Skip to content
Connor Pink

Generative AI for Medical Imaging

My thesis work on synthetic chest X-rays: when they help a tuberculosis classifier, is it because they add new variety, or just because there are more images? Part of it became a short paper at Canadian AI 2026.

year
2026
area
Healthcare AI
where
M.Sc. thesis, University of Guelph · paper at Canadian AI 2026
keywords
Generative AI, Medical Imaging, Diffusion Models, Evaluation

Read the paper ↗ Source on GitHub ↗

Screenshot or figure from Generative AI for Medical Imaging.

This repository is where my thesis work on synthetic chest X-rays lives. Part of it became a short paper at the 39th Canadian Conference on Artificial Intelligence.

The paper

Connor Pink and Ed Sykes. Disentangling dataset size from synthetic diversity in tuberculosis chest X-ray classification. Proceedings of the 39th Canadian Conference on Artificial Intelligence, PMLR 318:954–961, 2026.

PMLR page ↗PDF ↗

I compared two generators on the Tuberculosis Chest X-ray Database: a 256×256 WGAN-GP and a 512×512 latent diffusion model fine-tuned from RoentGen-v2. The diffusion model wins on image-quality metrics (lower FID and Rad-Dino FID). But when the synthetic images were used to train a DenseNet-121 classifier, neither generator beat a control that just duplicated real images up to the same count.

The best synthetic setting was diffusion at a low ratio (0.25 synthetic images per real one). Larger synthetic shares made things worse, especially with the WGAN-GP. Better-looking images don’t guarantee a better classifier, and claims about synthetic data need a count-matched control.

Pipeline and results

Pipeline diagram: real normal and TB X-rays, generative models, preprocessing and splits, three experiment designs with different real-to-synthetic ratios, and the evaluation metrics.
Fig. 1. Experiment pipeline. 4,200 real X-rays (3,500 normal, 700 TB) plus synthetic images feed three designs: fixed size (E1), augmented fixed budget (E2), and augmented scaled budget (E3).
Four line charts of test AUC, test accuracy, TB recall and TB precision against synthetic ratio for each experiment and generator.
Fig. 2. Test AUC, accuracy, TB recall and TB precision as the synthetic share grows, for each design and generator.