A Deep Bottleneck U-Net Combined With Saliency Map For Classifying Diabetic Retinopathy In Fundus Images

Abstract

Early detection of retinopathy plays an important role in the care of people with diabetes. Classification of diabetic retinopathy in fundus images is very challenging because the blood vessels in the retinal images are too small. Morphology of objects with multi-level saliency is the recent choice because of the activation of feature extraction. However, the challenges of the input models are very complex with the blood. The color, lighting or context can become the reasons that create the decline of the primary key for training. This paper proposes a method for classification of diabetic retinopathy using saliency and shape detection of objects based on a deep Bottleneck U-Net (DbU-Net) and support vector machines  in retinal blood vessels. The proposed method includes four stages: preprocessing, feature extraction using DbU-Net, saliency prediction and classification based on the support vector machine. To evaluate this method, its results are compared to the results of the other methods by using the same datasets of STARE and DRIVE for testing with evaluation criteria such as sensitivity, specificity, and accuracy. The accuracy of the proposed method is about 97.1% in these datasets. To assess the levels of diabetes, the diagnostician must initially identify the retinal image with diabetes or not. The result of this paper may help the diagnostician to easily do this.