Comparative Skull Stripping Techniques on Pediatric Magnetic Resonance Imaging

Skull stripping (SS) is a well-known preprocessing step commonly adopted in Magnetic Resonance (MR) scans to eliminate non-brain tissues. While it is well-established for mid-adults, there is limited experimentation with the pediatric population. The continuous development of the pediatric brain presents challenges for existing SS techniques. This paper aims to compare different SS techniques for a pediatric population quantitatively and qualitatively. Due to the absence of manual brain segmentation (ground truth), the quantitative analysis focuses on pairwise Dice Similarity Coefficient (DSC) comparisons of SS techniques, while qualitative analysis was performed by visual inspection. The work also intends to detect potential outliers by thresholding DSC and visualizing SS results. We performed experiments with five techniques, Freesurfer (FS) versions 6.0.0, 7.1.1, 7.4.1, FSL-BET, and SynthStrip, using 279 T1-weighted MR scans from pediatric subjects. We observed the highest agreement between FSL-BET and SynthStrip, confirmed by the visual inspection. The different FS versions showed inconsistent and poor agreement with other techniques. Furthermore, we noticed complementary information for different SS techniques. For instance, when FSL-BET failed to remove the scalp, Synth-Strip performed well in removing non-brain tissues and viceversa. This suggests the potential benefits of combining these techniques into a consensus approach, such as ensembles in machine learning. In conclusion, the quantitative and qualitative analyses of different SS techniques for the pediatric population highlight the superiority of FSL-BET and SynthStrip and allow the identification of prospective outliers. The future exploration of ensemble techniques and tailored technique adaptations could lead to specialized SS techniques and more accurate brain analyses in pediatric subjects.