Keywords:
Artificial Intelligence, MR, Segmentation, Image guided radiotherapy
Authors:
J. Korte, N. Hardcastle, S. P. Ng, B. Clark, T. Kron, P. Jackson; Melbourne, VIC/AU
DOI:
10.26044/ranzcr2021/R-0365
Results
The parotid gland auto-segmentation performance (dice: 0.860±0.067, mean surface distance: 1.33±0.40 mm) on the RT-MAC images with institutional segmentations was higher than previously reported MRI methods [7-9]. An example of auto-segmentations from each of the network architectures is shown in Figure 2. The submandibular gland performance (dice: 0.830±0.032, mean surface distance: 1.16±0.47 mm) is the first reported on MRI images. In the comparison of CNN architectures we demonstrate that by cascading a localiser CNN with a cropped high-resolution CNN it is possible to generate higher resolution auto-segmentations with improved geometric performance for parotid and submandibular glands (see Figure 3.). No performance gain was observed with the prior-knowledge method.
When applying the auto-segmentation methods to our institutional MRI dataset we observed reduced auto-segmentation performance for all organs at risk; only the parotid gland auto-segmentations were considered clinically feasible for manual correction (dice: 0.775±0.105, mean surface distance: 1.20±0.60 mm).