3D Unet Generative Adversarial Network for Attenuation Correction of SPECT Images

Attenuation correction (AC) images are essential to perform qualitative and quantitative analysis in myocardial perfusion tomography (MPI) using a single-photon emission tomography (SPECT) scanner. Hybrid SPECT/CT systems were born by a combination of a CT scanner and a regular SPECT machine to directly measure photon degradation in the body. However, in reality, these systems are often very expensive and are not available in most health centers in Vietnam. In this paper, we propose the method of filtering attenuated noise by using 3DUnet-Gan network. Experimental results show that with non-attenuation correction (NC) images as the input, our proposed model can generate AC images with SSIM similarity = 0.945%, NMAE error = 0.034 compare to true AC. The results of this study show the prospects of correcting NC images as well as and other medical images using Deep Learning models. As it will not only save time for diagnosis but also avoid exposing patients to large radioactive doses.