Direct 3D PET image reconstruction into MR image spaceDirect 3D PET image reconstruction into MR image space
Faculty of Medicine and Health Sciences
New York, N.Y. :IEEE, 2011[*]2011
Engineering sciences. Technology
18th IEEE Nuclear Science Symposium/Medical Imaging Conference (NSS/MIC), International Workshop on Room-Temperature Semiconductor X-Ray and Gamma-Ray Detectors, October 23-29, 2011, Valencia, Spain
A method which includes both the motion correction and image registration transformation parameters from PET image space to MR image space within the system matrix of the MLEM algorithm is presented. This approach can be of particular significance in the fields of neuroscience and psychiatry, whereby PET is used to investigate differences in activation patterns between groups of participants (such as healthy controls and patients). This requires all images to be registered in a common spatial atlas. Currently, image registration is performed post-reconstruction. This introduces interpolation effects in the final image and causes image resolution degradation. Furthermore, motion correction introduces a further level of interpolation and possible resolution degradation. To include the transformation parameters (both for motion correction and registration) within the iterative PET reconstruction framework (through iterative use of actual software packages routinely applied after reconstruction) should reduce these interpolation effects and thus improve image resolution. Furthermore, it opens the possibility of direct reconstruction of the PET data into standardized stereotaxic atlases, e.g. ICBM152. To validate the proposed method, this work investigates registration, using 2D and 3D simulations based on the HRRT scanner geometry, between different image spaces using rigid body transformation parameters calculated using the mutual information similarity criterion. The quality of reconstruction was assessed using bias-variance and mean absolute error analyses to quantify differences with current post-reconstruction registration methods. We demonstrate a reduction in bias and in mean absolute error in reconstructed mean ROI activity when using the proposed method.