Diabetic Retinopathy Grading Based on a Hybrid Deep Learning Model
Eman AbdelMaksoud, Sherif Barakat, Mohammed Elmogy
Abstract
Diabetic retinopathy (DR) is a dangerous disease that may cause blindness suddenly without any indications. Therefore, it is necessary to continuously screen and audit the disease progress from the early to severe stages. By nature, the color fundus image may facilitate many lesion types that lead to diagnosing different DR grade. In this respect, deep learning achieved great success in medical image analysis, especially in multi-label (ML) classification. In this paper, we present a novel hybrid, deep learning technique for diagnosing different DR grades, which is called the E-DenseNet model. The proposed technique is a hybrid model of the EyeNet and DenseNet using transfer learning. We got benefits from combining the two models as we customized the EyeNet and embedded dense blocks. The proposed computer-aided diagnosis (CAD) system with E-DenseNet can diagnose different DR grades (normal, mild, moderate, severe, and proliferative DR (PDR)) from various ML color fundus images accurately with minimal training time and memory space. The proposed CAD system gives promising results in diagnosing different DR grades from two benchmark datasets. The proposed system achieved an average accuracy (ACC) equals 91.6%, the Dice similarity coefficient (DSC) equals 92.45%, and the Kappa score equals 0.883.