20% off on annual plans this month
code SMILE2020% off on annual plans this month
code SMILE20Spotting the radiopaque outline of a fixture takes an AI model a couple of seconds. What actually changes a patient's outcome is what happens after — tracking bone-to-implant contact and crestal bone levels across every recall film that patient ever takes.
What this article covers
Where implant detection accuracy on OPG actually stands today, what “beyond detection” means in practice — bone-to-implant contact, peri-implant crestal bone tracking across serial films — and the specific places a flat 2D panoramic film still can't do a probe's job.
A dentist in Kanpur pulled up a five-year-old OPG last month and couldn't tell, from the film alone, whether the bone around a lower-left implant had actually dropped or the exposure angle had just shifted between visits. That's the real problem AI is being asked to solve now — not whether it can spot a screw-shaped radiopacity, because it can, reliably, at scale.
Detecting an implant on a panoramic radiograph is, at this point, a solved capability. A 2025 study out of Sivas Cumhuriyet University trained a YOLOv8 model on 32,585 panoramic radiographs and hit 93.1% F1 on implant segmentation. The same model numbered fixtures by jaw position with F1 scores between 0.917 and 0.966 depending on the region. Detection is done. Numbering is close behind it. What's still catching up is the part that matters two years, five years, and ten years after the fixture goes in.
What AI Detects Today
What AI Tracks Over Time
Where It Still Needs a Clinician
An OPG's angulation isn't a footnote — it's the whole measurement problem
A panoramic film is a single 2D projection of a 3D jaw, and the angle shifts slightly almost every time a patient sits down for one. Any tool measuring bone level across serial films has to correct for that drift first, or the “change” it reports is really just positioning noise.
Two different claims get bundled into one headline here. “AI can spot an implant on a film” has strong, repeated published support. “AI can tell you whether that implant is losing bone” is a newer, separate claim — and the evidence for it is thinner but growing fast.
93.1%
F1 score, implant detection & segmentation on OPG
32,585 panoramic radiographs, 2025
AUC 0.95
Pooled diagnostic performance peri-implant bone loss detection
Meta-analysis, 12,545 radiographs
98.1%
Detection precision, bone-loss severity grading model
780 radiographs, 1,210 implants
That middle number is worth sitting with a little longer. A 2026 systematic review pooling five studies and 12,545 periapical and panoramic radiographs found deep learning models detecting peri-implant marginal bone loss at 88% pooled sensitivity and 91% specificity. Translation: for every 100 implants with real bone loss, the model catches roughly 88 — and it isn't crying wolf on the healthy ones either, most of the time.
Three separate jobs, three separate accuracy bars
How to read thisDetection
Is there an implant here, and where? Mature. Works reliably across 2D OPG and periapical film.
Quantification
How much bone is around it, and is that changing? Newer, strong pooled numbers, but dataset sizes are still small compared to detection studies.
Diagnosis
Is this peri-implantitis, and does it need treatment? Not attempted by any current AI product — this stays a clinical call.
Peri-implantitis rarely announces itself. Bone loss around a fixture can progress for months — sometimes longer — before a patient feels anything or a dentist notices it by eye on a busy recall day. That's exactly the gap bone-to-implant contact (BIC) tracking is built to close: instead of reading one OPG in isolation, the model measures the bone level against the implant thread pattern on every film that patient has ever taken, and plots the trend.
Three things happen once that trend line exists. First, a single stable reading tells you almost nothing on its own — it's the second and third follow-up scan that turn a static image into a signal. Second, small changes that a busy clinician would round off to “looks fine” become visible as a slope, not a guess. Third — and this is the part that actually saves chair time — flagging happens automatically at the recall visit, before probing depth has degraded enough to be obvious without measuring it.
Grading, not just flagging
A 2025 study on 780 periapical radiographs covering 1,210 implants used a YOLOv8-based model to both localize peri-implant bone and grade its severity into three tiers — mild, moderate, severe — hitting 98.1% detection precision, 96.0% sensitivity, and 99.1% specificity. That's the shift from “an implant is here” to “here's roughly how much trouble it's in.”
| Task | AI Software Today | Manual Chairside Read | Where the difference comes from |
|---|---|---|---|
| Spotting implants and counting fixtures | Strong | Strong | Both reliable on a clear, well-angled film |
| FDI numbering across the arch | Strong | Strong | Routine pattern recognition either way |
| Identifying implant brand or system | Early-stage | Strong | AI models trained on 5–7 systems max; humans check records |
| Bone-to-implant contact quantification | Strong | Subjective | AI measures pixel distance consistently; eye estimation varies reader to reader |
| Trend across multiple follow-up films | Strong | Rarely done | AI compares serial images automatically; side-by-side manual review is uncommon in practice |
| Peri-implantitis diagnosis & treatment call | Not applicable | Strong | Needs probing depth, bleeding on probing, clinical judgment |
None of this is an argument against using AI for implant monitoring. It's an argument for knowing precisely where a panoramic film's limits sit, regardless of who or what is reading it.
A 2D projection can't show buccal or lingual bone loss. An OPG flattens a three-dimensional ridge into one plane. The model can read the mesial-distal crestal line perfectly and still miss a defect on the buccal or lingual surface that a CBCT would catch immediately.
Angulation drift between visits skews raw measurements. Unless a model explicitly corrects for magnification and patient positioning differences between exposures, a "bone loss" trend can partly just be geometry.
A bone-level trend is a radiographic proxy, not a diagnosis. Peri-implantitis is confirmed clinically — probing depth, bleeding on probing, suppuration — not from a film in isolation, however clean the trend line looks.
Brand identification is narrow by design. Most published implant-brand classifiers are trained on five to seven systems. An older or less common fixture can return no match at all, which isn't the same as the software being wrong — it's outside what it was ever trained on.
No AI bone-loss output carries regulatory sign-off. It's decision support that flags a trend worth a closer look — not a report a clinician can skip reading behind.
Medecro's AI X-Ray Analyzer runs this exact split on OPG and RVG — flagging every implant on a scan, tracking bone-to-implant contact and crestal bone level across each patient's recall history, with confidence scores and one-click override built into the read you're already doing. It's built to catch a bone-level slope worth a second look before probing depth confirms it, not to hand down a peri-implantitis diagnosis on its own.
Yes, reliably — published models detect and number implants on OPGs with F1 scores above 90%. But detection is only the starting point. The more useful capability tracks bone-to-implant contact, peri-implant bone levels, and crestal bone loss across every follow-up scan a patient takes, which is what actually matters for long-term implant survival, not just a one-time read.
Medecro AI X-Ray Analyzer
Implant detection, numbering, and crestal bone-level trends on OPG and RVG, with confidence scores and one-click override — inside your existing clinic workflow. No claim that it diagnoses peri-implantitis on its own.
Book a Demo — See It LiveSources & references
