Hybrid Deep Learning and YOLO-Based Lesion Detection for Diabetic Retinopathy

Abstract

In order to increase diagnostic accuracy and usability, this study proposes an expanded lesioncentric diabetic retinopathy (DR) detection system that combines cutting-edge deep learning, machine learning, and real-time object identification algorithms. The upgrade includes the Xception model for improved feature extraction in addition to hybrid CNN-based classification utilizing ResNet and GoogleNet. Optimized machine learning classifiers and an ensemble voting technique are then used to increase robustness. Multiple YOLO-based detection models (YOLOv5x6, YOLOv5s6, YOLOv8n, and YOLOv9n) are used to precisely localize retinal lesions such microaneurysms, hemorrhages, and exudates, allowing for quick and reliable anomaly identification. Additionally, to facilitate user interaction, testing, and result display, a Flask-based web interface with secure user authentication is created. The expanded system is appropriate for affordable diabetic retinopathy screening in embedded and resource-constrained healthcare settings because to its enhanced accuracy, real-time performance, and practical application.