Enhancing Early-Stage Diabetic Retinopathy Detection Using a Weighted Ensemble of Deep Neural Networks

Diabetic Retinopathy (DR) is one of the biggest reasons for vision loss. It is a fatal eye disease damaging the retina, which is the light-sensitive tissue in the rear of the eye. Ophthalmologists use fundus images to capture retinal inner structures to find broken blood vessels and scars. To detect DR on time, early diagnosis is very important which is often not possible due to complex procedures. Therefore, automation of DR detection can solve this problem. Accessibility to regular examinations and specialized eye care remains a challenge, especially in underserved areas, due to late diagnosis, a lack of healthcare infrastructure, and other factors. Although automated detection and grading of diabetic retinopathy from retinal images has shown promising results, challenges arise to accomplish high accuracy, particularly in hidden or early-stage DR situations. One of the major limitations in developing a detection model is the lack of imaging datasets as it requires a large number of images to train the model more accurately. Deep transfer learning-based models have shown promising results especially when datasets are not very large. This study used a weighted average ensemble approach to combine three different deep learning models: inception-v3, VGG16, and a custom-built convolutional neural network. The proposed weighted average ensemble approach achieved an accuracy of 95.06%, a precision of 87.88%, a recall of 83.78%, f1-score of 85.69%, and a 98.10% area under the curve which is higher compared to other pre-trained models. A comprehensive comparative analysis is done to compare the proposed approach with other state-of-the-art methods. The proposed system is efficient, considerably accurate, and can aid as a clinical assistant to detect and grade diabetic retinopathy.