HYBRID CONVOLUTIONAL NEURAL NETWORK MODEL FOR AUTOMATIC DIABETIC RETINOPATHY CLASSIFICATION

Abstract

Diabetic Retinopathy (DR) is one of the most common complications of diabetes and remains a leading cause of preventable blindness globally. If not diagnosed and treated promptly, it can result in progressive vision loss due to damage to the retina's small blood vessels. Early detection is critical, yet traditional manual screening and diagnosis methods, which rely on the examination of fundus images by ophthalmologists, are often time-consuming, subjective, and susceptible to human error due to the complex and subtle patterns associated with DR, especially in its initial stages.To overcome these limitations, the medical imaging and computer vision communities have explored a wide range of automated diagnostic systems. However, many of the techniques that are currently in use are still unable to adequately capture the complex and diverse characteristics of DR at various severity levels. An innovative and effective deep learning-based automated detection framework for improving diagnostic accuracy and reliability is presented in this study. Our method for robust feature extraction combines the advantages of two powerful convolution neural network architectures, ResNet50 and InceptionV3. These models are pre-trained and fine-tuned to identify relevant patterns from retinal fundus images. The features extracted from both models are fused at the feature level, combining the complementary information captured by each architecture. This fused feature representation is then fed into a custom-designed CNN classifier for final DR classification. Using a publicly accessible data set of labeled fundus images, we trained and validated our model to ensure reproducibility and compare it to previous approaches. The experimental results demonstrate the effectiveness of our proposed framework, achieving notable performance metrics: an accuracy of 96.85%, sensitivity of 99.28%, specificity of 98.92%, precision of 96.46%, and an F1score of 98.65%. Our model has the potential to be used in real-world clinical settings because these results significantly outperform several current cutting-edge methods, particularly in the early detection of DR.