Performance evaluation of a smoothed total variation regularization algorithm

Total variation (TV) regularization algorithm is effective in image noise reduction and lesion quantification accuracy improvement. However, TV reconstructed image tends to be patchy if the detected counts are low. Low-count scenarios are common for PET/MR imaging, since a large number of emitted counts are attenuated by MR coils. In this work, an improved regularization method with a new smoothed total variation (STV) prior was proposed. It was initialized by a few iterations of TV regularization reconstruction, and followed by STV regularization reconstruction until the image converges. OSEM, TV and TV-STV reconstruction algorithms were compared through phantom and clinical studies. A numerical observer was designed to evaluate the lesion detectability in the phantom study. TV-STV achieved the highest area under LROC curve in both 30 s and 60 s reconstructions, and had the highest contrast recovery coefficients for cold spheres. For the patient study, TV-STV generated images with reduced hot spots and improved lesion contrast. And the image texture was similar to that of OSEM images.