Vessel Preserving CNN-Based Image Resampling of Retinal Images

High quality resolution enhancement of eye fundus images is an important problem in medical image processing. Retinal images are usually noisy and contain low-contrast details that have to be preserved during upscaling. This makes the development of retinal image resampling algorithm a challenging problem. The most promising results are achieved with the use of convolutional neural networks (CNN). We choose the popular algorithm SRCNN for general image resampling and investigate the possibility of using this algorithm for retinal image upscaling. In this paper, we propose a new training scenario for SRCNN with specific preparation of training data and a transfer learning. We demonstrate an improvement of image quality in terms of general purpose image metrics (PSNR, SSIM) and basic edges metrics-the metrics that represent the image quality for strong isolated edges.