ResCoWNet: A deep convolutional neural network with residual learning based on DT-CWT for despeckling Optical Coherence Tomography images

Abstract

Various retinal illnesses can be diagnosed and treated using the non-invasive diagnostic technology called OCT (Optical Coherence Tomography). However, the processes used to create OCT images generate speckle noise which considerably reduces the quality of the images, and the subsequent disease diagnosis is negatively impacted by such poor-quality images. The despeckling is effected using conventional methods such as spatial/transform domain filtering, dictionary learning, or a combination of these methods. Deep Convolutional Neural Networks (CNN) have improved the ability to exploit spatial correlations and extract data at various resolutions by adopting a hierarchical network structure, thereby making image-denoising methods more reliable. In deep networks, the Discrete Wavelet Transform (DWT) and its inverse are typically used in place of pooling operations. Nevertheless, the Dual-Tree Complex Wavelet Transform (DT-CWT) can provide a significant performance enhancement compared to deep network models for noise reduction based on DWT pooling. Hence to despeckle OCT images, this paper proposes a novel deep convolutional neural network (ResCoWNet) with residual learning based on DT-CWT. The input noisy image is transformed using DT-CWT into various subbands and is made to pass through several convolutional as well as residual blocks at three different levels in ResCoWNet. The denoised versions of OCT images produced by our model outperform the experimental results both quantitatively and subjectively.