S2280
Physics - Machine learning and AI algorithms
ESTRO 2026
with only 150 R-patients available was highly similar to the prediction accuracy with 750 R-patients (Fig. 1, orange vs green). Figure 2 displays mean absolute errors (MAE) for dose distributions, confirming the large impact of anatomical augmentation.
Digital Poster 517
Anatomical augmentation and automated planning for high accuracy deep learning dose prediction with few training patients. Joep van Genderingen, Hazem Nomer, Franziska Knuth, Linda Rossi, Sebastiaan Breedveld, Ben Heijmen Department of Radiotherapy, Erasmus MC Cancer Institute, University Medical Center Rotterdam, Rotterdam, Netherlands Purpose/Objective: Dose prediction accuracy generally increases with the number of training patients, but the actual number of available patients is often limited. In this study, we introduce ‘anatomical augmentations’ to increase the number of training patients. Augmented patients (A- patients) were generated by expanding or shrinking target volumes of real patients (R-patients), followed by automatic generation of dose distributions. Material/Methods: The study cohort of R-patients consisted of 1250 H&N cancer patients with various tumor locations. The cohort was split into 750 training/validation patients and 500 test patients, using stratified sampling. Three different sets were used to train an HD U-Net for dose prediction: (1) 150 R-patients, (2) 3691 patients, with 150 R-patients in (1) + 3541 A-patients derived from them, and (3) all 750 R-patients. For generation of the A-patients in (2), for each R-patient, the two targets were isotropically expanded or shrunk by combinations of margins of 0, ±5 or ±10 mm, resulting in a maximum of 24 (25-1) feasible A-patients per R- patient. For all 1250 R-patients and 3541 A-patients, ground truth (GT) dose distributions were generated with automated planning. For sets (1) and (3), the models were trained on 125 and 625 R-patients, respectively, with validation sets of 25 and 125 R- patients. For training set (2), the 3541 A-patients were used for training and the 150 R-patients for validation. All training strategies applied the same data augmentation during training with random flipping (x- and/or z-axis) and rotations (0°, 90°, 180° or 270°), both with an independent 50% probability. All training strategies were evaluated on the same test set of 500 R-patients, which were not seen during training. Two- sided Wilcoxon signed rank tests (p < 0.05) were used to assess statistical significance of differences between training strategies. Results: Training with 150 R-patients + 3541 A-patients significantly improved accuracy of predicted dose metrics compared to using only the 150 R-patients for training (Fig. 1, blue vs orange): medians, IQRs and whiskers were on average reduced by 34%, 31% and 30%. By adding A-patients, the prediction accuracy
Made with FlippingBook - Share PDF online