Deep Learning Based Detection of Diabetic Retinopathy using Retinal Fundus Images

Diabetic retinopathy is a type of diabetes which affects the eye by causing damage to retinal blood vessels. It may have no symptoms at first or cause diminished vision problems. As the condition deteriorates, it affects both eyes, leading to partial or complete loss of vision. This is especially so when blood sugar levels are uncontrollable. As a result, the diabetic patient is at greater risk for developing this condition. The risk of complete and permanent blindness can be avoided if an early detection is made. As a result, effective screening method is required. In this paper the four salient features microaneurysms, blood vessels, hemorrhages and exudates are drawn out from the unprocessed images using image-processing techniques and convolutional neural network is used for automatic identification and it implements fundus images classification of Diabetic Retinopathy. The pre-trained CNNs use DenseNet-169. The transferred CNNs are then fine-tuned using the fundus images. Pre-trained CNN models were considered as feature extractors for fundus pictures. As features, the outputs of the final fully connected layers are used. By using DenseNet-16 the highest accuracy is obtained compared to remaining models. The ensuing result displays visual examples as well as images of the corresponding DRIVE database basic facts. The model is further trained with a Conv2 layer with 128 filters to improve accuracy, and greater integration is used to obtain an accuracy of 80%.