Low Dose Myocardial Perfusion SPECT Denoising Using an Attention-Based Generative Adversarial Network

Deep learning-based denoising is promising for low dose (LD) myocardial perfusion (MP) SPECT. However, conventional convolutional neural network (CNN)-based methods used fixed-sized convolutional kernels to convolute one region within the receptive field at a time, which would be ineffective for learning the feature dependencies across large regions. Attention scheme is able to leverage voxels with stronger correlations beyond the kernel receptive field. In this study, we propose a 3D attention-based conditional generative adversarial network (AttcGAN) for denoising LD MP-SPECT images on both projection- and reconstruction-domain. Forty-six patients underwent ~1184 MBq
^99m
Tc-sestamibi stress SPECT/CT scan were retrospectively recruited. Sixty projections were acquired from right anterior oblique to left posterior oblique positions and the acquisition time for each projection was 10 s for the full dose (FD) mode. LD MP-SPECT images (1/10 to 7/10 of FD) were generated from the FD list mode data. cGAN and AttcGAN were implemented using Tensorflow with the Adam optimizer running up to 200 epochs. FD and LD MP-SPECT pairs of 30 patients were used for training the cGAN and AttcGAN separately for each dose level, while 6, and 10 patients were applied for validation and testing. Normalized mean square error (NMSE), normalized standard deviation (NSD), and perfusion defect size (PDS) were analyzed. The results indicate that denoising on the projection-domain outperforms reconstruction-domain for both networks, and the AttcGAN is superior to cGAN on both projection- and reconstruction-domain on all dose levels. We conclude that AttcGAN denoising on projection-domain can further improve the denoising performance for LD MP-SPECT with better defect quantification accuracy.