A lightweight CNN for Diabetic Retinopathy classification from fundus images

Diabetic Retinopathy (DR) is a complication of diabetes mellitus that damages blood vessel networks in the retina. This is a serious vision-threatening issue in most diabetic subjects. The DR diagnosis by color fundus images involves skilled clinicians to recognize the presence of lesions in the image that can be used to detect the disease properly, making it a time-consuming process. Effective automated detection of DR is a challenging task. The feature extraction plays an excellent role in effective automated disease detection. Convolutional Neural Networks (CNN) have superior image classification efficiency in the present scenario compared to earlier handcrafted feature-based image classification techniques. This work presents a novel CNN model to extract features from retinal fundus images for better classification performance. The CNN output features are used as input for different machine learning classifiers in the suggested system. The model is evaluated through various classifiers (Support Vector Machine, AdaBoost, Naive Bayes, Random Forest, and J48) by using images from generic IDRiD, MESSIDOR, and KAGGLE datasets. The efficacy of the classifier is evaluated by comparing the specificity, precision, recall, False Positive Rate (FPR), Kappa-score, and accuracy values for each classifier. The evaluation results indicate that the proposed feature extraction technique along with the J48 classifier outperforms all the other classifiers for MESSIDOR, IDRiD, and KAGGLE datasets with an average accuracy of 99.89% for binary classification and 99.59% for multiclass classification. Furthermore, for the J48 classifier, the average Kappa-score (K-score) is 0.994 for binary classification and 0.994 for multi-class classification.