HP-Skin-01: Epidermis Segmentation Project

This project implements deep learning models for epidermis segmentation in skin whole slide images using U-Net and DeepLabv3+ architectures.

Quick Start

1. Setup Environment

# Create and setup conda environment
bash setup_environment.sh

# Activate environment
conda activate hp-skin-01

2. Run Complete Pipeline

# Run preprocessing and training
bash run_training.sh

# Evaluate trained models
bash run_evaluation.sh

Dataset Structure

Place your datasets in the following structure:

dataset/
├── Histo-Seg/
│   ├── WSI/          # .jpg files (20x resolution)
│   └── Mask/         # .jpg files (multiclass masks)
└── Queensland/
    ├── WSI/          # .tif files (10x resolution)
    └── Mask/         # .png files (multiclass masks)

Pipeline Overview

1. Binary Mask Generation: Extracts epidermis pixels from multiclass masks

   • Histo-Seg: RGB(112, 48, 160)
   • Queensland: RGB(73, 0, 106)

2. Patch Extraction: Creates 384×384 patches with tissue segmentation

   • Non-overlapping patches
   • Tissue segmentation to remove background
   • Paired WSI-mask patches

3. Model Training: Trains U-Net models with different encoders

   • ResNet50 encoder
   • EfficientNet-B3 encoder
   • Pure Dice loss
   • Wandb integration for experiment tracking

4. Evaluation: Computes Dice, IoU, and other metrics on test set

Models

• U-Net with ResNet50: Balanced performance
• U-Net with EfficientNet-B3: Higher accuracy, more parameters
• DeepLabv3+: Coming soon

File Formats

• Histo-Seg: .jpg for both WSI and masks
• Queensland: .tif for WSI (requires OpenSlide), .png for masks

Results

• Models are saved in experiments/*/checkpoints/
• Evaluation results in evaluation_results/
• Training progress tracked on Weights & Biases

Troubleshooting

OpenSlide Installation (for .tif support)

# Ubuntu/Debian
sudo apt-get install openslide-tools

# macOS
brew install openslide

GPU Memory Issues

• Reduce batch size in configs/training_config.yaml
• Enable gradient accumulation
• Use mixed precision training

Environment Issues

# Verify installation
python -c "import torch; print(f'PyTorch: {torch.__version__}')"
python -c "import segmentation_models_pytorch as smp; print('SMP installed')"

Documentation

See CLAUDE.md for detailed project documentation and implementation details.

License

This project is for research purposes only.