Joint Sparse Coding-Based Super-Resolution PET Image Reconstruction

This paper presents a comparative study of the effects of using joint sparse coding (JSC) for regularized super-resolution (SR) PET reconstruction. With an assumption that a limited number of high-resolution (HR) PET images are available for a joint training dataset for JSC, we attempt to improve the accuracy of sparse coding (SC) based SR reconstruction in conventional non-HR PET imaging. Here we also assume that the anatomical (CT or MR) and PET images acquired from the same patient lie in coupled feature spaces. The images in one feature space can be transformed into corresponding images in the other feature space by a common mapping function. In this case, the images in the coupled feature spaces have a common sparse representation in terms of the specific dictionaries that are jointly trained, which is the main key to the JSC method. We implemented the penalized-likelihood SR reconstruction algorithm whose penalty term is modeled as JSC and compared with our previous method using the single dictionary-based SC penalty. The experimental results demonstrate that our proposed JSC method clearly outperforms the standard SC method by more accurately restoring the fine details that are often missed by the standard SC method.