Skip to content
David Lonski
← Back to research

layer: archive · research log

Medical Image Synthesis

Undergraduate research with Dr. Ning Yu · 2025

GANs for synthetic medical imagery, built before generative AI went mainstream. 2nd place at SUNY Brockport's annual AI competition.

  • Python
  • PyTorch
  • DC-GAN
  • W-GAN
  • Kaggle

what I did

  • 01Implemented DC-GAN for stable convolutional training and W-GAN to reduce mode collapse.
  • 02Trained for roughly 450,000 iterations, tracking generator and discriminator loss to catch instability.
  • 03Benchmarked the synthetic images against real OCT scans from Kaggle.
  • 04Stress-tested LLMs on high-resolution images to find edge cases and hallucination triggers, the same failures I watch for in agent systems today.

from the work

A grid of real OCT eye scans next to a larger grid of generated scans
Real OCT scans (left) beside scans generated by the GAN (right).
Line chart of generator and discriminator loss over about 450,000 iterations
Generator (G) and discriminator (D) loss across training.
David Lonski standing next to his research poster
Presenting the poster at SUNY Brockport Scholars Day 2025.

presented

Poster #35 at SUNY Brockport Scholars Day 2025: “A Low-Cost Approach for Fake Medical Image Synthesis.”

next log

OCR & Object Recognition →