Evaluating the effectiveness of CNN-based models for diabetic retinopathy detection

Abstract

One of the known eye conditions that affect human retinal blood vessels is diabetic retinopathy (DR). People with diabetes are typically at significantly increased risk for this. The blood vessels in the retina are damaged when blood sugar levels in the body increase. Due to the possibility of blindness, people should take precautions and prioritize early detection. As a result, it is a serious condition because it can impair vision. It has several stages, including normal, mild, moderate, severe, and proliferative DR, where it can be quickly determined how severely it has damaged the retinal blood vessels and the area surrounded by the optical disc. Highly qualified specialists typically review the colored fundus photos to diagnose this fatal condition. Clinicians struggle to diagnose this condition accurately, and it takes time. Therefore, several computer vision-based techniques are used to recognize DR and its various stages from retinal scans automatically. These methods, however, can only very roughly categorize the various stages of DR because they are unable to capture the underlying complex properties. However, it is hypothesized that computerized diagnostic systems using intricate Deep Learning (DL) and convolutional neural network (CNN) structures present a compelling approach to learning about different patterns of Diabetic Retinopathy (DR) from fundus images, enabling the precise assessment and categorization of the disease’s severity. This study highlights the performance summary of CNN-based models EfficientNetV2B3, EfficientNetV2S, Inception-RestnetV2, MobileNetV2, a fusion model that combines all of these models, and a KNN classifier that uses all of these features that were extracted from each model to improve the classifications of the stages of DR from these optical fundus images. This will consequently give the model’s accuracy and a confusion matrix. In addition, we provide an accurate explanation of the performance of the models using ExplainableAI. Here, LIME is used for this purpose.