Enhanced Early Detection of Diabetic Retinopathy Based on Transfer Learning using DenseNet121

Abstract

Diabetic retinopathy (DR) is a complication of diabetes that impacts the eyes. It damages the blood vessels of the light-sensitive tissue at the rear of the eye (retina). Initially, diabetic retinopathy may be asymptomatic or may produce mild vision problems. Nevertheless, it can lead to vision loss. Additionally, the decreasing ratio of ophthalmologists per population and the growing number of diabetes patients have extended waiting times for patients needing help. Thus, the early detection of diabetic retinopathy is more effective in facilitating diagnoses and preventing the progression of the illness to more severe stages. In recent years, deep learning has become a popular approach for enhancing performance in diverse fields, with a particular emphasis on medical image classification. Transfer learning models are extensively utilized as a deep learning technique and demonstrate considerable efficacy. This paper aims to use various deep-learning disciplines to detect diabetic retinopathy by employing specialized fundus imaging cameras. Our research is based on binary and quinary classification of DR with a handcrafted CNN model, transfer learning using DenseNet121, and this latter model in different tree-based algorithms and SVM classifiers. The proposed DenseNet121 model achieved better performance than other approaches with an accuracy of 99% on the binary and 81% on the quinary classifications. Finally, our model is compared with recent and pertinent research.