Deep learning–based denoising of low-dose SPECT myocardial perfusion images: quantitative assessment and clinical performance

• DOI: 10.1007/s00259-021-05614-7
• Publication Date: 2021-11-15T03:02:36Z
• Authors: Narges Aghakhan Olia; Alireza Kamali-Asl; Sanaz Hariri Tabrizi; Parham Geramifar; Peyman Sheikhzadeh; Saeed Farzanefar; Hossein Arabi; Habib Zaidi
• Keywords: Tomography, Emission-Computed, Single-Photon; Denoising; Myocardial perfusion imaging; Low-dose; Deep learning; info:eu-repo/classification/ddc/616.8; Signal-To-Noise Ratio; ddc:616.0757; ddc:616.8; 3. Good health; Perfusion; info:eu-repo/classification/ddc/616.0757; 03 medical and health sciences; Deep Learning; 0302 clinical medicine; SPECT; Image Processing, Computer-Assisted; Humans; Original Article; Retrospective Studies

This work was set out to investigate the feasibility of dose reduction in SPECT myocardial perfusion imaging (MPI) without sacrificing diagnostic accuracy. A deep learning approach proposed synthesize full-dose images from corresponding low-dose at different levels projection space.Clinical SPECT-MPI 345 patients acquired on a dedicated cardiac camera list-mode format were retrospectively employed predict standard-dose half-, quarter-, and one-eighth-dose levels. To simulate realistic projections, 50%, 25%, 12.5% events randomly selected data through applying binomial subsampling. generative adversarial network implemented non-gated space Well-established metrics, including peak signal-to-noise ratio (PSNR), root mean square error (RMSE), structural similarity index metrics (SSIM) addition Pearson correlation coefficient analysis clinical parameters derived Cedars-Sinai software used quantitatively assess predicted images. For evaluation, quality evaluated by nuclear medicine specialist using seven-point (- 3 + 3) grading scheme.The highest PSNR (42.49 ± 2.37) SSIM (0.99 0.01) lowest RMSE (1.99 0.63) achieved half-dose level. coefficients 0.997 0.001, 0.994 0.003, 0.987 0.004 for levels, respectively. Using as reference, Bland-Altman plots sketched exhibited remarkably less bias variance compared with all reduced Overall, considering assessment performed specialist, 100%, 80%, 11% clinically acceptable respectively.The noise effectively suppressed network, comparable reference half- quarter-dose However, recovery underlying signals/information beyond quarter standard would not be feasible (due very poor ratio) which will adversely affect interpretation resulting