Introduction
Computed Tomography (CT) is a cornerstone of modern medical diagnostics, yet the ionizing radiation associated with high-dose scans remains a significant health concern. Sparse-view CT, which utilizes fewer X-ray projections, offers a path toward dose reduction. However, reconstructing high-quality images from sparse data is an ill-posed inverse problem, typically resulting in severe streak artifacts and noise when using traditional methods like Filtered Back Projection (FBP).
Existing solutions generally fall into two categories: traditional iterative methods, which are reliable but often lack detail in extremely sparse scenarios, and supervised deep learning models. While supervised models achieve impressive results, they require vast amounts of labeled, high-dose ground-truth data—a resource that is often scarce or unavailable in clinical practice. Furthermore, these "black-box" models can fail when encountering out-of-distribution data. This paper introduces the Deep Radon Prior (DRP), a fully unsupervised framework that leverages the structural bias of neural networks to reconstruct CT images without any prior training data.
Methodology
The DRP framework is inspired by the concept of Deep Image Prior (DIP), where the architecture of a convolutional neural network (CNN) acts as an implicit regularizer, favoring natural image structures over noise. Unlike standard DIP, DRP integrates this prior directly into the CT reconstruction pipeline by optimizing in the Radon domain.
The core of DRP is an untrained encoder-decoder network embedded within an iterative reconstruction process. The model employs a cross-domain gradient feedback mechanism: it calculates the error between the network's current projection (via the Radon transform) and the actual sparse measured data. This error is then backpropagated to update the network parameters ($\theta$) and the latent representation ($z$). The optimization follows an alternating scheme that refines the image pixels and network weights simultaneously, ensuring the final output adheres to the physical laws of X-ray geometry.
Experimental Setup
The performance of DRP was evaluated using two open-source datasets: the CHAOS (Combined Healthy Abdominal Organ Segmentation) dataset and the Lung-PET-CT-Dx dataset. It is important to note that DRP is entirely unsupervised; these datasets were used strictly for testing. The model reconstructs each image from scratch using only its specific sparse-view projection data, demonstrating its independence from large-scale training sets.
Results and Discussion
A key insight of this work is the observation of information entropy during reconstruction. Traditional iterative methods often start with a noisy initialization and attempt to remove information to find the underlying structure. In contrast, DRP begins with a simple, low-frequency "halo" and progressively adds structured, high-frequency information.
Quantitative results demonstrate that DRP significantly outperforms traditional methods. In a 30-view configuration, DRP achieved a Peak Signal-to-Noise Ratio (PSNR) of approximately 35.1 dB, compared to 20.3 dB for FBP and 30.2 dB for ADMM-TV. Remarkably, DRP also surpassed supervised models like FBPConvNet in certain metrics, achieving a Structural Similarity Index (SSIM) of 0.941 versus FBPConvNet's 0.838 in specific test cases. Qualitatively, DRP excels at preserving fine edges and maintaining homogeneity in soft tissues, avoiding the over-smoothing common in other unsupervised approaches.
Conclusion
Deep Radon Prior represents a significant step forward in unsupervised CT reconstruction. By treating a neural network as a hyperparametric iterative operator, the framework achieves high-fidelity results that rival or exceed supervised methods without the need for training data. While the iterative nature of DRP currently results in higher computational costs per scan compared to pre-trained models, its superior generalization and physics-integrated approach make it a robust solution for low-dose imaging. Future work will explore Network Architecture Search (NAS) to optimize network structures for specific hardware and extend the framework to address limited-angle CT and metal artifact removal.