AI Based Diabetic Retinopathy Detection and Severity Classification System || VIP || IEEE || HYD

TO PURCHASE OUR PROJECTS IN ONLINE (OR) OFFLINE
CONTACT:VENKAT INNOVATIVE PROJECTS
NAME: VENKATARAO GANIPISETTY
Mobile & WhatsApp :+91 9966499110
Mobile & WhatsApp :+91 9573201550
Email :venkatjavaprojects@gmail.com
Email :venkatinnovativeprojects@gmail.com
website:https://venkatinnovativeprojects.com/
ABOUT PROJECT:
Diabetic Retinopathy (DR) is a serious eye disease caused by long-term diabetes. High blood sugar levels damage the small blood vessels in the retina (the light-sensitive part at the back of the eye). This damage leads to abnormalities such as microaneurysms, hemorrhages, and exudates. If not detected early, DR can progress through stages — Mild, Moderate, Severe, and Proliferative — and may eventually cause permanent vision loss. Early detection and regular monitoring are very important to prevent blindness.
In Existing system traditional healthcare systems, DR is diagnosed manually by ophthalmologists by examining retinal fundus images. This process is time-consuming, requires expert knowledge, and may not be accessible in rural or underdeveloped areas.
In recent research papers, automated systems have been proposed using machine learning algorithms such as:
• Support Vector Machine (SVM)
• K-Nearest Neighbors (KNN)
• Random Forest
• Basic Convolutional Neural Networks (CNN)
Some systems also use feature extraction techniques (like texture or vessel segmentation) before classification. However, these methods have limitations:
• Lower accuracy in multi-class severity grading
• Poor performance on imbalanced datasets
• Heavy dependence on handcrafted features
• Limited capability for disease progression comparison
Many earlier models focused only on binary detection (DR or No DR) instead of full severity classification.
To overcome the limitations of existing systems, our proposed system uses Deep Learning-based automated severity grading. Instead of manually extracting features, we use advanced Convolutional Neural Networks (CNN) and EfficientNet
• It is pre-trained on ImageNet (large dataset)
• It extracts deep hierarchical features automatically
• It provides better accuracy for medical image classification
• It handles complex patterns like microaneurysms and hemorrhages effectively
We also apply:
• Data augmentation
• Class balancing techniques
• Fine-tuning of last layers
• Confusion matrix and performance metrics analysis
Unlike many existing systems, our model performs 5-stage severity classification (0–4) and not just binary detection. Additionally, we implemented a Disease Progress Monitoring Module, where two retinal images taken at different times are compared to determine whether the condition has improved, remained stable, or worsened.
Implementation of Modules Description:
1) Admin & User Authentication Module
This module controls secure access to the system. Users can register, login, and upload images for prediction. Admin login is separate and used to manage dataset upload, preprocessing, training, and monitoring results. User details are stored in MySQL, and session handling is used to restrict pages (admin-only and user-only). This ensures only authorized people can access training and dataset functions.
2) Dataset Upload Module (Admin)
In this module, the admin uploads the DR dataset as a ZIP file (Disease Grading). The system stores it in the server (media folder) and automatically extracts it into a fixed directory. After extraction, the system reads the training labels CSV and displays dataset information such as total images, stage-wise class distribution, and a small CSV preview. This helps verify dataset integrity before preprocessing/training.
3) Preprocessing Module (Admin)
This module prepares the retinal fundus images for deep learning. It reads the CSV label file, matches each image name, assigns its class label (0–4), and creates the required folder structure (train/0..4 and val/0..4). All images are resized to a fixed input size (224×224), converted to RGB, and saved into the correct class folders. A sample preview image is displayed after preprocessing so the admin can confirm that preprocessing worked correctly. This step ensures the dataset is ready for training using flow_from_directory().

4) Model Training Module (Admin)
This is the main AI module. It trains two models: a basic CNN (baseline) and an advanced EfficientNetB3 (transfer learning). The training module loads the preprocessed folders, applies augmentation (rotation, flip, zoom), and trains the selected model. During training, the system saves performance graphs (accuracy and loss) and generates evaluation outputs such as confusion matrix and classification report. The best performing trained model is stored in the server as a .h5 file (example: EfficientNetB3_best.h5), which is later used for prediction.

Видео AI Based Diabetic Retinopathy Detection and Severity Classification System || VIP || IEEE || HYD канала Venkat Innovative Projects