RADDINO Feature Extraction Pipeline

This repository contains Jupyter notebooks for extracting deep learning features from medical images (specifically chest X-rays) using a pre-trained RADDINO model.

Notebooks

1. Global Feature Extraction (Global_feature_extraction.ipynb) - Extracts global image features using the RADDINO model.

2. Patch Feature Extraction (Patch_feature_extraction.ipynb) - Extracts patch-level features from medical images using the RADDINO model.

Both notebooks implement an efficient pipeline leveraging:

• PyTorch and PyTorch Lightning for deep learning operations
• MONAI (Medical Open Network for AI) for medical image-specific processing
• Parallel execution and persistent caching for performance optimization
• GPU acceleration for faster processing

Dependencies

Dependencies are managed via uv, with a lock file included to ensure reproducible environments.

# Install uv if you haven't already
pip install uv

# Clone the repository
git clone https://github.com/f10409/RAD-DINO_Embedding_Extractor.git
cd RAD-DINO_Embedding_Extractor

# Create a virtual environment and install dependencies using uv
uv sync

Usage

1. Update the BASE_PATH variable in the get_data_dict_part() function
2. Configure parameters like cache settings, batch size, and GPU selection
3. Provide a CSV file with image paths in the ImagePath column
4. Run the notebook to extract and save features to the specified output directory

Pipeline Workflow

1. Data loading from CSV containing image paths
2. Medical image-specific preprocessing
3. Dataset preparation with persistent caching
4. Model initialization
5. Validation through visual spot-checking
6. Feature extraction using PyTorch Lightning

The extracted features are saved to disk for downstream tasks like classification or clustering.