# Skull stripping tools in pediatric T2-weighted MRI scans: a retrospective evaluation of segmentation performance

## TL;DR

This study evaluates seven skull stripping tools for pediatric T2-weighted MRI scans, finding SynthStrip to be the most accurate and efficient.

## Contribution

The study provides a comprehensive evaluation of skull stripping tools specifically for pediatric T2-weighted MRI scans.

## Key findings

- SynthStrip achieved the highest median Dice score (0.96) among evaluated tools.
- Preprocessing improved performance for BET, HD-BET, and HD-BET-fast but not for other models.
- SynthStrip had the fastest computation time (7 seconds per scan) while maintaining high accuracy.

## Abstract

For brain maturity assessment of infants aged above 6 months, T2-weighted MRI scans are recommended. Prior to automated brain tissue analysis, skull stripping is typically applied. However, most skull stripping tools neither focus on T2-weighted scans nor on pediatric cohorts. Here, we present the evaluation results of seven common skull stripping tools in a comparably large pediatric cohort.

This study is based on 199 T2-weighted scans of children under the age of 5 years retrospectively acquired from the clinical routine at Hannover Medical School. We established a manually labeled ground truth under quality control of a senior neuroradiologist specialized in pediatric neuroradiology and evaluated seven skull stripping tools (BET, ROBEX, HD-BET, HD-BET-fast, SynthStrip, SynthStrip-noCSF and d-SynthStrip). Segmentation performance (Dice score, 95th percentile Hausdorff distance, sensitivity, specificity) and computation time were assessed on non-preprocessed and preprocessed scans (zero padding, contrast enhancement, artifact removal and normalization) as well as in different brain regions. For the best performing model, we manually assessed the top and bottom quartile of segmentations with respect to the integrity of different anatomical brain structures.

Only BET, HD-BET, HD-BET-fast profited from data preprocessing. Considering this, all models had median Dice scores between 0.88 and 0.96, with SynthStrip performing best. All models segmented most accurately in the middle axial slices of the brain. Resampling lowered the performance of all models, except ROBEX. Mean computing times ranged from 2 s (BET) to 132 s (HD-BET) with SynthStrip requiring 7 s. per scan. SynthStrip was prone to not entirely including the Sinus sagittalis superior, the upper Cerebrum, the temporal pole, the Cerebellum and the Chiasma opticum/pituitary gland. In contrast, the petrous bone and the skull in the middle axial slices have often been partly included.

Due to its robustness and quick computation time, we recommend SynthStrip for skull stripping of pediatric T2-weighted MRI scans. We attribute the observed segmentation errors to the partial volume effect, which should be addressed in future research. Limitations of our study include the monocentric setting, the exclusion of pathological cases and the skewed age distribution in our cohort.