Torso synthetic CT generation by integrating deep learning and segmentation for FDG-PET/MR attenuation correction

Introduction.Positron Emission Tomography/Magnetic Resonance (PET/MR) offers benefits over PET/CT including simultaneous PET and MR acquisition, intrinsic spatial registration accuracy, MR-based functional information, and superior soft tissue contrast. However, accurate attenuation correction (AC) for PET remains challenging as MR signals do not directly correspond to attenuation. Using deep learning algorithms that learn complex relationships, we generate synthetic CT (sCT) from MR for AC. Our novel method for AC, merges deep learning with threshold-based segmentation, to produce an AC map for the entire torso from Dixon MR images, which heretofore has not been demonstrated.Method. Twenty-nine prospectively collected, paired FDG-PET/CT and MR datasets were used for training and validation using the U-net Residual Network conditional Generative Adversarial Network integrated with tissue segmentation (URcGAN_mod) from Dixon MR data. Our application focused on torso (base of the skull to mid-thigh) AC, a common but challenging field of view (FOV). Performance was compared to that of 4 previously published methods.Results. Using 15 paired datasets for training and 14 independent datasets for testing, the URcGAN_modgenerates an accurate torso sCT with a mean absolute difference of 32 ± 4 HU per voxel. When applied for AC for FDG images, and considering evaluable (SUV ≥ 0.1 g/ml) voxels across all regions of interest, absolute values of the differences were within 4.4% from those determined using the measured CT for AC. Reproducibility was excellent with less than 3.5% standard deviation. The results demonstrate the accuracy and precision of URcGAN_modmethod for torso sCT generation for quantitatively accurate MR-based AC (MRAC), exceeding the comparison methods.Conclusion. Combining deep learning and segmentation enhances MRAC accuracy in torso FDG-PET/MR, improves SUV accuracy throughout the torso, achieves less than 4.4% SUV error, and outperforms comparison methods. Given the excellent sCT and SUV accuracy and precision, our proposed method warrants further studies for quantitative longitudinal multicenter trials.