Deep residual-convolutional neural networks for event positioning in a monolithic annular PET scanner

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Monolithic scintillator PET scanners offer potential for superior performance, including high spatial resolution and detection sensitivity.
• The unique geometry of monolithic PET systems, such as the AnnPET scanner, presents challenges in accurately estimating event positions due to complex optics.
• Traditional methods like the center-of-mass algorithm may not fully leverage the capabilities of advanced PET detector designs.

Purpose of the Study:

• To evaluate the efficacy of deep-residual convolutional neural networks (DR-CNN) for estimating the 3D position of photon interactions in the AnnPET scanner.
• To compare the performance of DR-CNN-based event positioning with the conventional center-of-mass (COM) algorithm.

Main Methods:

• Monte Carlo simulations were employed to model the AnnPET scanner, including its physical and optical properties.
• A ten-layer DR-CNN was developed and trained using simulated data to estimate interaction positions.
• Performance was assessed by comparing mean absolute errors (MAE) in event positioning and reconstructed spatial resolutions (FWHM) between DR-CNN and COM methods.

Main Results:

• The ten-layer DR-CNN achieved significantly lower MAEs for event positioning (0.54 mm axial, 0.42 mm transaxial, 0.45 mm depth) compared to COM (1.22 mm, 1.04 mm, 2.79 mm).
• Reconstructed spatial resolutions using DR-CNN data were superior: 0.8 mm radial, 0.7 mm tangential, 0.71 mm axial (FWHM).
• COM-derived data resulted in lower spatial resolutions: 1.15 mm radial, 0.96 mm tangential, 1.14 mm axial (FWHM).

Conclusions:

• Deep-residual convolutional neural networks demonstrate a strong potential for enhancing event positioning accuracy in monolithic scintillator PET scanners.
• The DR-CNN approach offers significant improvements over standard analytical methods like COM, leading to better image quality and spatial resolution.
• This AI-driven method is well-suited for optimizing the performance of novel PET detector geometries like AnnPET.