Multi‐dimensional attention‐enhanced reconstruction for sparse‐view CBCT

BACKGROUND: Cone-beam computed tomography (CBC T) is a high-resolution 3D imaging modality that has been widely applied clinically. However, it faces limitations due to the lag effect and concerns about radiation dose. While sparse-view CBCT can alleviate these problems by reducing the detector readout frequency and the exposure time, it introduces streak artifacts that can degrade image quality. Recent research has explored techniques for sparse-view CBCT reconstruction. Nevertheless, it remains an ongoing challenge to achieve improved image quality. PURPOSE: The purpose of this study is to eliminate the streak artifacts and recover image details with dual-domain deep learning networks. METHODS: This study proposed a Multi-dimensional Attention-Enhanced reconstruction (MAE-Recon) algorithm for sparse-view CBCT. MAE-Recon consists of four key components: a linear interpolation module, a projection domain network, an FDK operator, and an image domain network. Firstly, the linear interpolation module generates an initial full-view projection. Subsequently, the projection domain network predicts high-quality projection data. The FDK operator then reconstructs CBCT images. Finally, the image domain network refines the reconstructed images. Considering the long-range dependency among projections and the information redundancy within features, this study introduces two plug-and-play modules for enhancement. RESULTS: , respectively, and the SSIM and PSNR scores were improved by 0.066, 0.009, and 4.62 dB, 1.44 dB on the two datasets. An ablation study was also performed. The improved RMSE, SSIM, and PSNR scores illustrate that the modules are practical and efficient. CONCLUSIONS: The MAE-Recon proposed in this study obtains high-quality images in sparse-view CBCT reconstruction, demonstrating the application potential.