A Parallel Cross-Domain Reconstruction for PET

This paper introduces a novel Parallel Cross-Domain Reconstruction Deep Neural Network architecture for PET imaging. Unlike conventional PET reconstruction methods, which often operate exclusively within the image or sinogram domain or attempt a joint approach without fully harnessing their intrinsic correlations, our approach effectively combines complementary information from both domains by integrating multiple stages and incorporating cross-attention mechanisms. Evaluations conducted on synthetic data illustrate that our approach surpasses two deep-learning algorithms, including the proposed architecture lacking the cross-attention mechanism and the LPD algorithm, both in qualitative and quantitative assessments.