CarinaNet

Automatic detection of carina and ETT (endotracheal tube) in chest X-rays using deep learning. CarinaNet provides a simple, pip-installable package for medical image analysis with state-of-the-art accuracy.

Features

• 🔬 Medical AI: Specialized for carina and ETT detection in chest X-rays
• 🚀 Easy to Use: Simple Python API and command-line interface
• 🎯 High Accuracy: Based on RetinaNet architecture optimized for medical imaging
• 📦 Pip Installable: Just pip install carinanet and you're ready to go
• 🔧 Flexible Input: Supports various image formats (JPG, PNG, DICOM)
• 💾 Bundled Model: Pre-trained weights included - no separate downloads needed

Installation

Install CarinaNet with pip:

pip install carinanet

Quick Start

Python API

import carinanet

# Simple prediction on an image file
result = carinanet.predict_carina_ett("chest_xray.jpg")
print(f"Carina: {result['carina']}")
print(f"ETT: {result['ett']}")
print(f"Carina confidence: {result['carina_confidence']}")
print(f"ETT confidence: {result['ett_confidence']}")

# Using the model class for multiple predictions
model = carinanet.CarinaNetModel()
result = model.predict("chest_xray.jpg")

# Works with PIL Images and numpy arrays too
from PIL import Image
import numpy as np

image = Image.open("chest_xray.jpg")
result = carinanet.predict_carina_ett(image)

# Or with numpy arrays
image_array = np.array(image)
result = carinanet.predict_carina_ett(image_array)

Command Line Interface

# Predict on a single image
carinanet predict chest_xray.jpg

# Get package information
carinanet info

# Show help
carinanet --help

Output Format

CarinaNet returns predictions as dictionaries with the following structure:

{
    'carina': (x, y),           # Carina coordinates
    'ett': (x, y),              # ETT coordinates  
    'carina_confidence': 0.95,  # Confidence score (0-1)
    'ett_confidence': 0.87      # Confidence score (0-1)
}

Coordinates are in image pixel space (x, y) where (0, 0) is the top-left corner.

Requirements

• Python 3.8+
• PyTorch 1.9+
• See requirements.txt for full dependency list

Model Information

CarinaNet uses a RetinaNet-based architecture fine-tuned specifically for carina and ETT detection in chest X-rays. The model:

• Architecture: RetinaNet with ResNet backbone
• Input Size: 640x640 pixels (automatically resized)
• Output: Bounding box coordinates converted to center points
• Training Data: Curated medical imaging dataset
• Performance: High accuracy on clinical validation sets

Use Cases

CarinaNet is designed for:

• 🏥 Clinical Decision Support: Assist radiologists in identifying anatomical landmarks
• 📊 Research: Batch processing of chest X-ray datasets
• 🎓 Education: Teaching anatomical landmark identification
• 🔬 Quality Control: Automated verification of ETT placement

Device Support

CarinaNet automatically detects and uses:

• GPU: CUDA-enabled GPUs for faster inference
• CPU: Fallback to CPU if GPU not available
• Apple Silicon: Optimized for M1/M2 Macs

Contributing

We welcome contributions! Please see our contributing guidelines for details.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Citation

If you use CarinaNet in your research, please cite:

@misc{heiman2025factchexckermitigatingmeasurementhallucinations,
      title={FactCheXcker: Mitigating Measurement Hallucinations in Chest X-ray Report Generation Models}, 
      author={Alice Heiman and Xiaoman Zhang and Emma Chen and Sung Eun Kim and Pranav Rajpurkar},
      year={2025},
      eprint={2411.18672},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2411.18672}, 
}

Support

• 📖 Documentation: See this README and inline documentation
• 🐛 Issues: Report bugs on GitHub Issues
• 💬 Discussions: Join our GitHub Discussions

Disclaimer: CarinaNet is intended for research and educational purposes. It should not be used as the sole basis for clinical decisions. Always consult with qualified medical professionals.