NeuroVFM: Multiscale Neurovascular & Imaging Models

• NeuroVFM is a dual framework that models neuro-glial-vascular interactions and learns volumetric representations from extensive clinical neuroimaging data.
• It employs a multiscale, biophysically informed architecture combining 3D, 1D, and 0D models to simulate coupled neuronal, glial, and vascular dynamics.
• Its foundation imaging model uses a 3D vision transformer trained on millions of scans to achieve high diagnostic accuracy and generate reliable clinical outputs.

NeuroVFM refers to two distinct but conceptually related frameworks for modeling, synthesizing, and interpreting neurovascular and neuroimaging data. The term is primarily used to denote (1) mechanistic, multiscale computational models of the neuro-glial-vascular unit (NGVU), as in the context of synaptic and vascular dynamics in the dorsal vagal complex (DVC) (Hermann et al., 3 Apr 2025), and (2) foundation models for volumetric medical imaging, built using predictive self-supervision on large clinical neuroimaging datasets (Kondepudi et al., 23 Nov 2025). Both utilize multilevel abstraction, system identification, and biophysically informed architecture to integrate neuronal, glial, and vascular dynamics or to learn generalizable visual representations from high-dimensional clinical data.

1. Multiscale Architecture of the Neuro-Glial-Vascular Model

The original NeuroVFM, as formulated for the DVC, incorporates a hierarchical fusion of physical domains and cellular biochemistry. The architecture partitions dynamics into three primary scales:

• 3D capillary microcirculatory network (Ω): Parametrizes pressure and flow-conductivity across a spatial microdomain corresponding to the capillary bed, coupled to a 1D capillary network.
• 1D macrocirculatory network (Λ): Represents large arterial trees that convey pulsatile flows, coupled to the 3D block via terminal branches and to Windkessel elements for boundary resistance.
• 0D quadripartite synaptic units: Encapsulate pointwise dynamics of excitatory/inhibitory synapses, astrocytic processes, smooth muscle cells (SMC), and endothelial cells (EC).

This architecture is designed to capture the feedback loops between neuronal action potentials, neurotransmitter and gliotransmitter release, astrocyte Ca²⁺/IP₃ signaling, and vascular smooth muscle contractility. The synchronization of these biophysical processes underlies neurovascular coupling and functional hyperemia (Hermann et al., 3 Apr 2025).

2. Dynamical Equations and Biophysical Couplings

The governing equations are grounded in conservation laws and detailed kinetic models:

• 1D macrocirculation is governed by mass and momentum conservation,

$$\partial_t A + \partial_z Q = 0, \qquad \partial_t Q + \partial_z\!\bigl(\tfrac{Q^2}{A}\bigr) + \tfrac{A}{\rho}\,\partial_z P = -2(\bar\gamma+2)\,\frac{\mu(2R)}{\rho}\frac{Q}{A},$$

with vessel area-pressure relations enforcing vessel compliance.

• 3D–1D microcirculatory coupling imposes transmural exchange between centerlines and embedding tissue by

$$-\pi R^2\,\partial_s\bigl(K_1\,\partial_s p_1\bigr) + 2\pi R\,L_p\bigl(p_1-\Pi_Ω p_3\bigr)=0$$

for 1D, and a corresponding source in 3D.

• SMC crossbridge dynamics are modeled by a four-state kinetic system (Hai–Murphy), where Ca²⁺ concentration gates contractility, and circumferential vessel wall mechanics are represented via Kelvin–Voigt stress balance.
• Neuronal and glial units use Hodgkin–Huxley ODEs, Tsodyks–Markram short-term plasticity, Li–Rinzel IP₃/Ca²⁺ models for the bouton and astrocyte, and explicit vesicle kinetic schemes.
• Nitric oxide (NO) signaling further modulates SMC relaxation and is produced in both neurons (nNOS) and ECs (eNOS), with compartmental mass balance and diffusion across cellular interfaces.

This results in tightly coupled ODE–PDE systems spanning spatial scales from subcellular compartments to the vascular tree and tissue blocks.

3. Numerical Implementation and Boundary Conditions

To simulate the above system, a hybrid discretization strategy is used:

• 1D vasculature: Finite-volume or finite-difference upwind methods for transport, with explicit–implicit time integration.
• 3D tissue: Mixed-dimensional finite elements to handle spatially distributed pressure and nutrient/oxygen diffusion.
• 0D units: Adaptive Runge–Kutta or implicit backward differentiation formula (BDF) solvers for stiff ODEs representing HH kinetics, Ca²⁺/IP₃ dynamics, crossbridge activation, and NO signaling.
• Coupling: Operator splitting propagates updates across domains: (i) vessel radii, (ii) flows, (iii) solute transport, (iv) 0D biochemical integration.
• Boundary conditions: Inlet flow at the aorta; Windkessel terminal resistance; O₂ inflows at capillary inlets; pressure pinning; and Neumann or Dirichlet conditions as appropriate for tissue and vascular subdomains (Hermann et al., 3 Apr 2025).

4. Physiological Insights and Model Outputs

Simulations using the NeuroVFM framework reproduce empirically observed features of neurovascular dynamics:

• Synaptic activity: Neurotransmitter (glutamate, GABA) pulses following spiking events, triggering astrocytic IP₃/Ca²⁺ transients via metabotropic signaling.
• Astrocytic and vascular response: Astrocyte Ca²⁺ exhibits biphasic responses (fast spikes, plateaus) that gate gliotransmitter release, in turn modulating SMC Ca²⁺, vessel tone, and local diameter via direct and NO-mediated mechanisms.
• Hemodynamic outcomes: Vessel radius changes of ±10–20% over tens of seconds (time-to-peak dilation ~2–5 s, decay ~20 s), and redistribution of microvascular flow and oxygen tension within targeted tissue regions.
• Feedback loops: The system establishes causal chains—synaptic activity drives astrocyte and vascular response, which feeds back onto neural activity (via pressure reflexes and vasculo-neuronal signaling), recapitulating aspects of functional hyperemia and metabolic coupling (Hermann et al., 3 Apr 2025).

The modeling platform exposes points of dysfunction (e.g., disruption of astrocytic or SMC pathways) and permits modular extension to additional cell types (pericytes, microglia) or metabolic processes (glucose, lactate).

5. Parameterization and Extensibility

Parameters are sourced from validated literature—vascular compliance from Alastruey et al. (2007), viscosities from Pries et al. (1996), neurophysiological kinetics from Li–Rinzel, Tsodyks–Markram, and De Pitta et al. (2009), among others.

• Boundary inputs allow model instantiation for specific anatomical regions and subject data.
• Scalability: The mixed-dimensional architecture supports straightforward tiling for larger tissue domains, integration with patient-specific macrovascular trees, or networked astrocyte–synaptic layers.
• Modularity: The design enables incorporation of additional pathways (gap-junction coupling, BBB integrity, inflammation, metabolic byproducts) with minimal modification to the existing solver infrastructure (Hermann et al., 3 Apr 2025).

6. NeuroVFM as a Visual Foundation Model in Clinical Imaging

The NeuroVFM acronym also denotes a scalable, joint-embedding predictive foundation model for medical neuroimaging (Kondepudi et al., 23 Nov 2025). This system is characterized by:

• Volumetric Joint-Embedding Predictive Architecture (Vol-JEPA): A 3D ViT backbone tokenizes DICOM volumes into patches, self-supervised via latent representation prediction rather than voxel-level reconstruction.
• Training: 5.24 million head and brain MR/CT volumes from clinical PACS, with random context/target masking, orientation augmentations, and EMA teacher–student embeddings.
• Emergent properties: Patch embeddings form a latent neuroanatomic atlas, with zero-shot anatomical retrieval and instance attention maps highlighting diagnostically relevant regions (algorithmic pointing accuracy >0.85 on tumor/stroke/MS lesion tasks).
• Clinical downstream tasks: Multi-label diagnosis (82 CT, 74 MRI labels, AUROC ~92.5–92.7%), radiology report generation, and acuity triage (triage accuracy ∼85%, outperforming GPT-5 and Claude 4.5 by ≥10–25 points).
• LLM integration: Visual encoding is coupled to open-source LLMs (Qwen3–14B) via a minimal connector-MLP and Perceiver resampler for radiology report generation with low hallucination and error rates.
• Health system learning paradigm: Trains directly on uncurated, real-world clinical data streams with minimal filtering, enabling extensibility, robustness, and seamless integration into agentic AI pipelines (Kondepudi et al., 23 Nov 2025).

Aspect	DVC NGVU Model (Hermann et al., 3 Apr 2025)	Foundation Model (Kondepudi et al., 23 Nov 2025)
Problem domain	Multiscale neuro-glio-vascular physics	Clinical MRI/CT representation
Modeling principle	Biophysical PDE/ODE coupling	Predictive joint embedding, 3D ViT
Scales	3D micro, 1D macro, 0D cell	3D volume (patch tokens)
Outputs	Pressure, flow, O₂ maps, vessel radius	Diagnosis, report, grounded regions

This table presents the distinctions and commonalities across NeuroVFM usages.

7. Limitations and Future Directions

Assumptions in the NGVU instantiation include stationary 3D–1D flow per timestep, homogeneous tissue conductivity, and well-mixed compartments within 0D units. Physiological subject-specificity can be enhanced by individualized macrovascular modeling or connecting adjacent astrocyte–neuron clusters via spatial gap junction networks. Extensions include explicit pericyte and microglial modules, metabolic fluxes (glucose, lactate, ROS), and synaptic plasticity mechanisms, as well as experimental integration with in vivo imaging and next-generation transcriptomic/epigenomic data.

This suggests that NeuroVFM frameworks, whether focused on mechanistic simulation or large-scale self-supervised learning, are converging toward unified, extensible models capable of supporting both fundamental neurovascular research and clinical decision support pipelines (Hermann et al., 3 Apr 2025, Kondepudi et al., 23 Nov 2025).