CADRE: Contour-Guided ADMM-Optimized Deep Radon Enhancement for Unsupervised CT Reconstruction from Incomplete Data

In clinical practice, reducing radiation dose in Computed Tomography (CT) is of paramount importance. However, low-dose protocols, such as sparse-view or limited-angle scanning, result in incomplete projection data, which severely degrades reconstructed image quality and can compromise diagnostic accuracy. While supervised deep learning methods have shown promise, their heavy reliance on large-scale paired training datasets limits their clinical applicability and generalizability. Unsupervised methods like Deep Radon Prior (DRP) circumvent the need for training data but often suffer from convergence instability and fail to preserve fine anatomical details. To address these limitations, this paper introduces CADRE (Contourguided ADMM-optimized Deep Radon Enhancement), a novel unsupervised framework that enhances DRP through a synergistic integration of three innovations: (1) an ADMM-inspired optimization strategy to stabilize convergence and regularize the solution space; (2) a contour-guided mechanism to enforce structural integrity using explicit geometric priors; and (3) an attention-enhanced network architecture with CBAM to improve feature extraction. Extensive experiments on medical datasets demonstrate that CADRE achieves state-of-the-art reconstruction quality from incomplete data. Its robustness is further validated across industrial and cultural heritage applications, offering a powerful and generalizable solution for high-fidelity imaging without paired training data.