This project aims to detect and grade the level of Diabetic Retinopathy (DR) in retinal images using advanced machine learning techniques. DR is a complication of diabetes that affects the eyes and can lead to blindness if not detected and treated early. It is one of the leading causes of vision impairment among working-age adults globally.
- 0: No DR (No Diabetic Retinopathy)
- 1: Mild DR (Mild Non-Proliferative Retinopathy)
- 2: Moderate DR (Moderate Non-Proliferative Retinopathy)
- 3: Severe DR (Severe Non-Proliferative Retinopathy)
- 4: Proliferative DR (Proliferative Retinopathy)
- Prevalence: Diabetic Retinopathy affects nearly one-third of the diabetic population, with millions at risk of vision loss.
- Early Detection: Early detection through regular screening can prevent 90% of diabetes-related vision loss.
- Impact: DR can lead to severe complications, including blindness, if left untreated. It significantly affects the quality of life and imposes a substantial economic burden due to treatment costs and loss of productivity.
- Diabetics: All individuals with diabetes are at risk, making up a significant portion of the global population.
- Elderly: Older adults with diabetes are at a higher risk due to the prolonged duration of diabetes.
- Underserved Communities: Populations with limited access to healthcare services are less likely to undergo regular eye screenings, increasing their risk.
The project uses a pre-trained InceptionV3 model to classify the DR levels into five categories: No DR, Mild DR, Moderate DR, Severe DR, and Proliferative DR. The project encompasses the following steps:
We used the APTOS dataset, which contains labeled images indicating different levels of Diabetic Retinopathy (DR).
- Images are resized to 256x256 pixels.
- Normalized to a [0, 1] range.
- Enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE).
- Rotations (90, 120, 180, and 270 degrees).
- Horizontal flips to increase the dataset size and improve model generalization.
- Initial Training: Used InceptionV3 pre-trained on ImageNet, added custom layers, and trained with the base layers frozen for 5 epochs using RMSprop optimizer.
- Fine-Tuning: Unfroze all layers and continued training for 4 more epochs using the SGD optimizer with a learning rate of 0.0001 and momentum of 0.9.
The trained model was saved to Google Drive for persistent storage using model.save('my_model').
- Flask Web Application
- Home Page: Users can upload images.
- Backend: The uploaded image is sent to a backend server hosted on Colab via a POST request.
- Model Loading: The Colab server loads the model from Google Drive, predicts the DR level, and sends the result back to the Flask app for display.
- Ngrok
- Exposing Local Server: Used Ngrok to expose the local Flask app to the internet, enabling user access.
git clone https://github.com/ShanmukhiKairuppala/Diabetic_Retinopathy_Detection.git
cd Diabetic_Retinopathy_Detection- Train the model using
HyperParameterTuning_%2B_CLAHE_on_DeepVision.ipynband save it to Google Drive.
- Open
Mini_Project_DR_Flask_App.ipynbin Google Colab. - Run the notebook to start the backend server and obtain the Ngrok link.
- Copy the Ngrok link and paste it into the COLAB_SERVER_URL variable in main.py.
- Run the Flask app using the command:
python main.py- Access the app via http://localhost:5000.
- Upload an image via the Flask app.
- The Flask app sends the image to the Colab server for prediction.
- The Colab server processes the image, predicts the DR level, and returns the result.
- The Flask app displays the prediction result to the user.
This project provides an accessible tool for early detection of Diabetic Retinopathy, potentially aiding in timely medical intervention and reducing the risk of vision loss in diabetic patients.
Contributions are welcome! Please open an issue or submit a pull request for any improvements or bug fixes.