📄 中文摘要
计算机断层扫描（CT）探测器中的低效像素（LPP）会导致重建图像中出现环状和条纹伪影，严重影响图像的临床可用性。针对这一问题，近年来涌现出多种解决方案，主要分为图像域校正和正弦图域校正，大多采用监督式深度学习方法。然而，这些方法普遍面临训练数据收集成本高昂的挑战。此外，现有方法在处理复杂或多变的LPP缺陷模式时，往往难以达到理想的校正效果。针对这些局限性，一种新的基于展开网络（unrolled network）的LPP校正框架被提出，该框架结合了合成数据训练策略，旨在克服真实数据稀缺的问题。通过将迭代重建算法中的步骤展开成神经网络层，可以充分利用先验信息和数据驱动的学习能力。
📄 English Summary
Low performing pixel correction in computed tomography with unrolled network and synthetic data training
Low performance pixels (LPPs) in Computed Tomography (CT) detectors are a significant source of ring and streak artifacts in reconstructed images, severely compromising their clinical utility. While various solutions have been proposed, primarily leveraging supervised deep learning in either the image or sinogram domain, these methods often necessitate expensive and dedicated datasets for training. Furthermore, existing approaches frequently struggle to achieve optimal correction performance when confronted with complex or diverse LPP defect patterns. Addressing these limitations, a novel LPP correction framework is introduced, employing an unrolled network architecture coupled with a synthetic data training strategy to mitigate the reliance on scarce real LPP data. By unfolding iterative reconstruction algorithms into neural network layers, the framework harnesses both prior knowledge and data-driven learning capabilities. A synthetic data generation module is designed to simulate various types and degrees of LPP defects, enabling effective model training without extensive real LPP data. This module generates artifact-ridden sinograms by modeling the impact of LPP defects on projection data, pairing them with pristine, artifact-free sinograms to construct large-scale synthetic training datasets. The unrolled network structure allows for end-to-end optimization of correction performance while maintaining model interpretability. Specifically, the network may incorporate data consistency layers to ensure corrected data adheres to physical models, and regularization layers to introduce image priors for noise and artifact suppression. This hybrid approach, combining physical modeling with deep learning, holds promise for improving correction accuracy while reducing dependence on real LPP training data.