Case study · Images & video
Generative AI for knee X-rays
A research collaboration with a university hospital in Seoul on generative AI for knee radiographs: re-generating X-ray images with state-of-the-art models to support earlier disease detection, fewer imaging errors and enhanced X-ray quality.
Business result
Together with a university hospital, we tested whether generative models can produce realistic knee radiographs. That groundwork supports better image quality and more training data for under-represented conditions.
Client type
Research with a university hospital in Seoul
Where this applies for you
Healthcare and imaging teams short on labeled data or dealing with variable image quality.
Technical result
- 3generative model familiesStyleGAN3, Stable Diffusion and Pix2Pix GAN
- 3clinical aimsEarlier detection · fewer imaging errors · enhanced X-rays
The problem
Radiograph quality varies with equipment, positioning and exposure, and some conditions are under-represented in the data. The clinical team wanted to know whether generative models could produce realistic, useful knee images.
What we built
We experimented with three generative families on knee radiographs: style-based generation (StyleGAN3), diffusion (Stable Diffusion) and paired image-to-image translation (Pix2Pix GAN), and compared what each does well for re-generating knee images.
Workflow
How it works, step by step.
The pipeline in 4 stages, from curate to compare. Each stage has a clear input, a clear output and a reason to exist.
- 01
Curate
Knee radiographs prepared with the hospital research team.
- 02
Model
StyleGAN3, Stable Diffusion and Pix2Pix GAN trained and adapted.
- 03
Generate
Knee images re-generated by each model.
- 04
Compare
Outputs compared across models.
Stack & techniques
- StyleGAN3
- Stable Diffusion
- Pix2Pix GAN
- Generative AI
- Medical imaging
Delivered by members of the Saerosense founding team. Client names are withheld unless public. References are available on request.
Related service
Start a project
Have a similar problem?
Send a few lines about your data and the outcome you want. We reply within 1 business day with an honest read and a proposed scoping plan.





