Diabetic Retinopathy Detection

A hybrid deep learning framework for automated diabetic retinopathy detection with uncertainty quantification and explainability.

📋 Table of Contents

• Overview
• Features
• Model Architecture
• Dataset
• Installation
• Usage
• Results
• Future Work
• References
• Contributing

🔍 Overview

Diabetic Retinopathy (DR) is a diabetes complication that affects the eyes and can lead to blindness if left untreated. Early detection is crucial for effective treatment, but manual screening by ophthalmologists is time-consuming and subject to variability.

This project implements a hybrid deep learning approach that:

1. Accurately classifies retinal images into 5 severity levels of DR
2. Quantifies uncertainty in predictions using Bayesian methods
3. Explains decisions through gradient-based visualization techniques
4. Addresses class imbalance through focal loss and weighting strategies

✨ Features

Feature	Description
📊 Preprocessing Pipeline	Ben Graham's technique with green channel extraction, CLAHE, and denoising
🧠 Hybrid Architecture	EfficientNetB0 + Swin Transformer for improved feature representation
🔍 Bayesian Uncertainty	Monte Carlo Dropout for confidence estimation and uncertainty quantification
👁️ Explainable AI	Grad-CAM visualizations showing which retinal regions influence decisions
⚖️ Class Imbalance Handling	Focal Loss and class weighting techniques to handle unbalanced datasets

🏗️ Model Architecture

Our hybrid architecture combines:

• EfficientNetB0: Pre-trained CNN for efficient feature extraction
• Swin Transformer: Attention-based refinement of features with hierarchical window partitioning
• Monte Carlo Dropout: Bayesian approximation for uncertainty estimation
• Grad-CAM: Class activation mapping for model explainability

📊 Dataset

The model is trained and evaluated on the APTOS 2019 Diabetic Retinopathy Detection dataset, which contains retinal fundus photographs labeled with DR severity levels:

Class	Severity Level	Description	Visual Signs
0	No DR	No signs of diabetic retinopathy	Normal retina
1	Mild	Microaneurysms only	Small red dots
2	Moderate	More than microaneurysms but less than severe	Red lesions, hard exudates
3	Severe	Extensive hemorrhages and venous beading	Cotton wool spots, venous beading
4	Proliferative	Abnormal blood vessel growth and potential retinal detachment	Neovascularization, preretinal hemorrhage

🔧 Installation

# Clone the repository
git clone https://github.com/romilagarwal/diabetic_retinoplasty.git
cd diabetic_retinopathy

# Create and activate virtual environment
python -m venv env
source env/bin/activate  
# On Windows use: env\Scripts\activate

# Install dependencies from requirements
pip install -r [requirements.txt]

Additionally, Graphviz must be installed on your system for model visualization.

🚀 Usage

1. Data Preprocessing

python pre_process_with_dataset_download.py

Preprocessing Details

This script:

Downloads the dataset (if not already present) Applies Ben Graham's preprocessing technique with green channel extraction Enhances images with CLAHE (Contrast Limited Adaptive Histogram Equalization) Applies denoising filters Resizes images to 224×224 Organizes processed images into class folders

2. Base Model Training

python efficientnet_model.py

Trains a baseline EfficientNetB0 model with transfer learning from ImageNet weights.

3. Hybrid Model Training

python train_hybrid_model.py

Training Parameters

The hybrid model training uses:

· Focal loss for class imbalance · Mixed precision for memory efficiency · Class weighting for balanced learning · Learning rate scheduling · Early stopping to prevent overfitting

4. Model Evaluation

# Test the base model
python testing_efficientnet_model.py

# Test the hybrid model
python test_hybrid_model.py

5.Bayesian Uncertainty Estimation

python bayesian_inference.py

Uncertainty Metrics

The Bayesian component performs:

· Monte Carlo Dropout inference with multiple forward passes · Confidence score calculation · Uncertainty estimation (standard deviation of predictions) · Predictive entropy calculation · Reliability diagram generation

6.Explainable AI Visualizations

python explainable_ai.py

Generates Generates Grad-CAM visualizations highlighting regions that influence the model's decisions.

7. Generate Visualizations for Publication

python generate_all_visualizations.py

Creates comprehensive visualizations for research papers or presentations.

📈 Results

Performance Metrics

Model	No DR	Mild	Moderate	Severe	Proliferative	Average
EfficientNet	0.76	0.70	0.72	0.65	0.63	0.69
Hybrid Model	0.82	0.75	0.79	0.73	0.71	0.76

Key Improvements · +7% Average F1 Score improvement over baseline EfficientNet · Better Generalization across all DR severity classes · Enhanced Performance on minority classes (Severe and Proliferative) · Reduced Uncertainty in predictions compared to baseline

👁️ Visualizations

Grad-CAM Explainability

Uncertainty Analysis

🔮 Future Work

1. DR-GAN++: Implementation of Generative Adversarial Networks for synthetic data generation to further address class imbalance
2. Ensemble Methods: Combining multiple models for improved performance
3. Clinical Integration: Development of a user-friendly interface for clinical use
4. Mobile Deployment: Optimization for edge devices to enable screening in remote areas Multimodal Learning: Integrating patient metadata with retinal images

📚 References

1. Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional networks. Proceedings of the IEEE conference on computer vision and pattern recognition, 4700-4708.

2. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., ... & Guo, B. (2021). Swin transformer: Hierarchical vision transformer using shifted windows. Proceedings of the IEEE/CVF International Conference on Computer Vision, 10012-10022.

3. Gal, Y., & Ghahramani, Z. (2016). Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. International conference on machine learning, 1050-1059.

4. Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. Proceedings of the IEEE international conference on computer vision, 618-626.

5. Lin, T. Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal loss for dense object detection. Proceedings of the IEEE international conference on computer vision, 2980-2988.

👥 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.