Image-space compensation of inter-crystal scattering in PET using a neural network based filter

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Computational Imaging

Background:

• Positron emission tomography (PET) gamma photons can scatter between crystals (inter-crystal scattering, ICS).
• ICS causes incorrect line of response (LOR) assignments, leading to spatial blurring and noise in reconstructed PET images.
• Current ICS correction methods often operate in the LOR domain, posing challenges for iterative reconstruction algorithms like ordered-subset expectation-maximization (OS-EM).

Purpose of the Study:

• To introduce a novel image-domain technique for correcting inter-crystal scattering (ICS) in PET.
• To develop a neural network-based approach for efficient and effective ICS compensation.
• To integrate ICS correction seamlessly into the OS-EM reconstruction workflow.

Main Methods:

• A neural network predicts spatially-varying 3D filter kernels to model ICS blurring.
• These kernels are applied to the PET image estimate before each OS-EM forward projection.
• Two network variants were evaluated: ICS-Net-direct (full 3D kernels) and ICS-Net-skewnorm (compact skew normal representation).

Main Results:

• Both ICS-Net variants significantly enhanced spatial resolution and image contrast.
• The methods demonstrated computational efficiency.
• ICS-Net-skewnorm excelled with symmetric objects, while ICS-Net-direct performed better in complex, real-world scenarios.

Conclusions:

• The proposed image-domain ICS correction is efficient and effective for OS-EM reconstruction.
• The choice between ICS-Net variants depends on the specific imaging application.
• This technique offers a practical solution for mitigating ICS artifacts in PET imaging.