Training Strategies for Increased Temporal Resolution in Clinical Cardiac-gated SPECT Studies with Deep Learning Denoising

Cardiac-gated SPECT studies are limited to the use of a small number of gated frames due to high imaging noise in clinical acquisitions. Here we investigate three strategies for training a deep learning (DL) network for suppressing the elevated imaging noise when the number of gate frames is doubled from conventional clinical studies. These training strategies are aimed at addressing the challenge of a lack of ground truth images in clinical acquisitions for use as the learning target by a DL network. In the experiments we investigated these strategies with a set of 237 clinical acquisitions. The results indicate that the trained DL network can effectively suppress both the spatial and temporal variability in the reconstructed myocardium in the gated frames. Moreover, surrogate training with gated data and noise-to-noise (N2N) training can yield better preservation of perfusion defects than with ungated data.