Computerized Attenuation Correction of Liver SPECT Images Using Extended 3D-pix2pix Model with Bidirectional Learning

This study aims to develop a computerized attenuation correction method for liver single photon emission computed tomography (SPECT) images using an extended 3D-pix2pix model with bidirectional learning. Our database consisted of liver SPECT images obtained from 809 patients. For each patient, SPECT images were reconstructed both with and without CT-based attenuation correction. In this study, the conventional 3D-pix2pix architecture is extended by incorporating a conditional encoder and two decoders, allowing for mutual transformation between SPECT images before and after attenuation correction. The conditional encoder takes an input image along with a class label indicating whether the input image is pre- or post-attenuation correction. This shared encoder facilitates the extraction of consistent features from both pre- and post-attenuation corrected SPECT images. The two decoders generate virtual post- or virtual pre-attenuation corrected image, respectively, from the input pre- or post-attenuation corrected image. With the proposed method, the root mean squared error (RMSE) and structural similarity index measure (SSIM) were 15.94 and 0.916, respectively, showing higher fidelity compared to the conventional 3D-pix2pix model (16.81 and 0.913). The proposed attenuation correction method for liver SPECT images achieved high fidelity and shows potential usefulness as an attenuation correction method without CT.