Abstract:
The majority of people worldwide have the risk of getting the eye disease called
Diabetic Retinopathy (DR). This is a major problem that could impair vision for
most of diabetic patients. High glucose levels in retinal blood vessels are the main
reason for this. Color fundus images are used to diagnose DR from databases like
IDRiD, KAGGLE, MESSIDOR and DIARETDB1 etc. The traditional method re quires trained doctors to determine the presence of lesions in the image, that may
be utilised to effectively detect the illness, making it a time-consuming process. The
effective classification of the disease relies heavily on feature extraction. Due to the
superior image grading efficiency of deep learning approach, computer diagnosis of
DR has developed into a viable tool for rapid detection and evaluation of the severity
of DR. In this work, different types of CNN architectures are used to extract the
features. The CNN output features are used as input for different types of machine
learning methods (Support Vector Machines, Decision Tree, Random Forest, Naive
Bayes Classifier and K-Nearest Neighbour). Deep learning techniques like (ResNet 50, VGG-16, MobileNet, EfficientNet-B3, Inception-v3, and Self-designed models) are
also implemented for classification. Here performed a performance comparison of dif ferent techniques using different datasets. The models are evaluated using various
evaluation metrices such as accuracy, precision, recall and auc etc. The model having
highest value for evaluation metrices is selected as the best model.
