ESTRO 2025 - Abstract Book

S3322

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2025

uncertainty. Per type of classification (i.e., true positive, true negative, false positive, and false negative), the distribution of the PE was analyzed to find a threshold value. Predictions of segmentations with PE values above this threshold value were allocated to the ‘uncertain group’ and those below to the ‘certain group’. Dice scores were computed for both groups as a performance metric, using manual segmentations as ground truth. Results: The segmentation model obtained Dice scores of 85.7% for the CTV, 94.8% for the bladder, and 86.6% for the rectum. The highest PE values were found at the segmentation borders, but high PE values were occasionally found elsewhere. Distinct uncertainty distributions were found for correct and incorrect predictions. For over 99% of all correct predictions, the PE values were near zero, whereas for incorrect predictions, the PE values were more commonly around 0.35. See Figure 1 for the PE distribution of incorrect predictions for the CTV. The Dice scores for the two groups per PE threshold are shown in Table 1. Higher threshold values resulted in better separation between the two groups. With a threshold of 0.36, the Dice scores of the certain group were near the mean Dice scores, while the Dice scores of the uncertain group were near zero. This shows the ability of detecting incorrect predictions with uncertainty estimation.