Dose-prediction generalization: Evaluating CNNs on previously unseen patient anatomies and radiation field geometries

Convolutional neural networks (CNNs) can predict patient-specific radiation dose distributions, but their generalization to unseen beam geometries and anatomical sites remains challenging. Five CNN models (standard U-Net, U-Net++, U-Net3+, Attention U-Net, and Trans U-Net) were trained and evaluated using 750 datasets comprising computed tomography (CT) images, binary radiation-field masks, and corresponding dose distributions. Data were divided into training, validation, and test subsets. Performance was assessed using gamma pass rates (GPRs) and qualitative comparisons of isodose distributions and dose-volume histograms. On the test set, Attention U-Net achieved the highest GPR under the 3%/3 mm criterion (88.0 ± 5.8%). GPR denotes the percentage of evaluated points with a gamma index ≤ 1. Wilcoxon signed-rank tests with Bonferroni adjustment identified significant differences between selected models and standard U-Net, but not across all comparisons. In single-patient evaluations of previously unseen conditions, GPRs were lower than in the representative test case. Standard U-Net achieved the highest GPR for both the new beam-angle case and the case involving previously unseen head anatomy and beam geometry, whereas U-Net++ performed best in the representative case. These results indicate that CNNs may retain some dose-prediction capability under conditions absent from the development dataset. However, each previously unseen condition was evaluated in only one patient; therefore, these findings do not establish generalizability across diverse patients or anatomical sites.