ss25_Hierarchical_Multiscale_Image_Classification

HiPAC — Hierarchical Patch-based Adaptive Classifier

Repository to detect cancer metastasis on lymph node

---

Architechture for MIL Pooling Classification using a residual feature extractor

Installation

Clone the repository and install dependencies:

git clone https://github.com/yourusername/ss25_Hierarchical_Multiscale_Image_Classification.git
cd ss25_Hierarchical_Multiscale_Image_Classification
pip install -r requirements.txt

---

Usage

All commands are run from the root of the repository:

python src/main.py [OPTIONS]

---

Show CLI Options and Flags

• --download: Download the CAMELYON16 dataset.
• --base_dir BASE_DIR: Set the base directory for downloaded files (default: ./data).
• --remote: Download all files (default downloads only a subset for testing).
• -p, --patch: Extract patches from WSIs.
• --patch_level LEVEL: WSI level for patch extraction (0, 1, 2, 3, or 'all').
  
  • Level 0: 1792x1792
  • Level 1: 896x896
  • Level 2: 448x448
  • Level 3: 224x224
  • Example:
    python src/main.py --patch --patch_level 0
python src/main.py --patch --patch_level all
• -prep, --prepare: Prepare data (create validation set, extract masks, etc).
• -val, --validation: Create a validation set (5 normal + 5 tumor images).
• -train, --train: Train a ResNet18 classifier on extracted patches (default, weighted loss for class imbalance).
• --train_strategy: Train a ResNet18 classifier with a specific strategy. Use with --strategy.
• --strategy STRATEGY: Training strategy for ResNet classifier. Options:
  • self_supervised: Use SimCLR pretraining for feature extraction.
  • balanced: Balance the number of tumor and normal patches in the training set.
  • weighted_loss: Use weighted loss for class imbalance (default for --train).
  • Example:
    python src/main.py --train_strategy --strategy balanced
python src/main.py --train_strategy --strategy self_supervised
python src/main.py --train_strategy --strategy weighted_loss
• If you encounter CUDA errors or want to debug GPU operations, you can run with:
  CUDA_LAUNCH_BLOCKING=1 python src/main.py --train_strategy --strategy self_supervised
  This will force synchronous CUDA execution and provide more informative error messages.
• --extract_features: Extract feature vectors from patches using ResNet18.
• --check_structure: Check if the directory structure is correct.

---

Example Workflows

Download a small subset for testing:

python src/main.py --download

Download the full dataset:

python src/main.py --download --remote

Extract patches at a level:

python src/main.py --patch --patch_level 1

Extract patches at all levels:

python src/main.py --patch --patch_level all

Prepare data (validation set, masks):

python src/main.py --prep

Create validation set only:

python src/main.py --val

Train ResNet18 classifier:

python src/main.py --train

Extract features from patches:

python src/main.py --extract_features

Check directory structure:

python src/main.py --check_structure

---

📁 Directory Structure

data/
└── camelyon16/
    ├── train/
    │   └── img/
    ├── val/
    │   └── img/
    ├── test/
    │   └── img/
    ├── masks/
    │   ├── lesion_annotations.zip
    │   └── annotations/
    └── patches/
        └── level_0/
            ├── normal_001/
            ├── tumor_001/
            └── ...
        └── level_1/
        └── level_2/
        └── level_3/

---

⚙️ Configuration

• Modify src/config.py to adjust paths, hyperparameters, and experiment settings.

📄 Citation

If you use this codebase, please cite the repository and the CAMELYON16 dataset.

📝 License

This project is licensed under the MIT License. See LICENSE for details.