Channel Attention-Enhanced Deep Learning Network for Low-Dose PET Reconstruction

To address the risks associated with high radiation exposure in high-dose PET imaging, low-dose PET reconstruction has emerged as a critical area of research. However, low-dose PET images are frequently compromised by substantial noise, which severely diminishes the quality of the reconstructed images. In response to this challenge, we propose an efficient low-dose PET denoising model that leverages a redesigned deep Convolutional Neural Networks (CNNs) as the backbone for feature extraction, incorporating an advanced convolutional block based on HighRes-net to enhance image representation. Additionally, we introduce a novel Channel Attention Denoising (CD) module, specifically designed to reduce noise and improve the quality of image reconstruction. To further refine the model’s performance, we employ Huber loss, which effectively highlights key features while mitigating the impact of large errors. The experimental results demonstrate that this method improves both the resolution and diagnostic accuracy of the reconstructed PET images.