Vertex and edge labelling strategies for graph-based computed tomography image denoising

Abstract

Low-dose computed tomography (LDCT) reduces radiation exposure but introduces elevated noise and streak artifacts that degrade structural fidelity. This paper proposes a graph-based LDCT denoising framework that stabilizes graph construction through explicit vertex and edge labelling guided by paired full-dose CT (FDCT) data. The overlapping LDCT patches are modeled as vertices, and FD-guided affinities are used to build a structurally consistent adjacency matrix and a Laplacian spectrum that are less sensitive to noise. Denoising is performed by spectral filtering via spectral graph wavelet transform (SGWT), followed by overlap–add patch aggregation for image reconstruction. Experiments on paired LDCT/FDCT slices (318 pairs) show that FD-guided labelling improves denoising quality compared with conventional filters and non-guided graph baselines. Quantitative results demonstrate higher peak signal-to-noise ratio (PSNR)/structural similarity index measure (SSIM) with improved edge and feature preservation, indicating better structural boundary retention under noise suppression.