Advancing diabetic retinopathy classification using ensemble deep learning approaches

Abstract

Diabetic retinopathy is a condition that significantly weakens diabetic individuals, identified by impairment of the blood vessels in the retina. Successful treatment requires early diagnosis and categorization using retinal image segmentation and classification. This study proposes a hybrid pre-trained convolutional neural network (CNN) and recurrent neural network (RNN) architecture to categorize the severity levels of diabetic retinopathy accurately. The proposed model capitalizes on the feature extraction capabilities of CNNs and the spatial dependencies captured by RNNs to achieve higher classification accuracy. The CNN is trained on a generous dataset and optimized on the retinal dataset to extract salient features specific to the task. The RNN then utilizes these features to create a final classification by discovering their spatial relationships. The proposed hybrid pre-trained CNN-RNN model outperforms existing leading-edge approaches on an openly accessible DR dataset, reaching a precision of 0.96. The promising results reveal the potential of the proposed model to accurately and efficiently categorize the severity levels of diabetic retinopathy, which could ultimately improve the diagnosis and intervention. By facilitating early detection and treatment, the model can potentially decrease the threat of severe vision loss and blindness, enhancing patient outcomes and quality of life.