Laparoscopy Image Defogging AI
An AI-powered laparoscopic image enhancement system designed to remove fog, haze, and surgical smoke from minimally invasive surgical imagery using deep learning and image restoration techniques.
This repository contains pretrained Pix2Pix UNet-256 generator weights for real-time laparoscopic image defogging and enhancement.
Model Details
Model Description
This model is designed to improve the visual clarity of laparoscopic surgical images by removing:
- Lens fogging
- Surgical smoke
- Haze
- Low contrast artifacts
The system combines:
- Pix2Pix GAN image translation
- Dark Channel Prior (DCP)
- Guided filtering
- CLAHE enhancement
- Contrast restoration
- Sharpening and post-processing
The model aims to enhance visibility in minimally invasive surgical environments for research and educational applications.
- Developed by: Vishnu Das
- Model type: Pix2Pix GAN / UNet-256 Generator
- Framework: PyTorch
- Language(s): English
- License: MIT
- Task: Image-to-Image Translation / Medical Image Enhancement
Model Sources
- Repository: https://github.com/YOUR_GITHUB_USERNAME/YOUR_REPOSITORY_NAME
- Model Repository: https://huggingface.co/vishnudaspk/Laparoscopy-Image-Defogging-AI
Uses
Direct Use
This model can be used for:
- Laparoscopic image enhancement
- Surgical smoke removal
- Fog removal
- Medical imaging research
- Computer vision experimentation
- Deep learning demonstrations
Downstream Use
Possible downstream applications:
- Real-time surgical visualization systems
- AI-assisted medical imaging pipelines
- Surgical simulation environments
- Medical video enhancement workflows
Out-of-Scope Use
This model is NOT intended for:
- Clinical diagnosis
- Real surgical deployment
- Medical decision-making
- Autonomous healthcare systems
Outputs should always be reviewed by qualified professionals.
Bias, Risks, and Limitations
- Performance depends heavily on image quality and training distribution.
- The model may produce artifacts under severe smoke or lighting conditions.
- Results may vary across different laparoscopic devices and environments.
- This system is intended for research and educational purposes only.
Recommendations
Users should:
- Validate outputs before use
- Avoid clinical reliance
- Test across multiple datasets
- Use CUDA-enabled GPUs for best performance
How to Get Started with the Model
Installation
pip install -r requirements.txt
Recommended:
- NVIDIA GPU
- CUDA-enabled PyTorch
Place Model Weights
Place:
best_net_G.pth
inside:
scripts/checkpoints/pix2pix_laparoscopy_dc/
Run Application
python app.py
Open:
http://127.0.0.1:5000
Training Details
Training Data
The model was trained on custom laparoscopic imagery containing varying levels of:
- Surgical smoke
- Fogging
- Low visibility
- Illumination artifacts
Data preprocessing included:
- Resizing
- Contrast normalization
- Paired image generation
Training Procedure
Preprocessing
- Image normalization
- CLAHE enhancement
- Resizing
- Data augmentation
Training Hyperparameters
- Architecture: Pix2Pix UNet-256
- Framework: PyTorch
- Training regime: Mixed precision CUDA training
- Loss Functions: GAN Loss + L1 Loss
Evaluation
Testing Data, Factors & Metrics
Testing Data
Custom laparoscopic test imagery.
Metrics
Evaluation metrics include:
- PSNR
- SSIM
- Visual perceptual quality
Results
The model demonstrated:
- Improved image clarity
- Reduced haze and smoke artifacts
- Enhanced contrast and edge visibility
Environmental Impact
Training performed on:
- Hardware Type: NVIDIA RTX 4060 GPU
- Framework: PyTorch CUDA
Technical Specifications
Model Architecture and Objective
- Pix2Pix GAN
- UNet-256 Generator
- Image-to-image translation objective
Compute Infrastructure
Hardware
- NVIDIA RTX 4060 Laptop GPU
Software
- Python
- PyTorch
- OpenCV
- NumPy
- Flask
Citation
If you use this project in research or educational work, please cite the repository appropriately.
More Information
This project was developed as a deep learning and medical imaging research initiative focused on improving surgical visualization quality using AI-powered enhancement techniques.
Model Card Authors
Vishnu Das
Model Card Contact
For questions or collaboration: