Abstract: Sparse CT reconstruction can reduce patient radiation dose and is of great significance for clinical diagnosis. In deep learning-based image reconstruction tasks, classic networks such as Uformer, Restormer, and AST fail to consider multi-scale and directional information in images, neglecting the balance between local details and global structure, resulting in limited artifact suppression. To address these issues, a Multi-scale Attention Adaptive Fusion Transformer (MAAF-Transformer) network was proposed. A parallel attention fusion strategy was adopted by this network, which was combined with a multi-scale channel-directional awareness module to capture artifact features. The weights were dynamically adjusted by utilizing an adaptive spatial channel attention module. Finally, effective information was filtered through a gated deep convolutional feedforward network. Experimental results show that, under 60 sparse views, MAAF-Transformer achieves 0.7141 dB higher peak signal-to-noise ratio (PSNR), 0.33%higher structural similarity (SSIM), and 7.76% lower root mean square error (RMSE) than classic Uformer, demonstrating excellent performance in terms of visual effects as well. Evidently, MAAF-Transformer possesses higher sparse reconstruction accuracy and stronger artifact suppression capabilities.

摘要： 稀疏CT重建能够降低患者辐射剂量，对临床诊断具有重要意义。在基于深度学习的图像重建任务中，经典Uformer、Restormer、AST等网络未考虑图像多尺度与方向信息，忽略了局部细节和全局结构的平衡，伪影抑制效果有限。针对上述问题，提出一种多尺度注意力自适应融合Transformer网络(MAAF-Transformer)。该网络采用并行注意力融合策略，结合多尺度通道方向感知模块捕获伪影特征，并利用自适应空间通道注意力模块动态调节权重，最后通过门控深度卷积前馈网络筛选有效信息。实验结果表明，在60个稀疏视角下，提出的MAAF-Transformer比经典Uformer在峰值信噪比(PSNR)提高0.7141dB、结构相似性(SSIM)提高0.33%、均方根误差(RMSE)降低7.76%，并在视觉效果上同样表现优异。可见，MAAF-Transformer稀疏重建精度更高，抑制伪影能力更强。