artifact reduction (MAR) algorithms are therefore employed that however despite many years of development remain inadequate. This thesis investigates a promising way to increase the eﬃcacy of MAR by using the superior anatomical information in artifact corrupted regions that is available in a co-registered Magnetic Resonance Image (MRI) that in RT may have been acquired for tumor delineation. This is challenging due to the inherently weak correlation between conventional-sequence MR intensities and CT values, as well as the potentially imperfect co-registration of the images. We develop a novel, Bayesian generative model of MRIs and CTs that is suitable to handle these challenges, and use it with promising results for MAR by image inpainting in the corrupted regions; as a prior for sinogram inpainting of metal-aﬀected projections; and for prior modelling in model-based iterative reconstruction (MBIR). We ﬁnally validate our methods in the context of head-and-neck RT, ﬁnding dosimetric as well as image space improvements compared to a standard approaches in clinical use.
|Publisher||Technical University of Denmark|
|Number of pages||156|
|Publication status||Published - 2019|