Fig. 4
From: Automated ventricular segmentation and shunt failure detection using convolutional neural networks
Examples of barriers that need to be overcome for accurate segmentation. Notably, all of these scans are just above the inter-reviewer threshold for inclusion (i.e., Dice score > 0.7 between two human reviewers) and are included in our final training and test sets. Excluded scans have even lower inter-reviewer Dice scores. (a) An intraparenchymal hemorrhage obscuring and distorting portions of both lateral ventricles. (b) Dysmorphic and slit ventricles exacerbate any small differences in accuracy between labeling systems. (c) Cystic changes, dysmorphic ventricles, and infarcted tissue all can blur the borders between parenchyma and ventricle. (d) Streak artifact from an aneurysm clip can make certain axial slices of a scan unusable.
