Deep learning for anterior segment OCT angiography automated denoising and vascular quantitative measurement

Abstract

Anterior segment optical coherence tomography angiography (AS-OCTA) has superior advantages in objective assessment of anterior segment (AS) vessels. Resolving noise interference in image is necessary to optimize the application of AS-OCTA. The study aimed to explore an automated denoising algorithm based on deep learning (DL). The algorithm was built through 21,000 pairs of images, and tested with 30 healthy eyes, 47 preoperative eyes with glaucoma, and 30 eyes undergone trabeculectomy (Trab). The real pure noise images were acquired by artificial simulation of eye movements through AS-OCTA. The algorithm included deep convolutional generative adversarial network (DCGAN), Res-Unet and Otsu. ImageJ software quantified vessel density (VD) and vessel diameter index (VDI). Images after noise reduction had relatively satisfactory peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Indiana bleb appearance grading scale (IBAGS), Kronfeld grading system (KGS) and intraocular pressure (IOP) used for Trab analysis. The DL method was superior to the conventional methods (PSNR = 16.45, SSIM = 0.52, both P < 0.001), and the denoising reduced measurement error of VD and VDI (P < 0.001). The denoising methods enabled the differentiation of V2 in IBAGS from V0 and V1 (P < 0.001) or that of II in KGS from I (P = 0.020). VD and VDI could better reflect IOP after noise reduction (R2 increased from 0.25 to 0.63, 0.14 to 0.41, both P < 0.001). Our research offered a DL denoising algorithm which improved the quality of AS-OCTA and the accuracy of AS vessel analysis.