RETINOPATHY IMAGE ANALYSIS FOR EARLY DETECTION OF DISEASE USING MACHINE LEARNING ALGORITHMS

Abstract

Worldwide, diabetes is associated with a number of visual impairments including blind spots, the most common of which is diabetic retinopathy (DR). In order to reduce the likelihood of long-term vision loss and the progression of the disease, early detection is crucial. However, ophthalmologists' knowledge and experience are heavily relied upon in the laborious and time-consuming process of manually examining the retinal fundus pictures. In response to these issues, this research presents a thorough framework for automated diabetic retinopathy detection utilizing retinal fundus photographs, which is based on machine learning. Image resizing, median filtering, Contrast Limited Adaptive Histogram Equalization (CLAHE), color normalization, illumination correction, and supervised machine learning algorithms K-Nearest Neighbors (KNN), Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) are all part of the proposed framework. Then, features are optimized using Principal Component Analysis (PCA) and Recursive Feature Elimination (RFE). Finally, the images are subjected to comparative classification using six supervised machine learning algorithms: LR, RF, SVM, DT, and XGBoost. We used Grid Search with five-fold cross validation to optimize the model's hyperparameters and improve its generalizability and classification performance.  Under the same experimental circumstances, the proposed framework was tested on the publicly available EyePACS retinal fundus picture set using a training-testing split of 80:20. Accuracy, Precision, Recall, F1-score, ROC, AUC, sensitivity, specificity, and confusion matrix analyses were used to evaluate the model's performance. When compared to other machine learning models, XGBoost performed the best in terms of classification accuracy (93.4%), precision (93.1%), recall (94.0%), F1-score (94.0%), sensitivity (93.5%), specificity (93.5%), and area under the curve (0.96). A potential computer-aided decision support system for automated retinal screening in resource-limited health care settings, the suggested framework is accurate, computationally economical, and clinically interpretable in detecting diabetic retinopathy, according to experimental data.