20% off on annual plans this month
code SMILE2020% off on annual plans this month
code SMILE20CEJ-to-crest measurement, horizontal vs. vertical defect classification, and AAP/EFP Stage I–IV staging — from a single scan, without manual calipers.
A periodontist can glance at a radiograph and get a rough read on where the bone has pulled back. An AI model does something closer to a Schei ruler — except it draws that ruler on every tooth in the image, in a fraction of a second, and it doesn't get sloppy by the fortieth film of the afternoon.
Three things happen, in order. The model finds two fixed points on each tooth. It measures the gap between them. Then it sorts what it sees into a bucket a dentist already recognises — horizontal, vertical, or somewhere on the AAP/EFP Stage I to IV scale. None of it touches a periodontal probe. All of it depends on how clean the anatomy looks on the film.
The measurement underneath almost every AI periodontal tool is radiographic bone loss, or RBL — the distance between the cementoenamel junction (CEJ) and the marginal alveolar bone (MAB) on a 2D projection. Because 1–2mm of non-mineralised supracrestal tissue normally separates the two, a small gap is expected. Anything meaningfully past that threshold is what gets flagged.
Two developments changed how well this actually works. First, tooth-level segmentation got good enough to isolate a single tooth cleanly from a crowded panoramic image. Second — and this is the part most explainer content skips — the better systems stopped training purely against radiographic landmarks and started training against real clinical outcomes from periodontal probing, which is a messier but far more honest ground truth.
94.2%
vs 85.6% for periodontal specialists
AUROC, Stage II–IV detection
Li et al., npj Digital Medicine, Nov 2025
0.967
AUROC across 4 independent centers (760 OPGs, multinational validation)
Same study, external dataset II
0.02s
Average read time per image
vs. 30–62 seconds for a human panel
Li et al., 2025
That 94.2% vs. 85.6% gap is the headline. The quieter detail — worth sitting with for a second — is that the three specialists reading the same 760 images only agreed with each other 65% of the time (Fleiss' kappa 0.64). A single OPG read is genuinely noisy, even among people who've done this for years. The AI isn't beating a stable benchmark; it's beating a benchmark that disagrees with itself.
Not all bone loss looks the same on film, and the shape matters clinically. Horizontal loss drops evenly across the crest, roughly parallel to a line joining the CEJs of neighbouring teeth. Vertical — also called angular or infrabony — cuts obliquely into one surface of a single tooth, and it's the pattern most associated with the disease getting genuinely aggressive.
| Pattern | Horizontal bone loss | Vertical / infrabony defect |
|---|---|---|
| Radiographic appearance | Even, crest-parallel reduction across adjacent teeth | Angular, isolated to one tooth surface |
| Associated stage | Stage I–II | Stage III–IV, often with furcation involvement |
| Training data availability | Abundant — the dominant pattern in most datasets | Scarce — a genuine minority class |
| Detection maturity | Well established, high sensitivity | Improving, but still the harder call |
The scarcity problem isn't a modelling failure — it's a reflection of how rare these defects actually are on a typical film. A 2026 feasibility study out of the University Hospital of Cagliari pulled 7,464 periapical radiographs from routine archives and found only 581 contained an identifiable infrabony or furcation defect worth annotating. That's roughly one in thirteen. Vertical-defect classifiers aren't starved for effort; they're starved for examples, because most mouths simply don't produce many.
How the 2018 framework links pattern to stage
Quick referenceStage I–II
Horizontal bone loss, 1–4mm clinical attachment loss, no furcation involvement
Stage III–IV
Vertical/angular defects, ≥5mm CAL, frequently paired with furcation defects and a higher tooth-loss risk
Turning a millimetre figure into an AAP/EFP stage isn't a lookup table, even though it's often described as one. The most rigorously validated system to date — HC-Net+, tested across four dental centres in 2025 — sorts cases into three practical buckets rather than four: periodontal health/gingivitis/Stage I, Stage II on its own, and Stage III–IV combined. That grouping isn't a shortcut. It's an honest response to how blurry Stage II calls get on a single image, even for people trained to make them.
Missed diagnoses cluster almost entirely at the edges of this scale. In the same 2025 multicentre trial, localised Stage II periodontitis had a 20.6% miss rate for the AI model — better than the specialist panel's 25.4%, but nowhere close to solved. General dentists missed 44.4% of the same cases; dental students missed 88.9%. The pattern holds everywhere this gets tested: early, localised disease is where every reader — human or model — struggles most.
None of this is presented as a solved problem, and it shouldn't be marketed that way either.
2D projection limits. A panoramic or periapical film can't show buccal-lingual defect depth. Suspected vertical defects that will inform surgical planning still need CBCT confirmation before a flap is opened.
Image quality dependency. Radiographs are taken by general dentists, not radiologists, and quality varies widely — only a minority of routine films meet an "optimal" standard. Model accuracy tracks image quality closely, especially for borderline Stage II cases.
False positives from foreign bodies. Implants, crowns, and dense restorations get mistaken for bone loss more often than any other single error type — accounting for close to two-thirds of false positives in one 2025 multicentre analysis.
Localised disease hides. When fewer than 20% of teeth show attachment loss, both AI and human readers under-call the case. Full-mouth context still beats a single image for genuinely localised periodontitis.
Where does Medecro sit in this? As an AI infrastructure layer, with dental as the validated entry point — the periodontal bone-loss flag runs through the same OPG and periapical pipeline used for caries and periapical lesion detection. The model measures and flags; the treating clinician reviews the film and confirms the stage before anything reaches a treatment plan. That's not a hedge — it's how every serious system in this space is actually built and validated.
No. AI measures radiographic bone loss, which correlates with disease but isn't the diagnostic standard. The 2018 AAP/EFP framework defines periodontitis using clinical attachment loss and probing depth — an exam a radiograph can't replace. AI here is a screening and triage layer, not a diagnosis.
Book a walkthrough with your OPGs and periapical films — no free trial, just a session with your actual radiographs.
Sources
