A Multiscale Convolutional Neural Network for Speckle Noise Removal in Optical Coherence Tomography Images

Abstract

Optical Coherence Tomography (OCT) is proficient in imaging micrometric resolution of retinal cross sections to identify ophthalmology diseases. However, the OCT images are corrupted by multiplicative speckle noise, generates severe degradation in its quality. The challenge is to remove these inherent noises to avoid deceptive results in fringe areas. Recent advances in discriminative learning can promisingly remove complex noises in SD-OCT images. This paper proposes a methodology for effective denoisingmodel which deploys convolutional network with multiscaleoperation. Ahugeamount of local features can be generated with applying dissimilar filter sizes parallelly.The performance of the proposed approach is assessed on Duke (SD-OCT) dataset. The suggested technique is evaluated against conventional speckle denoising methods on parameters namely PSNR, SSIM,AD, LMSE, NK and NAE.Experimental analysis show that our method can effectively minimize noise and preserve the retinal structuresthan other traditional methods in terms of both quantity and quality assessment.