Related Experiment Video
Updated: Sep 30, 2025
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
228
High-resolution knee plain radiography image synthesis using style generative adversarial network adaptive
Gun Ahn1,2, Byung Sun Choi2, Sunho Ko3
1Interdisciplinary Program of Bioengineering, Seoul National University, Seoul, Korea.
Summary
Generative adversarial networks (GANs) can create realistic knee X-ray images, aiding in medical data augmentation for arthritis research. These synthetic images are indistinguishable from real ones, addressing data scarcity and class imbalance in datasets.
Area of Science:
- Medical imaging
- Artificial intelligence
- Data augmentation
Background:
- Medical datasets often suffer from scarcity and class imbalance, hindering AI model development.
- Generating realistic medical images is crucial for training robust diagnostic tools.
Purpose of the Study:
- To evaluate the efficacy of generative adversarial networks (GANs) in synthesizing realistic knee X-ray images.
- To assess the potential of GANs for addressing data scarcity and class imbalance in medical imaging datasets.
Main Methods:
- Utilized deep convolutional GAN (DCGAN) and StyleGAN2-ADA for image generation from 10,000 anteroposterior knee radiographs.
- Conducted a Visual Turing test with experts (computer vision, orthopedic surgeons, radiologists) to assess image realism.
- Employed Fréchet inception distance (FID) and principal component analysis (PCA) for quantitative evaluation.
Main Results:
- Generated images accurately reproduced key osteoarthritic features like osteophytes, joint space narrowing, and sclerosis.
- Experts could not reliably distinguish generated images from real ones (classification accuracy ranging from 34% to 57%).
- Achieved a low FID score of 2.96, significantly better than existing medical datasets (BreCaHAD=15.1), with no significant difference in PCA between real and generated images (p>0.05).
Conclusions:
- Generative models like GANs can synthesize high-fidelity knee X-ray images, effectively mimicking arthritis progression.
- This approach offers a viable solution for data augmentation, overcoming limitations of scarcity and class imbalance in medical datasets.
- The generated images are realistic enough to be used in training AI models and for research purposes, indistinguishable from real radiographs by human experts.
