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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Magnetic resonance-based attenuation correction (MRAC) in PET/MRI faces challenges with artifacts and MRI contrast limitations.
• Accurate bone delineation, density, and air/bone separation remain critical for MRAC.

Purpose of the Study:

• To develop a novel Bayesian deep convolutional neural network for improved pseudo-CT generation and uncertainty estimation in MRAC.
• To integrate uncertainty estimates with maximum-likelihood estimation of activity and attenuation (MLAA) for robust attenuation map generation.

Main Methods:

• Proposed a Bayesian deep convolutional neural network to generate pseudo-CT and uncertainty maps from MRI data.
• Combined pseudo-CT uncertainty estimates with MLAA reconstruction using PET emission data.
• Evaluated the approach (UpCT-MLAA) in pelvic lesion imaging and in patients with and without metal implants.

Main Results:

• UpCT-MLAA demonstrated accurate PET uptake estimation in pelvic lesions and successful recovery of metal implants.
• In non-implant patients, UpCT-MLAA showed slightly higher RMSE compared to other methods against CTAC.
• While MLAA recovered implants, Dixon MRI-based pseudo-CT inaccurately estimated metal implant attenuation as air.

Conclusions:

• The proposed UpCT-MLAA method effectively estimates attenuation coefficients for metal implants and provides accurate anatomical depiction outside implant regions.
• This approach addresses key limitations in MRAC, offering improved quantitative accuracy in PET/MRI.
• Uncertainty estimation in pseudo-CT generation is crucial for robust attenuation correction in the presence of artifacts.