Assessment of Defect Detection in Post-Filtering and Deep Learning Denoising Strategies for Reduced Dose Myocardial Perfusion SPECT Employing Human and Polar Map Observers

We have shown that administered activity levels can be reduced using optimized reconstruction strategies, personalized injected dose models, and denoising deep learning networks. In this study we continue our task-based assessment of reduced dose reconstruction and denoising strategies using receiver operator characteristic (ROC) methods. We also introduced remote observer participation over a wide range of viewing stations employing a virtual private network (VPN), or in some cases make use of an Amazon Web Server (AWS) to facilitate observer access.Seventy-eight patients with a normal stress Tc-99m cardiac perfusion SPECT outcomes were selected for the insertion of artificial cardiac perfusion defects, while a further 78 patients with normal cardiac perfusion SPECT distributions were selected to complete the test set. Some of these patients were used to train observers. From list mode acquisitions, dose was incrementally reduced and processed using previously optimized parameters. Six strategies were evaluated, using ordered-subset expectation maximization (OSEM) reconstructed data with attenuation compensation (AC), distance dependent resolution compensation (RC), and scatter compensation (SC), while another only included RC. Gaussian post filtering and a three-dimensional convolutional autoencoder (CAE) model were used for denoising. Both human observers and a polar map clinical observer were used to rank the strategies.Ranking of the processing strategies with the polar map clinical observer revealed that the OSEM strategy without dose reduction with all physics (AC, RC, SC) ranked first, while it was not so clear for human observers. In general, ranking follow the amount reduced dose, except for the strategy where only RC was used without reducing dose.The deep learning approach showed promise, however more work and more data are needed to come to a definitive conclusion.