Parallax Correction for Pet Imaging Based on Improved Unet Combined With Transfer Learning

In PET imaging, gamma photons may penetrate multiple crystals before depositing energy, leading to the mispositioning of the LOR. This phenomenon is known as the parallax error, which significantly degrades PET imaging quality. We treat the parallax error as a special type of image degradation and implement end-to-end parallax correction based on deep learning in the image domain. In response to the serious artifacts in PET imaging due to the parallax error, we propose PCNet, an improved UNet that combines spatial self-attention and channel self-attention, capturing long-range dependencies between pixels and integrating global and local features. To reduce the model's dependency on specific datasets, we propose a transfer learning strategy based on multi-branch and multi-scale feature fusion. This strategy involves pre-training multiple encoders on a general dataset and fine-tuning the decoder with a small amount of target domain data. We evaluate the performance of PCNet on a generic dataset (e.g., basic shapes) and assess the effectiveness of the multi-branch and multi-scale feature fusion strategy on a dataset generated by the ROBY phantom. The results show that, compared to traditional methods and general state-of-the-art image restoration models, our PCNet achieves the best PSNR and SSIM metrics on five types of general datasets. Additionally, compared to non-transfer learning strategies, our transfer learning strategy improves the correction quality on the ROBY phantom dataset. We also validate our method on a real PET system and prove its effectiveness.