Inside the project
Prepared with the project’s development agent from source and retained evidence. This is a project-owner assessment, not an independent customer review. Dated checks were not rerun for this website.
Computational Radiology is the analysis layer under the NeuroAtlas DICOM workstation. It is built to show its working, not to hand down a verdict. Around 19 engine cores (tissue posterior, texture, symmetry, Frangi vesselness, uncertainty budget, counterfactual and self-critique reasoners) feed an Evidence Explorer, where you can drill from a report into the graph, the image overlay and the 3D anchor. In October the vessel workup gained a branch graph. A vessel mask is thinned to a skeleton and split into branches and bifurcations, using verified kernels from the owner's SemanticCompute library. Each branch is measured on its own, so a narrowing on one daughter vessel is not missed by a single end-to-end path. Junctions that topology cannot separate are flagged rather than guessed. The workup now picks a vessel unattended and saves its graph with the session. Where a study includes a vendor's own vessel-analysis centrelines, the app compares its centrelines with them. All output is a candidate research measurement, never a diagnosis.
What stands out
- Branch-by-branch measurement: in a synthetic Y-vessel test, a 50% narrowing on one daughter was found, while the old single-path method followed the other branch and reported 8%
- Honest topology: crossings, touching vessels and trifurcations are flagged 'not separable'
- Typed refusals: sub-series placement says why a half cannot be placed (half a voxel off, rotated, overhanging) rather than rounding it into place
- Independent cross-check: the app's centrelines compared against the scanner vendor's own centrelines
- Parity-verified compute: thinning, the skeleton graph and lattice congruence come from SemanticCompute, which has CPU/GPU parity tests
Current limits
- Not run in the app on a study yet: verification is probe, harness, mutation testing and builds
- Consistency checks only: no clinical reference standard; no stenosis is ever called a NASCET grade on the workstation path
- Measures the segmentation mask, not the anatomy: a leak joins vessels and a gap splits one
- Thinning large crops is CPU work, capped at 48M voxels
- Not released; not in the Integration branch
Features & evidence
Vessel branch graph + per-branch stenosis
component-testedThinning, then a skeleton graph with pruning, then each branch measured with perpendicular sections
Recorded evidence
Harness 32/32 at SemanticCompute 9d27451; mutation checks 11/11; combined app build green 2026-10-02
Unattended vessel choice
component-testedAfter segmentation, picks a named artery by HU and size rules and says which and why
Recorded evidence
Probe VesselSegmentChoice 5/5; wiring mutants 5/5 (2026-10-05)
Saved branch graph
component-testedReused after reload only for the identical mask (SHA-256 bound)
Recorded evidence
Probe VesselGraphStore 2/2 (2026-10-05)
Named-vessel lane per branch
component-testedBranch check per carotid label from saved segmentation models
Recorded evidence
Probe NamedVesselBranches 5/5 (2026-10-03)
Vendor centreline comparison
component-testedThe study's own vessel-analysis paths compared with the lane's carotids
Recorded evidence
Opt-in real-data suite 4/4 with controls (2026-10-03); results private
Sub-series placement (lattice congruence)
component-testedPlaces split series halves on the whole-series grid, or refuses with a typed reason
Recorded evidence
Harness 54/54 at SemanticCompute b698a69; mutants 5/5 (2026-10-03)
Evidence Explorer + 19 engine cores
component-testedDrill-down report ⇄ graph ⇄ MPR overlay ⇄ 3D
Recorded evidence
Per the status doc (2026-07); doc self-rates 88%; runtime profiling still owed
Review updates & corrections
- New: vessel branch graph, bifurcations, per-branch lane, vendor comparison, unattended choice, persistence, assistant tool dicom_vessel_branches (Oct 2–5)
- Correction: an older note said the vendor's centrelines were discarded on export; they exist in the study's vessel-analysis preprocess series
- The bio's 'peak-HU lower bound' and 'Evidence Explorer ~80%' statements are unchanged by this work; the status doc now says 88%
Workflow walkthrough
A guide to the native workflow. These steps describe the app; only the recordings and images embedded on this page are available as website demonstrations.
- Load a contrast CT (for the site, use the harness's synthetic Y-tube with a waist)
- Let segmentation finish: the status line names the vessel chosen and why
- Open the vascular card's 'Branches & bifurcations' list and locate the narrowest point on each branch
- Compare with the single-path figure shown above it
- On a study that includes vendor vessel analysis, run the whole-study preset and read the 'vendor centrelines' check row
Planned captures · not yet available
- Synthetic Y-vessel: per-branch narrowing found where the single path misses it