AI in the clinic
Why we'll never let the AI sign a note
Human-in-the-loop isn't a compliance checkbox, it's the design decision everything else follows from.
8 min read
The question people ask first about imaging AI is whether it beats a radiologist. We are not going to answer that, and the reason is written out further down: we have no figure of our own to offer, and a figure borrowed from someone else's population is not an answer.
What we can describe precisely is what our own models read, what they return, and where they stop. That is the subject here — the output on the screen, not a claim about the literature.
What the models return
Take the chest film, because it is the clearest case. The model localises and ranks findings — nodule, consolidation, effusion, pneumothorax, cardiomegaly — and each one arrives drawn as a bounding box with a calibrated confidence score beside it. Not a label floating over the study: a region, and a number attached to it. It never writes a diagnosis. It points; the radiologist decides.
The same shape repeats across the imaging and signal models. Each does one clinical job, returns something a clinician can argue with, and flows into the report draft rather than around it.
| Model | What it reads | What it returns |
|---|---|---|
| Chest X-ray detection | Chest film | Bounding boxes and confidence |
| Dermatology lesion triage | Lesion photo, and priors | Ranked differentials and a change map |
| ECG pre-read | 12-lead trace | Ranked differentials |
| Retina / fundus grading | Fundus image | Severity grade |
| Dental caries detection | Bitewing, panoramic | Per-tooth flags |
| Spirometry interpreter | Flow-volume loop | Pattern and severity |
Every row is decision support, to be confirmed clinically. Nothing in that column writes to a record on its own.
Ranked, not decided
A ranked list is a different object from a verdict, and the difference is not rhetorical. A verdict invites you to accept or reject it. A ranked list with a confidence score and a visible region invites you to check it — you can see what the model looked at and how sure it was, and disagree with a specific part rather than with the whole.
That is why the output is a draft like every other AI output on the platform. It reaches the report the doctor reads, edits and signs; until that signature there is no clinical authority in it and no silent write, a rule set out in full in why the AI never signs a note. The rest of the model library, and what the model never sees, sits on the AI page.
Why there is no accuracy figure here
This is the part of the article that would normally carry a percentage, and it does not. Three reasons, and they are ours rather than a disclaimer inherited from somewhere else.
We have not published a trial of these models, so we have no number that is honestly about them. Sensitivity and specificity reported elsewhere are properties of a particular population, a particular scanner fleet and a particular reference standard — rescaling them onto a Tunisian practice produces a figure that looks like evidence and is not. And a claim that imaging AI matches specialists is a claim about a whole field, which no vendor is positioned to make on the field's behalf.
What replaces the number is a test you can run. Bring your own films to a demonstration, including the awkward ones, and watch what comes back. Thirty minutes of your own studies will tell you more about fit than any published percentage.
What the model never sees
An image is not a neutral input. A film can carry a name in its burned-in annotation and a face in its reconstruction, which makes the handling question sharper here than it is for text. Three commitments govern it, and they are three rather than one.
Personal identifiers are stripped before any prompt leaves Doctopass servers for AI processing — the model sees the clinical picture, not the patient's identity. All AI processing runs on a model self-hosted by the processor within the applicable jurisdiction, with no personal data transmitted to any external artificial-intelligence provider. And model improvement uses only irreversibly anonymised data: the published privacy policy states that identifiable personal data is not used, and that patient images containing identifying features are not used, to train models.
Where it meets a specialty
The models are not offered as one general-purpose reader. They are applied through each specialty's module, so the AI a cardiologist works with is not the AI a dermatologist works with — twenty specialty modules, each carrying four AI tools, four note templates and three or more clinical calculators tuned to that practice. The full list is on the specialties page, and how the clinical layer is put together for Tunisian practices is the subject of AI for Tunisian doctors.
The AI layer starts on Pro; Essential carries the record, the calendar, prescriptions and manual invoicing with no AI drafting at all. Specialty templates are an add-on on every plan.
Worth remembering
The imaging models localise, rank and score — a region and a number, never a diagnosis. The output arrives as a draft inside the report you sign. We publish no accuracy figure because we have none of our own, and a borrowed one would be decoration.
Frequently asked
Does the model write anything into the record?
No. It contributes to a draft. The draft is visually distinct until a named physician reviews and signs it, and there is no automatic finalisation and no silent write anywhere on the platform.
Can I see why it flagged something?
That is the design intent. Findings come back localised, with a confidence score, so the reasoning is inspectable rather than asserted. A model that returns only a label gives you nothing to check.
Which plan carries the imaging models?
Pro, at 159 DT, is where the AI layer begins; Essential, at 69 DT, has none of it, and Premium, at 269 DT, raises the scribe allowance and adds CNAM auto-coding, handwriting OCR and audit log search. Per doctor, per month, billed annually — the row-by-row comparison is on the pricing page.
The honest headline is not that the AI reads the scan. It is that the AI points at a region, says how sure it is, and hands the study back to the person who signs. To test that on your own images rather than on ours, request a 30-minute demo and bring the difficult files.