There's a quiet assumption underneath most medical imaging AI: that the way to make a model good is to show it more. More films, more labels, more compute, more parameters. The field has largely accepted that the road to a better reader runs through scale — bigger datasets, bigger networks, and the data centre and connection that both require. The models we build are made on a different assumption, and this piece is about what that assumption is and why we think it's the more honest one for the places we work.
The thing everyone already knows
Start with something uncontroversial. Human chest anatomy is remarkably consistent. Two lungs, a mediastinum down the middle, a heart that sits and points a particular way, a diaphragm below, a ribcage around. The proportions vary — some chests are tall, some are broad — but the arrangement does not. Across sex, across ancestry, across the vast majority of adults who will ever have a film taken, healthy structure looks like healthy structure. A radiologist doesn't relearn what a normal lung field is for every patient. They carry a model of health in their head, refined over years, and read each new film against it. What draws their eye is the departure — the place where this film stops looking like health.
This is not a small point. It is most of what reading a chest film is. The knowledge of what healthy tissue and healthy structure look like, region by region, is stable, well-characterised, and — this is the part that matters — already known. It does not need to be rediscovered.
If the anatomy of health is already known, why should a model spend enormous data and compute learning it from scratch?
What scale actually buys, and what it costs
The statistical-learning approach — the deep networks that define most of the field — answers a chest film by having been shown a very large number of them. The model has no prior notion of a lung or a heart. It is handed pixels and labels and left to discover, over millions of examples, whatever statistical regularities separate one label from another. Health is not something it's told; it's something it has to infer, indirectly, from the accumulated shape of the data.
This works. It genuinely works, and the results over the last decade are real. But look at what it costs. To learn health from scratch, the model has to be large enough to hold everything it discovered — millions of parameters, often far more. To train it, you need the massive labelled dataset and the compute to match. And to run it, at the quality the training implied, you increasingly need the hardware and the connection to reach it. The intelligence ends up living somewhere with power and bandwidth — a server, a cloud — and the clinic becomes a terminal that sends an image away and waits for an answer to come back.
For a well-resourced hospital in a well-connected city, that's a manageable arrangement. For the clinics we build for — where the internet is intermittent and the power is limited — it is often the difference between a tool that works and one that doesn't. The cost of learning everything from scratch isn't only computational. It's architectural. It decides where the model can physically live, and scale pushes that answer steadily away from the edge and toward the data centre.
The economy of scale is not free
"Economies of scale" is usually invoked as an unqualified good: more data, better model, the flywheel turns. And in learning what's genuinely unknown, scale earns its keep. But there's a version of this argument that quietly overreaches — the assumption that if some data is good, more is always better, and that the path to a better model is always to feed it more.
The trouble is that a lot of what a from-scratch imaging model spends its scale on is not the hard, unknown part. It's re-deriving the known part. A large fraction of the parameters and the compute go toward reconstructing, statistically and approximately, a model of normal anatomy that was already available to be stated directly. The model arrives, expensively, at a fuzzy internal picture of what a healthy lung field looks like — a picture that anatomy could simply have handed it.
That's the overreach. Scale applied to the unknown is leverage. Scale applied to the already-known is just weight. It produces a clunkier model — larger, heavier, more dependent on the infrastructure that keeps it standing — not because the problem demanded it, but because the design chose to learn from zero what it could have been told.
The other direction: tell it what's known
There is a class of AI that takes the opposite starting point. Rather than learn everything from data, it is given the structure it can be given — the stable, known facts of the domain — and reserves its learning for the part that genuinely varies. In the literature this sits under the broad heading of neuro-symbolic AI: systems that combine explicit, structured knowledge with learned components, instead of forcing everything through statistical learning alone. The engine our models are built on, DRF-MWU, belongs to this family in spirit. A model derived from it reads a film against a known, fixed model of health, region by region, and spends its effort on one question: where does this film depart from health?
We're deliberately not going to open the engine's mechanism here — that's the part we protect. What matters for this argument is the property, not the machinery: health is told, not learned; the model's effort is deferred to pathology — the departure — rather than spent reconstructing the norm.
And the consequence of that choice is not subtle. Take our chest model, Rad — the current version, Rad 1.1, is the reader that flags effusion, consolidation, and an enlarged heart on a chest film. Because it isn't storing a from-scratch statistical model of all of anatomy, it doesn't need the parameters to hold one.
That is not a typo and it is not a stripped-down demo; it is the working reader, and the reason it can run offline, on the ordinary computer a clinic already owns, with no server behind it and no connection required to produce a read. That number is the whole thesis in one figure. A model that is told the anatomy instead of learning it doesn't just save some compute at the margin — it changes class. It becomes small enough to live where the patient is.
One engine, many readers
It's worth being precise about what's small here, because it's easy to blur. Rad is a model. GRX Edge is not. GRX Edge is the app a clinician actually uses — the imaging platform that runs on the device and lets them pick the right reader for what's in front of them. Rad reads chest films today. Mammography is a separate model, still ahead of us, and it will have its own size and its own training once we build it. The way you move between them is the way you'd switch between models in an app you already know: you choose the one that fits the read. If you've used a system like Claude, where the same interface lets you switch between different models for different needs, the shape will be familiar.
The reason more than one reader can share one small footprint is the engine underneath. DRF-MWU is domain-agnostic. It isn't a chest model with the chest parts filed off; it's a general engine for reading structure against a known reference, and a specific model is what you get when you point it at a specific problem with a specific training and viewing objective. Rad is that engine pointed at chest films. Mammo will be the same engine pointed at breast imaging. Each derived model inherits the same starting advantage — told the known, learning only the departure — which is why each can be lean by design rather than lean by compromise. The engine is the constant; the models are what it becomes when it's aimed.
Why "where it lives" is the point
Everything about a deployment follows from where the model physically runs. A from-scratch model heavy enough to need the cloud makes the clinic dependent on a connection it may not have, in exactly the settings where a second read is most scarce. When the link drops, the tool goes dark.
A model small enough to run on the device inverts that. The reading happens locally, in the room, whether or not there's internet that day. Power draw is low enough to matter where power is not guaranteed. The film never has to leave the machine to be read — which is also, incidentally, the cleanest possible answer to where patient images go: nowhere.
This is why we keep saying GRX Edge is built to run where the patients are, and not as a slogan. It's a direct consequence of the architectural choice. Deferring learning to pathology is what makes a model like Rad small; being small is what lets it live on the device; living on the device is what makes it usable in the clinics that the cloud-shaped approach reaches last.
The honest limits
A measured argument has to state its own edges, so here are three.
First, telling a model the known structure only helps to the extent the structure really is known and really is stable. Chest anatomy is a good case — consistent, well-characterised, decades of prior. Not every imaging problem has that. Where the "normal" is genuinely various or poorly understood, the from-scratch approach's willingness to learn everything is a real advantage, and we wouldn't pretend otherwise.
Second, a lean, told-the-anatomy model still has to learn the part that is genuinely unlearnable in advance: what the many forms of pathology look like, and how to tell them apart. That learning is real work, it improves with verified data over time, and being small doesn't exempt it from needing to be right. Leanness by design is a starting advantage, not a finished product.
Third, none of this is a claim that scale is wrong. It's a claim that scale is a tool, not a virtue — and that pointing it at the already-known is where a lot of the field's weight, cost, and cloud-dependence quietly comes from. Spend scale on the unknown. Structure away the known. That's the whole of it.
Where this goes
The next generation of edge imaging AI, we think, will be judged less by how much a model was shown and more by how little it needs to carry to do the reading. Not because bigger models don't work, but because a model that has been told what can be told is free to be small, and small is what it takes to put a second read in a clinic that doesn't have a radiologist and doesn't have a reliable connection. The consistency of human anatomy isn't a limitation to be trained around. It's a gift the architecture can accept — and every parameter a model saves by accepting it is a parameter that clinic no longer has to power, store, or stay connected to use.
That's the direction we're building for. One agnostic engine; lean models derived from it, each pointed at a real read; and an app, GRX Edge, that puts the right one in a clinician's hands wherever they are. Rad is the first. It's pointed at the one question that actually needs the learning: not what is a lung — we know that — but what, in this film, has gone wrong.