ESTRO 2024 - Abstract Book

S2992

Physics - Autosegmentation

ESTRO 2024

Models trained on real data ( U-Net-real ) tend to achieve better performance than models trained on synthetic data ( U-Net-syn ), see Table 1. Despite the usage of off-the-shelf data generation models with default hyperparameters, the gap is relatively small. Secondly, U-Net-syn-all (trained on synthetic data from both hospitals) on average outperforms U-Net-syn trained on synthetic data from only one hospital. Thus, pulling synthetic data together improves performance (though not as much as pulling real data for U-Net-real-all ).

Thirdly, U-Net-syn-real outperforms U-Net-real , showing that fine-tuning on real data after pretraining on shared synthetic data is a realistic way of improving performance.