A Deep Learning Approach for Noise Suppression in Cardiac-gated SPECT Studies

Cardiac-gated SPECT images are prone to elevated imaging noise in clinical acquisitions. This study investigates the use of a deep learning (DL) network for noise suppression in electrocardiogram-gated cardiac studies.

We adapted a 3D convolutional autoencoder network, pre-trained on reduced-counts ungated SPECT image data, for processing cardiac-gated images in standard-dose acquisitions. The network was trained on image data from 862 ungated studies (acquired with Philips BrightView SPECT/CT). We tested this approach on a set of 50 cardiac-gated studies (from the same imaging system). We quantified the denoising performance in terms of the noise level in the myocardium and the functional measures (wall thickening and volumes) of the left ventricle (LV). We also tested this approach on a different imaging system (GE StarGuide SPECT/CT) using 20 acquisitions from a second clinical site.

With DL denoising the image uniformity of the LV wall was improved by 20.6% on average over individual gate frames in normal studies, and the temporal image variability of the myocardium was reduced by 47.4%. There was a high degree of agreement in the functional measures of the LV obtained before and after DL denoising, with the Pearson correlation-coefficient r between the two all above 0.94 in the measurements (LV wall thickness, chamber volumes, and EF). Similar performance was also observed for the StarGuide image data.

The proposed DL approach can effectively suppress the imaging noise in cardiac-gated SPECT images without inducing distortion in LV function measurements, with potential applicability across different imaging platforms.