Personalized predictions and non-invasive imaging of human brain temperature

The study addresses the need for accurate, non-invasive assessment of human brain temperature, a critical parameter for brain health, function, and post-injury recovery. Brain temperature is influenced by metabolic heat production and dissipation via cerebral blood flow; disruptions in metabolism or hemodynamics after stroke, traumatic brain injury, or cardiac arrest are associated with worse outcomes, and even small increases (~1 °C) correlate with tissue damage and mortality. Body temperature is an inadequate surrogate because brain–body temperatures can decouple in disease and injury, and intracerebral spatial variations are clinically relevant. The research question is whether a first-principles, mass- and energy-conserving, subject-specific biophysical model informed by individual MRI-derived tissue and vascular structure can accurately predict local brain temperature and reproduce spatial variations observed with MR thermometry.

Multiple MR-based thermometry techniques exploit temperature-sensitive parameters (magnetization transfer, diffusion, proton density, T1/T2, and proton resonance frequency). PRF-based thermometry via phase-difference mapping is effective for relative temperature changes (e.g., thermal ablation monitoring), while chemical shift thermometry with spectroscopy (water–NAA shift) supports absolute temperature estimates; recent EPSI approaches enable whole-brain mapping. However, routine clinical use and direct validation in healthy volunteers remain limited. Biophysical models (e.g., Pennes’ bioheat equation, discrete vasculature models like DiVa, and the VaPor model) have advanced spatially resolved brain temperature predictions, incorporating vasculature and anatomy. Yet many models lack strict adherence to mass and energy conservation and often use simplified or empirical closures, limiting generality and predictive value. Importantly, prior approaches have not accounted for individual variability in metabolism, vessel structure, and flow, which is necessary given observed inter- and intra-regional brain temperature differences.

The authors developed a steady-state, first-principles biophysical model that enforces local mass and energy conservation across three interacting domains: arteries, tissue (tissue plus capillaries, assumed in local thermal equilibrium), and veins. Three heat transfer modes are included: conduction, advection (fluid transport), and convection (fluid-to-wall heat transfer). Governing equations (Eqs. 1–3) are derived via control-volume analysis for each domain, explicitly conserving energy for intra- and inter-domain transport, and correcting prior formulations by: (1) conservative treatment of advection in vessel domains, and (2) mirroring inter-domain advection proportional to origin-domain temperature to ensure energy conservation. Boundary and inlet conditions: scalp Dirichlet temperature of 33.5 °C (except adiabatic base), arterial inlet temperature 37 °C at right/left internal carotids and basilar artery; overall inflow split 40%/40%/20% and two venous outlets 50%/50%; porous tissue flow uses no-penetration and free-slip boundaries. Model inputs are individualized from each subject’s MRI: (i) Tissue probability maps from T1-weighted images (SPM12 segmentation into gray matter, white matter, CSF, soft tissue, skull, background) provide voxel-wise properties via volume-averaging and set metabolic heat generation using gray/white matter probabilities (rates: gray matter 16,700 W m−3; white matter 4,175 W m−3). (ii) Arterial and venous geometries segmented from MRA (3D TOF) and MRV (2D TOF) using neuTube to obtain node locations, segment lengths, and diameters. Because small vessels are under-resolved in non-contrast MRA/MRV, the vascular trees are augmented stochastically with a rapidly exploring random tree (RRT) algorithm guided by an ideal CBF prior: CBF(v) = 80·Prob_GM(v) + 20·Prob_WM(v) mL·100 g−1·min−1, generating terminal segments up to 3 mm length and ≥10 μm diameter. RRT iterations tested: 5,000 to 500,000; metrics computed include terminal segment lengths/diameters and mean arteriole–venule terminal distances (surrogate for capillary transport path). (iii) Inter-domain blood exchange rates from arteries→tissue (M12) and tissue→veins (M23) are proportional to terminal segment lengths; total flow F_tot is constrained by the sum of ideal CBF across voxels. Vessel flow rates follow mass conservation from terminal segment exchange, enabling vessel diameter estimates. Tissue voxel blood velocity field is solved as porous flow with source/sink terms from vessel exchange, enforcing local continuity. Voxel size sensitivity analysis compares baseline 2.6×2.6×2.0 mm³ (matching terminal segment length scale) to refined 1.3×1.3×1.0 mm³ (T1 resolution) and coarsened 5.2×5.2×4.0 mm³. MR thermometry for validation: Whole-brain EPSI (MIDAS processing) computes voxel-wise absolute temperature from the water–NAA chemical shift difference: T (°C) = 313.37 − 102.76·(ω_water − ω_NAA), with susceptibility correction for gray/white matter content, and quality control thresholds (metabolite linewidth <13 Hz, water linewidth <12 Hz, CRLB water <2%, frequency shift <20 Hz). Thermometry maps are registered to T1 space; Gaussian smoothing (σ=1) for visualization. Core body temperature for comparison is estimated as axillary +1.1 °C. Quantitative comparison includes voxel-wise absolute differences and within-threshold maps using ±0.8 °C tolerance (based on prior phantom work). Data acquisition details: 3T Siemens PrismaFIT, 32-channel head coil; MPRAGE T1 (TR/TI/TE 2300/900/3.39 ms, flip 9°, FOV 256×256 mm², 192×192, 160 slices, 1 mm), MRA TOF 3D (TR/TE 22/3.86 ms, flip 15°, FOV 200×200 mm², 256×256, 0.62 mm), MRV TOF 2D (TR/TE 18/3.79 ms, flip 60°, FOV 220×220 mm², 256×256, 3.0 mm), EPSI (TR1/TR2/TE 1551/511/17.6 ms, flip 71°, FOV 280×280 mm², interpolated 64×64×32). Subjects: 3 healthy adults (1 male, 2 females; ages 36, 28, 26).

- RRT augmentation: Increasing RRT iterations (5,000→500,000) reduced terminal vessel diameters and lengths, and yielded more physiologically realistic CBF distributions. Maximum CBF decreased from 3901.3 (5,000 iters) to 270.4 mL·100 g−1·min−1 (500,000 iters); ≤20,000 iters produced unrealistically high gray matter CBF (>140 mL·100 g−1·min−1) and under-supplied white matter. At 500,000 iters, mean CBF matched literature: gray matter 83.5 and white matter 21.3 mL·100 g−1·min−1; whole-brain flow 50–65 mL·100 g−1·min−1. - Terminal segment metrics at 500,000 iters: Terminal arterial diameters median 51.2 μm and 95th percentile 147.8 μm; terminal venous segments were roughly twice arterial sizes (95th percentile <400 μm). Median arteriole–venule terminal distances: gray matter 667.0 μm, white matter 1025.6 μm; whole brain median 683.9 μm, indicating higher terminal density in gray matter. - Voxel size sensitivity: Compared to baseline 2.6×2.6×2.0 mm³, coarsened 5.2×5.2×4.0 mm³ produced RMSE 0.07 °C; refined 1.3×1.3×1.0 mm³ vs baseline RMSE 0.05 °C. Temperature patterns were similar when voxel size ≤ maximum terminal segment length (3 mm); coarsening reduced local thermal variation. - Subject-specific maps: Personalized metabolic heat, CBF, and temperature maps showed distinct spatial patterns across the three subjects, reflecting individual anatomy and vasculature. All maps showed higher temperatures in white matter than gray matter, but with subject-specific spatial variation not captured by a generic atlas-based model. - Validation with MR thermometry: Model-predicted whole-brain temperatures were higher than axillary-based core body estimates and within ≤0.5 °C of core for all subjects; MR-measured brain temperatures were 0.5, 0.1, and 0.1 °C above core for subjects 1–3. Voxel-wise differences between MR and model had 95% CIs of [−0.86, 0.70), [−0.75, 0.71], and [−0.86, 0.58] °C for subjects 1–3. Most voxels agreed within ±0.8 °C; discrepancies were notable in frontal regions where MR coverage was limited. Model-predicted temperature histograms were narrower (no measurement noise) yet fell within MR-measured distributions without significant bias.

The findings demonstrate that a first-principles, mass- and energy-conserving biophysical model can generate accurate, personalized predictions of spatial brain temperature when driven by subject-specific tissue and vascular data. By explicitly conserving inter- and intra-domain energy transport and modeling blood–tissue exchange along terminal vessel lengths, the model captures local temperature variations governed by the interplay of metabolic heat production and perfusion-mediated heat removal. The RRT augmentation strategy was critical to construct physiologically plausible microvascular distributions and CBF, with 500,000 iterations yielding terminal segment scales and CBF values consistent with literature and realistic arteriole–venule transport distances. Voxel size analysis indicates that resolving tissue at or below the terminal vessel length scale preserves spatial temperature heterogeneity. Agreement with whole-brain MR chemical shift thermometry within ±0.8 °C for most voxels supports the model’s capacity to reproduce absolute temperature levels and spatial patterns, despite known MR thermometry limitations. Clinically, individualized thermal maps may improve interpretation of temperature–perfusion coupling in health and disease and support precision temperature management strategies in neurocritical care and thermal therapies.

This work introduces a rigorously conservative, subject-specific biophysical model that predicts 3D brain temperature distributions using each individual’s MRI-derived tissue composition and vasculature. The model reproduces key spatial features and absolute levels observed with whole-brain MR thermometry and outperforms generic atlas-based predictions in capturing inter-subject variability. Contributions include: (1) a corrected energy-conserving formulation for coupled arterial–tissue–venous heat transfer; (2) integration of stochastic vascular augmentation to achieve physiologic CBF and microvascular length scales; and (3) empirical validation against in vivo MR thermometry. Future directions include modeling transient dynamics of perfusion and metabolism (e.g., post-injury changes), refining inter-hemispheric and meningeal heat exchange and capillary-scale representations, quantifying parameter uncertainty impacts, leveraging improved MR thermometry hardware and shimming, and conducting larger studies to generalize and calibrate personalized temperature prediction across populations and pathologies.

- Validation is limited by MR thermometry constraints: B0 inhomogeneity (notably near sinuses), susceptibility effects, spectral resolution and SNR limits, and incomplete brain coverage (e.g., parts of frontal lobe), potentially biasing comparisons. - Steady-state modeling excludes transient hemodynamic and metabolic dynamics relevant to acute injury or therapy. - Capillary network is homogenized within the tissue (porous medium approximation), not explicitly resolved; inter-hemispheric and meningeal boundaries may allow minor heat transfer inconsistent with dura physiology. - Parameter and input uncertainties (e.g., metabolic rates, tissue properties, vessel diameters) are not fully propagated; small sample size (n=3) limits generalizability. - Computational cost increases nonlinearly with RRT iterations and voxel refinement, constraining resolution and scalability. - No gold standard for absolute non-invasive brain temperature; core body temperature was estimated from axillary measures with a correction, introducing potential bias.