ESTRO 2024 - Abstract Book

S3619

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

submandibular glands, total brachial plexuses, and optics (optic nerves, orbits, and optic chiasm). For the loss function, a combined L1 loss and hinge loss was used.

To evaluate the performance of the PatchGAN, we compared the prediction accuracy to the same model without a discriminator. Global performance was evaluated using gamma pass rate (criterion = 3%/2mm), mean absolute error (MAE), mean error (ME), mean absolute percent error (MAPE), and percentage error (PE). PTV coverage was evaluated using D 95 , D 98 , D 99 , D mean , and D max . OAR doses were evaluated using the minimum dose to the hottest 1 cm3 of the volume (D1cc) and mean dose error over the whole volume.

Results:

No statistically significant differences were identified between the two models evaluated. For the PatchGAN and Cascade U-Net models, respectively, the mean absolute error was 4.26±1.28 Gy and 4.25±1.33 Gy, the gamma passing rate was 52.00±8.18% and 51.32±8.74% , and the mean percent error was 21.63±6.12% and 21.28±6.05% (Figure 1). Similarly, the dose to different OARs and target volumes was the same for both models. Dose volume histograms (DVH) for individual PTVs and OARs show congruency between predictions and true doses used for treatment (Figure 2).