Novel deep learning-based CT synthesis algorithm for MRI-guided PET attenuation correction in brain PET/MR imaging

MRI-guided synthetic CT (sCT) generation is one of the main challenges hampering quantitative PET/MR imaging as well as MRI-only radiation planning. Deep learning-based approaches have recently gained momentum in a variety of medical imaging applications. In this work, a novel synthetic CT generation algorithm based on deep convolutional neural network is proposed for MRI-guided attenuation correction in PET/MRI. The proposed algorithm (AsCT) exploits adversarial semantic structure learning implemented as a CT segmentation approach to constrain the adversarial synthetic CT generation process. The proposed technique was trained using 50 pairs of CT and MR brain scans under a two-fold1 cross validation scheme. The AsCT method was compared to an atlas-based method (Bone-Atl), previously developed for MRI-only radiation planning, as well as the commercial segmentation-based approach (2-class) implemented on the Philips TF PET/MRI system. The evaluation was performed using clinical brain studies of 40 patients who have undergone PET/CT and MRI scanning. The accuracy of the CT value estimation and cortical bone identification were assessed for the three different methods taking CT images as reference. Bias of tracer uptake (SUV) was measured on attenuation corrected PET images using the three techniques taking CT-based attenuation corrected PET as reference. Bone-Atl and AsCT exhibited similar cortical bone extraction (using an intensity threshold of 600 Hounsfield Unit (HU)) resulting in Dice coefficient (DSC) of 0.78 +/- 0.07 and 0.77 +/- 0.07, respectively. Bone-Atl method performed slightly better in terms of accuracy of CT value estimation where a mean absolute error of 123 +/- 40 (HU) was obtained for the whole head region while AsCT and 2-class methods led to 141 +/- 40 and 230 +/- 33 (HU), respectively. Quantitative analysis of brain PET images demonstrated competitive performance of AsCT and Bone-Atl methods where mean relative errors of 1.2 +/- 13.8% and 1.0 +/- 9.9% were achieved in bony structures, respectively, while the 2-class approach led to a mean SUV error of -14.7 +/- 8.9%. The proposed AsCT algorithm showed competitive performance with respect to the atlas-based method and outperformed the segmentation-based (2-class) method with clinically tolerable errors.