Deep learning-guided attenuation correction in the image domain for myocardial perfusion SPECT imaging

Abstract We investigate the accuracy of direct attenuation correction (AC) in image domain for myocardial perfusion SPECT (single-photon emission computed tomography) imaging (MPI-SPECT) using residual (ResNet) and UNet deep convolutional neural networks. MPI-SPECT 99mTc-sestamibi images 99 patients were retrospectively included. ResNet networks trained non-attenuation-corrected as input, whereas CT-based attenuation-corrected (CT-AC) served reference. Chang’s calculated AC approach considering a uniform coefficient within body contour was also implemented. Clinical quantitative evaluations proposed methods performed CT-AC 19 subjects (external validation set) Image-derived metrics, including voxel-wise mean error (ME), absolute error, relative structural similarity index (SSI), peak signal-to-noise ratio, well clinical relevant indices, such total deficit (TPD), utilized. Overall, generated learning exhibited good agreement with images, substantially outperforming method. The models resulted an ME −6.99 ± 16.72 −4.41 11.8 SSI 0.99 0.04 0.98 0.05, respectively. led to 25.52 33.98 0.93 0.09, Similarly, evaluation revealed TPD 12.78 9.22% 12.57 8.93% models, respectively, compared 12.84 8.63% obtained from images. Conversely, 16.68 11.24%. have potential achieve reliable imaging.