Local models and local convergence for maximum likelihood iterative reconstruction

Iterative methods are being increasingly used forreconstruction of tomographic data as in computed tomography (CT) and positronemission tomography (PET). These methods often yield better reconstructionscompared to analytical methods. An important reason for this is that iterativereconstruction has a more flexible data acquisition model which allows for amore accurate simulation of the true acquisition during the reconstruction.Moreover, a noise model can be included in the reconstruction algorithm.This thesis investigated several remaining problemsfor iterative reconstruction in CT and PET based on the principles of localmodeling and local convergence improvements.For CT, iterative metal artifact reduction wasstudied. Metals attenuate the majority of the photons which makes that smalldeviations in the reconstruction model can have a large influence and introducesevere artifacts. In the current state of the art methods, data points with lowphotons are removed and replaced with artificial data points that do not causeartifacts. This reduces the most severe artifacts but by the removal of thedata points also contrast in the regionsaround and between the metals isremoved. Alternatively, iterative methods can be used. These methods model theeffects that cause the artifacts and do not remove data points. The existingiterative methods do not reduce the artifacts sufficiently and they are oftentoo slow for possible clinical usage. To reduce computation time, a local model schemewas introduced. This approach uses the most accurate (and computationintensive) model for the metals in the image and a simpler model elsewhere.This local model structure is called the patchwork structure and was able toreduce the computation time without losing image quality. With the patchworkstructure it was also possible to improve the convergence of the metal regionslocally, leading to an improved artifact reduction. This method has beenapplied on simulated and measured phantom and patient scans. The artifacts inthe reconstructed images were clearly reduced without loss of contrast aroundthe metals. Some artifacts, however reduced, are still present. The origin ofthese artifacts remains unclear.The local convergence improvements that can beobtained by the patchwork structure were also evaluated for convergenceimprovements for the entire image volume. Hence, the volume was divided in aset of equal small image blocks that are separately updated. This alternativepatchwork structure has been applied on both CT and PET data. For bothapplications, a considerable increase in convergence rate was obtained,yielding the same image quality in a shorter time, which is important forclinical application of iterative methods.In clinical environments, PET is usuallyreconstructed with iterative reconstruction. A problem of the applied iterativemethod is that it tends to produce bias when the data are very noisy and incold regions. This bias may cause problems for tracer kinetic analysis. Originsof this bias are the strong asymmetry of the employed Poisson distribution incase of low count data and the inherent nonnegativity constraint of the algorithm.A modified algorithm has been developed that switches from the Poissondistribution to the Gaussian distribution for low count data points. This newmethod lowers the bias significantly and gives results comparable to other biasreduction methods. An advantage of this new method is that the distribution islocally changed and that it has only a single parameter, the value of which iseasily determined. Normal 0 21 false false false NL-BE X-NONE X-NONE /* Style Definitions */ table.MsoNormalTable {mso-style-name:Standaardtabel; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""; mso-padding-alt:0cm 5.4pt 0cm 5.4pt; mso-para-margin-top:0cm; mso-para-margin-right:0cm; mso-para-margin-bottom:10.0pt; mso-para-margin-left:0cm; line-height:115%; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Calibri","sans-serif"; mso-ascii-font-family:Calibri; mso-ascii-theme-font:minor-latin; mso-hansi-font-family:Calibri; mso-hansi-theme-font:minor-latin; mso-fareast-language:EN-US;}status: publishe