Sinogram-Free Low-Dose CT Reconstruction via Differentiable Radon-Regularized Optimization Unrolling.
Authors
Affiliations (3)
Affiliations (3)
- Biomedical Engineering , Central University of Rajasthan, CURAJ, Ajmer, RJ, 305817, India.
- Department of Biomedical Engineering, Central University of Rajasthan Ajmer Rajasthan, CURAJ, Ajmer, 305817, India.
- Department of Biomedical Engineering, Sathyabama University, OMR, Chennai, 600119, India.
Abstract
Reducing X-ray dose in computed tomography (CT) is essential for patient safety; however, low-dose acquisitions introduce noise and artifacts that degrade diagnostic image quality. Most existing physics-informed reconstruction methods rely on raw projection-domain measurements to enforce data consistency, although sinogram data are rarely retained in routine clinical workflows. To address this limitation, we propose the Unrolled Radon Consistency Network (URCN), a reconstruction framework that incorporates physically motivated priors without requiring access to projection measurements.
URCN embeds approximate projection-domain structure as a differentiable spatial regularizer within a learned unrolled optimization architecture. At each reconstruction stage, a parallel-beam Radon surrogate computes a projection residual between the current reconstruction estimate and the low-dose input. The backprojected residual generates a spatially structured gradient that guides iterative image refinement. Unlike conventional inverse-problem formulations, this mechanism operates as a geometry-aware image-domain regularizer rather than an exact data-consistency constraint, enabling physics-informed reconstruction directly from reconstructed DICOM images.
Evaluated on 30 held-out Mayo Clinic patients, URCN achieved a PSNR of 39.4 ± 0.4 dB, SSIM of 0.929 ± 0.013, RMSE of 0.035 ± 0.003, and LPIPS of 0.098 ± 0.011. Compared with SwinIR-CT, the strongest sinogram-free baseline, URCN produced a statistically significant PSNR improvement of 0.7 dB (p < 0.01; Cohen's d = 1.1). Task-based analysis confirmed improved spatial resolution preservation alongside reduced noise magnitude. Direct evaluation on the AAPM dataset without retraining maintained a 0.6 dB advantage, demonstrating robustness to acquisition-domain shift. These results establish that structured differentiable gradient injection provides a practically viable and physically grounded pathway for low-dose CT reconstruction in retrospective clinical settings where raw projection data are unavailable.
.