AI-Based Detection of Diabetic Retinopathy Severity Levels

Diabetic Retinopathy (DR) is an eye complication due to high sugar levels leading to vision loss (partial or permanent) in diabetic patients. Early detection can help in eradicating vision loss and worsening the situation. Manual screening is tedious and expensive to carry out on a large scale. Accurate automated detection of the ailment becomes the need of the hour. This work proposes DR severity detection using deep Convolutional Neural Network (CNN) model utilizing the state-of-art transfer learning approaches. The pretrained models(DenseNet121, ResNet50, VGG19, EfficientNetV2B3 and MobileNetV2) are trained and tested on publicly available dataset - APTOS 2019 dataset. The images are pre-processed to increase the accuracy of feature extraction. Class imbalance is resolved using class weights methodology. Data augmentation and callback scheduler techniques are used for preventing overfitting of the five selected pretrained models used in our experimentation. It is observed that DenseNet121 is highly effective due to its dense pattern of capturing accurate information, it gives a high quadratic weighted kappa score of 0.8945 compared to other models when used for multi-class classification, confirming DenseNet121 is the most effective and accurate model for early diagnosis of DR for automated detection.