2D Versus 3D Diffusion for In Silico Training of Interventional X-ray AI Models

This study introduces two methods for synthesizing training data for interventional X-ray AI models: a 3D conditional latent diffusion model that generates CT volumes and a 2D diffusion model that produces synthetic X-ray images. Experiments reveal that models trained on synthetic 2D X-rays can achieve performance comparable to those trained on real X-ray data for anatomical landmark detection. This approach potentially alleviates the bottleneck of obtaining annotated high-resolution anatomical models, offering a scalable solution for generating diverse datasets necessary for robust AI model development in medical imaging.