Automated Diabetic Retinopathy Prediction Using a Deep Learning Based Modified DenseNet

Diabetic Retinopathy (DR) is a form of diabetes caused by damage to the retinal blood vessels. Early identification and treatment of this condition are facilitated through retinal screening. DR progresses through five phases: usual, malignant, acute, moderate and slight. To enable early detection and intervention, computer-aided diagnostic methods are essential. Numerous deep learning-based automated approaches have demonstrated their effectiveness in accurately grading DR. In this paper, a deep leaning technique called modified Densenet-169 is proposed for early detection of DR. Initially, the retinal images are pre-processed using Gabor filter to enhance the visibility of blood vessels for easy detection of DR. Then the size of the dataset is increased by using image data generator. Finally, the images are processed in modified Densenet-169 which classifies retinal images as absence of DR, light DR, beginner DR, serious DR, and progressive DR. The performance of DenseNet is enhanced by including four pairs of dense blocks and transition layers. Dense block in DenseNet combines feature maps by concatenation, while the transition layer reduces the dimensions via a composite of average pooling, 1x1 convolution, and batch normalization. The empirical observations showcase that the modified Densenet169, as proposed, achieves superior levels of accuracy, precision, and recall in DR detection than cutting-edge methods.