Analysis and Detection of Diabetic Retinopathy using Deep Learning Techniques

Abstract

newlineDiabetic retinopathy (DR) is a severe complication of diabetes that affects the eyes, leading to potential vision loss if not detected and managed early. This thesis compares four innovative methods for diagnosing DR, each leveraging advanced machine learning and image processing techniques to enhance diagnostic accuracy and efficiency. Method 1 employs an AdaBoost meta classifier combined with various base classifiers to improve classification performance. It uses a Fuzzy Opponent Histogram Filter for effective feature extraction from retinal images. Among the evaluated base classifiers, the Decision Stump achieved the highest accuracy of 75.39% and a ROC score of 0.759. Other classifiers like Decision Table and JRip also showed strong precision and recall metrics, while the KNN classifier had the lowest performance across all metrics. Method 2 focuses on image segmentation using an Improved Residual U-Net (IRU-Net). This enhanced U-Net framework includes three U-paths, residual blocks, and a channel attention module to better capture and fuse relevant features from fundus images. The method also incorporates a revised weighted focus loss function to address class imbalance in the dataset. Using the DRIVE and IDRiD image libraries, IRU-Net achieved an accuracy of 79%, demonstrating its potential for clinical imaging segmentation with high sensitivity, Dice Similarity Coefficient (DSE), and Intersection over Union (IoU) metrics. Method 3 combines deep learning-based segmentation and classification to assess DR severity. It involves pre-processing, segmentation with DenseNet-169, feature extraction using Enriched Haralick and SURF techniques, and optimal feature selection via the tuna swarm optimization algorithm. Classification is performed using the Deep AlexNet algorithm, which categorizes the disease into Non-proliferative vi DR (NPDR), mild, and severe cases. The model achieved the highest validation accuracy of 91.82% at epoch 10, with notable improvements observed by epoch 3. Method 4 utilizes