🩺 Kvasir-HealthAI: Advanced U-Net Pipeline for Gastrointestinal Polyp Segmentation

---

📘 Overview

Kvasir-HealthAI is an end-to-end deep learning pipeline for gastrointestinal polyp segmentation using the Kvasir-SEG dataset.
It features a modular design with preprocessing, training, and explainability stages — enabling reproducible, high-performance medical segmentation.

---

⚙️ Preprocessing Pipeline

On-the-fly image enhancement improves illumination uniformity and texture contrast.
Each input frame passes through:

Stage	Technique	Purpose
1️⃣	Specular highlight removal	Reduces glare reflections
2️⃣	Homomorphic filtering	Corrects uneven illumination
3️⃣	Guided filtering	Smooths noise while preserving edges
4️⃣	CLAHE	Local contrast enhancement
5️⃣	Retoning	Normalizes brightness and tone range

Example

Original	After Preprocessing

Effect on training

• Higher Dice & IoU
• Fewer false negatives in low-contrast areas
• Faster convergence and more stable learning

---

🧠 Training Setup

python -m src.training.train

Default Hyperparameters

Parameter	Value
Image Size	256×256
Batch Size	8
Epochs	40
Learning Rate	1e-3 (Adam)
Loss Function	BCE + Dice
Scheduler	Cosine Annealing + Early Stopping
Device	CUDA / CPU auto-detect

Best Result

Performance Summary

🎯 Dice: 0.8554  📈 IoU: 0.7838  ⚙️ Learning Rate: 3.75e-05

---

📊 Results

Metric	Value
Dice	0.8554
IoU	0.7838
Validation Loss	0.1603
Train Loss	0.0416

Training Curves

Loss Curves	Metrics (Dice / IoU)

---

🔍 Explainability (XAI)

SegGrad-CAM is used to visualize the model’s focus regions during segmentation.
It highlights the areas contributing most to the final prediction, enabling interpretability and reliability in medical AI systems.

Visualization Example

Input	Grad-CAM	Overlay

---

🧰 Folder Structure

Kvasir-HealthAI/
│
├── src/
│   ├── dataset/                  # Dataset loading & augmentation
│   ├── models/                   # U-Net model definitions
│   ├── preprocessing/            # Image preprocessing pipeline
│   ├── training/                 # Loss, training loop, early stopping
│   ├── explainability/           # Grad-CAM integration
│   ├── utils/                    # Visualization & metrics
│
├── assets/
│   ├── images/                   # Raw sample visuals
│   ├── results/                  # Training logs, plots, checkpoints
│   ├── xai_visualizations/       # Grad-CAM outputs
│
├── examples/
│   ├── inference_example.py
│   ├── explainability_demo.py
│
├── README.md
├── setup.py
├── requirements.txt
└── LICENSE

---

🧪 Quick Inference

from src.models.unet import UNet
from src.utils.visualization import overlay_mask
import torch, cv2

model = UNet(in_ch=3, out_ch=1, base=64)
model.load_state_dict(torch.load("assets/results/unet/best_unet.pt"))
model.eval()

img = cv2.imread("assets/images/test_sample.png")
x = torch.from_numpy(img.transpose(2,0,1)).unsqueeze(0).float()/255.
with torch.no_grad():
    mask = torch.sigmoid(model(x))[0,0].numpy()

🧭 Roadmap

• Add Attention U-Net and U-Net++ variants
• Deploy inference as FastAPI + ONNX Runtime
• Integrate Weights & Biases (wandb) for experiment tracking
• Enable mixed precision for Jetson / Edge AI deployment
• Expand preprocessing for multi-center datasets
• Publish pretrained models on Hugging Face Hub

---

🧑‍💻 Author

Ahmet Yasir Duman
Computer Engineer — Healthcare AI & Computer Vision
📧 ahmetyasirduman@gmail.com
🔗 LinkedIn • GitHub

---

https://github.com/ahmetduman23/Kvasir-HealthAI

🩶 Citation

If you use this repository, please cite it as:

@misc{duman2025kvasirhealthai,
  author       = {Ahmet Yasir Duman},
  title        = {Kvasir-HealthAI: Advanced U-Net Pipeline for Gastrointestinal Polyp Segmentation},
  year         = {2025},
  url          = {https://github.com/ahmetduman23/Kvasir-HealthAI}
}

## 📜 License

This project is licensed under the **MIT License**.  
You are free to use, modify, and distribute the code with proper attribution.  
See the [LICENSE](LICENSE) file for complete details.

---

> **Kvasir-HealthAI** — bridging *Medical Imaging* and *Explainable Deep Learning.*

© 2025 Ahmet Yasir Duman — All rights reserved.