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

In positron emission tomography (PET), gamma photons arriving at the detector ring may undergo one or more Compton scattering events, potentially reaching a different scintillation crystal than its initial interaction point. This phenomenon, known as inter-crystal scattering (ICS), can lead to incorrect line of response (LOR) assignments, introducing spatial blurring and noise in the reconstructed image. Existing ICS correction methods include energy-based heuristics, Compton kinematics, Monte Carlo simulation, and deep learning to recover the first interaction position of the photons in the detector. These corrections are performed in the LOR domain, since that is where ICS manifests. However, in ordered-subset expectation-maximization (OS-EM) reconstruction, only a subset of LORs is processed at each iteration, making LOR-domain corrections difficult to apply on-the-fly.

We propose a novel ICS correction technique performed entirely in the image domain. A neural network is trained to predict spatially-varying 3D filter kernels that model the blurring effect of ICS across the image volume. These kernels are applied to the PET image estimate prior to each OS-EM forward projection. We evaluated our method on simulated and real PET data using two network variants: one predicting full 3D kernels (ICS-Net-direct), and another using a compact skew normal representation (ICS-Net-skewnorm).

Both ICS-Net-direct and ICS-Net-skewnorm significantly improved the spatial resolution and the contrast of the output image, while also being computationally efficient. We found that ICS-Net-skewnorm is better suited for structural, symmetric reconstruction objects, while ICS-Net-direct performs best in complex, real-world scenarios.

The proposed image-domain ICS correction technique enables efficient and effective compensation of inter-crystal scattering for OS-EM reconstruction. Network selection should be guided by the target imaging scenario to maximize performance.