Melanoma Segmentation Using Pseudo Annotations

A deep learning approach for melanoma segmentation using pseudo annotations and semi-supervised learning. This project is implemented as a Kaggle notebook utilizing the ISIC 2017 dataset.

Overview

This notebook implements a semi-supervised learning pipeline for melanoma image segmentation, leveraging both labeled and unlabeled data to improve model performance.

Key Features

• Hair removal using DullRazor algorithm
• U-Net architecture for initial training
• Pseudo-labeling for unlabeled data
• High-confidence filtering
• Combined training approach

Dataset

The project uses the ISIC 2017 dataset, which should be structured in your Kaggle environment as follows:

../input/isic-2017/
    ├── ISIC-2017_Training_Data/
    │   └── ISIC-2017_Training_Data/
    └── ISIC-2017_Training_Part1_GroundTruth/
        └── ISIC-2017_Training_Part1_GroundTruth/

Implementation Details

The notebook contains:

1. Data preprocessing using DullRazor
2. U-Net model implementation
3. Semi-supervised training pipeline
4. Pseudo-label generation and validation
5. Metrics calculation and visualization

Usage

1. Upload the notebook to Kaggle
2. Add the ISIC 2017 dataset to your notebook
3. Run all cells sequentially

Requirements

pip install requirements.txt

License

MIT License

Contact

For questions or support, please open an issue in the repository.