ESTRO 2023 - Abstract Book

S1363

Digital Posters

ESTRO 2023

DL Pipeline: The method consists of a UNet CNN for predicting dose distributions, followed by post-processing of the UNet output dose based on DVH metrics and dose mimicking optimization to generate deliverable plans. Model development: Using a set of 100 patients (90 for training, 10 for validation) previously treated at RUH Bath, a UNet model was trained to predict dose distributions based on 3D binary representations of the bladder, rectum and target volumes. The resulting CNN was included in the prediction-postprocessing-mimicking pipeline, and settings for postprocessing and mimicking were manually configured using an independent set of 10 patients. The aim of the configuration was to achieve deliverable plans fulfilling the CHHiP dose constraints and having preferable or equal properties compared to the clinically used benchmark plans. The configuration of the settings was iterated two times based on clinical input from RUH Bath, where the DL plan generation was tested on a set of independent patients. Finally, the model was evaluated on a held-out test set of 10 patients. A Wilcoxon signed-rank test was used for comparison with the clinical plans

Results

For the first version of the settings, all dose constraints defined in the CHHiP trial for the bladder, rectum and PTVs were satisfied for all test patients. Upon clinical evaluation performed by RUH Bath, however, two issues were found with the spatial properties of the dose. These two issues were remedied in a second and third iteration of the configuration, see Figure 1. For the third version, statistically significant (p < 0.05) improvements compared to the clinical plan was observed for most of the dose constraints considered, see Table 1.