A multi-view pyramid network for skull stripping on neonatal T1-weighted MRI

Skull stripping or brain extraction on magnetic resonance imaging is a crucial step for structure analyses. In spite of good performances of conventional methods on adult brains, the skull stripping for T1-weighted imaging (T1WI) images on the neonatal brain remains a challenge because of the low image contrast. Therefore, this paper proposes a multi-view pyramid skull stripping network (PSSNet) for neonatal T1WI. To achieve superior skull stripping performance, the conventional pyramid scene parsing network was modified through (1) adding the spatial information of raw feature maps by squeezing the channel information during the feature extraction; (2) increasing the receptive field and adding boundary repair block instead of direct up-sampling; (3) obtaining the final mask through a fusion module on multi-view 2D slices. The 3D skull stripping problem was decomposed into multi-view 2D segmentation tasks to improve the efficiency. We enrolled T1WI images of 70 neonates from the local hospital and 7 infants from the publicly available dataset NeuroBrainS12 (MICCAI 2012). Images of 51 and 26 subjects were used for model training and validation. We compared the proposed method with 7 commonly used methods by using the Dice ratio, sensitivity, specificity, and efficiency. The proposed multi-view PSSNet with the highest Dice ratio (95.44-97.33%) was superior to other methods. Meanwhile, the sensitivity (93.19-97.02%), specificity (97.52-99.68%), and efficiency (8.59-9.30 s per subject) of the proposed method were comparable with the state-of-the-art method. In conclusion, the proposed skull stripping network was robust on neonatal T1WI datasets and feasible in clinical applications.