Conception et évaluation d'un nouvel algorithme de reconstruction itérative en tomodensitométrie à faisceau conique implanté sur matériel graphique

Authors: Matenine, Dmitri
Advisor: Després, Philippe; Goussard, Yves
Abstract: This thesis relates to the field of medical physics, in particular, three-dimensional (3D) imaging and 3D dosimetry for radiotherapy. The global purpose of the work was to design and evaluate a new fast iterative reconstruction algorithm for cone beam computed tomography (CT), an imaging technique used to create 3D maps of subject densities based on measurements of partial attenuation of a radiation beam. This algorithm was implemented for graphics processing units (GPU), a highly parallel computing platform, resulting in original optimization strategies. First, a new iterative regularized statistical method, dubbed OSC-TV, was designed and implemented for the GPU. It was evaluated on synthetic and clinical X ray cone beam CT data. The proposed algorithm yielded improved image quality in comparison with similar methods for low-dose acquisitions, as well as reconstruction times compatible with the clinical workflow. The main impact of this work is the capacity to reduce ionizing radiation dose to the patient by a factor of two to four, when compared to standard imaging protocols. Second, this algorithm was evaluated on experimental data from a cone beam optical tomography device, yielding one of the first studies of this kind. The spatial resolution of the resulting 3D images was improved, while the noise was reduced. The spectral properties of the light source were shown to be a key factor to take into consideration to ensure accurate density quantification. The main impact of the study was the demonstration of the superiority of iterative reconstruction for data affected by aberrations proper to cone beam optical tomography, resulting in a potential to improve 3D radiochromic gel dosimetry in radiotherapy. Third, different methods to handle an exact thin-ray system matrix were evaluated for the cone beam CT geometry. Using a GPU implementation, a fully pre-computed and stored system matrix yielded the fastest reconstructions, while being less flexible in terms of possible CT geometries, due to limited GPU memory capacity. On-the-fly ray-tracing was shown to be most flexible, while still yielding reasonable reconstruction times. Overall, the three studies resulted in the design and evaluation of the proposed reconstruction method for two tomographic modalities, as well as a comparison of the system matrix handling methods.
Document Type: Thèse de doctorat
Issue Date: 2017
Open Access Date: 24 April 2018
Grantor: Université Laval
Collection:Thèses et mémoires

Files in this item:
33839.pdf7.26 MBAdobe PDFView/Open
All documents in CorpusUL are protected by Copyright Act of Canada.