🧠 NeuroSentinel — from Tale Research

The brain, measured.

NeuroSentinel reads a single 3D T1 MRI and returns a structured volumetric read: 104 brain regions, age- and sex-matched percentiles, asymmetry indices, and a structural signature — in the time it takes to get coffee. Built for neurosurgical triage, neurological follow-up, and the research questions that sit between them.

✉️ Request a demo 🔬 How it works

A neurosurgeon has minutes to read a scan. We built NeuroSentinel so they don't have to measure 104 structures by hand.

— Tale Research

⚙️ How it works

One scan in. A quantitative read out.

A single 3D T1-weighted MRI enters the pipeline and exits as a structured report a clinician can act on. Every step is deterministic, every number is traceable to the image it came from, and nothing leaves the local network.

NeuroSentinel MRI analysis pipeline: Upload MRI, Parcellation, Normalize Volumes, Signature Analysis, Visualize Results, Narrate Findings
The six-step MRI analysis pipeline · detailed below

Upload

A standard 3D T1-weighted MRI in NIfTI format (.nii.gz). Vendor-agnostic — 1.5T or 3T scanners, Siemens, GE, or Philips. No special sequences, no proprietary file formats.

Parcellation

FreeSurfer's SynthSeg --parc produces a 104-structure Desikan-Killiany parcellation — contrast-agnostic, robust to clinical-quality scans, and the same engine used in FreeSurfer research workflows.

Normalize

Per-structure volumes are corrected for measured total intracranial volume, then converted to age- and sex-matched Z-scores and percentile ranks against a 205-subject open normative cohort. Structures below the 25th percentile are flagged as atrophic.

Signature

A Raji-style structural discriminant runs over the Z-scores and left-right asymmetry indices, surfacing patterns consistent with post-traumatic structural change — ventral diencephalon and pallidum loss with relative hippocampal sparing — distinct from neurodegenerative signatures.

Visualize

An interactive WebGL viewer renders the original scan with the segmentation overlay. Click any flagged structure to jump the crosshair; toggle the percentile heatmap to see the whole brain colored by deviation from normal.

Narrate

A local vision-language model reads the heatmap image together with the structured findings and returns a short clinician-facing paragraph. Entirely on-premises — no PHI ever leaves the network, no cloud LLM in the loop.

End-to-end runtime per scan: a few minutes on commodity CPU. Nothing in the path touches an outside service.

⚡ The instrument

Built for the reading room, not the data center.

NeuroSentinel is a clinical-grade measurement tool for neuroradiologists, neurosurgeons, and neurology researchers. It sits next to the reading workflow, makes the quantitative pattern visible, and leaves every clinical decision with the physician.

📐

104-structure parcellation

Every scan is segmented into the full Desikan-Killiany atlas — cortical lobes, subcortical nuclei, ventricles, and cerebellum — with volumes reported in millilitres. The read that would take an afternoon by hand, computed in minutes.

SynthSeg Desikan-Killiany Contrast-agnostic
📊

Age- and sex-matched percentiles

Structures are compared against a 205-subject open normative cohort with per-structure regression against total intracranial volume. Each region comes back with a Z-score, a percentile, and a clear atrophy flag when it falls below the 25th percentile.

205 subjects mTIV-normalized IXI reference
🧭

TBI structural signature

Based on the Raji et al. volumetric findings in the Journal of Alzheimer's Disease, NeuroSentinel evaluates each case against a structural fingerprint of traumatic injury — asymmetric pallidum, ventral diencephalon loss, hippocampal sparing — distinct from neurodegenerative patterns.

Raji discriminant Asymmetry-aware
🖥️

Interactive WebGL viewer

A browser-based viewer renders the original MRI with the segmentation overlay and a percentile heatmap. Click a flagged structure in the findings panel and the viewer jumps the crosshair to that region. No desktop software, no licensing per seat.

Niivue Click-to-jump Browser-native
✨

Local-first AI narrative

A vision-language model hosted on-premises reads the heatmap together with the structured findings and returns a short paragraph describing the dominant pattern. Every inference happens locally — no cloud LLM, no PHI over the wire.

On-prem VLM No PHI transmitted
🔒

On-premises by design

NeuroSentinel is built to run inside the hospital firewall. No cloud dependency, no third-party telemetry, no vendor-held imaging. Every scan, every prediction, and every audit trail stays where it was acquired.

Dockerized HIPAA-aware audit Role-based access
🔬 The science

Open tools. Cited methods. No black boxes.

NeuroSentinel is assembled from peer-reviewed open-source components. Every number the platform produces can be traced back to the image and the reference cohort it was derived from. No proprietary black-box classifier, no undisclosed training data.

Parcellation engine SynthSeg · FreeSurfer

FreeSurfer's SynthSeg runs in contrast-agnostic mode with the --parc flag, producing a full Desikan-Killiany cortical and subcortical parcellation (~104 labeled structures) that is robust to clinical-quality MRI acquired across vendors and field strengths.

Normative reference 205 subjects · IXI cohort

Per-structure Z-scores and percentiles are computed against a 205-subject normative cohort drawn from the open IXI dataset. Volumes are corrected for measured total intracranial volume before comparison, and age-matched within a sliding window for stratified analysis.

TBI structural signature Raji et al. · J. Alzheimer's Dis.

The TBI discriminant is grounded in Raji et al.'s published volumetric findings separating traumatic structural change from neurodegenerative atrophy — ventral diencephalon and pallidum loss with relative hippocampal sparing, distinct from the temporal-lobe-dominant Alzheimer pattern.

Local-first AI On-prem VLM · no PHI transmitted

All generative inference — narrative summaries of the volumetric read — runs through a locally-hosted vision-language model. Nothing in the imaging or text path touches an external service. Every hospital deployment is self-contained behind the firewall.

Structures measured
104
per T1 scan
Desikan-Killiany atlas

⚠️ What NeuroSentinel is not

NeuroSentinel is a research platform and a measurement instrument. It is not a diagnostic device. It does not read, triage, or prescribe — it makes quantitative structural information available to the clinician who does. Every institution deploying NeuroSentinel remains fully responsible for clinical decisions, regulatory compliance, informed consent, and the interpretation of every result. Tale Research builds the instrument; the practitioner brings the judgment.

Evaluating NeuroSentinel for your institution?

NeuroSentinel is designed for neurosurgical groups, academic imaging centers, and neurology research labs that need an on-premises volumetric read. If what you've seen on this page is what your reading room or your research question needs, we'd like to hear from you.

✉️ Request a demo 🔬 Revisit the pipeline

neurosentinel@taleresearch.com