Predict Diabetic Retinopathy in Early-Stages: A Novel Ensemble Model using Efficient nets and an Automated System to Detect the Disease

Abstract

Diabetic Retinopathy is eye condition caused by high sugar levels inside the blood, which is the origin of excessive pressure inside blood vessels inside the eye, with the smallest vessels being the most vulnerable. This condition does not appear suddenly; rather, it develops gradually over time. After the disease progress, it can show symptoms like blurry vision, changes in vision from blurry to clear, and vice versa, blackspots or dark areas in the vision, poor night vision, fading out of colours, etc. Therefore, pre-emptive identification of disease is one of the beneficial tactics to prevent or get cured of this disease. This technique is also susceptible to human misjudgement, which exists in many clinical diagnoses. An Image Classification Model can accelerate the process of blindness detection in patients. We accomplish this by constructing a classifier using transfer learning that can extract key features from pictures and categorise them into separate stages. This work focused on making an efficient classifier with high accuracy and providing the patient with advance notice of their disease using an easy-to-use mobile application. Our model gave a 0.907 quadratic weighted kappa (QWK) score on independent test dataset and 93.2% accuracy on test time augmented data in multi-class classification. Furthermore, providing the necessary use cases with which the patient can track the diabetic retinopathy screening diagnosis