20% off on annual plans this month
code SMILE2020% off on annual plans this month
code SMILE20Per-tooth analysis covering caries depth staging, periapical bone changes, PDL space widening, and post/crown/core detection — the same output a dentist would build up manually, generated the moment the RVG image loads.
What this article covers
The four things AI actually checks on every tooth in an RVG periapical shot — caries depth, periapical bone changes, PDL space widening, and existing posts, crowns, or cores — how each one gets detected under the hood, what the published accuracy looks like, and where a dentist's eye still has to close the loop.
Look at an RVG periapical film the way a dentist actually does, and it's never one question. It's four, fired at the same tooth in the space of a few seconds — is there caries, and how deep. Does the bone around the apex look right. Has the ligament space widened. Is there already a post, crown, or core in there that changes what you can even do next. A tool that returns a single “normal / abnormal” tag on the image throws away three of those four answers before the dentist even opens the chart.
Per-tooth analysis is the fix for that. Instead of one verdict on the whole film, the same four checks run against every tooth visible in the shot, and each finding gets tagged to a specific tooth number using FDI notation — not “somewhere in the lower left.”
None of this happens as one black-box guess. It's closer to three handoffs, each narrower than the last.
1. Image Intake & Segmentation
2. Per-Tooth Feature Detection
3. Structured Output
That per-tooth crop step matters more than it sounds. Run a model on the whole film at once and a hairline PDL change on tooth 46 competes for attention with a metal crown on tooth 36 glowing white a few millimetres away. Crop first, then analyze — the same reason a dentist leans in on one tooth at a time under the loupe instead of eyeballing the whole arch in one pass.
Caries depth staging
Dentin — moderate
Periapical bone change
Radiolucency present
PDL space
Widened, mesial root
Post / crown / core
None detected
A caries flag on its own isn't much use to a dentist mid-consult. What changes the conversation is depth — enamel-only caries is a watch-and-remineralize case, dentin involvement usually means it's time to prep, and anything reaching the pulp reopens the endodontic-versus-extraction question. The staging model works off radiographic radiolucency depth relative to the enamel-dentin junction and the pulp chamber outline, then bins the finding into one of those three tiers per tooth.
Published numbers back the underlying detection task fairly well — 91.7% sensitivity for AI caries detection against 83.4% for conventional visual-radiographic assessment, with specificity following the same pattern (89.2% vs. 81.6%).1 Depth staging specifically is a newer, less-studied layer on top of that — a 2024 periapical segmentation study reported a dice similarity score of 0.75 for caries region boundaries, which is solid but not perfect agreement with a human-drawn outline.2 Worth knowing before you treat a “moderate dentin” tag as gospel.
This is the layer most directly tied to whether a tooth needs endodontic attention. The model looks for three things at the apex specifically — a radiolucent zone (classic sign of periapical pathology), sclerotic bone (a defensive, denser response), and continuity breaks in the lamina dura, the thin radiopaque line that should trace the root outline cleanly.
On periapical films specifically, CNN-based models land in the 0.82–0.85 accuracy range with AUCs above 0.88 — described in one 2025 systematic review as “comparable to clinician performance,” which is a fair, not a triumphant, way to put it.3 AI sensitivity for periapical lesion detection specifically has been measured at 93.5%, against 84.8% for conventional assessment.1 A separate CBCT-trained model (PLA-Net) pushed AUC to 0.98 — but that's volumetric data, a different imaging modality with far more information than a single 2D periapical shot gives you.4
91.7%
AI sensitivity, caries detection
vs. 83.4% conventional exam
LMJHCR, 2025 (n=240 teeth)
93.5%
AI sensitivity, periapical lesions
vs. 84.8% conventional exam
Same study, periapical arm
0.82–0.85
CNN diagnostic accuracy
Periapical bone-loss models
Systematic review, 2025
By the time a periapical radiolucency is visible, the disease process has usually been running a while. A widened periodontal ligament space is often the earlier tell — the ligament reacting to inflammation at the apex before enough bone has resorbed to show up as a clear lucency. Several apical-lesion detection studies actually build this staging into the reference standard itself: a three-point ordinal scale where 0 is no finding, 1 is “widened PDL, uncertain,” and 2 is a clearly detectable lesion.5 AI models trained against that scale are, in effect, being asked to catch the ambiguous, early-stage case — not just confirm what's already obvious.
This is also the finding most sensitive to image geometry. A slightly off angulation on the RVG shot can make a normal PDL space look artificially widened, or hide a genuine one. It's the single biggest reason this particular output tends to carry a wider confidence band than the other three — worth flagging to the dentist reviewing it, not smoothing over.
Before any treatment plan gets written, someone needs to know what's already in the tooth. A post-and-core rebuild changes the restorative options completely compared with a virgin tooth, and a crown margin sitting subgingivally is a different conversation than one sitting clean above the gumline. Radiopaque materials — metal posts, cores, crown copings — actually make this one of the more reliable detections in the pipeline, since the contrast against tooth structure is stark to begin with.
Segmentation studies on periapical radiographs specifically report F1 scores of 1.0 for crown detection and 0.98 for root canal filling detection, with filling material close behind at 0.95.6 Those numbers run higher than the softer-tissue calls like caries depth or PDL widening — radiopaque metal against tooth structure is, frankly, an easier visual problem than judging a millimetre of ligament space.
Why the four outputs don't share one accuracy number
How to read thisPost/crown/core
High-contrast radiopaque material — the easiest of the four calls, F1 scores near or at 1.0 in periapical-specific studies.
Caries/bone
Mid-range accuracy, 0.82–0.93 depending on the study and dataset — solid, not infallible.
PDL widening
The most geometry-sensitive output — angulation shifts move the reading more than for the other three.
A single 2D projection has real limits. RVG periapical films compress a 3D root and bone structure into one flat image — overlapping anatomy or superimposed structures can mimic or mask a genuine finding, on a human read or an AI one.
Radiopaque restorations can interfere with adjacent readings. A large metal crown or post can cast enough scatter to make caries staging or bone assessment on the same tooth, or the one next to it, less reliable — the same failure pattern shows up in restoration-detection literature on panoramic films.7
Depth staging and PDL widening are still comparatively young outputs. Detection of caries and periapical lesions as binary findings is well-validated; staging depth into three tiers, or catching a subtle PDL change before a lesion forms, is a newer and less studied layer on top of that same detection.
Image angulation changes the read. A cone-cut or a steep vertical angle can distort apparent PDL width and root length enough to shift a finding either direction — retaking the shot is sometimes the right call before trusting the output.
None of this is a diagnosis on its own. A structured per-tooth output is a starting point for the dentist's read, not a replacement for it — every finding carries a confidence score and a one-click override for exactly that reason.
Medecro's AI X-Ray Analyzer runs this exact four-part per-tooth breakdown on RVG and OPG images inside your existing clinic workflow — confidence-scored, tagged to FDI tooth numbers, with a one-click override on every finding. It's built to make the first pass faster, not to write the treatment plan on its own.
Per-tooth, not film-wide. The image is segmented into individual teeth, and four checks run on each one — caries depth staging, periapical bone changes, PDL space widening, and post/crown/core detection — with a confidence score attached to every finding and tagged to the specific FDI tooth number.
Medecro AI X-Ray Analyzer
Confidence-scored, FDI-tagged findings on OPG and RVG, with one-click override, inside your existing clinic workflow — built to speed up the first pass, not replace the read.
Book a Demo — See It LiveSources & references
