Neural MLAA for PET-enabled Dual-Energy CT Imaging

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Dual-energy CT (DECT) imaging on PET/CT scanners typically requires a separate x-ray CT scan.
• Reconstructing 511 keV gamma-ray attenuation images from PET data using Maximum-Likelihood Attenuation and Activity (MLAA) often results in noisy images.

Purpose of the Study:

• To develop a noise-suppression technique for PET-enabled DECT imaging.
• To improve the quality of gamma-ray CT attenuation images reconstructed from PET emission data.

Main Methods:

• A neural network approach was proposed for MLAA reconstruction, incorporating x-ray CT images as anatomical priors.
• The reconstruction problem was solved using an iterative algorithm involving PET activity and attenuation image updates, and neural network learning with a weighted mean squared-error loss.
• Optimization transfer was employed to ensure monotonic increase of data likelihood.

Main Results:

• Computer simulations demonstrated significant improvement in gamma-ray CT image quality compared to existing algorithms.
• The proposed neural MLAA algorithm effectively suppresses noise in reconstructed attenuation images.
• Anatomical priors from x-ray CT images were successfully integrated into the PET attenuation image reconstruction.

Conclusions:

• The neural network-based MLAA reconstruction offers a promising solution for high-quality PET-enabled DECT imaging without a second CT scan.
• This method enhances the diagnostic utility of PET/CT scanners by providing clearer attenuation images.
• The integration of anatomical priors via neural networks represents a significant advancement in PET image reconstruction.