20% off on annual plans this month
code SMILE2020% off on annual plans this month
code SMILE20Spotting a radiolucency at the apex is the easy part. Working out whether it's a granuloma, a cyst, or an abscess — from a flat, 2D image — is where most AI models start running into the same limits a general dentist does.
What this article covers
How AI models are trained to flag periapical radiolucencies on OPG and RVG, what happens when you ask them to go one step further and classify the lesion as a granuloma, cyst, or abscess, and the specific points — small lesions, anatomical mimics, texture overlap — where that second step still falls short of a histopathology report.
Same shadow on the film. Three different treatment plans. A well-defined apical radiolucency on a lower molar could mean:
A 2D X-ray was never built to make that call with certainty on its own.
Periapical Granuloma
Radicular Cyst
Periapical Abscess
Same radiolucency, three different biological realities
Border sharpness and size give hints — cysts trend larger with a well-corticated border, abscesses trend diffuse — but none of those features are diagnostic alone. That ambiguity exists for a human reader before AI ever enters the picture.
Two different claims get collapsed into one in most marketing:
Detection numbers, pooled from a November 2025 systematic review across 28 studies (2000–2025):
That's a wide spread — wider than the 2022 literature suggested. Worth remembering before any single AUC headline does all the talking.
70–99.65%
Accuracy range
Lesion detection, 28 pooled studies
Alaqla et al., Front. Dent. Med., Nov 2025
65–100%
Sensitivity range
Across detection studies
Same systematic review
97% / 88%
AUC — cysts / granulomas
Classification, single study
Ver Berne et al., 249 images
That third card needs unpacking before it goes anywhere near a claim. The model wasn't equally good at both lesions:
Three different questions, three different accuracy bars
How to read thisDetection
Is there a periapical radiolucency here at all? Well studied, strong published numbers, works across OPG and RVG.
Segmentation
Where exactly are the lesion's borders? Reasonably mature — pixel-level accuracy above 0.97 in recent Mask R-CNN work.
Classification
Granuloma, cyst, or abscess? The thin part of the literature — only 3 of 28 studies in the most recent systematic review even attempted it.
The table below reflects directional patterns from the published detection literature — exact numbers by view, specific to granuloma-cyst-vs-abscess classification, aren't publicly established yet.
| Task | RVG (Periapical) | OPG (Panoramic) | Where the difference comes from |
|---|---|---|---|
| Detecting a periapical radiolucency | Strong | Strong | Both well represented in training data |
| Assessing border definition | Better | Fair | RVG resolution captures cortication more clearly |
| Avoiding anatomical false positives | Better | Weaker | Sinus floor, foramina overlap roots on OPG |
| Granuloma vs. cyst differentiation | Early-stage | Under-studied | Limited to small research datasets on either view |
| Catching early / small lesions (<5mm) | Sensitivity drops | Sensitivity drops further | Lower resolution and superimposition on OPG |
The model reads texture and geometry — not biology
An AI classifier learns statistical patterns in pixel intensity, border shape, and density that correlate with a label in its training set. It has no access to what's actually happening at the cellular level. And in a meaningful share of published datasets, that training label was itself a radiographic read rather than a confirmed histopathology result — so the “ground truth” the model is learning from can carry the same ambiguity a human reader would have.
None of this is a case against using AI here. It's a case for knowing exactly which parts of the read to trust unreviewed and which parts still need a clinician's eyes on the image before anything gets decided.
Small and early lesions get missed. Sensitivity across published detection studies fell as low as 0.44 in some datasets — almost always concentrated in early-stage or sub-5mm lesions, exactly the cases where catching something early matters most.
Accuracy swings hard by tooth type. A 2025 study testing a commercial AI system (Diagnocat) on 357 panoramic radiographs found overall sensitivity of 0.78 and specificity of 1.00 — solid numbers. But sensitivity for canines dropped to just 0.27, driven by projection geometry and superimposition specific to that tooth position on a panoramic image. A single overall accuracy figure can hide a tooth-by-tooth performance cliff.
Granuloma-cyst texture overlap is a real, unsolved problem — not just an AI limitation. Older texture-analysis research comparing lesion classification against histopathology found 6 out of 25 lesions misclassified even with dedicated feature extraction. The overlap exists in the biology, not just in the model's training data.
Acute abscess without bone change yet. Bone destruction takes time to show up radiographically. A patient can be in genuine acute distress — pain, swelling, a tooth that's suddenly mobile — while the X-ray still looks close to normal. No amount of model sophistication reads a change that the bone hasn't made yet.
Classification research still runs on small datasets. The strongest published granuloma-vs-cyst numbers come from a couple hundred images at most. Promising, worth watching — but not yet the kind of large, multi-center validation that detection tasks have already been through.
Medecro's AI X-Ray Analyzer is built around exactly that division of labor — confidence-scored detection on OPG and RVG with one-click override, while the differential diagnosis stays with the dentist looking at the whole clinical picture, not just the film.
Not reliably enough to treat as final. A 2023 study reported 97% AUC for identifying cysts — but only 88% AUC for granulomas, missing about 1 in 4 actual granulomas. Radiographic size and border definition give a useful lean, not a diagnosis. Anything going to surgery still needs histopathology.
Medecro AI X-Ray Analyzer
Confidence-scored detection, lesion annotation, and one-click override, built into your existing clinic workflow. No separate login, no standalone app — and no claim that it replaces your clinical judgment on what the lesion actually is.
Book a Demo — See It LiveSources & references
