PET Image Reconstruction from Under-sampled Data

To decrease the positron emission tomography (PET) scanning system cost, image reconstruction algorithms for under-sampled data have been extensively studied, for example, total variation (TV) minimization regularization was nested in the Maximum likelihood-expectation maximization (MLEM) method to suppress noise (MLEM-TV). However, the gradient operator of TV cannot distinguish true structures from noise image, that is, the image reconstructed by MLEM-TV algorithm may lose some fine features. Recently, we developed a feature refinement approach for statistical interior computed tomography (CT) reconstruction. Its potential usefulness have been demonstrated for feature preservation of interior tomography with truncated projection measurements. Inspired by this study, the feature refinement (FR) step is applied to PET imaging for recovery the fine features. In the implementation of the proposed method, the filtered back-projection (FBP) reconstruction is used as the initial guess to mitigate artifacts. Moreover, the feature refinement step, as an important processing operation is performed after each MLEM-TV iteration to extract the desired structure information lost in the TV minimization. Here, a feature descriptor is specifically designed to distinguish structure from noise and artifacts. A modified steepest descent method is adopted to minimize the objective function. To study the MLEM -TV-FR algorithm and evaluate its performance, we simulated realistic brain PET imaging data using an emission activity phantom and compared to FBP, MLEM and MLEM-TV. The reconstruction results show that the presented method implements better than other methods in suppressing noise, reducing artifacts and preserving features from under-sampled data.