An ensemble-based approach for precise diabetic retinopathy classification using contour-guided ROI isolation and multi-blur augmentation

Abstract

Diabetic Retinopathy (DR) is a leading cause of vision loss worldwide, necessitating early and accurate diagnosis for effective treatment. This research proposes a robust ensemble-based approach for DR classification, incorporating advanced data pre-processing and optimization techniques to enhance model performance. The study introduces a contours-based cropping method that isolates the retinal region, minimizing background noise and emphasizing the Region of Interest (ROI). A novel data augmentation strategy combining Gaussian and Bilateral Blurs is employed to simulate real-world imaging conditions, increasing dataset diversity and model resilience. Feature extraction uses pre-trained VGG16 and VGG19 models, which transfer high-level visual knowledge to the DR classification task. Feature selection and dimensionality reduction are achieved through ElasticNet and Principal Component Analysis (PCA), optimizing model interpretability and efficiency. Multiple classifiers, including Gradient Boosting and XGBoost, were evaluated to assess model performance, with the Stacking ensemble achieving the highest accuracy of 94.4%. Hyper parameter tuning using Optuna further refined the model, underscoring the importance of optimization in achieving reliable classification results. The findings highlight the effectiveness of ensemble methods combined with sophisticated pre-processing techniques, presenting a valuable tool for automated DR diagnosis. Future work aims to expand the dataset, explore advanced ensemble techniques, and integrate model interpretability methods to enhance clinical applicability. This research contributes to developing accurate, transparent, and resilient models for DR detection, supporting early diagnosis and improved patient outcomes.