Diabetic Retinopathy Detection using Deep Learning

A comprehensive deep learning project for automated classification of diabetic retinopathy stages from retinal fundus images. This research implements and compares multiple CNN architectures, transfer learning approaches, regularization techniques, and ensemble methods to achieve robust medical image classification.

---

📋 Table of Contents

• Overview
• Dataset
• Project Structure
• Installation
• Usage
• Models & Results
• Key Findings
• Documentation
• Citation

---

🔍 Overview

Diabetic retinopathy (DR) is a leading cause of vision loss among adults with diabetes. This project develops an AI-powered diagnostic system capable of classifying DR severity into 5 stages:

• Class 0: No DR (Healthy)
• Class 1: Mild Non-Proliferative DR
• Class 2: Moderate Non-Proliferative DR
• Class 3: Severe Non-Proliferative DR
• Class 4: Proliferative DR

Key Features

✅ Custom CNN Architecture: Baseline model built from scratch
✅ Transfer Learning: EfficientNet-B0 & ResNet-50 with fine-tuning strategies
✅ Advanced Regularization: Dropout, Weight Decay, Label Smoothing, Data Augmentation
✅ Ensemble Learning: Soft voting across 5 best models
✅ Unsupervised Analysis: Convolutional Autoencoder for anomaly detection & latent space visualization
✅ Comprehensive Evaluation: ROC-AUC, Confusion Matrix, Learning Curves, t-SNE visualization

---

📊 Dataset

Source: Kaggle - Diabetic Retinopathy Detection

• Total Images: 2,750
• Image Size: 256×256 pixels (RGB)
• Classes: 5 (staged severity levels)
• Split: 70% train / 15% validation / 15% test

Class Distribution

Class	Label	Count
0	Healthy	~550
1	Mild DR	~450
2	Moderate DR	~600
3	Severe DR	~400
4	Proliferative DR	~500

Note: Dataset not included in repository. Download from Kaggle and place in data/ directory.

---

📁 Project Structure

DiabeticRetinopathy(CVproj)/
├── README.md
├── requirements.txt
├── checkpoints/                      # Saved model weights
│
├── data/                             # Dataset (not included)
│   ├── Healthy/
│   ├── Mild DR/
│   ├── Moderate DR/
│   ├── Severe DR/
│   └── Proliferate DR/
├── docs/                             # Research documentation (PDF report)
├── exploration/                      # Jupyter notebooks for experiments and visualization
│   ├── baseline.ipynb
│   ├── transfer.ipynb
│   ├── regularization.ipynb
│   └── autoencoder.ipynb
├── models/                           # Model architectures
│   ├── __init__.py
│   ├── baseline_cnn.py
│   ├── transfer_models.py
│   ├── ensemble.py
│   └── autoencoder.py
├── scripts/                          # Training scripts
│   ├── train_baseline.py
│   ├── train_transfer.py
│   ├── train_ensemble.py
│   └── train_autoencoder.py
├── training/                         # Training utilities
│   ├── config.py
│   ├── data_processor.py
│   ├── train_pipeline.py
│   ├── early_stopping.py
│   └── check_gpu.py
└── results/                          # Output metrics

---

🚀 Installation

Prerequisites

• Python 3.8+
• CUDA-capable GPU (recommended)
• 8GB+ RAM

Setup

# Clone repository
git clone https://github.com/deyme17/Diabetic-Retinopathy-CV.git
cd Diabetic-Retinopathy-CV

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

# Install dependencies
pip install -r requirements.txt

# Verify GPU availability
python training/check_gpu.py

Download Dataset

1. Download from Kaggle
2. Extract to data/ directory maintaining folder structure

---

💻 Usage

Training Models

1. Baseline CNN

python scripts/train_baseline.py

2. Transfer Learning (EfficientNet-B0)

python scripts/train_transfer.py --model "efficientnet_b0" --freeze_until 2

3. Transfer Learning (ResNet-50)

python scripts/train_transfer.py --model "resnet50" --augmentation 1

4. Ensemble Model

python scripts/train_ensemble.py --models "checkpoint1.pth" "checkpoint2.pth"

5. Autoencoder (Anomaly Detection)

python scripts/train_autoencoder.py --epochs 50

Jupyter Notebooks

Explore experiments interactively:

jupyter notebook exploration/

• baseline.ipynb - Custom CNN development
• transfer.ipynb - Transfer learning experiments
• regularization.ipynb - Dropout, weight decay, augmentation analysis
• autoencoder.ipynb - Unsupervised learning & t-SNE visualization

---

📈 Models & Results

Performance Comparison

Model	Val Accuracy	Test Accuracy	Parameters	Training Time
Baseline CNN	64.00%	63.55%	~3.67M	21 epochs
EfficientNet-B0 (freeze_until=2)	72.00%	69.57%	~4.0M	9 epochs
ResNet-50 (regularized)	62.00%	65.66%	~23.5M	17 epochs
Ensemble (Soft Voting)	70.00%	70.57%	Combined	N/A

Key Metrics (Best Ensemble Model)

• Accuracy: 70.57%
• Precision: 0.6537
• Recall: 0.6557
• F1-Score: 0.6559

Class-wise Performance (AUC-ROC)

Class	AUC
Healthy	0.99
Mild DR	0.92
Moderate DR	0.82
Severe DR	0.92
Proliferate DR	0.90

Autoencoder Results

• Anomalies Detected: 17/299 test images
• Detection Threshold: μ + 2.3σ (MSE-based)
• Latent Space Dimensionality: 16×16×256

---

🔬 Key Findings

1. Transfer Learning Superiority

• EfficientNet-B0 achieved +6% accuracy over baseline CNN
• Compact architecture (4M params) outperformed ResNet-50 (23.5M params)
• Partial freezing (freeze_until=2) optimal for medical imaging domain adaptation

2. Effective Regularization

• Dropout (0.5) + Data Augmentation extended training epochs
• Weight Decay (1e-4) reduced validation loss
• Label Smoothing proved ineffective (decreased accuracy by 2-3%)

3. Ensemble Benefits

• Soft Voting increased accuracy by +1% while improving stability
• Critical for Proliferate DR detection
• Reduced false negatives in high-risk classes

4. Unsupervised Insights

• t-SNE visualization revealed continuum between DR stages (no clear cluster boundaries)
• Healthy class forms distinct cluster, but Mild/Moderate DR overlap significantly
• Autoencoder successfully identified low-quality images and severe anomalies

5. Challenges

• Class imbalance required weighted loss functions
• Moderate DR (Class 2) most difficult to classify (often confused with Mild/Severe)
• Medical domain gap: ImageNet pretraining requires careful fine-tuning

---

📚 Documentation

Full research report available in docs/DiabeticRetinopathyReport.pdf (Ukrainian)

Contents:

1. Problem Statement & Dataset Analysis
2. Theoretical Background (CNNs, Transfer Learning, Autoencoders)
3. Baseline Model Development
4. Transfer Learning Experiments
5. Regularization Techniques
6. Ensemble Methods
7. Autoencoder Analysis
8. Comparative Results & Conclusions