Comparison of quantitative accuracy of three edge-preserving image reconstruction strategies

In this paper, we compare three strategies for improving accuracy of activity distribution in a region of interest (ROI) using the information about the ROI boundaries. Specifically, two reconstruction-based methods with anatomic constrains (maximum a-posteriory (MAP) algorithm and penalized least squares (PLS) solution), as well as a post-reconstruction mask-based technique are examined in terms of their ability to accurately reconstruct activity distribution in an ROI. Imitating a clinical oncology situation, we assume that only the boundaries of this ROI are available and that activity distributions both inside and outside this ROI remain completely unknown. We modeled eight different scenarios by altering (i) camera resolution; (ii) number of counts in projections; and (iii) activity in the neighboring organ (liver). The created models were blurred by a Gaussian operator and Poisson noise was added to the projections. For each experiment, 10 individual noise realizations were processed. Our analysis shows that the activity distribution in inhomogeneous ROIs reconstructed in SPECT or PET images can be substantially improved if the ROI boundaries are available. For all eight experiments, three investigated anatomy-guided strategies improved both visual appearance and quantitative parameters of conventional images reconstructed using MLEM algorithm. For the improvement of the activity quantification in currently generating images, post-reconstruction mask-based processing can be recommended. Both the total activity and activity distribution inside ROI obtained after such processing are comparable to, or even better than, the ones achieved by edge-preserving reconstructions. No tuning parameters are required by this mask-based technique. However, the noise in the background is not suppressed by this method. For the improvement of both visual and quantitative characteristics of the image (including ROI and background), the implementation of the edge-preserving reconstruction can be recommended. The anatomy-guided AMAP method provides the user with a balanced improvement of both quantitative accuracy and noise level, especially for experiments with relatively high camera resolution. However, to achieve the best results, the case-specific optimization of penalty factor and number of iterations are necessary.