Confidential · Partnership Overview
A commercially-clean, edge-deployable chest X-ray reader
An independently-validated system that reads chest X-rays with a panel of specialist readers,
runs without a GPU, and was built from the ground up to be transparent and licence-clean — designed for the
settings incumbents reach least.
12
findings, each with named operating points
CPU
first — no GPU required
6,000
unseen multi-site studies stress-tested
0.92
lead-finding AUC on the unseen cohort*
The thesis
Real-world chest X-ray performance is governed by label quality and cross-site generalisation —
not model size. Building on that, we assembled a system whose accuracy holds across hospitals, whose components
are individually auditable, and whose licensing is clean enough to become a product. The hard science is done and
de-risked; what remains is the capital-intensive productisation a partner is meant to fund.
What is proven today
- A twelve-finding reader panel, each returning a named operating point with measured sensitivity and specificity.
- A high-confidence "normal" rule-out advisory — safely deprioritises a large share of normal studies, with a multi-reader safety veto.
- Visual localisation — bounding boxes for the highest-value lesion class, plus a coarse region-of-interest overlay on any flagged finding, so a positive is shown where, not just that it is present.
- Independently validated on a large unseen multi-site cohort, using a rigorous protocol that separates true performance from measurement noise.
Why it is different
- Edge-ready. Runs CPU-first, at a fraction of the deployment cost of GPU/cloud systems — the decisive advantage for high-volume screening and low-resource settings.
- Commercially clean by construction. The core engine is permissively licensed; we deliberately avoided research-only and copyleft components.
- Transparent and safe. A weighted consensus of independent readers with per-finding provenance, and a fail-open rule — a missing reader forces human review, never a silent "normal."
- Unusually self-critical validation. Our method repeatedly caught its own weaknesses and measurement artifacts before they could become false claims — the property that survives clinical scrutiny.
What a partner gets. A de-risked, independently-validated interpretation engine and a rigorous,
reproducible methodology for judging label quality and generalisation — where the expensive dead-ends have already
been walked and documented. The remaining scope is exactly what capital and a clinical partner unlock: regulatory
clearance, prospective validation, clinical-system integration, and a clean-data provenance pass.