Deep Feature Extraction and Classification of Diabetic Retinopathy Images using a Hybrid Approach

Hybrid approaches have improved sensitivity, accuracy, and specificity in Diabetic Retinopathy (DR) classification. Deep feature sets provide a more holistic analysis of the retinal images, resulting in better detection of premature signs of DR. Hybrid strategies for classifying DR images combine the strengths of extracted deep features using pre-trained networks and Machine Learning (ML)-based classifiers to improve classification accuracy, robustness, and efficiency. Perfect pre-trained networks VGG19, ResNet101, and Shuffle Net were considered in this work. The networks were trained using a transfer learning approach, the pre-trained networks were chosen according to their classification accuracy in conjunction with the Softmax layer. Enhanced characteristics were extracted from the pre-trained networks' last layer and were fed to the machine learning-based classifier. The feature reduction and selection methods are essential for accomplishing the desired classification accuracy. ML-based kNN classifier was used to classify DR and a PCA-based feature reduction approach was utilized to obtain optimized deep feature sets. The extensive experiments revealed that ResNet101-based deep feature extraction and the kNN classifier delivered enhanced classification accuracy. It is concluded that combining deep features and the ML-based classifier employing a hybrid method enhances accuracy, robustness, and efficiency. The hybrid approach is a powerful tool for the premature identification of DR abnormalities. The PCA-kNN classification algorithm, which employs features obtained from the ResNet101, attained a peak classification accuracy of 98.9%.