GeoRadX reads chest X-rays for signs of disease in seconds, on the hardware your clinic already runs. It calibrates to the population it sits in and keeps adapting as local data accumulates or the device moves.
The app installs freely. Activation requires clinic authentication and approval.
Most imaging AI is shown many films and learns whatever separates them, then returns one label for the whole image. GeoRadX divides the chest cavity into a 14×14 grid, sixteen pixels a side, and measures every region against its own baseline first. What a finding means depends on where it sits.
The crop matters: the frame is the chest cavity, not the film. Neck, shoulders, arms and abdomen never enter the read.
Film: NIH Clinical Center, ChestX-ray14. Select any region of the grid.
Directional grain falls away in the lung field when architecture is lost — 1.41 below baseline. In the mediastinum, where filling occurs around the dense central structures, the same measurement rises — 1.00 above. One rule cannot describe both, so GeoRadX does not use one.
Every region carries its own baseline, measured from healthy tissue. Rows are normalised per patient: row 0.00 is that patient's lung apex and row 1.00 their diaphragm, so the same row means the same anatomy on a tall chest and a short one.
At the chest wall the signature inverts: the texture measurement runs the opposite way, so a rule built in the lung field would give the wrong answer there. In the peripheral lung too few findings were recorded to derive a signature that held up under review. Both stay empty in version 1.
How it works →GeoRadX runs on DRF-MWU, a hybrid transformer that fuses self-attention over grid cells with deterministic field physics and rigid attractor mechanics. Three primitives: attention coupled to physical attractor dynamics, a static reference against an evolving read, and vector representation governed by discrete topological rules.
The deterministic half is why the engine can be inspected rather than probed. It is also why it trains on hundreds to low thousands of films instead of hundreds of thousands, and why it runs on a phone.
Learn more →A chest X-ray arrives as an image. GeoRadX converts it into a spatial problem before it reads anything: the film becomes vectors and tensors positioned in the grid, and the engine moves through that structure rather than across pixels.
What it compares against is the Master Grid, which holds the known variations of health and pathology for every cell position — and which grid it uses depends on where the device sits. Move the device and it detects the move and fetches the grid for its new location.
Building a grid for a new region takes around 350 films. That number is why this is possible at all: an architecture needing hundreds of thousands could only ever ship one global model. A single busy diagnostic centre reaches 350 in weeks.
Because both are topology, the engine compares them cell by cell instead of asking a clinician to hold two images side by side and remember the difference. What changed, where, and by how much.
Two ways in.
Register a government-approved clinic. We verify it exists and is licensed, then grant it.
Its admin authorises your account and sends us that approval before we license your device.
Inside a clinic, the admin decides who has access. We don't.
Your images never do. Films stay on your hardware. What reaches our servers is vectors and tensors under a de-identified hash — the measurements, not the picture. They can't be turned back into an X-ray, and we hold no patient records.
Reads run on the device. No lag waiting on a server, no stable connection required, low enough power draw to run on a phone.
Connection is needed for what sits on top: querying a region and getting it explained, comparing a read against the local or wider population, reading the engine's nuances for research, geolocation and version updates. These run through AIDA in our cloud. The engine underneath keeps reading either way.
No per-person subscriptions. No export fees. Patient records, follow-up tracking, updates and support are included for everyone.
One meter. That's the whole pricing model.
You buy usage credit. Spend it on readings, on AIDA, or on both — it's the same pool. And a paying clinic's readings are what fund free access for the clinics that can't pay.
* High-volume centres can run a Master Grid built for their own hardware and patients. Learn more about custom Master Grids →
A small download that runs on the computer you already own. Available from this site only.
GeoRadX is not yet cleared for clinical use. You can download and register now — accounts open when regulatory approval is granted (expected December 2026).
Half the world cannot readily get a radiologist's read. Why we built GeoRadX the other way around.
Read the missionAre you a licensed clinician? Join the paid reviewer network and help train the model that reads for clinics like yours.
Join the network Separate serviceLongitudinal population data built on AIDA — regional disease trends, cohort comparison and progression signal, for governments, research institutions and pharmaceutical partners.