U-Net based Segmentation and Transfer Learning Based-Classification for Diabetic-Retinopathy Diagnosis

Diabetic retinopathy (DR) diagnostic automation using AI is becoming more common. Diseases affecting the blood vessels in the retina, such as those produced by diabetes, are a primary cause of sightlessness and visual impairment worldwide. As a result, early screening and treatment of DR would benefit substantially from automated DR detection systems, preventing visual loss caused by DR. Over the last several years, many methods for identifying anomalies in retinal pictures have been presented by researchers. Traditional automated approaches for detecting diabetic retinopathy relied on manually extracted features from retinal pictures and a classifier for final classification. Diabetic retinopathy may be detected and classified in fundus pictures with the help of a deep learning approach suggested in this study. In this method, the network makes a prediction depending on the quality of the dataset it was trained on. In the first stage, OD (eye) and BV (blood vessel) segmentation are performed using separate U-Net models. The second stage involves applying transfer learning to the deep learning models in order to fine-tune them for improved performance in both the training and validation phases. We use partial data augmentation methods to evenly expand our training dataset. Compared to the sum of the separate models, the suggested weighted classifier performs the best. Additionally, the model is evaluated using the open-access dataset, which consists of 3662 pictures. The APTOS 2019 dataset served as the basis for the five groups. The suggested technique significantly recovers the presentation of DR detection for fundus pictures as assessed by a number of different metrics and compared to current methods to DR identification.