AI in Dentistry: What X-Ray Analysis Can and Cannot Do

In short: AI radiograph tools are most useful as a second read that catches early interproximal caries a tired eye misses. They over-flag, cannot judge film quality, and do not transfer liability. Read the film first, treat output as a prompt to verify clinically, and never let an overlay alone justify treatment.

AI radiograph analysis has moved from conference demo to something you can actually run on an IOPA in a general practice. It's genuinely useful. It's also being marketed with claims that don't survive contact with a busy clinic. This is the honest split.

What it does well

Second-look detection

The strongest use case. You read the radiograph, form your opinion, then let the model flag what it sees. On interproximal caries in particular, a second pass catches early lesions that a human eye scanning quickly at the end of a long day will miss. Not because the model is smarter — because it doesn't get tired and doesn't anchor on the tooth the patient complained about.

Consistency

Your reading of a borderline lesion at 10am and at 7pm is not the same reading. The model's is. For longitudinal monitoring — is this lesion progressing? — consistency is worth more than peak accuracy.

Patient communication

Underrated and immediately valuable. A radiograph with clear visual annotation is dramatically more persuasive to a patient than a dentist pointing at a grey area on a screen. Case acceptance improves because the patient can actually see what you're describing.

Teaching

For students and new graduates, comparing your own read against a model's output is a fast feedback loop that used to require a senior standing next to you.

What it does badly

False positives, and their cost

Models flag things. Some of them are normal anatomy, artefacts, or lesions too early to warrant intervention. In the hands of an experienced clinician this is a minor annoyance. In the hands of an inexperienced one, or a commercially motivated one, it becomes justification for treating teeth that didn't need treating. This is the single biggest risk in the category.

Context

The model sees pixels. It does not know the patient is 78 with a lesion that has been static for six years, or that this tooth is a bridge abutment, or that the patient's caries risk is low and the sensible plan is monitoring. Treatment decisions require the patient, not the image.

Image quality dependence

A poorly angulated, overlapped or under-exposed radiograph produces unreliable output — often confidently. Garbage in, confident garbage out. The model rarely tells you the film wasn't diagnostic. See the systematic approach to reading radiographs, where film quality assessment comes first for exactly this reason.

Anything beyond its training

Most tools are trained on caries, bone levels and periapical radiolucencies. Uncommon pathology, developmental anomalies and unusual presentations are where models are least reliable and where a clinician is most needed.

How to use it without losing your judgement

1. Read first, then reveal. Form your own opinion before you see the overlay. If you look at the AI output first, you'll anchor on it — and you'll stop developing your own reading skill.
2. Treat output as a prompt, not a diagnosis. Every flag gets clinically verified.
3. Never let it justify treatment on its own. If you wouldn't have treated it before the overlay appeared, ask why the overlay changed your mind.
4. Check the film quality yourself. The model won't.
5. Be careful with the patient framing. "The computer found six cavities" is both misleading and, if it drives unnecessary treatment, an ethical problem.

How these models actually work

Almost every dental radiograph tool on the market is a convolutional neural network trained on annotated films — typically tens to hundreds of thousands of bitewings and periapicals, each marked up by dentists. The model learns the visual signature of a lesion from those annotations. It is pattern recognition against a labelled corpus, not reasoning about the tooth.

Two consequences follow from that, and they explain most of what the tool gets right and wrong.

It is only as good as the corpus. If the training set was annotated mostly on well-angulated bitewings from one sensor type, output degrades on a film that doesn't resemble them. Indian general practice sees a wider spread of equipment than most training sets represent — older sensors, phosphor plates, occasionally scanned film. Performance on your radiographs is not the performance in the vendor's paper.

Annotator disagreement sets the ceiling. Two dentists shown the same borderline interproximal lesion agree less often than you would hope. When the labels themselves carry that disagreement, no amount of training removes it — the model inherits the ambiguity. This is why "accuracy" figures above the human inter-rater rate should be read sceptically rather than as progress.

Why it over-flags, in numbers

Every detection model sits on a trade-off between sensitivity — how much real disease it catches — and specificity, how much healthy tissue it leaves alone. You cannot maximise both. Vendors almost always tune for sensitivity, because a missed lesion is a visible failure and a false alarm looks like diligence.

That choice has a consequence people underestimate, and it is arithmetic rather than opinion. Suppose a model runs at 90% sensitivity and 90% specificity, which would be a strong result. Run it across a hundred surfaces where ten genuinely have early caries:

• It catches 9 of the 10 real lesions.
• It also flags 9 of the 90 healthy surfaces.

So half of everything it flags is wrong — not because the model is bad, but because the condition is uncommon in the population being screened. That is the positive predictive value problem, and it gets worse the healthier your patient base is. A low-caries-risk adult recall list is exactly where the flags are least trustworthy.

This is the technical reason behind the clinical rule above: the overlay is a prompt to look, never a reason to drill.

What the image pipeline needs

Output quality is bounded by what reaches the model, and this is where clinics lose accuracy without realising it.

Resolution and bit depth

Sensors capture at high dynamic range — commonly 12 to 16 bits per pixel, meaning thousands of grey levels. A screenshot, a photo of the monitor, or an 8-bit export collapses that to 256 levels. Early demineralisation is a subtle density change; flatten the greyscale and the signal it depends on is gone before the model sees the file.

Compression

JPEG is lossy, and its artefacts cluster at exactly the high-contrast edges a caries detector is reading. Re-saving a radiograph through two or three systems compounds this. Send the original export from the sensor software wherever possible, not a file that has been through a chat app.

Angulation

Overlapping contacts are the single most common reason an interproximal read is unreliable. The model will still return output on an overlapped film, usually without warning you — which is why film-quality assessment stays a human step, first, every time.

Where the radiograph goes

If analysis happens off-device, a patient's radiograph leaves your clinic. That is a processing decision you are accountable for, so it is worth being able to answer three questions about whichever tool you use:

1. Where is it processed, and is the image retained after? Inference does not require storage. Ask whether retention is for model training, and whether you can decline.
2. What identifiers travel with it? A radiograph tied to a name and date of birth is health data about an identifiable person. The same image with the identifiers stripped is a materially different disclosure.
3. What did the patient consent to? Consent to a radiograph is not automatically consent to send it to a third party for processing.

India's Digital Personal Data Protection Act treats health data as sensitive, and the obligations sit with whoever determines how it is used — the clinic, not the vendor. Get the answers in writing before the first film is uploaded, and revisit them when a vendor changes its terms.

Fitting it into a real appointment

The workflow question that decides whether any of this survives contact with a busy clinic is simple: does the read finish before you need it?

An analysis that returns in a couple of seconds fits inside the moment you are already spending on the film — you take the radiograph, you read it, the overlay is there when you look up. One that takes thirty seconds does not; the appointment has moved on, and the tool becomes something you check afterwards, if at all. EnamDoc's own figure is roughly 2.4 seconds per scan, which is the range where the result arrives inside the clinical moment rather than after it.

The same logic governs charting. A tool that writes its findings into the patient record saves genuine minutes per patient. One that shows an overlay you then retype into your notes has moved the work rather than removed it — and in a clinic seeing thirty patients a day, that difference decides whether the tool is used in month three.

Regulatory and record-keeping notes

Clinical decisions remain the dentist's responsibility — AI output does not transfer liability. Note in your records what you diagnosed and why, not merely that a tool flagged something. If radiographs leave your clinic for processing, understand where patient data goes and what consent you need. Treat this as a live area and check current requirements rather than assuming.

Where this is heading

The realistic near-term direction is narrow and useful: better caries and bone-level detection, automated charting that saves genuine minutes per patient, and treatment-plan drafting a clinician edits. Autonomous diagnosis is not on the table, and the vendors claiming otherwise are selling ahead of the evidence.

EnamDoc includes AI X-ray analysis for dentists alongside appointments, digital records and payments — see what's included.