Attenuation Coefficient Estimation for PET/MRI With Bayesian Deep Learning Pseudo-CT and Maximum-Likelihood Estimation of Activity and Attenuation

A major remaining challenge for magnetic resonance-based attenuation correction methods (MRAC) is their susceptibility to sources of resonance imaging (MRI) artifacts (e.g., implants and motion) uncertainties due the limitations MRI contrast accurate bone delineation density, separation air/bone). We propose using a Bayesian deep convolutional neural network that in addition generating an initial pseudo-CT from MR data, it also produces uncertainty estimates quantify data. These outputs are combined with maximum-likelihood estimation activity (MLAA) reconstruction uses PET emission data improve maps. With proposed approach prior robust MLAA (UpCT-MLAA), we demonstrate uptake pelvic lesions show recovery metal implants. In patients without implants, UpCT-MLAA had acceptable but slightly higher root-mean-squared-error (RMSE) than Zero-echo-time Dixon Deep when compared CTAC. recovered implant; however, anatomy outside implant region was obscured by noise crosstalk artifacts. Attenuation coefficients were normal anatomy; estimated have air. alongside anatomic depiction regions.