Ophthalmic Segmentation v1.0.0

Ophthalmic Segmentation v1.0.0

First official release of the high-accuracy ophthalmic fundus image segmentation project.

🎯 Key Features

• Val Dice = 0.9554 (V16 with EfficientNet-B4)
• Complete V16 training script with Combined Loss (Focal + Dice + Lovasz)
• 4-direction Test-Time Augmentation (TTA)
• Hot-start training support
• Real model inference scripts
• Example segmentation results

📦 Included

• scripts/train_v16.py - Complete V16 training script
• scripts/predict.py - Real model inference
• ophthalmic_segmentation/ - Python package with model, data, losses
• examples/ - Example segmentation results (Original | Prediction | Overlay)
• docs/ - Training report and results analysis

🔬 Model Architecture

UNet++ with EfficientNet-B4 encoder, trained with:

• Combined Loss: 0.3×Focal + 0.4×Dice + 0.3×Lovasz
• CosineAnnealingWarmRestarts scheduler
• Hot-start from V13 pretrained weights

📊 Performance

Metric	Score
Val Dice (no TTA)	0.9554
Val Dice (4x TTA)	0.9398
ROI Dice	0.9282
Reflex Dice	0.9660

📖 Usage

pip install ophthalmic-segmentation

Or use the training/inference scripts directly.

📄 License

MIT License