How the AI identifies caries

The detection doesn't happen all at once. Understanding why the pipeline is staged helps explain both what it catches and where its limits are.

Preprocessing

The raw OPG gets normalised, contrast-enhanced, and brought to a consistent resolution. Beam angle variation and positioning artifacts — both common in clinical OPGs — are corrected through geometric normalisation before any detection runs. This step matters more than it sounds: an unnormalised image will produce inconsistent sensitivity across the arch.

Zone segmentation

A dedicated segmentation model isolates all 32 tooth regions, plus supporting bone and sinus anatomy. Each zone is then passed independently to the caries model. This is why the model can give you tooth-specific findings rather than a general 'caries present' flag — it's working at the zone level, not the whole-arch level.

Classification

A multi-label CNN scores each isolated zone for caries probability across the four depth levels. Zones that exceed the confidence threshold get a bounding region annotation and a severity grade in the report. Zones that don't, don't.

How does the AI reporting workflow work? →

What the AI flags

Not all caries looks the same on a panoramic radiograph, and the model treats each depth level separately. Detection rates vary by level — partly because of what OPG geometry can and can't resolve.

ENAMEL CARIES

~71% detection rate — Radiolucency confined to enamel. Usually shows as a faint interproximal shadow. The lowest detection rate of the four levels — OPG resolution is the constraint here, not the model. Consistent with what a trained human reader achieves on the same modality.

DENTIN CARIES

92–96% detection rate — Extension into dentin. This is the clinically most important intercept point for conservative treatment, and it's where the model performs most reliably. If you're using Medecro.AI as a screening pass, dentin-level caries is where it earns its value.

PULPALLY INVOLVED

High confidence threshold — Radiolucency approaching or breaching the pulp chamber. When flagged, the report auto-suggests an RCT assessment note. The model uses a higher confidence threshold here — the cost of a false positive at this level is higher than at enamel.

PERIAPICAL COMPLICATION

Co-detected alongside caries — Associated periapical changes are flagged alongside the carious lesion when present. The combination of caries + periapical change is suggestive of pulp necrosis or chronic apical periodontitis — the report surfaces both findings together.

Can AI detect periapical lesions on OPG and RVG? →

Clinical significance

OPG isn't the gold standard for caries detection. Bitewing radiographs give better resolution for interproximal surfaces, and that's a physical geometry problem, not something an AI model can fix. What Medecro.AI does is improve the consistency of what you get from an OPG — which, for many Indian clinics, is the scan being taken anyway.

Four areas where it adds something measurable:

Missed occlusal and buccal/lingual caries — These are consistent blind spots on manually read OPGs, especially under time pressure. The model doesn't get tired at film 20. It reads the same zones with the same sensitivity regardless of where in the session the scan came in.
Baseline documentation — When the model flags an enamel-level lesion at recall, you have an objective record of the initial status. That changes the monitor-versus-treat conversation — both for you and for the patient.
Report auto-population — Dental charting fields in the clinical report are populated automatically from the AI findings. In practice, that saves 3–6 minutes of transcription per patient — which, across 20 patients, is a real chunk of the day.
Patient communication — The model generates patient-readable language for each finding. Not clinical notation — plain language that maps to what the patient sees on the annotated image. That matters for case acceptance, particularly for patients who would otherwise leave with a verbal recommendation they'll second-guess at home.

Clinical bottom line: Medecro.AI on OPG is a first-pass screening layer. It is not a substitute for bitewing radiographs on cases where interproximal caries is your primary concern. Use it for what it's good at — consistent, fast, documented screening on the scan you're already taking.

Accuracy benchmarks

The figures come from a hold-out dataset of 3,200 OPGs — scans the model hadn't seen during training. Ground truth was consensus from two oral radiologists.

READER	SENSITIVITY	SPECIFICITY	CONTEXT
Medecro.AI (OPG)	94.7%	91.3%	Hold-out set, 3,200 OPGs, 2024
General dentist (OPG)	82–87%	79–86%	Published literature range, 2019–2023
Bitewing (human + AI)	96–98%	93–96%	Different modality — included for reference only

The bitewing row is in there for context, not comparison. It's a different modality with better interproximal resolution. Comparing OPG AI to bitewing performance is like comparing a chest X-ray to a CT — the better number doesn't mean the first tool is useless.

Known limitations

These aren't caveats that might apply. They apply. Every limitation below is clinically relevant and documented in the validation data.

OPG resolution ceiling — Early enamel-only caries drops to ~71% detection. That's not a model failure — it's consistent with what a trained human achieves on the same modality. If early enamel caries is what you need to find, bitewing is the right tool.
Superimposition artifacts — Overlapping roots or existing restorations can mask lesions. The model flags an image quality warning when significant superimposition is detected, but it can't manufacture information that isn't visible in the image.
Image quality dependency — Under-exposed or motion-blurred OPGs reduce sensitivity significantly. Each scan gets a quality score before analysis runs — if the image is poor, the report says so. Don't rely on findings from a scan that failed the quality threshold.
No substitute for clinical examination — Clinically detectable smooth surface or occlusal caries that doesn't have a radiographic correlate won't be flagged. The model works from what the X-ray shows. If it's not visible on the film, the AI won't find it.

Quick answers

← Back to FAQ Library Book a Demo →