Global dictionary guided brain PET image reconstruction

Objectives The MR assisted wavelet-based joint entropy (WJE) and dictionary learning (DL) maximum a posteriori (MAP) algorithms showed promises in improving PET image reconstruction. However, the corresponding MR images are needed to create an anatomical prior. The goal is to study the global dictionary guided brain PET image reconstruction and to quantitatively evaluate the reconstructed images.

Methods We developed a global dictionary based approach to incorporating the sparse representation prior into the MAP image reconstruction algorithm. Specifically, a brain PET image is divided into 3D overlapping patches which are sparsely represented over a redundant global dictionary learned using the singular value decomposition (K-SVD) algorithm. Using the BrainWeb phantoms, we simulated PET data at different noise levels. The 3D natural images (Ohio University Campus taken two frames per second) and the image of a man-made hollow sphere were applied in the DL process. A patient Florbetapir PET dataset with corresponding T1-MPRAGE MRI images were also studied. Using the noise versus bias (mean value) tradeoff, we evaluated the reconstructed images from using the global dictionaries, compared with those from using corresponding MR image formed priors.

Results The DL-MAP algorithms using different dictionaries performed similarly to one another in brain PET image reconstruction in terms of noise versus bias (mean value) tradeoff. In both simulated low noise level and high noise level cases, the global DL-MAP algorithms showed clear improvement over the maximum likelihood (ML) algorithm with a little higher noise compared to the WJE algorithm in the gray matter (GM) region. In the case of real patient study, the global DL-MAP algorithm performs comparable with the WJE-MAP algorithm.

Conclusions The brain PET images, the corresponding MR images, the natural and man-made images can be patch-wise sparsified under a common dictionary. Achieving robust performance in various noise level simulation and in real patient study, the global DL-MAP algorithm presents its potential in clinical quantitative PET imaging, especially when the corresponding MR images are not available.

Research Support NSF grant ECCS 1228091