Automatic Generation of MR-based Attenuation Map using Conditional Generative Adversarial Network for Attenuation Correction in PET/MR

Attenuation correction is an important correction for quantitative PET image reconstruction. Current PET/MR attenuation correction methods involve segmenting MR images acquired with zero-time echo (ZTE) or Dixon sequences and assigning known attenuation coefficients to different tissues. This work builds upon our previous work where we explore a novel deep learning method of attenuation map (µ-map) generation using a conditional generative adversarial network (cGAN) that allows for continuous attenuation coefficients [1]. We develop the use of a cGAN network to directly convert MR images to CT images (pseudo CT) through registered training data. A straightforward bilinear conversion can be applied to the pseudo CT images to obtain attenuation maps at 511keV for PET attenuation correction of the head and neck region, including brain. The overall average MAE of the pseudo CT compared to the real CT test images was found to be 88.2 ± 32.7 HU. Future work includes applying the correction on PET data and comparing the reconstructed PET image with CT-based attenuation correction at 511keV as the gold standard.