Attention-recon: Exploring backprojection image reconstruction with attention mechanisms

In the field of computed tomography (CT) image reconstruction, analytical reconstruction algorithms are computationally efficient but lack robustness against noise and incomplete data, especially in sparse-view scenarios. Iterative reconstruction algorithms can handle sparse-view data and suppress artifacts, yet suffer from high computational complexity, slow speed, and strong dependence on regularization parameters. Most existing deep learning methods for sparse-view CT reconstruction are based on these two algorithms and follow a staged image optimization paradigm. To address these limitations, this study proposes an end-to-end deep learning method based on attention mechanism for direct reconstruction of tomographic images from sparse projection data. First, by integrating attention mechanism with an image-to-image network architecture, a small-resolution 64×64 Recon-Transformer network is constructed to explore the ability of attention mechanism in capturing the mapping relationship between projection and image domains. Second, based on the small-resolution framework, we extend to 128×128 and higher resolutions and design the Attention-Recon (AttnRecon) network to investigate the reconstruction performance of local sinusoidal attention in high-resolution sparse projection scenarios. Finally, extensive experiments are conducted to verify the effectiveness of the proposed method. Experimental results demonstrate that the proposed method achieves significant performance improvements over conventional FBP and ART algorithms at 64×64 resolution, and also outperforms the deep learning method Glimpse on the out-of-distribution (OOD) medical dataset. At 128×128 resolution, the local sinusoidal attention mechanism adopted in AttnRecon still maintains advantages in in-distribution fitting compared with FBP and SART, but its advantage in cross-distribution generalization is notably weakened.