Detection of Diabetic Retinopathy from fundus images using Resnet50

High blood glucose levels cause lesions on the retina of the eye, resulting in a degenerative condition known as diabetic retinopathy (DR), which impacts vision and can cause irreversible vision loss. The most common cause of blindness in diabetic people is thought to be diabetic retinopathy. Early diagnosis of diabetic retinopathy is essential to efficiently maintaining the patient’s vision. We attempted to give first-hand verification to this fundamental problem of DR detection to save time, money and efforts of ophthalmologists. The latter also proved to be more challenging, especially early on in the disease, when disease characteristics are less obvious in the fundus images. Deep learning algorithms and machine learning-based medical image analysis have aided in the early identification of diabetic retinopathy along with the evaluation of retinal fundus images. This paper attempts to preprocess and binary classify fundus images from the famous Aptos dataset using finetuned ResNet50 as well as features extraction from ResNet50 and later classifying using machine learning models. We have achieved an accuracy of 0.9802, an AUC score of 0.9937, F1 score of 0.9870, a precision of 0.9890, a recall as 0.9845 and kappa score of 0.9481 on the evaluation data by fine-tuning of ResNet50.