S1804
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
fixed training, validation and test sets, and to increase the effective size of the test set. For each N, training was performed for 8 cohorts using different, fold- specific validation and test sets of 100 and 150 patients. This resulted in an effective test set of 8x150 patients, none of them seen in their respective fold training.For each experiment, the learning rate was tuned using a grid-search. Data augmentation with random flipping (x- and/or z-axis) and rotations (0°, 90°, 180° or 270°) was performed during training, both with an independent 50% probability. In each experiment, the network with the lowest validation loss was used for evaluations. Statistical significance of differences in prediction accuracy was assessed using two-sided Wilcoxon signed rank tests (p<0.05). Results: For N ≥ 200, only 1 of the 1200 test patients had a hard constraint violation in predicted dose; for N<200, 5 patients had violations. Median errors in predicted dose metrics with IQR and whiskers decreased with increasing training set size (Figure 1). In line with dose metrics, mean absolute errors (MAE) improved with increasing N (Figure 2). For the vast majority of dose metrics, the decrease in prediction errors when going from N=500 to N=1000 was statistically significant (Figure 1), showing that convergence with training set size was not yet reached.
Conclusion: Hypofractionated breast radiotherapy induces modest early pulmonary function changes without clinical relevance, predominantly affecting FEV1 in relation to low-dose lung exposure, with minimal long-term impact on lung volumes. These findings support the pulmonary safety of modern hypofractionated regimens and highlight the potential relevance of low- dose lung exposure in early functional decline. Keywords: Pulmonary Function, Hypofractionated Breast RT Instantaneous treatment planning for head-and- neck cancer patients – impact of training set size on deep learning dose prediction accuracy Joep van Genderingen, Hazem Nomer, Franziska Knuth, Linda Rossi, Ben Heijmen, Sebastiaan Breedveld Department of Radiotherapy, Erasmus MC Cancer Institute, University Medical Center Rotterdam, Rotterdam, Netherlands Purpose/Objective: Our center is developing instantaneous treatment planning (IP), where a 3D deep learning predicted dose is used by the clinician for plan evaluation and approval, immediately after contouring. The predicted dose is then off-line automatically converted into a deliverable treatment plan. For a practically feasible IP workflow, predicted dose must closely match ground truth (GT) planned dose for the majority of Mini-Oral 519 patients.We have systematically increased training set size for head-and-neck patients up to N=1000 and assessed impact on prediction accuracy. Material/Methods: 1250 head-and-neck patients with two dose levels were included. GT dose distributions were generated with autoplanning.An HD U-Net was trained with N=50/100/200/500/1000 patients using 8-fold nested cross validation. The aim of the nested cross validation was to avoid potential bias in splitting the cohort in
Made with FlippingBook - Share PDF online