Fig. 5
From: Domain-specific cues improve robustness of deep learning-based segmentation of CT volumes
The influence of data augmentation on segmentation quality. Typical examples of low-quality segmentation results of a 2D U-Net trained with (a–d) MIA and (e–h) CTIA. The arrows in the magnified results in (c) highlight incorrect tissue classifications obtained by training with MIA. The reference segmentation is depicted in (d) for comparison. In contrast, segmentation errors in the pipeline trained with CTIA are typically limited to the size of the segmented region (g) as compared to the reference shown in (h). Both examples demonstrate cases with an accuracy of less than one standard deviations below the mean for the respective pipeline. The examples were selected randomly from a pool of examples of the same quality. The effects shown occur for 2D and 3D segmentation models alike (see Supplementary Fig. S1 online) for further examples.
