Brain-WM: Modeling, Imaging & Cognitive Insights

• Brain-WM is a framework that defines white matter as myelinated axon bundles segmented into distinct anatomical compartments using advanced neuroimaging and tractography.
• Deep learning approaches, such as 3D U-Net variants, enable high-accuracy segmentation and biomarker quantification, achieving Dice coefficients up to 0.94 in WM analysis.
• Integrated computational models link WM microstructure and mechanics to cognitive functions, facilitating precise prediction of age, sex, and disease progression.

Brain-WM encompasses the structural, functional, and quantitative modeling of brain white matter (WM) across domains ranging from neuroimaging and biomarker development to computational disease modeling and cognitive neuroscience. It includes both the neurobiological tissue substrate—comprised primarily of myelinated axons forming the connective architecture of the brain—and the algorithms, frameworks, and data representations used to segment, quantify, simulate, and analyze WM in health and disease.

1. White Matter: Structure, Parcellation, and Imaging

White matter in the CNS is constituted by bundles of myelinated projection, commissural, and association fibers. Spatially, WM is commonly segmented into deep, superficial, and cerebellar compartments, each with distinct anatomical and developmental trajectories. For quantitative study, tractography via diffusion MRI and intensity-based segmentation of T1w/FLAIR MR images are standard.

Comprehensive parcellation of WM enables across-cohort and developmental comparisons. The NABA pipeline generates a cross-population atlas using spectral clustering of 42 million streamlines from neonate and adult tractography, resulting in 800 clusters labeled into 78 major tracts (association, projection, commissural, limbic, cerebellar, superficial). Affine scaling, entropy-based tractographic registration, and robust multiscale alignment support tract matching across brains of widely different size and shape (Zhang et al., 23 Dec 2025). This approach facilitates harmonized, tract-specific developmental analysis and enables studies of sex, preterm birth, or disease-related WM growth curves.

2. Deep Learning-Based Segmentation and Biomarker Quantification

Deep learning models have achieved state-of-the-art accuracy in both global and regional WM segmentation tasks. For tissue-type segmentation, architectures such as NeuroNet (3D ResNet-FCN) and 3D U-Net variants are employed, with pre-training, data augmentation, and intensity harmonization driving Dice coefficients up to 0.94 for WM on standard datasets (Tushar et al., 2019). Domain-agnostic frameworks (e.g., WMH-SynthSeg) utilize synthetic domain randomization to allow robust deployment across clinical and portable MRI of arbitrary resolution and contrast, enabling rapid, vendor- and site-independent quantification of WM atrophy and white matter hyperintensity (WMH) burden (Laso et al., 2023).

Regional WMH mapping is a key innovation for dementia biomarker development. A dual-network 3D U-Net approach produces both WMH lesion probability maps and 34-region anatomical parcellations; regional lesion volumes $V^{(r)}_{\rm WMH}$ normalized by local region volume provide spatially resolved indices of disease (Machnio et al., 18 Nov 2025). Integrated with global and subcortical tissue volumes, these regional lesion features allow multivariate classifiers to achieve high diagnostic performance (e.g., AUC up to 0.97 for AD vs. CN).

Novel biomarker discovery in demographic or clinical risk cohorts is also possible with tract-based statistics of fractional anisotropy and similar measures, as demonstrated in the derivation of obesity-related WM indices using greedy-random variable selection from large multiplex tracts (Suárez-García et al., 2024).

3. Functional, Developmental, and Structural Heterogeneity of WM

Functional variability across WM is not captured solely by functional connectivity. The introduction of fALFF differential identifiability as a robust marker of resting-state BOLD signal variability in WM reveals a spatial gradient with higher functional variability in superficial and commissural regions than deep projection tracts (Chang et al., 8 Oct 2025). Using high-confidence tissue masks and region-wise Laplacian embedding, this approach demonstrates that commissural fibers exhibit maximal functional variability, with a continuous deep-to-superficial ascending gradient paralleling evolutionary and developmental expansion.

Developmentally, joint atlasing and GLM analysis across NABA epochs shows that neonates have rapid FA growth in long-range association tracts (e.g., arcuate fasciculus β ≃ 0.0228 per week), with slower maturation in intra-cerebellar pathways. Sex differences in neonates and compensatory acceleration in motor tracts following preterm birth are tractable with this tract-centric, harmonized framework (Zhang et al., 23 Dec 2025).

4. Brain White Matter in Disease Modeling and Prediction

Synthetic approaches now allow modeling and prediction of disease processes directly in WM. "Brain-WM: Brain Glioblastoma World Model" implements a Y-shaped Mixture-of-Transformers (MoT) dynamics architecture, unifying next-step treatment planning and future MRI prediction, thereby capturing bidirectional tumor–treatment evolution. A shared 3D latent space encodes spatiotemporal context; mask alignment grounds tumor substructure semantics. This framework achieves 91.5% accuracy in planning and SSIM ≃ 0.85 in future MRI across FLAIR, T1CE, T2W, outperforming conventional generative and discriminative segmentation pipelines (Wang et al., 8 Mar 2026).

For individual-specific morphometry, deep generative conditional models (CSegSynth) can synthesize anatomically faithful 3D WM segmentations using only demographic, interview, and neurocognitive feature vectors as conditional priors. This enables volume prediction with mean absolute errors <40 mL and mean Dice ≈ 0.87, surpassing traditional conditional VAE, GAN, or LDM methods (Wang et al., 15 Apr 2025).

5. Quantitative Prediction and Classification Across the Lifespan

Deep neural models leveraging convolutional and transformer backbones, tract-based parcellations, and ensembling offer exceptional performance in age, sex, and pathology prediction from WM features. On the HCP cohort (22–37 y), an ensemble 1D CNN achieves 94.82% accuracy in sex classification and mean absolute error of 2.51 years in age regression, with FA being the strongest sex-predictive and fiber-count the best age-predictive measure (He et al., 2022).

Deep WM clusters are most sensitive for age prediction (MAE ≈ 2.86 yr for thalamo-frontal tract), while superficial WM features have lower predictive value in this age band (Wei et al., 2022). Regionally specific biomarkers, tract smoothing, and explainability methods (e.g., variable importance, ablation) highlight distributed feature contributions and open the door to personalized prediction of neurodevelopmental and neurodegenerative trajectories.

6. Biophysical and Micromechanical Modeling of WM

The quantitative mapping of WM mechanics and microstructural properties is crucial for simulation, injury prediction, and device design. Atomic force microscopy-based measurements across CC, CR, FO reveal pronounced orientation dependence in viscoelastic moduli, with up to 17% greater stiffness for perpendicular than parallel indentation and relaxation timescales spanning ≈0.1–0.7 s (Jamal et al., 2021). Fitted Prony parameters for axons and ECM constituents enable region-specific input for finite element and bottom-up modeling of trauma or neurosurgical procedures.

Magnetic susceptibility mapping of WM is heavily influenced by the microstructural (mesoscopic) arrangement of axons. Modeling axons as multilayer concentric cylinders introduces a spatially varying demagnetization tensor term in the QSM forward model, correcting up to 25% error in white matter χ estimation and resolving confounds between structural and genuine susceptibility anisotropy (Sandgaard et al., 2022).

7. Working Memory, Cognitive Function, and White Matter

The term "Brain-WM" in cognitive neuroscience encompasses both macroscopic network dynamics (as in neural mass models for working memory, with explicit equations for firing rates, membrane potentials, and short-term synaptic plasticity (Taher et al., 2020)) and benchmarking frameworks (e.g., WorM: 1 million-trial, 10-task dataset) for evaluating AI and human working memory performance (Sikarwar et al., 2023). Deep learning architectures trained on WorM replicate primacy/recency, set-size, and integration effects seen in human WM, delineating computational bottlenecks and neuro-inspired model improvements.

Furthermore, electrophysiological signatures—as detected by EEG and advanced functional connectivity/synchronization metrics—are employed to classify cognitive impairment and stratify neurodegenerative conditions via WM-related network properties (Ranjan et al., 15 Oct 2025, Shin et al., 2021, Maimon et al., 2020). Cross-subject domain adaptation models using spatial–spectral–temporal EEG representations improve generalizability of WM load classification and help to identify the cortical correlates of cognitive demand (Chen et al., 2021).

---

Key References:

• Deep Learning-Based Regional White Matter Hyperintensity Mapping as a Robust Biomarker for Alzheimer's Disease (Machnio et al., 18 Nov 2025)
• Gradient of White Matter Functional Variability via fALFF Differential Identifiability (Chang et al., 8 Oct 2025)
• Brain-WM: Brain Glioblastoma World Model (Wang et al., 8 Mar 2026)
• Cross-Population White Matter Atlas Creation for Concurrent Mapping of Brain Connections in Neonates and Adults with Diffusion MRI Tractography (Zhang et al., 23 Dec 2025)
• Age Prediction Performance Varies Across Deep, Superficial, and Cerebellar White Matter Connections (Wei et al., 2022)
• Deep Generative Model-Based Generation of Synthetic Individual-Specific Brain MRI Segmentations (Wang et al., 15 Apr 2025)
• Model and predict age and sex in healthy subjects using brain white matter features: A deep learning approach (He et al., 2022)
• Exact neural mass model for synaptic-based working memory (Taher et al., 2020)
• Decoding the Enigma: Benchmarking Humans and AIs on the Many Facets of Working Memory (Sikarwar et al., 2023)
• Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI (Laso et al., 2023)
• Incorporating the effect of white matter microstructure in the estimation of magnetic susceptibility in ex-vivo mouse brain (Sandgaard et al., 2022)
• Microscale characterisation of the time-dependent mechanical behaviour of brain white matter (Jamal et al., 2021)
• Novel brain biomarkers of obesity based on statistical measurements of white matter tracts (Suárez-García et al., 2024)