Fundus Image Enhancement Method Based on CycleGAN

Abstract

In this paper, we propose a retinal image enhancement method, called Cycle-CBAM, which is based on CycleGAN to realize the migration from poor quality fundus images to good quality fundus images. It does not require paired training set any more, that is critical since it is quite difficult to obtain paired medical images. In order to solve the degeneration of texture and detail caused by training unpaired images, we enhance the CycleGAN by adopting the Convolutional Block Attention Module (CBAM). To verify the enhancement effect of our method, we not only analyzed the enhanced fundus image quantitatively and qualitatively, but also introduced a diabetic retinopathy (DR) classification module to evaluate the DR level of the fundus images before and after enhancement. The experiments show that our method of integrating CBAM into CycleGAN has superior performance than CycleGAN both in quantitative and qualitative results.