ESTRO 2021 Abstract Book

S1375

ESTRO 2021

delineations made by a team of oncologists and a nuclear medicine specialist were used as the ground truth. Three image augmentation schemes were used on the training data to increase image variations: none, affine transformations (random rotation, zooming, translation), both affine transformations and image intensity changes (random noise & blurriness, random brightness & contrast changes). The Dice similarity coefficient (DSC) was used to quantify segmentation performance. The original U-net with 64 filters in the first layer, depth four and no augmentation (italics in Table 1) was used as the reference model. The 2D and 3D models with the best validation performance were evaluated on the test set. Results Increasing the number of filters and model depth in the 2D U-nets had only minor impact on segmentation performance (Figure 1a). In contrast, image augmentation increased 2D model DSC significantly relative to the reference model and models without augmented input (Figure 1a). The 3D U-net performed on par with the best 2D U-net and better than the reference model (Table 1). Augmentation did not improve 3D U-net performance (Table 1). Overall, the 2D U-net was about ten times faster to train than the 3D U-net. Moreover, the 2D U-net required a GPU of 4 GB memory, whereas the 3D U-net needed a GPU of at least 48 GB. Models 2D-F32-D4 and 3D-F32-D4 (Table 1), both with augmented input, were selected for test set evaluation. Test set performance was slightly higher than on the validation set most likely due to the larger number of patients. Example segmentations are shown in Figure 1b.