medsynth 1.0.0

Medical Synthetic Data Generator with Privacy-Preserving Synthesis

MedSynth: Medical Synthetic Data Generator with Privacy-Preserving Synthesis

MedSynth is a Python package for generating synthetic medical images with configurable privacy protection. It creates realistic CT scans using statistical methods without requiring machine learning or GPU resources.

Key Features

✨ Three Generation Modes:

1. Pure Synthetic - Generate CT scans from scratch using procedural methods
2. Augmentation - Standard augmentation of real CTs (rotation, scaling, noise)
3. Privacy-Preserving Synthesis - Template-based synthesis with statistical privacy protection using Multi-Scale Statistical Texture Synthesis (MS-STS)

🔒 Privacy Protection (Mode 3 only):

• Low mutual information (empirically < 1.8 bits on tested datasets)
• Synthetic intensity remapping via gradient-preserving transformations
• Designed to reduce re-identification risk

📊 Multiple Output Formats:

• DICOM (medical imaging standard)
• NRRD (3D visualization)
• OMOP CDM (healthcare data standard)

⚡ Computational Efficiency:

• No GPU or deep learning required
• Runs on standard workstations
• Approximately 67 seconds per 178-slice CT volume (hardware-dependent)

---

Performance Benchmarks

Evaluated on NSCLC-Radiomics dataset (November 26, 2025):

Privacy-Preserving Synthesis Mode (Optimized Parameters)

Metric	Value	Interpretation
SSIM (body region)	0.7880	Moderate structural similarity
SSIM (lung region)	0.9527	High structural similarity
Mutual Information	1.08 bits	Low information leakage
Generation Speed	67 sec/volume	Single-threaded on M-series Mac
Voxel Remapping	100%	All intensities transformed

Optimized Parameters (15-iteration empirical search):

• Frequency domain cutoff: 0.5456
• Point spread function blur: 0.40
• Texture noise standard deviation: 5.5449
• Edge enhancement strength: 0.25

Note: Performance may vary on different datasets, scanners, and protocols. These benchmarks represent single-case optimization and should be validated on your specific data.

---

Installation

From PyPI (when published)

pip install medsynth

From Source

git clone https://github.com/ankurlohachab/medsynth.git
cd medsynth
pip install -e .

Development Installation

pip install -e ".[dev]"
pytest tests/

---

Quick Start

Command Line Interface

1. Pure Synthetic Generation

medsynth \
  --num-subjects 10 \
  --output-dir ./output/pure_synthetic/

2. Augmentation Mode

medsynth \
  --augment ./path/to/real/ct/dicom_folder/ \
  --num-subjects 5 \
  --output-dir ./output/augmented/

⚠️ Note: Augmentation mode does NOT provide privacy protection - use for data augmentation only.

3. Privacy-Preserving Synthesis Mode

medsynth \
  --privacy-synthesis \
  --augment ./path/to/real/ct/dicom_folder/ \
  --num-subjects 10 \
  --output-dir ./output/privacy_synthesis/

Python API

from medsynth.config import Config
from medsynth.pipeline import SyntheticCTPipeline

# Configure privacy-synthesis
config = Config(
    num_subjects=10,
    privacy_synth_mode=True,  # Privacy-synthesis using MS-STS
    augmentation_input="./path/to/real/ct",
    output_root="./output/privacy_synthesis"
)

# Generate dataset
pipeline = SyntheticCTPipeline(config)
pipeline.generate_dataset()

---

Generation Modes Comparison

1. Pure Synthetic

Purpose: Generate CT scans from procedural noise without real data input.

Characteristics:

• No real patient data required
• Full control over anatomical features and pathology
• Quality suitable for algorithm development and testing
• No privacy concerns (no patient data involved)

Limitations:

• May lack some real-world anatomical variations
• Texture patterns are synthetic

---

2. Augmentation

Purpose: Standard data augmentation for machine learning training.

Characteristics:

• Applies geometric transformations (rotation, scaling)
• Adds controlled noise
• High fidelity to original (SSIM typically > 0.95)
• Fast processing (~10 seconds per volume)

⚠️ Privacy Warning:

• Does NOT provide privacy protection
• Retains original voxel intensity patterns
• Should only be used when privacy is not a concern
• Not suitable for data sharing under HIPAA/GDPR

---

3. Privacy-Preserving Synthesis (Recommended for Sharing)

Purpose: Generate privacy-protected synthetic versions of real CT scans.

Method: Multi-Scale Statistical Texture Synthesis (MS-STS)

Process:

1. Separates low-frequency anatomy from high-frequency texture
2. Applies gradient-preserving intensity remapping to all tissue types
3. Replaces texture with synthetic scanner-realistic noise
4. Preserves anatomical structure while transforming intensities

Measured Performance (NSCLC-Radiomics test case):

• Body region SSIM: 0.7880 (moderate similarity)
• Lung region SSIM: 0.9527 (high similarity in diagnostic regions)
• Mutual Information: 1.08 bits (low compared to original)

Privacy Considerations:

• Aims to reduce mutual information below 1.8 bits
• All voxel intensities undergo statistical transformation
• Preserves anatomical topology
• Trade-off between privacy protection and diagnostic utility

Limitations:

• Privacy protection is empirical, not cryptographic
• May not prevent all re-identification attacks
• Should be combined with other de-identification methods
• Requires validation for your specific use case

---

Output Formats

DICOM

Standard medical imaging format compatible with PACS systems.

medsynth --generate-dicom --output-dir ./output/

NRRD

3D volumetric format for visualization tools (3D Slicer, ITK-SNAP).

medsynth --generate-nrrd --output-dir ./output/

OMOP CDM

Healthcare data standard for multi-institutional studies.

medsynth --generate-omop --output-dir ./output/

---

Configuration

Custom Parameters

from medsynth.config import Config, VolumeConfig

config = Config(
    num_subjects=50,
    random_seed=42,
    privacy_synth_mode=True,
    augmentation_input="./input/ct/",
    volume=VolumeConfig(
        volume_shape=(178, 512, 512),
        spacing=(5.0, 0.976, 0.976),
        hu_range=(-1024, 3071),
        # MS-STS parameters (optimized values)
        privacy_synth_freq_cutoff=0.5456,
        privacy_synth_psf_blur_sigma=0.40,
        privacy_synth_texture_noise_std=5.5449,
        privacy_synth_edge_enhancement_strength=0.25,
    )
)

---

Quality Metrics

from medsynth.metrics import evaluate_synthetic_ct

results = evaluate_synthetic_ct(
    original=real_ct_volume,
    synthetic=synthetic_ct_volume,
    body_mask=body_mask,
    lung_mask=lung_mask,
    spacing=(5.0, 0.976, 0.976)
)

print(f"SSIM (body): {results['image_quality']['ssim_body']:.4f}")
print(f"PSNR: {results['image_quality']['psnr_body']:.2f} dB")
print(f"Mutual Information: {results['privacy']['mutual_information_body']:.3f} bits")

Available Metrics:

• Image Quality: SSIM, PSNR, MS-SSIM, NMSE
• Clinical Utility: SNR, CNR, Edge Sharpness
• Privacy Analysis: Mutual Information, Histogram Distance, Texture Analysis

---

Examples

See examples/ directory:

• example_pure_synthetic.py - Generate from scratch
• example_augmentation.py - Augment existing CTs
• example_privacy_synthesis.py - Privacy-preserving synthesis

---

Testing

pytest tests/ -v
pytest tests/ --cov=medsynth --cov-report=html

---

Citations & Data Usage

Dataset Citation (Required)

This package was developed and tested using the NSCLC-Radiomics dataset. Users of this data must abide by the TCIA Data Usage Policy and include the following citation:

Aerts, H. J. W. L., Wee, L., Rios Velazquez, E., Leijenaar, R. T. H., Parmar, C., Grossmann, P., Carvalho, S., Bussink, J., Monshouwer, R., Haibe-Kains, B., Rietveld, D., Hoebers, F., Rietbergen, M. M., Leemans, C. R., Dekker, A., Quackenbush, J., Gillies, R. J., Lambin, P. (2014). Data From NSCLC-Radiomics (version 4) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/K9/TCIA.2015.PF0M9REI

Software Citation

If you use MedSynth in your research, please cite:

@software{medsynth2025,
  title = {MedSynth: Medical Synthetic Data Generator with Privacy-Preserving Synthesis},
  author = {Lohachab, Ankur},
  year = {2025},
  month = {11},
  version = {1.0.0},
  url = {https://github.com/ankurlohachab/medsynth}
}

---

Privacy & Security Disclaimer

Important: This software provides empirical privacy protection through statistical methods, NOT cryptographic guarantees.

• Privacy-preserving synthesis aims to reduce mutual information and re-identification risk
• Protection level depends on data characteristics, parameters, and adversary capabilities
• Should be combined with other de-identification methods (e.g., metadata removal, expert review)
• Not a substitute for proper de-identification workflows
• Users are responsible for compliance with applicable regulations (HIPAA, GDPR, etc.)
• Validation recommended for each specific use case and dataset

Not approved for clinical use. For research purposes only.

---

Requirements

• Python ≥ 3.8
• NumPy ≥ 1.21
• SciPy ≥ 1.7
• SimpleITK ≥ 2.1
• PyDICOM ≥ 2.3
• scikit-image ≥ 0.19
• Pandas ≥ 1.3
• Pydantic ≥ 2.0
• pynrrd ≥ 1.0

---

Roadmap

• Web interface
• Cloud deployment

---

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Ensure all tests pass
5. Submit a pull request

---

Development Transparency

AI-Assisted Development Disclosure: AI-assisted development tools were used exclusively for routine tasks such as syntactic error correction, formatting, and generating descriptive comments.

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions

---

Acknowledgments

• Optimization Dataset: NSCLC-Radiomics Collection from The Cancer Imaging Archive (TCIA)
• Method: Multi-Scale Statistical Texture Synthesis (MS-STS) with gradient-preserving remapping
• Optimization: 15-iteration parameter search conducted November 2025

---

Author: Ankur Lohachab

Affiliation: Department of Advanced Computing Sciences, Maastricht University

Contact: [e-mail]

Date: November 26, 2025

Version: 1.0.0