Journal of Engineering Research and Reports

Aims: The purpose of this study is to create a deep learning-based method for automatic classification of Diabetic Retinopathy (DR) with the help of convolutional neural networks (CNNs) to facilitate early detection and enhance treatment outcomes.

Study Design: Experimental study with model training and testing.

Place and Duration of Study: The study was carried out using a dataset of retinal images gathered from publicly available sources Kaggle Datasets.

Methodology: The dataset of 2,750 retinal images, labeled into five DR severity grades, was preprocessed using data augmentation methods and divided into training, validation, and test sets. The DenseNet model was trained and tested using performance metrics such as accuracy, precision, recall, and F1-score. Furthermore, a web application was implemented using Streamlit to provide easy-to-use real-time DR classification.

Results: Experimental outcomes show that DenseNet-121 exhibited high classification accuracy of 92, 88 percent of validation accuracy and 88.3 percent of recall and precision scores. hence is a robust solution for DR detection in an automated fashion. The efficacy of the model was confirmed using extensive evaluation metrics to ensure its robustness in real-world scenarios.

Conclusion: The suggested deep learning model, as part of a web-based application, presents a cost-effective and accessible early DR detection solution. The technique has the capacity to aid medical practitioners in early diagnosis, hence lowering the threat of vision impairment in diabetic patients. It is advisable to carry out further studies to improve the generalization and performance of the model on different datasets.