🫁 Endotracheal Tube (ETT) and Carina Segmentation using Deep Learning

---

🧠 Background

An endotracheal tube (ETT) is a flexible plastic airway device inserted through the mouth into the trachea and connected to a breathing machine to provide mechanical ventilation to the lungs.

• Optimal ETT positioning is essential for effective ventilation and patient safety.
• The tube tip should ideally be 5 ± 2 cm above the carina, the point where the trachea divides into the main bronchi.
• ETT positioning is typically evaluated using frontal chest radiographs.
• However, due to high clinical workloads, the average turnaround time (TAT) for radiographic ETT assessment can reach ~23 minutes, delaying treatment decisions.

---

🎯 Objective

⚕️ Develop an automated, accurate, and objective tool for the detection and localization of the ETT and Carina on chest radiographs.

Such a system can:

• Assist radiologists by providing rapid localization feedback.
• Reduce clinical turnaround time.
• Prevent critical misplacement of ETTs.

---

⚙️ Implementation Workflow

This repository includes the first four stages of a full ETT–Carina detection pipeline.

Stage	Description
1. Load Dataset	Import and organize X-ray images
2. Pre-processing	Resize, normalize, augment data
3. ETT Segmentation	Detect ETT using U-Net
4. Carina Segmentation	Detect Carina using U-Net
5. Distance Calculation	(To be added) Compute ETT–Carina distance

---

🖼️ Sample Images

---

🧩 Step 2: Pre-processing

Preprocessing improves model robustness and generalization using:

• Data Augmentation:

  • Horizontal Flip
  • Shift, Scale, Rotate
  • CLAHE (Contrast Limited Adaptive Histogram Equalization)

• Cropping:
  Focus on relevant thoracic regions to reduce background noise.

---

🧠 Step 3: ETT Segmentation using U-Net

The U-Net architecture is a fully convolutional encoder–decoder network designed for biomedical image segmentation.
It combines low-level spatial features with high-level semantic information through skip connections.

U-Net Structure

---

🩻 Step 4: Carina Segmentation using U-Net

The same U-Net structure is used for Carina segmentation, trained on separate labeled datasets.

This allows the model to detect:

• The branching point of the trachea (carina).
• The position of the ETT tip in relation to the carina for distance estimation.

---

📏 Step 5: Distance Calculation (Planned Extension)

After segmentation, the Euclidean distance between the ETT tip and the Carina center is calculated as:

$$ d(p, q) = \sqrt{(x_1 - x_2)^2 + (y_1 - y_2)^2} $$

where:

• $p = (x_1, y_1)$ → ETT tip location
• $q = (x_2, y_2)$ → Carina location

This metric can be later converted into centimeters using pixel spacing information.

---

✅ Final Result Example

---

📂 Repository Structure

ETT_Carina_Segmentation/
│
├── config/
│ ├── carina_config.yaml # Configuration for Carina segmentation
│ └── ett_config.yaml # Configuration for ETT segmentation
│
├── data/
│ ├── carina_v2/ # Carina dataset
│ └── ett_v2/ # ETT dataset
│
├── results/
│ ├── carina/ # Model results and visualizations
│ └── ett/ # ETT localization results
│
├── src/
│ ├── models/
│ │ └── model_factory.py # Model builder (U-Net++, U2Net, etc.)
│ │
│ ├── utils/
│ │ ├── dataset.py # Data loading and augmentation
│ │ ├── losses.py # Dice, BCE, Focal loss
│ │ ├── metrics.py # IoU, Dice metrics
│ │ ├── postprocess.py # Post-processing utilities
│ │ ├── trainer.py # Training and validation loops
│ │ └── visualization.py # Visualization and comparison tools
│ │
│ └── main.py # Main training script
│
├── requirements.txt # Dependencies
└── README.md # Project documentation

---

🧮 Technical Specifications

• Framework: PyTorch
• Model: U-Net / U-Net++
• Loss Function: Dice + Binary Cross-Entropy (BCE)
• Metrics: IoU, Dice Coefficient
• Augmentation: Albumentations & ImgAug
• Visualization: Matplotlib & OpenCV
• Optimizer: Adam

---

💻 Dependencies

torch torchvision segmentation-models-pytorch albumentations imgaug opencv-python numpy matplotlib pandas pyyaml scikit-learn

Install all dependencies using:

pip install -r requirements.txt

---

---

📂 Data Availability

The raw data supporting the conclusions of this study will be made available by the authors upon reasonable request, without undue reservation.
Please contact the author for data access and research collaboration inquiries.

---

🧩 Citation

If you use this work or code in your research, please cite our paper:

A Robust Approach for Endotracheal Tube Localization in Chest Radiographs
Chung-Chian Hsu, Rasoul Ameri, Chih-Wen Lin, Jia-Shiang He, Meghdad Biyari, Atefeh Yarahmadi, Shahab S. Band, Tin-Kwang Lin, Wen-Lin Fan
Frontiers in Artificial Intelligence, 2023, Vol. 6:1181812
https://doi.org/10.3389/frai.2023.1181812

@article{hsu2023robust,
  title={A robust approach for endotracheal tube localization in chest radiographs},
  author={Hsu, Chung-Chian and Ameri, Rasoul and Lin, Chih-Wen and He, Jia-Shiang and Biyari, Meghdad and Yarahmadi, Atefeh and Band, Shahab S and Lin, Tin-Kwang and Fan, Wen-Lin},
  journal={Frontiers in Artificial Intelligence},
  volume={6},
  pages={1181812},
  year={2023},
  publisher={Frontiers Media SA}
}

📫 Contact

Rasoul Ameri
📧 rasoulameri90@gmail.com
🔗 GitHub Profile