Contour Guided-Deep Radon Prior: an Unsupervised Framework for CT Reconstruction

Background: The inverse problem in computed tomography (CT) reconstruction has long been a critical research topic in medical imaging. Although deep learning-based approaches have achieved remarkable progress in solving CT reconstruction problems, most methods heavily depend on large-scale high-quality training datasets and lack interpretability. Method: This paper proposes a novel Contour-Guided Deep Radon Prior (CG-DRP), a fully unsupervised framework for CT reconstruction. Built upon the Deep Radon Prior (DRP), CG-DRP incorporates geometric contour information to guide the reconstruction process. Specifically, the proposed method learns projection priors in the Radon domain while integrating edge-aware constraints to effectively compensate for missing projection data. By embedding a neural network as an implicit prior into the iterative reconstruction process, CG-DRP enables synergistic optimization through gradient feedback in both the image and Radon domains. Results: Experimental results demonstrate the effectiveness of CG-DRP in addressing inverse problems such as limited-angle and sparse-view CT reconstruction. The contour-guided mechanism significantly enhances edge preservation and improves the clarity of structural boundaries in reconstructed images. Compared with existing supervised learning approaches, CG-DRP achieves comparable or superior reconstruction performance without requiring any training data. Conclusions: CG-DRP introduces a new paradigm for CT image reconstruction by leveraging deep self-correlation in the Radon domain and contour priors to facilitate effective collaboration between neural network manifolds and image reconstruction. This unsupervised learning framework not only enhances reconstruction quality but also improves algorithm interpretability, providing a scalable solution for medical image reconstruction.