Figure 2.

We explore transfer learning by data augmentation to address concerns of generalizability in the context of adapting SLANT whole brain segmentation to anatomical variation (e.g., adults versus children) and scanning protocol (e.g., non-contrast research T1w MRI versus contrast enhanced clinical T1w MRI). We begin with a model that has been pretrained on 5,111 subjects with a whole brain segmentation obtained from multi-atlas segmentation. Next, we consider three datasets for transfer learning: First, we identified 30 T1w MRI of children with manually corrected volumetric labels, and accuracy of automated segmentation was defined relative to the manually provided truth. Second, we used the original 45 manually traced atlas from adult brains used in SLANT to refine the initial pretrained segmentation. Third, we retrieved 36 paired datasets of pre- and post- contrast clinically acquired T1w MRI, and accuracy of the post-contrast segmentations were assessed relative to the pre-contrast automated assessment.