ESTRO 2024 - Abstract Book

S2991

Physics - Autosegmentation

ESTRO 2024

Afterwards, one U-Net is trained on the union of the real train data from both hospitals ( U-Net-real-all ) as an upper bound on performance if real data could be shared, and one U-Net is trained on the union of synthetic data if only synthetic data could be shared ( U-Net-syn-all ). To see if pretraining on synthetic data can be beneficial, for each hospital a U-Net-syn-real is trained on their real private data, but initialized with U-Net-syn-all . All U-Nets are trained for 200 epochs, with one epoch defined as processing 1,000 slices from the dataset. The amount of computational effort when training on real and synthetic data remains constant regardless of the actual dataset size (making comparisons fair). Hyperparameter settings are taken from [4]. StyleGAN2 is trained using default hyperparameters for small datasets [2]. The experiments are repeated 5 times with different random seeds and different hospital and train-validation-test data splits. In each repetition, we calculate the average per-patient Dice Coefficient and the 95th percentile of Hausdorff Distance on the test set. We report mean and standard deviation of the metrics across splits. To study memorization, we create embeddings of real and synthetic data via OpenCLIP-ConvNext-XXLarge [5], a general-purpose image embedding model. Afterwards, for each real image, we measure the L2 distance between its embedding and embeddings of all the synthetic images. We visualize real images whose minimal distance to the synthetic data is the smallest. Next to each real image, 3 closest synthetic examples are shown.

Results: