Structure-guided prior-driven implicit neural representations for time-of-flight PET image reconstruction

Positron emission tomography (PET) reconstruction is an ill-posed inverse problem, particularly under low-count conditions where noise severely degrades image quality and quantitative accuracy. Although supervised learning approaches have demonstrated strong denoising capability, their performance often depends on large paired datasets and may suffer from limited generalization. This work aims to develop an unsupervised reconstruction framework for time-of-flight PET (TOF-PET) that improves image quality while maintaining quantitative reliability.

We propose a TOF-PET reconstruction method based on implicit neural representations (INR). A differentiable forward projection model is incorporated to explicitly model TOF-PET imaging physics and enable reconstruction directly in the INR domain. To suppress noise and promote spatial smoothness, a ray-based total variation (TV) regularization is introduced. The reconstruction network combines a multi-resolution hash encoder with a prior-image encoder that injects structural image priors into the INR representation.

The proposed framework was evaluated using simulated brain and whole-body datasets as well as clinical TOF-PET scans. Results show improved noise suppression and contrast recovery compared with conventional iterative reconstruction algorithms and representative unsupervised approaches.

The proposed approach integrates implicit neural representations with physics-consistent modeling and prior-guided regularization, providing an effective unsupervised framework for TOF-PET reconstruction and highlighting the potential of neural field representations for tomographic imaging.