Carotid Ultrasound Image Segmentation using U-Net

This project performs automatic segmentation of carotid artery regions in ultrasound images using a U-Net deep learning model. The dataset contains real-world medical ultrasound images and their corresponding expert-annotated masks.

---

🧠 What is Carotid?

The carotid arteries are the two major blood vessels in the neck that supply blood to the brain, neck, and face. Detecting blockages or abnormalities in these arteries is crucial for preventing strokes and monitoring vascular health.

Medical ultrasound imaging is commonly used to scan these arteries because it is non-invasive, safe, and cost-effective.

---

🎯 Project Goal

To build an accurate semantic segmentation model using U-Net that can automatically detect and segment the carotid artery in grayscale ultrasound images.

---

🧰 Technologies Used

• Python 🐍
• TensorFlow / Keras
• OpenCV & PIL
• NumPy, Pandas, Matplotlib
• Sklearn (train/test split)
• KaggleHub (to download dataset)

---

📁 Dataset Info

• Source: Kaggle - Carotid Ultrasound Images
• Total Images: 1100
• Each image has a corresponding binary mask (0 = background, 1 = carotid artery).
• Image format: .jpg, Mask format: .png

---

📦 Steps Performed

1. Dataset Loading & Validation

• Downloaded the dataset using kagglehub.
• Matched each image with its corresponding mask.
• Ensured no missing or duplicate files.

2. Preprocessing

• Resized images and masks to 256x256 pixels.
• Normalized pixel values to [0, 1] range.
• Split data into 80% training and 20% testing sets.

3. U-Net Model

🔍 What is U-Net?

U-Net is a deep learning segmentation algorithm. It is based on CNNs, designed specifically for biomedical image segmentation. The name U-Net comes from its U-shaped architecture, which has: A contracting path (encoder): learns context using convolution and pooling. An expanding path (decoder): enables precise localization using upsampling and skip connections.

📌 Why U-Net for this project?

~Carotid artery segmentation is a pixel-wise classification task. ~U-Net performs well on small datasets and gives accurate segmentation masks for medical images.

4. Training

• Loss Function: binary_crossentropy
• Optimizer: Adam
• Metrics: accuracy
• EarlyStopping & ReduceLROnPlateau used to improve training.
• Trained for 5 epochs.

5. Results

• Achieved 99.34% accuracy on the test set.
• Model correctly segmented carotid arteries in unseen images.

---

📊 Visual Results

Examples of segmented outputs:

Original Image	Ground Truth Mask	Predicted Mask
✅ Img	✅ Mask	✅ Output

---

🧪 Future Improvements

• Use Dice Coefficient or IoU for better segmentation metrics.
• Train longer (more epochs) for enhanced performance.
• Use data augmentation to improve generalization.
• Try other models like ResUNet, Attention U-Net, or TransUNet.

---

🧑‍⚕️ Real-World Application

This model can be used in clinical support tools for radiologists or vascular surgeons, helping to:

• Identify blocked or narrowed carotid arteries.
• Monitor patients over time.
• Reduce diagnostic error.

---

🧠 Learnings from the Project:

Understood how U-Net is used for medical image segmentation, especially for segmenting carotid artery regions.

Learned the process of preprocessing medical images and masks for deep learning.

Gained hands-on experience in training deep learning models using PyTorch.

Learned to evaluate segmentation performance using metrics like IoU (Intersection over Union).

Understood the importance of visualizing predictions to check model accuracy in medical applications.

---

🤝 Acknowledgements

• Dataset by Orvile
• U-Net architecture based on the paper: U-Net: Convolutional Networks for Biomedical Image Segmentation by Olaf Ronneberger et al.

---

📌 Author

• ✍️ Developed by: Aarya Falle
• 💡 Guided by: U-Net paper and community notebooks