Deep learning-based attenuation and scatter correction in myocardial SPECT without using X-ray CT images.

The objective of this study was to correct for the effects of photon scattering and attenuation in myocardial SPECT imaging using convolutional neural networks (CNNs), without relying on X-ray CT images. In the proposed framework, CNNs were employed to estimate both the projection data of scattered photons and the distribution of attenuation coefficients (µ-map). For scatter estimation, the CNN was trained using as input the projection data of primary and scattered photons affected by attenuation, along with the projection data of attenuated primary photons, while the projection data of attenuated scattered photons served as the supervised data. The training dataset comprised 23,544 projection samples including data augmentation, and the testing dataset comprised 3,600 samples. For µ-map estimation, images reconstructed from projection data acquired at two low-energy windows were used as CNN inputs, with the µ-map at 140 keV as the supervised data. A total of 327 µ-maps were used for training and 20 for testing. All projection data required for these estimations were obtained by Monte Carlo calculations using the publicly available database. To evaluate the reproducibility of myocardial infarction, pseudo-infarct regions were introduced in 10 of the 20 test cases. The performance of the estimated images was assessed using the polar map, circumferential profile, and normalized mean squared error. Simulation results demonstrated that the proposed method accurately estimated both scattered-photon projection data and µ-maps, achieving scatter and attenuation correction performance comparable to conventional CT-based approaches. Furthermore, the method showed robust performance even in the presence of myocardial infarction.