Diabetic Retinopathy Detection

Automated detection of diabetic retinopathy from retinal images using deep learning. This project utilizes an ensemble of EfficientNet-B3 models trained with 5-fold cross-validation to achieve high accuracy and robustness.

Performance

• Model: EfficientNet-B3 (Ensemble of 5 folds)
• Input Resolution: 300x300
• Test Set QWK: 0.9091 (on held-out test split)
• Validation QWK: ~0.78 (Mean across 5 folds)

Setup

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

Data Structure

Ensure your data is organized as follows:

dr_detection/
├── dataset_3500/        # Directory containing all training images
├── trainLabels.csv      # CSV with 'image' and 'level' columns
├── src/                 # Source code
├── checkpoints/         # Saved models
└── outputs/             # Test splits and predictions

Usage

1. Training

The training script automatically handles 5-fold cross-validation and creates a held-out test split.

# Train with default settings (EfficientNet-B3, 300x300, 50 epochs)
python src/main.py --data-dir dataset_3500 --test-split-ratio 0.2

# Custom training parameters
python src/main.py \
    --data-dir dataset_3500 \
    --backbone efficientnet_b3 \
    --image-size 300 \
    --batch-size 8 \
    --epochs 50 \
    --test-split-ratio 0.2

2. Evaluation

Evaluate the trained ensemble on the held-out test set.

# Evaluate ensemble on the test split generated during training
python src/evaluate.py \
    --data-dir dataset_3500 \
    --test-file outputs/test_split.csv \
    --backbone efficientnet_b3 \
    --image-size 300 \
    --batch-size 8 \
    --ensemble

3. Utilities

Create Custom Split:

If you need to regenerate a test split manually:

python src/create_split.py --data-dir dataset_3500 --test-size 0.2

Cleanup Checkpoints:

To remove intermediate checkpoints and keep only the best models per fold:

python cleanup_checkpoints.py

Project Structure

• src/main.py: Main training orchestration (CV loops, logging).
• src/trainer.py: Training loop logic, checkpointing, and validation.
• src/model.py: Model architecture definitions (ResNet, EfficientNet).
• src/dataset.py: Data loading, preprocessing, and augmentation.
• src/evaluate.py: Evaluation scripts for single models and ensembles.
• src/args.py: Command-line argument definitions.