Enhancing Diabetic Retinopathy Detection with a Hybrid Model of DenseNet121 and Support Vector Machines

Abstract

Diabetic Retinopathy (DR) is a leading cause of vision impairment globally, emphasizing the need for efficient and accurate diagnostic solutions. This paper presents a hybrid model combining DenseNet121 for feature extraction and Support Vector Machines (SVM) for classification to detect and classify DR into five stages: No DR, Mild, Moderate, Severe, and Proliferative DR (PDR). The model addresses key challenges such as imbalanced datasets, medical image noise, and imaging conditions variability by incorporating advanced data preprocessing and augmentation techniques. Ten pre-trained models were evaluated on a curated in-house dataset created from IDRiD, Kaggle, and AEH sources. DenseNet121 demonstrated superior performance, achieving an accuracy of 92.83% and an F1-score of 87.73%, followed by EfficientNetB0 and Xception. The results underscore the importance of architecture selection and robust preprocessing in improving DR classification. The proposed framework provides a scalable and accurate solution for retinal screening, paving the way for early detection and treatment of DR, especially in under-resourced regions. Future work aims to optimize weaker models, integrate ensemble methods, and extend the approach to additional datasets for broader applicability.