Abstract
Brain extraction is essential in neuroimaging studies for patient privacy and optimizing computational analyses. Manual creation of 3D brain masks is labor-intensive, prompting the development of automatic computational methods. Robust quality control (QC) is hence necessary for the effective use of these methods in large-scale studies. However, previous automated QC methods have been limited in flexibility regarding algorithmic architecture and data adaptability. We introduce a novel approach inspired by a statistical outlier detection paradigm to efficiently identify potentially erroneous data. Our QC method is unsupervised, resource-efficient, and requires minimal parameter tuning. We quantitatively evaluated its performance using morphological features of brain masks generated from three automated brain extraction tools across multi-institutional pre- and post-operative brain glioblastoma MRI scans. We achieved an accuracy of 0.9 for pre- and 0.87 for post-operative scans, thus demonstrating the effectiveness of our proposed QC tool for brain extraction. Additionally, the method shows potential for other tasks where a user-defined feature space can be defined. Our novel QC approach offers significant improvements in flexibility and efficiency over previous methods. It is a valuable tool, targeting reassurance of brain masks in neuroimaging and can be adapted for other applications requiring robust QC mechanisms.