🩺 Polyp Segmentation using UNet++ with EfficientNet-B4

Automatic colorectal polyp segmentation from colonoscopy images using UNet++ with an EfficientNet-B4 encoder. This project leverages deep learning for semantic segmentation to accurately identify and localize polyps in endoscopic images, aiding computer-assisted diagnosis.

---

📌 Overview

Colorectal cancer is one of the leading causes of cancer-related deaths worldwide. Early detection and accurate segmentation of polyps during colonoscopy can significantly improve diagnosis and treatment outcomes.

This project implements a semantic segmentation pipeline using UNet++ with an EfficientNet-B4 backbone. The model is trained on the Kvasir-SEG dataset with extensive data augmentation and evaluated using Intersection over Union (IoU) and Dice Coefficient.

---

🚀 Key Features

• ✅ UNet++ architecture for semantic segmentation
• ✅ EfficientNet-B4 pretrained encoder
• ✅ Dice Loss optimization
• ✅ Albumentations-based data augmentation
• ✅ AdamW optimizer with Cosine Annealing Learning Rate Scheduler
• ✅ Automatic train/validation/test split
• ✅ Performance evaluation using IoU and Dice Score
• ✅ Visualization of segmentation predictions

---

🛠️ Tech Stack

• Python
• PyTorch
• segmentation-models-pytorch
• Albumentations
• NumPy
• Matplotlib
• scikit-learn
• Pillow

---

📂 Dataset

Dataset: Kvasir-SEG

The Kvasir-SEG dataset contains 1,000 colonoscopy images with corresponding manually annotated segmentation masks for colorectal polyps.

Dataset Link:

https://datasets.simula.no/kvasir-seg/

---

🧠 Model Architecture

• Architecture: UNet++
• Encoder: EfficientNet-B4
• Loss Function: Dice Loss
• Optimizer: AdamW
• Learning Rate Scheduler: Cosine Annealing LR
• Framework: PyTorch

---

📊 Training Performance

The training and validation metrics demonstrate stable convergence throughout training.

The model shows:

• Consistent decrease in training and validation loss
• Stable convergence during training
• High validation IoU and Dice Score
• Good generalization with only mild overfitting

---

🔍 Sample Predictions

⭐ Excellent Prediction

---

⭐ High Accuracy Prediction

---

⭐ Typical Prediction

---

⭐ Challenging Case

This example demonstrates a more difficult segmentation scenario where the model partially captures the target region. Including such examples highlights both the strengths and limitations of the model.

---

📈 Evaluation Metrics

The model is evaluated using:

• Dice Coefficient
• Intersection over Union (IoU)

Representative prediction results include:

Example	IoU	Dice
Best Prediction	0.984	0.992
High Accuracy	0.963	0.981
Typical Case	0.903	0.949
Challenging Case	0.553	0.712

These values correspond to representative test samples and illustrate the model's performance across varying levels of difficulty.

---

📁 Project Structure

Polyp-Segmentation/
│
├── notebook/
│   └── Polyp_Segmentation_UNetPlusPlus.ipynb
│
├── images/
│   ├── training_history.png
│   ├── prediction_best.png
│   ├── prediction_good.png
│   ├── prediction_normal.png
│   └── prediction_challenging.png
│
├── README.md
├── requirements.txt
└── .gitignore

---

⚙️ Installation

Clone the repository:

git clone https://github.com/anikitachi/Polyp-Segmentation.git
cd Polyp-Segmentation

Install dependencies:

pip install -r requirements.txt

Open the notebook:

notebook/Polyp_Segmentation_UNetPlusPlus.ipynb

Run all cells to train or evaluate the model.

---

🔮 Future Improvements

• Attention U-Net
• DeepLabV3+
• Transformer-based segmentation models
• Hyperparameter optimization
• Mixed precision training
• Model deployment using Streamlit or Flask

---

👨‍💻 Author

Anand Narayan

B.Tech Computer Science and Engineering
Amrita Vishwa Vidyapeetham

GitHub: https://github.com/anikitachi

---

⭐ If you found this project useful, consider giving the repository a star.