Back to projects

ProgressGAN-CXR

Severity-conditioned GANs that simulate COVID-19 chest X-ray progression, honestly benchmarked four ways.

SWE (AI/MLOps)GANsMedical ImagingGenerative ModelsPyTorchONNXMLOps

A full-stack medical AI product: a FastAPI backend serving ONNX-quantized versions of all four generators plus a ResNet18 severity classifier, all pulled from the HuggingFace Hub at runtime, containerized with Docker and monitored through Prometheus and Grafana. The React and Tailwind frontend, deployed on Vercel, ships four interactive features: a real-time severity slider that regenerates an X-ray as it moves, a diagnostic challenge game that pits the user against the classifier, a side-by-side four-model comparison, and a research dashboard surfacing the FID, Spearman, and SSIM findings directly.

Live demo: progressgan-cxr.vercel.app home page
Disease progression simulation: same noise vector, severity 0.0 (Normal) to 1.0 (COVID-19)
Four-model comparison: FID, severity disentanglement (Spearman r), and temporal consistency (SSIM)
Downstream classifier accuracy: real-only vs real-plus-synthetic training data
Severity disentanglement analysis: DCGAN vs WGAN-GP semantic severity encoding

What I built

  • Built a FastAPI backend serving ONNX-quantized inference for all four generators and a ResNet18 classifier, with every model pulled from the HuggingFace Hub at runtime rather than bundled locally.
  • Containerized the backend with Docker, added health checks, and instrumented Prometheus metrics on every endpoint, visualized through a Grafana dashboard.
  • Shipped a React and Tailwind frontend on Vercel with four product surfaces: a real-time severity slider, a diagnostic challenge game, a four-model side-by-side comparison, and a research dashboard rendering the FID/Spearman/SSIM plots.
  • Set up MLflow for model versioning and GitHub Actions CI/CD that runs tests before every deploy.
  • Designed the API and model-loading layer so all secrets and model sources are environment-driven, no hardcoded keys or local checkpoint paths.

Tech stack

SWE (AI/MLOps)GANsMedical ImagingGenerative ModelsPyTorchONNXMLOps

Key results

142.14

Best FID (DCGAN)

0.996

Best disentanglement (Spectral DCGAN, r)

0.963

Production model (KD) SSIM

95.28%

Classifier accuracy

Back to all projects