BISON: Brain tISue segmentatiON pipeline using T1-weighted magnetic resonance images and a random forests classifier

Introduction Accurate differentiation of brain tissue types from T1-weighted magnetic resonance images (MRIs) is a critical requirement in many neuroscience and clinical applications. Accurate automated tissue segmentation is challenging due to the variabilities in the tissue intensity profiles caused by differences in scanner models and acquisition protocols, in addition to the varying age of the subjects and potential presence of pathology. In this paper, we present BISON (Brain tISue segmentatiON), a new pipeline for tissue segmentation. Methods BISON performs tissue segmentation using a random forests classifier and a set of intensity and location priors obtained based on T1-weighted images. The proposed method has been developed and cross-validated based on multi-center and multi-scanner manual labels of 72 subjects aging from 5-96 years old, ensuring the generalizability of the results to new data from various age ranges. In addition, we assessed the test-retest reliability of BISON on 2 datasets; a. using 20 subjects that had scan/re-scan MRIs and manual segmentations available, and b. using a human phantom dataset including 90 scans from a single individual acquired across 10 years. Results The results of the proposed method were compared against Atropos, a commonly used tissue classification method from ANTs. The proposed method yielded cross-validation Dice Kappa values of κ GM = 0.88 ± 0.03, κ WM = 0.85 ± 0.03, κ CSF = 0.77 ± 0.11, outperforming ANTs Atropos (κ GM = 0.79 ± 0.05, κ WM = 0.84 ± 0.05, κ CSF = 0.64 ± 0.22) as well as test-retest Dice Kappa values of κ GM = 0.94 ± 0.006, κ WM = 0.92 ± 0.006, κ CSF = 0.77 ± 0.11 outperforming both manual (κ GM = 0.92 ± 0.01, κ WM = 0.91 ± 0.01, κ CSF = 0.74 ± 0.03) and ANTs Atropos (κ GM = 0.87 ± 0.001, κ WM = 0.92 ± 0.001, κ CSF = 0.79 ± 0.05). The human phantom dataset validations showed high generalizability for both Atropos (κ GM = 0.97 ± 0.01, κ WM = 0.96 ± 0.01, κ CSF = 0.93 ± 0.02) and BISON (κ GM = 0.95 ± 0.01, κ WM = 0.94 ± 0.01, κ CSF = 0.85 ± 0.03), while Atropos tended to consistently under-segment the cortical CSF. Finally, our assessment of BISON, Atropos, FAST from FSL, and SPM12 segmentations in presence of white matter hyperintensities (WMHs) showed that BISON outperforms the other three methods, correctly detecting WMHs as WM. Conclusion Our results show that BISON can provide accurate and robust segmentations in data from different age ranges and various scanner models, making it ideal for performing tissue classification in large multi-center and multi-scanner databases.