Deep learning-based image reconstruction for TOF PET with DIRECT data partitioning format.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Positron Emission Tomography (PET)

Background:

• Conventional Positron Emission Tomography (PET) image reconstruction relies on statistical iterative methods.
• Deep learning offers potential for accelerating PET image reconstruction.
• Existing deep learning methods often require complex, hardware-intensive fully connected networks.

Purpose of the Study:

• To propose a novel deep learning-based image reconstruction method for Time-of-flight PET (TOF PET).
• To develop a method that avoids fully connected networks, reducing complexity and hardware costs.
• To achieve 3D image reconstruction within current hardware limitations using a U-net architecture.

Main Methods:

• Utilized the DIRECT data partitioning method with a U-net structure composed solely of convolutional layers.
• Employed patch-based model training and testing for 3D reconstructions.
• Generated Time-of-flight (TOF)-histoimages from listmode data, using projection angles as input channels.
• Validated the method using 15 patient datasets (372 scans) with a leave-one-patient-out approach.

Main Results:

• The deep learning method produced reconstructed images with image quality comparable to conventional expectation-maximization (EM) methods.
• Minimal differences were observed when using simulated TOF-histoimages, indicating successful learning of the reconstruction process.
• Quantitative differences emerged with measured TOF-histoimages, suggesting the need for physics-based corrections.

Conclusions:

• A novel deep learning-based reconstruction method for TOF PET was successfully developed.
• The DIRECT data partitioning method enabled 3D reconstruction without fully connected layers, fitting current hardware constraints.
• Physics-based corrections remain essential for achieving optimal quantitative performance in deep learning-based PET reconstruction.