Approximating anatomically-guided PET reconstruction in image space using a convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Anatomically-guided PET reconstruction improves bias-noise characteristics in brain imaging.
• Clinical adoption of current methods is limited due to practical challenges.

Purpose of the Study:

• To investigate if a convolutional neural network (CNN) can achieve anatomically-guided PET reconstruction improvements solely in the image domain.
• To develop a fast, clinically applicable post-reconstruction method for enhanced PET imaging.

Main Methods:

• Developed an image-based CNN post-reconstruction approach.
• Utilized data augmentation with 16 [18F]FDG and 10 [18F]PE2I datasets for training.
• Approximated asymmetric Bowsher prior-based reconstruction using a shift-invariant CNN.

Main Results:

• The CNN successfully generated anatomically-guided PET images in near real-time.
• The trained CNN demonstrated robustness against different PET tracers, noise levels, and MRI contrast.
• Image quality from the CNN closely matched target reconstructions in regional recovery and structural similarity.

Conclusions:

• A purely image-based CNN can effectively approximate anatomically-guided PET reconstruction.
• This CNN approach offers a fast and practical solution for enhancing clinical PET imaging.
• The method shows potential for widespread clinical evaluation and application.