Overview
This project performs blood vessel segmentation from retinal fundus images using a custom U-Net model. It is trained on the DRIVE dataset and focuses on robust vessel extraction that can support early diabetic retinopathy analysis workflows.
Key Highlights
- Custom U-Net implementation in PyTorch with Conv + BatchNorm + ReLU blocks.
- Combined BCE + Dice loss for stable training on a small medical dataset.
- Includes both CLI inference and a Streamlit app for interactive testing.
- Achieves test performance around Dice 0.76 and IoU 0.62.
Technologies
- Core: Python, PyTorch
- Data Pipeline: Albumentations, NumPy
- App Layer: Streamlit
- Dataset: DRIVE