How early can proximal caries be detected on a bitewing?

There's no depth at which a proximal lesion switches from invisible to visible. There's a probability that climbs as the lesion gets deeper, and at the enamel-only stage that probability sits close to a coin flip. The more interesting part is how widely it varies between studies — and how much of what's known comes from extracted teeth on a bench rather than from patients in a chair.

What the question actually is

"How early can you see it?" hides three separate questions, and most of the disagreement in the literature comes from mixing them up.

Detected against what? A sensitivity figure only means something next to a reference standard. Some studies section the tooth and read the lesion under a microscope. Some use an expert panel reading the same radiograph. Some use operative excavation. These are not interchangeable.

Early meaning what? The usual threshold in this literature is the non-cavitated, enamel-only lesion — demineralisation that hasn't reached the dentinoenamel junction and hasn't broken the surface. That's the lesion you'd want to catch, because it's the one that can still be arrested.

On what, and by whom? Extracted premolars mounted in a jig under standardised exposure are a different problem from a patient's posterior interproximal caries on a bitewing, where overlap, angulation and soft tissue all get a say. Population descriptors carry most of the weight here. A number without one is close to meaningless.

What the evidence shows

The largest synthesis is the Cochrane review by Walsh and colleagues, which pooled 104 datasets from 77 studies covering 15,518 tooth sites, all of them looking specifically at non-cavitated enamel-only coronal caries. Across all imaging methods, the summary sensitivity was 0.47 (95% CI 0.40 to 0.53) and specificity 0.88 (95% CI 0.84 to 0.92). The review rates its own certainty of evidence as low.

The authors put that in concrete terms. In a modelled cohort of 1,000 tooth surfaces at 63% enamel caries prevalence, those figures produce 337 surfaces called disease-free when caries was truly present, and 43 called diseased when it wasn't.

Read that pair of numbers together, because they're the shape of the whole problem. When a radiograph says a proximal surface is carious at the enamel stage, it's usually right. When it says the surface is sound, that's much weaker information.

Sensitivity rises with depth. In the meta-analysis by Schwendicke, Tzschoppe and Paris — 117 studies, 13,375 teeth, 19,108 surfaces, mostly permanent teeth and mostly at high risk of bias — pooled sensitivity for proximal dentine lesions was 0.36 (95% CI 0.24 to 0.49), and for cavitated proximal lesions it rose above 0.60. Detection of any lesion, including initial ones, fell in the 0.24 to 0.42 range, with moderate to high specificities. That range spans occlusal and proximal surfaces together, so it can't be pinned to proximal alone.

The gradient isn't a quirk of human attention. In Ammar and Kühnisch's meta-analysis of AI models reading bitewings, pooled sensitivity was 0.84 (0.80 to 0.87) for dentin lesions and 0.71 (0.66 to 0.75) for enamel lesions — the same direction, from readers that don't get tired. Most of the included studies carried unclear or high risk of bias, so treat those as indicative rather than settled.

Dentists reading real bitewings land in the same region. The ADEPT study by Devlin and colleagues gave 23 dentists 24 clinical bitewings containing 65 enamel-only proximal lesions confirmed by an expert panel, alongside 241 healthy proximal surfaces. Without software assistance, they found 44.3% of the lesions and wrongly called caries on 3.7% of the healthy surfaces. That's one study with a panel-based reference standard rather than histology, but it's clinical images and qualified readers, which is rarer in this literature than you'd think.

Whatever depth you measure, the lesion is probably deeper. Syriopoulos and colleagues, radiographing 56 extracted unrestored premolars against a histological standard, reported that caries depth was underestimated. Jacobsen and colleagues measured 64 lesions in 177 extracted teeth against histological sections across four digital systems and found untrained observers underestimated depth in all of them — concluding that a high probability remains that any individual measurement is over- or underestimated. Both studies are in vitro. What they support is a direction, not a correction factor.

Does a digital sensor change the answer?

A little, and not in the way the marketing suggests.

Splitting the Cochrane review by modality gives analogue radiographs a sensitivity of 0.44 (95% CI 0.38 to 0.50) and specificity of 0.90 (95% CI 0.86 to 0.93); digital radiographs 0.49 (95% CI 0.42 to 0.56) and 0.87 (95% CI 0.82 to 0.91); and CBCT 0.60 (95% CI 0.51 to 0.68) and 0.81 (95% CI 0.73 to 0.88). The difference between methods was statistically significant on meta-regression (Chi²(4) = 32.44, P < 0.001), with the highest sensitivity for CBCT and the highest specificity for analogue radiographs. These are separate subgroup estimates rather than a head-to-head comparison, so the gap between analogue and digital here is not itself a finding about which one detects more.

Set against that, Syriopoulos and colleagues found no significant difference in diagnostic accuracy between the dental films and digital systems they tested in vitro, and concluded that the reader — not the imaging modality — was the main source of variation. In the same study, radiologists outperformed general practitioners whatever the recording system, and lesion depth significantly affected observer performance.

Both findings can hold. A meta-regression across 77 heterogeneous studies can detect a real modality effect that's small next to the spread between individual readers.

Where it's unsettled

The spread is the headline. Across the individual studies in the Cochrane review, sensitivity ranged from 0 to 0.96 and specificity from 0 to 1.00. None of the pre-specified sources of heterogeneity explained that variation. A pooled 0.47 sitting on top of a 0-to-0.96 range is a description of disagreement, not a resolution of it.

Most of the evidence is laboratory work. Only 17 of the 77 studies in the Cochrane review used an in vivo design in a clinical setting. In the proximal-specific meta-analysis by Janjic Rankovic and colleagues, 31 in vitro studies and 5 clinical studies made it into the analysis; the bitewing figures they report — in vitro AUC of 0.55 to 0.82 for caries detection and 0.81 to 0.92 for dentine caries detection — are explicitly the in vitro subset, and the authors state that clinical data are limited and should be interpreted with caution. The Cochrane authors downgraded certainty partly for indirectness arising from the in vitro studies.

This isn't two camps reporting opposite results. It's a split in what the evidence can support. And Gimenez and colleagues, reviewing 240 manuscripts, showed the stakes: the ranking of detection methods actually reverses between settings, with laser fluorescence beating visual inspection for all lesions on proximal surfaces of permanent teeth in the laboratory setting but for advanced occlusal lesions in the clinical setting.

Real mouths have restorations. The studies didn't. No study in the Cochrane review included restored teeth in its sample or reported the inclusion of sealants. So these figures say nothing about reading a proximal surface next to an existing restoration — which is where secondary caries and a restoration overhang do their damage. That's absence of evidence, not evidence that performance is the same.

There's no mineral-loss threshold to quote. A specific percentage of demineralisation required before a lesion becomes radiographically visible is a figure that gets repeated, but no retrievable source establishing one for proximal caries turned up here. The sourced answer is the gradient, not a cutoff.

And the false positives have their own sources. Sensitivity studies count missed lesions; they don't tell you what produced the calls that were wrong. Cervical burnout and the Mach band effect both generate radiolucencies that look like early proximal decay, and neither is a lesion.

What this means at the chairside

Stated conservatively, and only as far as the evidence goes:

• A positive is worth more than a negative. Specificity at the enamel stage sits around 0.88 in the Cochrane pooled estimate. Sensitivity sits around 0.47. A clean bitewing doesn't rule out an enamel-only lesion.
• Thresholds for acting are set by consensus, not by the sensitivity data. The EFCD-ORCA-DGZ expert Delphi consensus recommends that only cavitated proximal lesions — confirmed visual-tactile, or radiographically extending into the middle or inner dentine third — be treated restoratively, and that non-cavitated lesions radiographically confined to enamel may be arrested non-invasively in low-risk individuals. The same statement names avoiding false-positive detection and over-treatment as a priority.
• Practice doesn't track that threshold neatly. In a retrospective study of Finnish public-service records by Muhonen and colleagues, covering 394 lesions registered by 16 dentists, the predicted probability of restorative treatment was 8% for lesions reaching the dentinoenamel junction and 75% for those extending to the border between the outer and middle thirds of dentin. Of the lesions those dentists registered, 57% were radiographically confined to enamel, 27% reached the outermost third of dentin, 12% the middle third and 4% the inner third. That's the depth distribution of what got recorded — not of what was there.
• The reader is the biggest lever you control. Syriopoulos and colleagues named the dentist, not the modality, as the main source of variation. The ADEPT study makes the same point from the other direction: the same 24 bitewings yielded 44.3% detection in one arm and 75.8% in the other. The images didn't change.

That last point is why the enamel stage is worth training on rather than writing off. It's the depth where a lesion can still be arrested.

Numbers are one thing. Try three real bitewings and see where your own threshold sits.