ducknet-polyp-segmentation

🦆 Duck-Net: Polyp Segmentation Model

🎯 Model Description

This is a Duck-Net model fine-tuned for polyp segmentation in colonoscopy images. The model is based on a U-Net architecture with Duck-inspired multi-scale feature extraction blocks for superior medical image segmentation performance.

📊 Model Performance

• Validation Dice Coefficient: 92.88%
• Validation Jaccard Index: 48.92%
• Dataset: Kvasir-SEG (1000 colonoscopy images)
• Training Time: ~83 minutes on GPU
• Model Size: ~29.6 MB

🚀 Quick Start

from huggingface_hub import hf_hub_download
import torch
import torch.nn as nn

# Define the model architecture
class DuckNet(nn.Module):
    def __init__(self, img_size=(256, 256), num_classes=3):
        super(DuckNet, self).__init__()
        # ... (model definition)
        
    def forward(self, x):
        # ... (forward pass)
        return torch.sigmoid(output)

# Download and load model
model_path = hf_hub_download(
    repo_id="ibrahim313/ducknet-polyp-segmentation",
    filename="pytorch_model.bin"
)

model = DuckNet(img_size=(256, 256), num_classes=3)
model.load_state_dict(torch.load(model_path, map_location='cpu'))
model.eval()

# Inference
import albumentations as A
from albumentations.pytorch import ToTensorV2

transform = A.Compose([
    A.Resize(256, 256),
    A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ToTensorV2()
])

# Process image
transformed = transform(image=your_image)
input_tensor = transformed['image'].unsqueeze(0)

with torch.no_grad():
    prediction = model(input_tensor)
    binary_mask = (prediction > 0.5).float()

🏗️ Model Architecture

This Duck-Net implementation features:

1. Encoder-Decoder Structure: U-Net based architecture
2. Multi-scale Features: Duck-inspired feature extraction
3. Skip Connections: Direct feature transfer between encoder and decoder
4. Batch Normalization: Stable training and inference
5. Sigmoid Activation: Multi-class segmentation output

Key Features

• Input Size: 256×256×3 (RGB images)
• Output: 256×256×3 (Multi-class segmentation mask)
• Parameters: ~7,766,051
• Architecture: Duck-Net (U-Net variant)

📋 Training Details

Dataset

• Source: Kvasir-SEG dataset
• Total Images: 1000 colonoscopy images with polyp annotations
• Split: 800 training, 100 validation, 100 test
• Classes: Background, Polyp, Other tissue

Training Configuration

• Optimizer: Adam (lr=0.001)
• Loss Function: Jaccard Coefficient Loss
• Batch Size: 8
• Epochs: 50
• Hardware: Tesla P100 GPU
• Framework: Originally TensorFlow/Keras, converted to PyTorch

💻 Usage Example

import cv2
import numpy as np
from PIL import Image

def predict_polyp(image_path, model, threshold=0.5):
    # Load image
    image = cv2.imread(image_path)
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    
    # Preprocess
    transformed = transform(image=image)
    input_tensor = transformed['image'].unsqueeze(0)
    
    # Predict
    model.eval()
    with torch.no_grad():
        prediction = model(input_tensor)
        binary_mask = (prediction > threshold).float()
    
    return binary_mask.squeeze().numpy()

⚠️ Medical Disclaimer

Important: This model is for research and educational purposes only. It should not be used for clinical diagnosis or treatment decisions. Always consult qualified medical professionals for clinical applications.

📚 Citation

If you use this model in your research, please cite:

@misc{ducknet_polyp_2024,
  title={Duck-Net for Polyp Segmentation in Colonoscopy Images},
  author={Ibrahim313},
  year={2024},
  howpublished={Hugging Face Model Hub},
  url={https://huggingface.co/ibrahim313/ducknet-polyp-segmentation}
}

👥 Model Card Authors

• Developed by: ibrahim313
• Model Type: Convolutional Neural Network (Duck-Net)
• License: Apache 2.0
• Repository: https://huggingface.co/ibrahim313/ducknet-polyp-segmentation

🔧 Technical Specifications

🏥 Applications

• Colonoscopy image analysis
• Polyp detection and segmentation
• Medical imaging research
• Computer-aided diagnosis (research only)

🤝 Contact

For questions or issues, please open an issue in the repository or contact the model author.