Dual-domain dual-branch residual-learning network for fast noisy sparse-view ultra-low-dose CT reconstruction.
Authors
Affiliations (5)
Affiliations (5)
- School of Physics, Beihang University, No. 37 Xueyuan Road, Haidian District, Beijing, Beijing, Beijing, 100091, China.
- Image Processing Center, Beihang University, No. 37 Xueyuan Road, Haidian District, Beijing, Beijing, Beijing, 100191, China.
- Hangzhou International Innovation Institute, Beihang University, No. 166 Shuanghongqiao Street, Pingyao Town, Yuhang District, Hangzhou, Zhejiang Province, China, Hangzhou, Zhejiang, 311115, China.
- Beijing Hangxing Machinery Co., Ltd, No. 11 Hepingli East Street, Dongcheng District, Beijing, China., Beijing, 100741, China.
- Department of Physics, Beihang University, No. 37 Xueyuan Road, Haidian District, Beijing, Beijing, Beijing, 100083, China.
Abstract
Ultra-low-dose CT can be achieved by reducing the tube current and employing sparse-view projections, thereby improving patient safety. However, this acquisition strategy inevitably introduces aliasing artifacts and degrades the signal-to-noise ratio.To address the coexistence of undersampling-induced aliasing artifacts and noise, we developed a dual-domain dual-branch residual-learning network (D$^3$R-Net) for high-fidelity ultra-low-dose CT reconstruction.The proposed framework first performs edge-preserving sinogram restoration using an improved Directional Cubic Convolution (iDCC) interpolation, followed by a U-Net optimized with an inner-structure gradient loss to ensure precise preservation of critical edge-gradient details.
In the image-domain network, we design a dual-branch structure-infiltrated guidance network (DB-SiGN) to extract low- and high-frequency information from the reconstructed image obtained from the refined sinogram and from the original projections, respectively. Within DB-SiGN, the gradient information extracted from the low-frequency branch is used as structural guidance for the noisy high-frequency branch, enabling the network to distinguish true high-frequency anatomical structures from noise and artifacts. As a result, a more effective separation and preservation of meaningful high-frequency details is achieved. Both branches learn the residuals between the sinogram-domain refined reconstruction and the normal-dose CT (NDCT) image. Their outputs are subsequently fused through adaptive spatial attention weighting to generate the final high-quality reconstruction. D$^3$R-Net consistently outperformed all comparative methods in both quantitative metrics and visual quality across all simulated dose-reduction settings and real CBCT data. In addition, D$^3$R-Net achieved superior and robust segmentation accuracy across all evaluated scenarios, with its reconstructions aligning most closely with the NDCT reference. D$^3$R-Net establishes a robust and interpretable dual-domain framework that consistently yields state-of-the-art reconstructions under extreme dose reduction, providing a promising pathway toward safe, reliable, and clinically deployable ultra-low-dose CT imaging.