Offline edge AI · TB triage · adults 15+
ClaraSight reads a chest X-ray on site and returns a TB risk score, a heatmap of what drove it, and a referral recommendation — with the threshold calibrated to the local population. Nothing leaves the device.
01 — The problem
Chest X-ray can find them. The reader cannot be there.
In 2024 an estimated 10.7 million people fell ill with TB and 1.23 million died — the leading cause of death from a single infectious agent.1
8.3 million were newly diagnosed in 2024 — 78% of estimated cases. The remainder is the hole this is aimed at.1
The settings with the gap are the settings without the connection. Sending imaging out of the clinic also moves a data-protection problem.
A CAD score means different things in different populations. Across twelve products evaluated against a South African prevalence survey, specificity at 90% sensitivity ranged from 32.5% to 67.7% depending on the product.3 A version change alone can shift the distribution enough to force a new cut-off.4
A threshold set in the Netherlands is wrong in Namangan.
02 — Signal to referral
A digital X-ray or a photo of film goes in. A score, a heatmap and a recommendation come out. Every positive goes to a confirmatory test.2
A digital X-ray, or a photo of film at the light box, loaded on the attached display.
A quantised DenseNet-121 runs on the accelerator and produces an abnormality score.
A Grad-CAM overlay shows which regions drove the score, so the clinician can check it against their own reading.
The score meets the clinic's own calibrated threshold, and a referral is recommended.
Threshold calibration is established practice, not our invention — WHO/TDR publishes a toolkit for it.6 Our narrower claim: calibration built into an offline device, run by clinic staff against their own outcomes, with no vendor engineer and no connection.
03 — In the clinic
Not a new imaging suite. It plugs into a display the clinic owns and reads the film it already produces.
The clinician, medical officer or nurse at a rural clinic with an X-ray machine and no radiologist. Not a reporting workstation.
04 — The unit
An NPU rather than a GPU: the constraint is performance per watt on an unstable grid. The enclosure is parametric, so a site can reprint the base instead of waiting for one.
The configuration the cost and power story rests on. The Hailo-8L carries the model; the Pi runs the server and the interface.
For larger models and the multi-pathology roadmap. Same enclosure and software image, more headroom — at higher cost and power.
The wipe-down clinical surface and the top of the thermal path. Lifts off without disturbing the standoff stack.
Parametric, authored in build123d, field-reprintable. Carries the I/O cut-outs, the standoffs and the light seam.
A dual M.2 carrier stacks the accelerator and the NVMe under the host board. The RJ45 serves a local PACS — there is no outbound path.
05 — Software
Nothing to install, nothing to log into. The device serves its own interface to the attached display.
FastAPI and Vue, served locally, used through a browser on the attached display.
DenseNet-121 for chest X-ray abnormality scoring, INT8-quantised to run on the accelerator.
Signed, encrypted packages by USB, verified on the device. No network path in or out.
06 — Evidence & limits
Read this first. The target below is a design target we have not met, not a result.
WHO target product profile values for a TB triage test are 90% sensitivity and 70% specificity.3
An internal split is not evidence of transfer, and we would rather say so first.
On low image quality or a score near the threshold, it recommends referral. It fails toward the human, never toward "clear".
Prospective, multi-site, blinded to the AI output, with a WHO-recommended rapid molecular test as the reference standard.
In 2024 only 54% of newly diagnosed people got an initial rapid test.1 Reference-standard availability is a design problem at our sites, not an assumption.
Note. ClaraSight is decision support — not a diagnosis, not a replacement for confirmatory testing, not autonomous. WHO is explicit that anyone who screens positive must be confirmed before TB treatment starts.2
07 — Data & regulatory
Software as a Medical Device. We have not submitted it anywhere, and will not imply otherwise.
Structural, not procedural: there is no outbound path, so there is no transfer to govern.
WHO has recommended CAD for TB screening and triage in people 15 and older since 2021, and approved six products meeting its standards in June 2025.2
08 — Technical
Bench prototype. Measured values are marked; everything else is a target.
09 — About us
A student engineering project aimed at national TB programmes, NGOs and ministry procurement — not at hospitals, and not at replacing the WHO-approved products.
We are not claiming a better model. Six products already passed WHO evaluation on accuracy, and we do not expect to beat them on AUC. We intend to win on four narrower things:
If we cannot beat the incumbents on those four, we have no product — better to learn that at a design review than in a clinic.
Academic supervisor
School of Information Engineering (Big Data College)
Methodological oversight and the route to ethics approval.
Team lead · AI/ML engineer
Hardware specialist
Backend developer
Hardware team
Civil engineer
Mechatronics engineer
10 — Questions we expect
The questions a reviewer should ask, and our answers — including where the answer is a number we still owe you.
Not accuracy. Cost, true offline operation, clinician-run calibration on site, and a nine-language non-radiologist interface. Those four, or nothing.
None yet. Internal validation is under way; no field validation, no regulatory submission. The protocol is prospective, multi-site and blinded, with a WHO-recommended rapid molecular test as reference.
The threshold is referral-biased, so borderline and low-quality images go to a human. Every positive gets a confirmatory test. An immutable audit log allows retrospective review. The clinician holds the decision.
The strongest question against us, and it deserves a price and a power number, not a paragraph. If they come back unfavourable, the honest answer is to buy the approved product.
No. It is a triage instrument: it narrows the field, shows its working, and hands the decision to the clinician. WHO requires confirmatory testing before TB treatment starts.2
11 — Contact us
We want pilot sites, a reference-standard pathway, and reviewers willing to be hard on the evidence above.
Tell us where the unit would be used and which reference standard the site can reach. Procurement and clinical questions are answered in the same thread.
Photographs of the prototype, the enclosure revisions and the bench setup.
No regulatory approval in any jurisdiction, and not available for clinical use. Pilots are research-use-only, under local ethics approval.
Sources