Orthodontics used to run on goop, plaster, and educated guesswork. Patients bit into trays of alginate, gagged politely, and waited weeks while a lab turned the mold into a stone model. A doctor then eyeballed that model to plan two years of treatment. But that world is disappearing fast. Walk into a practice built around modern orthodontics today, and the first thing you’ll meet is a wand-shaped intraoral scanner. It feeds thousands of images per second into software that reconstructs your entire dentition in minutes. The plaster is gone. The guesswork is going next.
For anyone watching applied AI, the interesting part is what happens after the scan. Machine learning models now segment teeth from cone-beam CT volumes, predict how roots will move through bone, and flag skeletal growth patterns in children. They also simulate a patient’s finished smile before a single bracket goes on. In fact, orthodontics turns out to be a nearly perfect sandbox for computer vision and predictive modeling: bounded anatomy, huge labeled datasets, measurable outcomes, and treatment timelines long enough to make good forecasting genuinely valuable.
3D scanning killed the impression tray
Intraoral scanners like the iTero Element and 3Shape TRIOS capture up to 6,000 frames per second. They stitch those frames into a digital model accurate to tens of micrometers. Accuracy studies published in orthodontic journals consistently show digital impressions matching or beating conventional ones for full-arch scans. As a result, chair time is cut roughly in half.
The AI layer, however, sits inside the stitching. Real-time algorithms detect soft tissue, saliva, and scan artifacts, then remove them on the fly. So the operator sees a clean model instead of a noisy point cloud. Some systems even flag missed areas mid-scan and guide the assistant back to them. Five years ago, that cleanup meant manual mesh editing by a technician. Now it happens before the patient leaves the chair.
Computer vision reads the X-rays now
Cephalometric analysis, the measurement of angles and distances on skull X-rays, anchors every orthodontic diagnosis. Doing it by hand takes a trained clinician 15 to 20 minutes per film. Worse, it introduces landmark-placement variance between examiners. Deep learning changed that math. Convolutional networks trained on tens of thousands of annotated cephalograms now place anatomical landmarks in seconds. Moreover, a growing body of peer-reviewed comparisons finds their precision falling within the range of disagreement between two human experts.
Segmentation models do similar work on 3D data. Tools cleared for clinical use, like Relu and Diagnocat, separate individual teeth, roots, nerves, and bone from CBCT volumes automatically. That segmentation used to be the bottleneck of digital treatment planning. Automating it, though, collapsed a multi-hour task into a coffee break.
Prediction is the real prize
Scanning and diagnostics save time. Prediction, on the other hand, changes outcomes. The hardest question in orthodontics has always been “what will this mouth look like in 18 months,” and several classes of models now attack it directly:
- Tooth movement simulators forecast how each tooth responds to force
- Growth prediction models estimate jaw development in young patients
- Aligner staging algorithms plan week-by-week movement sequences
- Treatment duration estimators set realistic timelines from intake data
- Relapse risk models flag cases likely to shift after treatment
- Smile visualization tools render the projected result for patients
Align Technology’s ClinCheck software is the most visible example. Its machine learning is trained on data from over 19 million Invisalign cases. Each new patient’s plan, therefore, draws on patterns learned from how millions of previous teeth actually moved versus how software predicted they would. That feedback loop — prediction against ground truth at scale — is exactly the setup ML thrives on. Consequently, it explains why aligner planning has improved so visibly over the past decade.
Where the tech stands today
| Technology | What the AI does | Maturity |
|---|---|---|
| Intraoral 3D scanning | Artifact removal, real-time mesh cleanup, scan guidance | Mainstream, standard of care in digital practices |
| Automated cephalometric analysis | Landmark detection on X-rays in seconds | Clinically deployed, accuracy near expert level |
| CBCT segmentation | Separates teeth, roots, nerves, and bone in 3D | FDA-cleared tools in active use |
| Aligner treatment planning | Predicts tooth movement, stages force application | Mature, trained on tens of millions of cases |
| Remote monitoring | Analyzes patient phone scans, flags issues between visits | Growing fast, adopted widely since 2020 |
| Outcome and relapse prediction | Forecasts final results and post-treatment stability | Emerging, promising research stage |
Remote monitoring, meanwhile, deserves its own mention. Platforms like DentalMonitoring ship patients a cheek retractor. Patients then scan their own teeth weekly with a smartphone. Computer vision checks aligner fit and tracks movement against the plan. It alerts the practice only when something drifts off course. As a result, patients skip routine check-ins that would have confirmed everything is fine, and clinicians spend their chair time on cases that actually need hands.
What the models still get wrong
A tech publication owes its readers the caveats, and orthodontic AI has real ones. Prediction models trained mostly on data from one demographic generalize poorly to others. This is a bias problem familiar from every corner of medical ML. Root movement through bone, similarly, remains harder to forecast than crown movement. That’s because the biology of bone remodeling varies between patients in ways intake data can’t fully capture. And the accuracy studies behind many commercial claims come from the vendors themselves, with independent replication lagging years behind marketing.
Clinicians, still, push back on full automation for good reason. An algorithm can propose a staging sequence. It cannot, however, feel that a patient’s periodontal condition makes aggressive movement risky. Nor can it weigh a family’s preference for fewer office visits against a marginally better outcome. Every serious deployment, therefore, keeps the doctor as the final decision layer, reviewing and overriding machine output. The specialty seems to have internalized a lesson some fields learned painfully: AI works best here as a very fast, very consistent junior colleague.
The next five years
Three developments look close. First, foundation models trained across imaging modalities should merge scan, X-ray, and photo data into single diagnostic pipelines instead of separate tools. Next, generative simulation will let patients toggle between projected outcomes of different treatment options in the consult chair. This turns an abstract sales conversation into a visual comparison. Finally, insurers are already eyeing predictive duration models for preauthorization, which will drag the technology into policy debates about who controls treatment decisions.
The quiet story underneath all of it: a mid-sized medical specialty with a few tens of thousands of practitioners worldwide became one of the densest per-capita adopters of applied computer vision in healthcare. Nobody planned that. The data was simply there, sitting in millions of scans and models, waiting for someone to train on it.
Related: AI Is Quietly Fixing Healthcare’s Biggest Scheduling Problem in 2026
