Anatomy-guided U-Net deep learning for sparse-view chest CT reconstruction using spatially varying RTV regularization

Sparse-view computed tomography (CT) reconstruction remains a challenging inverse problem due to severe undersampling artifacts, noise amplification, and the difficulty of preserving anatomical structures while reducing image degradation. Conventional regularization-based methods often require manual parameter adjustment and rely on spatially uniform priors, whereas deep learning approaches may provide limited interpretability or insufficient integration of physical constraints. In this study, we propose SvRTV-UNet, a physics-guided deep learning framework that integrates a U-Net architecture with spatially variant Relative Total Variation (SvRTV) regularization for adaptive sparse-view chest CT reconstruction. The proposed framework simultaneously performs anatomical segmentation and predicts reconstruction parameters, including the regularization strength (λ) and tissue-specific weighting parameters (A 1 and A 2 ), directly from image quality characteristics. A homoscedastic uncertainty-based multi-task learning strategy is employed to automatically balance segmentation, parameter estimation, and reconstruction objectives without manual loss weighting. In addition, anatomy-aware confidence maps derived from segmentation accuracy and edge information enable tissue-dependent regularization for improved structural preservation. Compared with conventional iterative reconstruction methods, model-based regularization approaches, and representative deep learning methods, SvRTV-UNet demonstrated improved reconstruction performance across quantitative metrics. The quantitative and qualitative results obtained demonstrate the effectiveness of the proposed method compared with the evaluated approaches. By integrating physics-based constraints with anatomy-guided adaptive regularization, SvRTV-UNet offers an interpretable framework that adapts to varying acquisition conditions without view-specific retraining.