Classification of Diabetic Retinopathy via Vascular Removal

Abstract

Diabetic retinopathy is one of the disabling complications of diabetes mellitus that causes the loss of central vision if not recognized and cured at the earlier stage. Existing classification models cannot accurately distinguish early diabetic retinopathy due to the influence of vascular venation. We proposed an image enhancement method for removing blood vessels: use multiple reduced even convolution kernels for mean filtering to blur and shift the vascular features at different levels in the original image. Further, we use convolution block attention module and generative adversarial network in the model, so that the model can weigh the pathological feature weights of different channels in the feature map and has larger feature space. We evaluate the proposed method on EyePACS dataset. It could effectively improve the accuracy of model as compared to use the images without removing blood vessels. The experimental results show that this method can solve the classification difficulty of normal, mild and moderate non-proliferative diabetic retinopathy to some extent.