Synthetic Data for Healthcare AI
Medical AI models suffer from limited training data due to privacy regulations and annotation costs. This thesis explores the use of Generative Adversarial Networks (GANs) to synthesize realistic medical images, enabling data augmentation without compromising patient privacy.
Three architectures were implemented and compared: a standard GAN, a Deep Convolutional GAN (DCGAN), and StyleGAN — evaluated on image quality metrics including FID score, SSIM, and expert visual assessment.
Technologies Used
Research Results
Research Contributions
GAN Implementation
Built a standard GAN baseline trained on annotated medical imaging datasets for comparison benchmarking
DCGAN Architecture
Implemented Deep Convolutional GAN with batch normalization for improved stability and image coherence
StyleGAN Integration
Adapted StyleGAN2 for medical domain with custom progressive growing and style mixing techniques
Evaluation Pipeline
Automated evaluation using FID, SSIM and IS metrics with cross-validation on held-out test sets
Privacy Compliance
All synthetic images are verified to contain no patient-identifiable information — fully GDPR-safe
Web Demonstration
Built an interactive web showcase presenting the methodology, results and generated samples
Technical Implementation
Model Training
- TensorFlow 2.x with GPU acceleration
- Custom data preprocessing pipeline
- Adaptive learning rate scheduling
- Gradient penalty (WGAN-GP loss)
Evaluation
- Fréchet Inception Distance (FID)
- Structural Similarity (SSIM)
- Inception Score (IS)
- Expert radiologist visual review
Web Showcase
- Interactive results gallery
- Architecture diagrams
- Comparison sliders
- Deployed on Netlify