Blind deblurring of retinal OCT images using an adaptive graph total variation

We propose a blind deblurring method for retinal optical coherence tomography (OCT) images degraded by depth-dependent spatially variant blur. Our approach leverages an adaptive graph total variation (AGTV) prior, which dynamically adjusts regularization weights using local gradient statistics from the input image. AGTV autonomously enhances smoothing in severely blurred deep regions while preserving fine structures in shallow layers. Validated on microsphere images, en-face images, and B-scans, AGTV outperforms state-of-the-art methods in PSNR/SSIM metrics and significantly improves retinal layer segmentation accuracy-particularly for deep boundaries. This single-image framework requires no predefined PSF models or hardware modifications, offering a potential solution for clinical OCT enhancement.