Anatomical Skull-Stripping Template and Improved Boundary-Oriented Quantitative Segmentation Evaluation Metrics

Skull-removal is a standard preprocessing step for brain magnetic resonance image analysis. The gold standard of skull dissection are the anatomical templates of brain region after the manual removal of skull by experts. Manual segmentation is very time-consuming, but it is the evaluation basis of automatic brain extraction method. The existing skull stripping templates have some confused and unclear partitions around the brain marginal regions (sulci, gyrus) and skull base (brainstem). To this end, the study carried out three tasks: First, we proposed a brain extraction protocol that strictly followed human brain anatomy. Secondly, a more accurate skull stripping template was constructed on two open access datasets. Thirdly, we improved the commonly used segmentation evaluation metrics, such as Dice, Jaccard, sensitivity and precision, and proposed quantitative boundary-oriented evaluation indicators that accurately reflect the vulnerable error segmentation of sulci, gyrus and brainstem. The proposed brain template is widely compared with the existing segmentation template and various automatic brain extraction results.