Cardiac motion correction with a deep learning network for perfusion defect assessment in single-photon emission computed tomography myocardial perfusion imaging.

Background: In myocardial perfusion imaging (MPI) with single-photon emission computed tomography (SPECT), ungated studies are used for evaluation of perfusion defects despite motion blur. We investigate the potential benefit of motion correction using a deep-learning (DL) network for evaluating perfusion defects.

Methods: We employed a DL network for cardiac motion correction in ECG-gated SPECT-MPI images, wherein the image data from different cardiac phases are combined with respect to a reference gate to reduce motion blur. For training the DL network, 197 cases were used. Given the variability of gated images during the cardiac cycle, we investigated the detectability of perfusion defects in two distinct reference gates. To assess perfusion defect detection, we performed receiver-operating characteristic (ROC) analyses on the motion-corrected images using a separate test dataset of clinical 194 subjects, in which studies were created from actual patient data with inserted simulated-lesions as ground truth. The reconstructed images were assessed by the quantitative-perfusion SPECT (QPS) software. We also evaluated the performance on reduced-count studies (by two and four folds).

Results: The quantitative results, measured by area-under-the-ROC curve (AUC), demonstrated that DL motion correction improves the detectability of perfusion defects significantly on both standard- and reduced-count studies, and that the detectability can vary with reference cardiac phases. A joint assessment from two reference phases achieved AUC = 0.841 on the quarter-count data, higher than with ungated full-count data (AUC = 0.795, P-value = 0.0054).

Conclusions: DL motion correction can benefit assessment of perfusion defects in standard- and reduced-count SPECT-MPI studies. It can also be beneficial to evaluate perfusion images over multiple cardiac phases.