Mucosal thickening, fluid levels, and maxillary sinus opacification — the kind of incidental findings that prompt ENT referrals and, sometimes, save patients a trip to a specialist they didn't know they needed. Here's what the newest paired-modality data actually shows an AI model catching on a flat panoramic film, and where the projection simply runs out of information.
What this article covers
Which maxillary sinus findings a dental AI model can actually flag on a panoramic radiograph, what the 2026 paired OPG-versus-CBCT accuracy data shows, why a 2D projection struggles with a structure it was never built to isolate, and when an AI flag (or a clean read) should — and shouldn't — change what you do next.
Nobody orders an OPG to check the sinuses. They order it for the teeth. But the maxillary sinus floor sits directly above the upper molars and premolars — there's no cropping it out. Any model scanning that film for periapical lesions or bone loss is looking straight through sinus tissue whether it's trained to or not.
That overlap turns out to matter. A review of CBCT scans found sinus pathology in 15% of cases flagged as incidental — findings nobody was looking for, discovered because the imaging happened to include that anatomy. Some of those get an ENT referral. A few, per published case reviews, catch something that genuinely needed catching.
What AI Tends to Flag on OPG
What Still Needs CBCT or ENT
Where the Two Overlap Today
OPG wasn't built for this job
A 2019 evaluation concluded that panoramic radiography is of limited value for assessing maxillary sinus disease — sensitivity is poor against a CBCT reference, even though specificity holds up. The zygomatic process and hard palate sit on top of the sinus floor in a 2D projection, and that superimposition doesn't go away just because a model is reading the film instead of a person.
One study, published in Diagnostics in May 2026, did something nobody had done before: it ran the same commercial AI platform against paired OPG and CBCT scans from the same 166 patients, then checked both against a consensus CBCT read. Same patients. Same sinuses. Same software. Only the imaging modality changed.
50.6% vs 69.9%
AI accuracy on OPG vs. CBCT
Same patients, same platform
Diagnostics, May 2026
21.4%
AI recall (sensitivity) on OPG
Near coin-flip territory
332 sinuses, paired design
75.7%
A sinus-specific CNN's OPG accuracy
Purpose-built, single task
Imaging Sci. Dent., 2022
That last card is the one worth sitting with. A model trained on nothing but sinusitis detection hit 75.7% accuracy on the exact same imaging modality where a general-purpose dental AI — one built to flag 40-plus conditions across an entire OPG — landed at roughly 50%, barely better than a coin flip. Breadth costs depth. That's not a knock on any specific vendor; it's what happens when one model has to do a hundred jobs instead of one.
Three different sinus questions, three different accuracy bars
How to read thisDetection
Is something there at all? Moderate on CBCT for general-purpose tools, weak on OPG — recall around 21% in the largest paired study to date.
Classification
Mucosal thickening, cyst, or fluid level? Current tools output a binary "abnormality" flag, not a type. That call still needs a human read.
Referral decision
Does this need CBCT or ENT? Not something the AI attempts. Size and symptoms drive that call, not confidence scores.
Size matters — but only on the modality that can measure it properly. On CBCT, mucosal thickness above 5mm and polyp or cyst volume above 400mm³ both strongly predicted a correct AI call, with AUC values of 0.80 and 0.85. Reassuring, until you look at the same relationship on OPG.
| Sinus Finding | AI on OPG | AI on CBCT | Why the gap |
|---|---|---|---|
| Mild-to-moderate mucosal thickening | Weak | Moderate | Most common finding, but subtle — gets lost under overlapping bone on 2D |
| Extensive sinus opacification | Moderate | Strong | Large enough to survive superimposition even on a flat film |
| Air-fluid level | Weak | Strong | Highly dependent on patient head position during OPG capture |
| Polyp / retention cyst | Weak | Moderate | Below ~400mm³, missed on both modalities; larger ones favor CBCT |
| Sinusitis diagnosis (clinical) | Not attempted | Not attempted | Imaging alone — 2D or 3D — doesn't diagnose; needs symptom correlation |
Think of it as a conspicuity detector, not a sinus reader
The pattern across both modalities points to the same underlying behavior: this class of AI catches what's obvious and misses what's subtle. On CBCT, obvious correlates with clinically real. On OPG, obvious mostly correlates with luck — a well-positioned film, a large enough finding, or both at once.
A negative AI read on OPG proves close to nothing. With recall around 21%, the model misses roughly four out of every five sinus findings that CBCT would have caught. A clean flag is not a clean sinus.
Superimposition is a hardware problem, not a software one. The zygomatic process and hard palate physically overlap the sinus floor on every panoramic exposure — no amount of model tuning changes what light actually reached the sensor.
Generalist tools trade depth for breadth. A model built to check 40-plus conditions across an entire OPG will not match a single-task sinus model on the same film — 50% versus 75.7% accuracy, in the data above.
It's a binary alert, not a diagnosis. Current tools output "abnormality present" — not mucosal thickening versus cyst versus fluid level. Someone still has to look and decide what it actually is.
Even CBCT AI misses a substantial share of true findings. Recall topped out near 54% on the higher-resolution modality — better than OPG, still far from a rule-out tool on its own.
Medecro's AI X-Ray Analyzer flags visible findings across the full OPG field it's already scanning for periapical and bone-level review — including the sinus floor sitting right above the upper molars. It's positioned the way this data suggests it should be: a second look inside a workflow you already run, not a sinus-diagnostic claim. Anything it surfaces near the sinus still goes through the same clinical judgment — and, where warranted, the same CBCT or ENT referral — it would without AI in the loop.
No. Current tools output a general "abnormality present" flag, not a diagnosis. Sinusitis is a clinical diagnosis that needs symptom correlation, and imaging — 2D or 3D — only ever supports that call.
Medecro AI X-Ray Analyzer
Confidence-scored detection on OPG and RVG, with one-click override, inside your existing clinic workflow. No claim that it replaces a CBCT read or an ENT referral, because — as the data above shows — it can't.
Book a Demo — See It LiveSources & references