Increasing angular sampling through deep learning for GE Alcyone dedicated cardiac SPECT

Introduction: The GE Discovery 530/570c dedicated cardiac scanners consist of 19 modules each with pinhole collimator and CZT detectors to simultaneously acquire 19 projection angles over a 180-degree arch for image reconstruction without moving the scanner. In this study, we investigated ways to provide additional projection angles for improved image quality. A deep learning approach is also proposed to generate synthetic multi-angle-set images from a single set of static projections. Method: By rotating the fixed detectors with 19 projections in 5-degree intervals, 3 additional sets (total 4) of projection angles are acquired, yielding enhanced angular sampling for reconstruction with MLEM algorithm. Both computer simulation of SPECT perfusion using XCAT phantom and 4 99mTc-tetrofosmin perfusion studies in pigs were performed for evaluation. After the in vivo scanning of each pig, the hearts were excised and additional 4 sets of 19-projection data were also obtained of the excised hearts and reconstructed. As acquiring multiple sets of projections by rotating the scanner is not always possible in clinical practice, a deep learning network was developed to convert images reconstructed with only one set of projection data (19 angles) to those with four set of projections (76 angles). To provide sufficient data for network training, 260 volumes of XCAT phantoms were simulated and used for pretraining. After the pretraining stage, 3 pig scans were used to fine-tune the network, and the remaining pig was used to test the network. This “leave-one-out” process was repeated for all the 4 pig studies. The proposed network was trained with one-angle set images as input and four-angle set images as the label. A U-net-like structure with dense-net implemented in both 2D and 3D (Unet-2D and Unet-3D) is used in this work. The network fine-tuned with in vivo pig studies was also applied to three sample human dataset acquired with one angular set of 19 projections.

Results: For the XCAT phantom, and in vivo and ex vivo pig imaging studies, the reconstructed images from 2 angle sets had substantially better contrast and resolution when compared to those from a single set of projections. Reconstructed images from 4 angle sets were similar to those generated from 3 sets of angles. Both reconstructions are slightly better than those of 2 angle sets. Unet-3D consistently outperforms Unet-2D in all the datasets. Images of 1 angle set processed with Unet-3D are consistent with those of 4 angle sets. Comparing to 4 angle set images, the SSIM of one-angle set images with and without Unet-3D are 0.988+/-0.003 and 0.898+/-0.016 respectively in phantoms images, 0.924+/-0.041 and 0.911+/-0.033 respectively in pig images, 0.988+/-0.012 and 0.985+/-0.010 respectively in ex vivo pig heart images. In the human datasets, both 2D and 3D networks generate images with higher resolution and contrast. The myocardium to blood pool ratio increased by 112% on average across 3 subjects. Conclusion: Increasing angular sampling can substantially improve image quality for the GE Alcyone SPECT systems. Without acquiring additional angular projections, deep learning can be used to generate synthetic images of multiple angular samplings from those reconstructed with only one set of static 19 projection data.