More realistic low-resolution OCT image generation approach for training deep neural networks

Abstract

Super-resolution image reconstruction has become a hot topic with the development of deep learning methods, which have been applied in medical images and shown its great potential application. The available simple and uniform bicubic interpolation down-sampling cannot reflect the actual OCT image degradation. A more realistic low-resolution OCT image generation approach is proposed for training deep neural networks. OCT images with high and low resolutions by multiplying two different spectral widths of the light source are obtained. Three kinds of classical deep learning networks are trained to super-resolve OCT images, and the primary results prove their effectiveness. Super-resolution study for the more realistic low-resolution images is of significance for improving the resolution of OCT system in practice.