Improving Temporal Resolution in Clinical Cardiac-gated SPECT Studies via Surrogate Training in Deep Learning Denoising

Cardiac-gated studies in myocardial perfusion imaging with SPECT in practice are limited to the use of a small number of gate frames due to the low data counts acquired. We investigate the feasibility of expanding the number of gate intervals to achieve a higher temporal resolution by applying a deep learning denoising network for noise suppression. To deal with the lack of ground truth images in cardiac-gated clinical studies, we employ a surrogate training approach in which ungated imaging data from reduced-dose studies are utilized for network training. In the experiments we evaluated this approach on a set of 60 clinical acquisitions with 16 gate frames used. The results demonstrate that the proposed approach can effectively improve the image uniformity of the LV wall and suppress the increased temporal variability quantitatively by 13.2% and 39.1%, respectively, in the gated images while without incurring adverse effect on LV functional volume measurements.