Self-supervised denoising with Edge Perception in OCT images

Due to the low coherence interference characteristic of optical coherence tomography (OCT), the pathological edge structures such as ellipsoid zone are inevitably affected by multiple noises. Although various existing schemes have been proposed for OCT images denoising, they are subject to the requirements of ground images and are difficult to collect clinically. Therefore, we propose a Self-supervised Denoising approach with Edge Perception in OCT images (SDEP-OCT) to effectively reduce the noise level while maintaining tissue edge clarity. The main architecture includes the Spatial Difference Perception Module (SDPM) and Content-Aware Feature Reorganization (CAFR). The SDPM combines channel and spatial attention to effectively capture crucial features of the boundary region, leveraging multidimensional information to maintain the structural integrity of the diseased tissue. By employing multi-channel information prediction to mitigate the impact of high-frequency noise artifacts at lesion boundaries, CAFR proposes using high-dimensional feature recombination prediction to retrieve details of the lesion region. Furthermore, the Global-local Structure Mapping (GSM) loss is designed to enhance the correlation between global and local information to minimize residuals of transition information at boundary recognition points simultaneously. Experimental results demonstrate that our approach exhibits certain advantages in terms of edge clarity compared to existing denoising methods. Specifically, SDEP-OCT achieves at least a 0.2 improvement in PSNR and reduces training time per epoch by 47% compared to the renowned Blind2Unblind method across a spectrum of noise levels, without sacrificing the SSIM.