ESTRO 2024 - Abstract Book

S2993

Physics - Autosegmentation

ESTRO 2024

In Figure 1, we can see that synthetic images can indeed look similar to real images. However, closer examination demonstrates that low-level structure is different (which is quantitatively measured as non-zero L2 distances between embeddings).

Only two hospitals were emulated in our experiments, extending the results with real across-hospital data and with many hospitals is an intriguing research direction.

Conclusion:

We have demonstrated that learning generative AI models on private data, then pooling these models to subsequently train auto-contouring models, and finally fine-tuning the resulting models on private data improves performance even when standard models with default hyperparameters are used. A memorization experiment further demonstrated that synthetic data is not simply memorized real data, but rather a variation of it.

Keywords: deep learning, synthetic data

References:

[1] Karras, Tero, et al. "Analyzing and Improving the Image Quality of StyleGAN." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.

[2] Zhao, Shengyu, et al. "Differentiable Augmentation for Data-Efficient GAN Training." Advances in Neural Information Processing Systems 33 (2020): 7559-7570.

[3] Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015.