Classification of Diabetic Retinopathy through Deep Feature Extraction and Classic Machine Learning Approach

Diabetic Retinopathy (DR) is a complication of diabetes, the leading cause of vision loss in working-age adults. An ophthalmologist can carry out the diagnosis of DR by examining color fundus images. However, the fundus image analysis process takes a long time. Automatic detection of DR is achallenging task. One of the deep learning approaches, Convolutional Neural Networks (CNN), is efficient in image classification tasks. In this research, a CNN architecture is used, namely ResNet-50, as feature extraction and classification. The ResNet-50 feature output at the feature extraction stage is also used as input for machine learning classifiers such as Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbor (k-NN), and Extreme Gradient Boosting (XGBoost). The model works by using fundus images from the DIARETDBI dataset. Data augmentation and preprocessing are proposed in this study to facilitate the model in recognizing images. The performance of each classifier is evaluated based on accuracy, sensitivity, and specificity. The SVM classifier achieved 99% for accuracy and sensitivity in the 80:20 dataset composition. The k-NN classifier obtains the highest specificity for the same dataset's design by 100%.