It reads the imaging, molecular and surgical inputs a real case already produces, runs them through eleven subsystems on a shared substrate, and returns four artefacts: an integrated diagnosis aligned to WHO 2021, an invasive-margin field with calibrated uncertainty, a draft radiotherapy plan, and a list of eligible trials. Every output is explainable and every clinician override is logged.
A non-functional preview of the AURORA-GLIO surface. Composable panels, logged overrides, explainable end-to-end. Pilot sites tune this to their own workflow.
GLIO-AI Foundation Models92%GLIO-DIAG Integrated Diagnosis88%GLIO-INFIL Infiltration Field79%GLIO-RAD Radiotherapy Plan71%GLIO-CLIN Clinical Trials64%A short, plain account of glioma today — the parts of the pathway that are settled, the parts that are still institution-specific, and the seam AURORA-GLIO is built to fill.
Glioma is the most common malignant primary brain tumour in adults. The pathway has a well-established shape: imaging, biopsy or resection, an integrated histo-molecular diagnosis, then radiotherapy and temozolomide for glioblastoma — the Stupp protocol from 2005. Lower-grade gliomas follow a parallel pathway with longer surveillance. None of this is in dispute, and AURORA-GLIO does not try to rewrite it.
The central problem is heterogeneity — between patients, and inside a single tumor. Single-cell work over the last decade has shown that a glioblastoma is not one tumor cell-type behaving badly; it is a network of cellular states that drift, interconvert and re-form after treatment. The radiomic signature does not capture this. The surgical microscope does not capture this. The trial-arm allocation does not capture this. AURORA-GLIO is an attempt to put the cell-state map, the imaging signature, the surgical reality and the longitudinal outcome into one substrate that any hospital — not just a top-tier academic centre — can run.
Standard-of-care references — Stupp, MGMT, the WHO 2021 framework, the TCGA-era genomic landscape, Sanai on extent-of-resection — are the floor; they are not in dispute. What does not exist is a shared substrate that lets a hospital combine all of that with single-cell state maps, radiomic foundation models, and methylation classifiers without rebuilding the plumbing from scratch.
AURORA-GLIO is our working answer. Eleven subsystems sharing one federated runtime. The molecular layer talks to the imaging layer. The imaging layer talks to the surgical layer. The surgical layer talks to the trial-routing layer. Every output explains itself. Every override is logged. Patient data never leaves the institution. The methods, design choices and evaluation protocols are written up in our own working papers — drafted by the module team and pilot collaborators, listed further down this page.
We are not trying to "solve" glioma. We are trying to make sure that the next ten years of glioma work is additive — across institutions, across modalities, and across the decade.
AURORA-GLIO is the deepest pilot module not because gliomas are the easiest disease to address — they are arguably the hardest — but because the field's infrastructure debt is the most expensive. A federated, open, audit-first substrate has to prove itself on the disease that has historically refused to yield to closed, single-institution AI. If AURORA works for gliomas, the rest of the modules become straightforward port-and-tune work.
AURORA-GLIO models the Sanai-curve relationship between extent-of-resection and outcome, with the margin-field uncertainty propagated through. The current case sits on the curve as a band, never a point.
Every AURORA-GLIO subsystem can be adopted alone or as part of the bundle. The numbering reflects roughly the order they fire on a real case — from systems-atlas priors through to the organoid digital twin.
Multi-omic systems map of glial lineages, niches and signalling.
Built on TCGA-era genomic landscape work and single-cell state maps (Patel 2014, Neftel 2019); the priors most other GLIO subsystems consume.
Connects scanner-scale imaging with cell-scale histology features.
Macro-to-micro bridge: connects MRI radiomic signatures to histopathology features and spatial transcriptomics at the case level.
Immune-stromal-vascular crosstalk simulator with spatial priors.
Immune-stromal-vascular crosstalk simulator with spatial priors from public neuro-oncology atlases.
Pre-trained vision-language-omics transformer for glioma.
Per-disease vision-language-omics foundation model. Distilled to edge sizes so it runs on clinic hardware.
Diffusion-reaction PDE solver for invasive margins beyond contrast.
Reaction-diffusion PDE solver for invasive margins beyond the contrast-enhancing region. Differentiable inverse problem on patient data.
Counterfactual graphs over treatment, resection and progression.
Counterfactual graphs over treatment, resection extent and progression — explicit about which comparisons are confounded.
WHO-aligned molecular + histo + radio diagnostic synthesizer.
WHO-2021-aligned diagnostic synthesiser combining histology, molecular and imaging into one signed report.
Auto-contoured GTV/CTV/PTV with dose-response priors.
Auto-contoured GTV/CTV/PTV with dose-response priors; consumes the infiltration field directly.
Eligibility matcher and adaptive arm allocator for active studies.
Trial eligibility matcher and adaptive-arm allocator for active glioma studies.
Bias, consent and equity audit layer with provenance reports.
Bias, consent and equity audit layer with provenance reports. Runs inside the training loop, not after release.
Patient-derived organoid digital twin synced with bench data.
Optional patient-derived organoid digital twin, synced with bench data for consenting patients.
Inputs the module reads from your existing systems, outputs it returns to them, the protocols it speaks, and the lifecycle of one case as it moves through AURORA-GLIO.
Multi-omic systems atlas. Combines TCGA/CGGA cellular-state maps with single-cell drift trajectories from the Patel-Neftel lineage. Other GLIO subsystems consume it as a prior; it is not user-facing, but it is the reason their outputs are coherent across cases.
Macro–micro bridge. Takes the MRI signal at scanner scale and aligns it against histology features and spatial transcriptomics at the cellular scale. Surfaces a discrepancy when the macroscopic radiomic signal and the underlying micro-anatomy disagree — useful for catching atypical presentations.
Tumour microenvironment simulator. Models immune-stromal-vascular crosstalk under treatment-induced shifts. Outputs a TME state estimate that conditions both GLIO-DIAG (subtype) and GLIO-CLIN (immunotherapy eligibility).
Foundation model. Vision-language-omics transformer distilled to 7–13B parameters with two-pass quantisation; a single RTX-class GPU handles a case under the latency target. Produces a dense case representation that downstream subsystems consume — it does not produce a label.
Reaction-diffusion PDE solver. Differentiable inverse problem on patient anatomy. Outputs per-voxel infiltration probability beyond the contrast-enhancing volume, with calibrated σ that GLIO-RAD's contours respect.
Counterfactual graph engine. Treatment, resection extent, progression and outcome as nodes; the engine is explicit about which comparisons are confounded. Used by tumour boards to interrogate retrospective decisions, not to make new ones.
Integrated-diagnosis synthesiser. Fuses histology, molecular and radio into a single WHO-2021-aligned report with confidence per axis. Discrepancies surface as overrides; the pathologist is the resolver.
Radiotherapy planner. Auto-contoured GTV/CTV/PTV with dose-response priors that consume GLIO-INFIL's margin field directly — contours respect σ rather than treating it as a footnote. Plans round-trip into Eclipse / RayStation / Pinnacle without lossy resampling.
Trial routing. Eligibility matcher and adaptive-arm allocator for active glioma studies. Refreshes nightly; per-arm fit scores explain the rank. Surgeons can pin a trial for re-review at a future tumour board.
Equity-as-loss audit. Stratified performance across ancestry, sex, age band, socioeconomic proxy and access proxy. Runs inside the training loop; release-gates a module if any stratum is below threshold. Threshold revisable only by RFC.
Optional organoid digital twin. For consenting patients with a tissue bank arrangement, synchronises bench-derived organoid responses with the in-silico case. Off by default; opt-in per site, per cohort, per patient.
Drafts by the module team and pilot collaborators. Each is a working paper that documents one slice of the substrate — methods, evaluation protocol, lessons. They are slot-marked here; full PDFs attach as you provide them.
Surgical wins are not the goal — quality-adjusted years are. These are the endpoints AURORA-GLIO is built to measure across pilot deployments. The targets below are pilot goals, not retrospective results.
Target lift on partner-hospital cohorts versus pre-AURORA baseline, measured against the Sanai 2011 threshold.
From sample arrival to WHO-2021-aligned report, tumor-board ready. Baseline is two-to-three weeks in most centres.
Validated against post-resection histology of invasive margins on multi-site held-out cohorts.
GLIO-CLIN matches eligible patients to active trials earlier in their pathway.
Distinguishing true recurrence from treatment-effect on longitudinal scans.
Every clinician override of a model recommendation has a logged reason and explanation chain.
pip install to a hospital deploy in one afternoon.Available to pilot partners today on private registries. At public alpha (Q4 2026), the same images, weights and signatures ship under MIT on public registries. Pick your stack — same code, same model cards, same audit story.
# 1 · install pip install aurora-glio # just AURORA-GLIO # or pip install aurora-neuro[glio] # full stack with GLIO included # pilot sites use a private credential; public alpha (Q4 2026) opens the public registry.
# 2 · verify aurora doctor glio # → AURORA-GLIO ✓ python ≥3.11 ✓ torch ≥2.4 ✓ cuda 12.4 # → models ✓ glio-foundation-2026-Q1 ✓ glio-infil-pde-v3 # → datasets ✓ glio-bench (open splits) # → license MIT @ public alpha # → ready in target <150ms (median, RTX 4090)
# 3 · try it on a public case
from aurora.glio import load_case, diagnose
case = load_case("bench/glio/case-0007")
rep = diagnose(case)
print(rep.subtype) # → "GBM, IDH-wildtype"
print(rep.mgmt) # → "methylated, conf 0.83"
print(rep.margin.sigma_mm) # → 1.4# 4 · attach to your PACS (pilot sites)
from aurora.glio.pacs import attach
attach(
pacs_endpoint="dicom.hospital.local:11112",
aurora_node="aurora.hospital.local",
consent_policy="cohort/neuro-onco-2026",
)
# → bound 412 patients, 9,318 studies
# → federation node: aurora-glio.hospital.local:8443AURORA-GLIO is designed to run on hardware that already exists inside hospital networks. The foundation model is distilled to a 7B–13B parameter band with two-pass quantisation; a single recent RTX-class GPU handles a case in well under the latency target. CPU-only inference is supported for the lighter subsystems (DIAG, CLIN, ETHIC) so a workstation deploy without a GPU is realistic for low-volume sites.
Federated deploys do not require open inbound ports. The runtime opens an outbound mTLS connection to the federation control plane; model updates are signed, audited and pulled. Patient data never traverses the federation. Consent metadata is treated as a first-class object: subsystems with an explicit consent dependency (e.g. GLIO-ORG, GLIO-ETHIC equity stratification) refuse to run on records that lack the appropriate cohort policy.
Most pilot sites land on AURORA-GLIO with an existing imaging-AI vendor in place. AURORA does not replace that vendor in a single step. The recommended path is: (1) install AURORA alongside, (2) compare outputs on a held-out cohort for one quarter, (3) move read-only surfaces (margin field, integrated diagnosis) to the AURORA surface, (4) decide on the rest. Reversibility is a design goal — nothing in AURORA's install creates lock-in.
AURORA is in private pilot today and opens to the world at public alpha in Q4 2026 under MIT. If you carry the weight of these diseases — as a clinician, scientist, builder, patient or advocate — there is a seat at the table.