BISON

BISON performs tissue segmentation of T1-weighted magnetic resonance images to differentiate gray matter, white matter, and cerebrospinal fluid for quantitative brain imaging analyses across multiscanner and age-diverse datasets.

Key Features:

• Input modality: Operates on T1-weighted magnetic resonance images (T1W MRI) for voxelwise tissue classification.
• Classifier: Uses a random forests classifier for tissue labeling.
• Priors: Incorporates intensity and location priors derived from T1W MRI.
• Training and validation data: Developed and cross-validated using manually labeled multiscanner datasets comprising 72 subjects aged 5–96 years.
• Cross-validation performance: Achieved κ_GM = 0.88, κ_WM = 0.85, κ_CSF = 0.77 versus Atropos (κ_GM = 0.79, κ_WM = 0.84, κ_CSF = 0.64) on 72 volumes.
• Test-retest reliability: On a 20-subject scan/rescan dataset yielded κ_GM = 0.94, κ_WM = 0.92, κ_CSF = 0.77, compared with manual segmentations (κ_GM = 0.92, κ_WM = 0.91, κ_CSF = 0.74) and Atropos (κ_GM = 0.87, κ_WM = 0.92, κ_CSF = 0.79).
• Comparative evaluation: Evaluated against Atropos (Advanced Normalization Tools), FAST (FMRIB Software Library), and SPM12 using cross-validation dice kappa and test-retest reliability.
• Robustness to pathology: Demonstrated superior segmentation performance in the presence of white matter hyperintensities (WMHs) compared to Atropos, FAST, and SPM12.
• Additional test dataset: Assessed on 90 scans from a single individual for repeatability analysis.

Scientific Applications:

• Multicenter tissue classification: Enables standardized tissue classification across multiscanner and multicenter MRI datasets.
• Longitudinal and reliability studies: Supports test-retest and repeated-scan assessments for reliability and longitudinal analyses.
• Pathology-focused studies: Applicable to studies involving white matter hyperintensities (WMHs) due to demonstrated robustness in pathological cases.
• Lifespan imaging: Suitable for pediatric-to-elderly and lifespan studies given validated performance across ages 5–96 years.

Methodology:

BISON applies a random forests classifier combined with intensity and location priors derived from T1W MRI, and was developed and cross-validated on manually labeled multiscanner datasets (72 subjects) with evaluation using cross-validation dice kappa and test-retest reliability against Atropos (ANTs), FAST (FMRIB Software Library), and SPM12 on datasets including 20 scan/rescan subjects and 90 scans from one individual.