Abstract:175 Objectives: Quantitative PET/MR imaging is challenged by the lack of accurate and robust synthetic CT (sCT) generation techniques from MRI. Deep learning-based algorithms have recently gained momentum for a number of medical image analysis applications. In this work, a novel sCT generation algorithm based on deep learning adversarial semantic structure (DL-ASS) is compared to an atlas-assisted MRI-guided attenuation correction algorithm in brain PET/MRI. Methods: The DL-ASS algorithm exploits ASS learning framework to constrain the synthetic CT generation process to comply with the extracted structural features from CT images. This is implemented as a CT segmentation component from which a gradient is back-propagated to regularize the main synthetic CT generation process. This technique was evaluated through comparison to an atlas-based sCT generation method (Atlas), previously developed for MRI-only or PET/MRI-guided radiation planning. Moreover, the commercial segmentation-based approach (Segm) implemented on the Philips TF PET/MRI system was included in the comparative evaluation. Clinical brain studies of 40 patients who underwent PET/CT and MR imaging were used for the comparison of the different attenuation correction techniques using a two-fold cross validation scheme. The CT-based attenuation corrected PET images (PETCTAC) were used as reference for comparison. Results: The accuracy of cortical bone extraction and CT value estimation were investigated for the three different methods. Atlas and DL-ASS exhibited similar cortical bone extraction accuracy resulting in a Dice coefficient of 0.78±0.07 and 0.77±0.07, respectively. Likewise, DL-ASS and Atlas techniques performed similarly in terms of CT value estimation in the cortical bone region where a mean error (ME) of less than -11 HU was obtained. The Segm approach led to a ME of -1025 HU. Furthermore, the quantitative analysis of the corresponding PET images corrected for attenuation using the three approaches demonstrated comparable performance of DL-ASS and Atlas techniques with a mean standardized uptake value (SUV) bias less than 4% in 63 brain regions. In addition, less than 2% SUV bias was observed in the cortical bone when using Atlas and DL-ASS approaches. However, Segm resulted in 14.7±8.9% SUV underestimation in cortical bone. Conclusions: The DL-ASS approach demonstrated competitive performance with respect to the state-of-the-art atlas-based technique achieving clinically tolerable errors, thus outperforming the commercial segmentation approach used in the clinic.

Does Deep Learning Approaches Outperform Atlas-Guided Attenuation Correction in Brain PET/MRI?