Software is already reading dental X-rays in roughly one in five US practices. A new review asks what it has actually been shown to change, and finds an uncomfortable gap between the two.
If you’ve had a panoramic X-ray taken recently, there’s a fair chance software looked at it before your dentist did. Two German periodontists have just worked through a decade of research on what that software can do for gum disease, and their answer splits cleanly in half. At measuring bone loss on an image, it’s genuinely good. At changing what a dentist decides or how a patient’s gums end up years later, almost nobody has bothered to check.
Key takeaways
- Deep learning models measure periodontal bone loss on dental X-rays about as consistently as trained periodontists, and far faster.
- Almost no published research shows AI changing a dentist’s decisions or improving outcomes patients care about, such as keeping teeth. Accuracy and benefit are separate questions, and only the first has been studied.
- More than two in five US dentists now use AI for at least one task, but fewer than one in twenty use it for treatment recommendations.
- Nothing about your gum care should change on the strength of this. The software reads the picture; your dentist still makes the call, and remains responsible for it.
What Herz and Bartha actually did
Marco M. Herz of Tübingen University and Valentin Bartha of the University of Leipzig and Heidelberg University wrote what’s called a narrative review, and that distinction matters up front. They searched PubMed and followed citations from key papers between January 2015 and April 2026, then picked which studies to discuss on their own judgement of quality and relevance, rather than applying fixed rules and counting everything that qualified. English only. No new patients, no new images, no new numbers.
The interesting move is what they measured the field against: five reporting checklists borrowed from medical AI, with names like TRIPOD+AI and DECIDE-AI. Strip out the acronyms and those ask three blunt questions. Was the model described well enough that someone else could rebuild it? Was it tested on data from somewhere other than where it was built? And did anyone put it in front of real clinicians and watch what happened?
Good at reading the picture, silent on the outcome
On the first question the field does reasonably well. Bone loss detection on panoramic and periapical X-rays is the most developed corner of periodontal AI by a wide margin, and the models work — they flag bone loss, measure how far it extends, and sort teeth into severity bands that line up with clinical staging.
On the third question the field has essentially nothing. Herz and Bartha keep returning to one distinction, and it’s the spine of the paper:
“Technical performance should not be equated with clinical utility.”
Marco M. Herz and Valentin Bartha, writing in Dentistry Journal — https://doi.org/10.3390/dj14080531
A model that agrees with three periodontists on where the bone level sits has shown it can measure something. It has not shown that a dentist using it schedules better recall intervals, catches deterioration sooner, or keeps more teeth in more mouths. Those are the outcomes that would justify the technology, and the review found no trials measuring them.
Forecasting who will get worse, rather than describing damage already done, is where the authors see the real prize and the widest gap. That’s what would let a dentist tighten up maintenance for the patients heading for trouble and leave everyone else alone. The evidence is thin enough that they call the whole domain promising but under-validated.
| Where AI is being used | What it does | How far along |
| Bone loss detection on X-rays | Finds and measures periodontal bone loss on panoramic and periapical images | Works in research settings; rarely tested outside the data it was trained on |
| Automated staging | Sorts teeth into severity categories matching clinical staging | Same. Ground truth varies between studies; shortcut learning is a live risk |
| Predicting progression | Estimates risk of worsening or tooth loss from dental records | Early. Best-known model reached modest accuracy at a single dental school |
| Biomarker classification | Diagnoses from saliva or gum crevicular fluid samples | Exploratory. Small groups, batch effects, little replication |
| Chairside decision support | Suggests recall intervals, treatment escalation, add-on therapy | Concept stage. No prospective evidence of decision impact |
Adapted from Table 1 in Herz & Bartha, Dentistry Journal 2026, published open access under CC BY 4.0.
The field has been saying this for a while
This review isn’t breaking ranks. When Yahia Khubrani and colleagues at Cardiff University scored 30 studies of AI bone loss detection with a formal quality tool, they found pooled performance solid — the models caught most of the bone loss that was there — while research quality clustered stubbornly in the middle. Their verdict on reporting was direct:
“AI studies can lack transparency and reporting standards could be improved.”
Khubrani et al., Cardiff University School of Dentistry, Dentomaxillofacial Radiology 2025 — https://doi.org/10.1093/dmfr/twae070
The prediction side has the same shape. Jay S. Patel and colleagues at Temple University built the best-known periodontal risk model, pulling 74 variables from the records of more than 27,000 patients — not just bone levels and bleeding, but smoking, alcohol, dental anxiety, insurance status, distance from the clinic. It worked, modestly. They called it a pilot and warned the results may not generalise beyond the one dental school that produced the data. Shankargouda Patil and colleagues, reviewing the diagnostic side for JADA in 2023, described the evidence supporting AI in periodontal bone loss assessment as low quality and stopped short of recommending it clinically.
Adoption hasn’t waited. The American Dental Association’s Health Policy Institute reported in July 2026 that more than two in five US dentists use AI for something, and around one in five for imaging and diagnostics. But the same survey found them holding a firm line at the clinical edge: fewer than one in twenty use it for treatment recommendations, and four in five don’t plan to. The ADA described dentists as skeptical about AI’s clinical role — which, read alongside Herz and Bartha, looks less like technophobia than an accurate reading of the evidence.
What this means if you’re in the chair
If you’re a patient
Nothing here should change your gum care, and nothing here is cause for alarm. If your practice uses AI on radiographs it works as a second pair of eyes flagging what it sees; the dentist reviews the image regardless and owns the diagnosis. That matches what the evidence supports — good at measuring, unproven at deciding.
What AI can’t yet do is tell you whether your gums will get worse. If you’ve been told you have periodontitis, the things that move that needle are unglamorous and well established: showing up for maintenance at whatever interval you’ve been given, cleaning between your teeth, and stopping smoking if you smoke. No model beats those. If you’re curious what the software on your X-rays does, that’s worth raising at your next check-up.
If you’re a dentist
Imaging AI is defensible as a consistency tool and hard to defend as more. When a patient asks whether the computer found their gum disease, the honest answer is that it measured the bone level and you interpreted it. Note too that most published models were built on images from one centre and one set of machines, so performance on your equipment and your patients is unknown unless the vendor can show external validation.
The questions nobody has answered
No one has run a prospective trial where periodontal AI is the intervention and tooth retention is the outcome. No one has published a periodontal prediction model that has been externally validated on data from a different institution and shown to stay accurate. Nobody has measured whether AI flags change recall intervals in practice, or whether changed intervals change anything. And there is no published work on the failure mode that worries clinicians most — automation bias, where a confident wrong answer from software talks an experienced person out of a correct judgement.
The bottom line
Reading the field from 2015 to early 2026, two independently funded periodontists conclude that AI can reliably measure periodontal bone loss on dental X-rays in research conditions, and that essentially no evidence exists showing it improves clinical decisions or patient outcomes. That is a review of the literature rather than a new test of it — but the same gap turns up in the systematic reviews that came before it.
Common questions
Can AI detect gum disease from a dental X-ray?
It can detect and measure bone loss, which is one component of a periodontitis diagnosis. Across 30 studies, models caught roughly 87 percent of the bone loss present. But bone loss on an image isn’t the whole diagnosis — that also needs probing depths, bleeding, and attachment loss measured in the mouth.
Is AI better than a dentist at reading gum X-rays?
In head-to-head tests, models generally match periodontists and sometimes outperform general dentists, and they read images far faster. Only about half of published studies made that comparison at all, and no study has shown that the speed or the agreement translates into better results for patients.
Can AI predict whether I’ll lose a tooth to gum disease?
Not reliably. The largest published attempt drew on records from more than 27,000 dental patients and reached only modest accuracy at a single dental school. Its authors described it as a pilot and said the results might not transfer elsewhere. No such model is in routine clinical use.
Should I be worried if my dentist uses AI on my X-rays?
No. In dental imaging, AI works as a second reader flagging findings for the dentist, who stays clinically and legally responsible for the diagnosis. Fewer than one in twenty US dentists use AI for treatment recommendations. If you’re curious, ask what the software does at your next visit.
Sources Used
- Herz MM, Bartha V. Artificial Intelligence in Periodontology: From Automated Diagnosis to Prediction and Clinical Decision Support — A Narrative Review. Dentistry Journal 2026;14(8):531. https://doi.org/10.3390/dj14080531
- Published 20 August 2026. Open access under the Creative Commons Attribution (CC BY) licence, so the full text, figures and tables are free to read and reuse with attribution.
Leave a Reply