Semantic denoising autoencoders for retinal optical coherence tomography

Abstract

Noise in speckle-prone optical coherence tomography tends to obfuscate important details necessary for medical diagnosis. In this paper, a denoising approach that preserves disease characteristics on retinal optical coherence tomography images in ophthalmology is presented. We propose semantic denoising autoencoders, which combine a convolutional denoising autoencoder with a priorly trained ResNet image classifier as regularizer during training. This promotes the perceptibility of delicate details in the denoised images that are important for diagnosis and filters out only informationless background noise. With our approach, higher peak signal-to-noise ratios with PSNR = 31.0 dB and higher classification performance of F1 = 0.92 can be achieved for denoised images compared to state-of-the-art denoising. It is shown that semantically regularized autoencoders are capable of denoising retinal OCT images without blurring details of diseases.