Accurate skull stripping is a critical preprocessing step for reliable brain MRI analysis, yet the performance of classical algorithms remains highly sensitive to parameter selection. This study proposes a statistically reproducible, dataset-level parameter recombination framework for the classical Brain Surface Extractor (BSE), which avoids per-subject tuning by constraining the parameter space using dataset-derived statistical descriptors. The framework is designed to ensure reproducibility across heterogeneous datasets. Two classical skull stripping methods, Brain Extraction Tool (BET) and BSE, were evaluated on three public brain MRI datasets, LPBA40, NFBS, and CC359. BET was optimized by combining the fractional intensity threshold, while a novel statistical tuning strategy was developed for BSE by deriving and recombining its diffusion and edge parameters, which underlie the statistical configuration. A deep learning-based method (HDBET) was used as the primary benchmark, with SynthStrip included as a contemporary reference. Optimized BET achieved strong performance on homogeneous data (Dice Similarity Coefficient, DSC = 0.95), while optimized BSE demonstrated superior adaptability across heterogeneous and multi-scanner datasets (DSC 0.96), approaching deep learning performance within 1-3%. Importantly, the results reveal that high overlap accuracy does not necessarily correspond to accurate boundary delineation, as assessed using 95th percentile Hausdorff distance (HD95). The proposed framework improves robustness and reduces failure rates compared to random parameter sampling, particularly in heterogeneous datasets. The proposed BSE framework improves boundary stability compared to BET in heterogeneous datasets, although deep learning methods still achieve the lowest boundary errors. Classical methods required 0.8-1.0 minutes per subject on CPU, whereas deep learning methods benefit from GPU acceleration but introduce additional hardware dependencies, highlighting a trade-off between accuracy and computational accessibility. Overall, this study demonstrates that principled, dataset-level parameter optimization enables classical skull stripping methods to achieve competitive performance while maintaining reproducibility, transparency, and computational efficiency.
Aryani et al. (Mon,) studied this question.