Hybrid Deep learning framework for automated Diabetic retinopathy classification using rential fundus images

Abstract

DR is a major cause of blindness. It is particularly crucial to make an early and accurate diagnosis of DR. The manual examination of retinal fundus images is time-consuming, subjective and prone to error and hence there is a motivation for automated computer aided solutions. Kaggle APTOS 2019 is used as the dataset. Images are scaled, normalised, augmented and feature extracted in order to setup parallel pipelines for classification and detection tasks. The range of models used include basic models (ResNet50, MobileNetV2 and DenseNet121), proposed models (SVM-RBF, Modified EfficientNetV2, EfficientNet–SVM hybrid) and additional hybrid or ensemble models (Xception, NASNet-Large and voting classifiers using several models). We use YOLOv5, YOLOv8, YOLOv9 and YOLOv11 as detectors in this work. Accuracy, precision, recall and F1-score are used for classification. DenseNet121 was 99.6 % for all of them. YOLOv5 had the highest precision (0.713) and mAP (0.788) for object detection and YOLOv9 had the highest recall for localising lesions. XAI approaches such as Grad-CAM can point out the important regions inside an image where the AI model based its conclusion on, thereby increasing the interpretability of the model. The web interface developed using flask provides the end to end inference and displays the real-time classified results to the user along with the annotated photographs which have the bounding boxes. It is a very accurate and interpretable hybrid system to detect and classify DR.