Can AI Detect Gum Disease On Dental X-Rays?

AI periodontal disease detection is showing strong results on dental X-rays. A 2026 meta-analysis of 50,080 radiographic images reported 93% sensitivity and 88% specificity at the patient level. Spotting bone loss on an image is only part of diagnosing periodontitis, though. Here is what today's tools can reliably detect, and where clinical judgment still decides.

TL;DR

• A 2026 review pooled 14 studies covering 50,080 dental radiographs.

• AI models reached 93% sensitivity and 88% specificity at the patient level.

• AI flags radiographic bone loss and can support screening and referrals.

• Probing, clinical examination, and professional interpretation still make the diagnosis.

How Accurately Can AI Detect Periodontal Disease?

The newest evidence comes from a systematic review in the Journal of Dentistry, published online on September 17, 2026. It is the latest piece in Dental Reviewed's coverage of how AI is entering clinical dentistry. The authors pooled 14 studies and 50,080 radiographic images, according to the PubMed abstract. Seven studies judged disease at the patient level, and seven at the tooth or image level.

Each metric answers a different clinical question. Sensitivity is the share of diseased cases the model flags, so high sensitivity means fewer missed patients. Specificity is the share of healthy cases it correctly clears, so high specificity means fewer false alarms and fewer unnecessary referrals.

At the patient level, the model only needs to find one affected site to call a patient positive. That makes it easier to score high sensitivity there than at the tooth level, where every tooth is judged separately.

Pooled figures also hide where models struggle, and the pattern runs against what clinicians need most. In one study from the 2024 review, staging accuracy reached 94% for stage III but only 64% for stage I. In another, agreement with a dentist's bone-loss readings was good for incisors but poor for molars. It almost disappeared for angular defects (ICC 0.04, compared with 0.74 for horizontal bone loss). Advanced horizontal bone loss is the disease any clinician spots in seconds. Early lesions and vertical defects around molars are where a second reader would add the most, and where current models are weakest. Those vertical defects are also the sites a periodontist evaluates for regenerative surgery.

These are pooled estimates from research models, not performance figures for any commercial product. A tool can perform differently on another practice's sensors, exposure settings, and patient population. The pooled numbers show what well-built models can achieve under study conditions, and they set a benchmark for asking vendors about their own data.

What Can AI Detect On Dental X-Rays?

AI dental X-ray analysis for periodontal disease centers on the hard tissue a radiograph can show, mainly the alveolar bone that supports the teeth. Depending on the model, that includes:

• Bone loss around individual teeth

• Bone height relative to landmarks such as the cementoenamel junction and root apex

• Horizontal and vertical (angular) bone-loss patterns

• Teeth flagged as periodontally compromised

• Changes that warrant a closer clinical look, such as furcation radiolucencies or calculus

Those capabilities fall into three technically different tasks. Segmentation outlines structures such as teeth, bone, and the cementoenamel junction pixel by pixel. Measurement then uses those outlines to calculate distances, such as millimeters from the cementoenamel junction to the bone crest, or bone loss as a percentage of root length. Classification assigns a label, such as "bone loss present," "healthy," or a stage.

Each task needs its own validation. A model can outline the crest accurately yet misjudge stage. Another can classify patients well without producing reliable measurements. When a vendor quotes accuracy, ask which task the figure describes.

There is a further mismatch between what most tools report and what staging requires. The 2017 classification stages radiographic bone loss as a share of root length, under 15%, 15–33%, or extending into the middle third and beyond. Bitewing-based tools report millimeters from the cementoenamel junction to the crest. Pearl's documentation, for example, color-codes readings as 0–2.5 mm, 2.5–4 mm, and over 4 mm. The same 4 mm reading is a larger share of a short incisor root than of a long canine root, so it can fall into a different stage. In the independent Overjet study, researchers had to convert millimeter readings to percentages using textbook average root lengths, which the authors flagged as a source of error.

Percentages matter for grading too. Grade uses the ratio of bone loss (as a percentage at the worst tooth) to the patient's age. A ratio above 1.0, such as 50% bone loss in a 40-year-old, points to grade C, rapidly progressing disease. Software that measures percentage bone loss on periapicals could calculate this index automatically, a practical use most marketing overlooks.

How Does AI Identify Periodontal Bone Loss?

Most systems use deep learning, usually convolutional neural networks trained on thousands of labeled radiographs. Experts mark the cementoenamel junction, bone crest, and root apex on each training image. The model learns to reproduce those labels on new images, then converts its landmarks into measurements or a classification.

The quality of those labels caps the quality of the model. Many studies use expert annotation of the X-ray as the reference standard, while fewer compare against a full clinical periodontal diagnosis.

Even a perfect copy of the expert's reading inherits the X-ray's own error. In a study of 331 defects in 100 patients, radiographs showed about 5.1 mm of interproximal bone loss where surgery revealed 6.1 mm. Another surgical comparison found radiographs underestimated bone loss by 1.5–2.5 mm. When a model "matches expert annotation," it matches a view that already tends to understate disease.

Image type changes the task. A bitewing shows the crest clearly but not the apex, so it cannot give a percentage of root length. Periapical films show the whole root. A panoramic X-ray covers both jaws in one image, but distortion and overlapping anatomy make fine measurements less reliable.

Other factors also affect results. Restorations and crowns can hide the cementoenamel junction. Overlapping contacts blur the crest, and sensor type and image processing vary between practices. Each product makes its own choices about these problems, so performance does not transfer automatically from one tool to another. Our broader explainer on dental imaging AI covers how these models are built.

Projection geometry matters more than most software menus suggest. Horizontal bitewings often cut off the crest once bone loss is advanced, which is why periodontists prefer vertical bitewings or periapicals for these patients. Beam angulation also shifts readings. Research on intrasurgical comparisons found that the further the beam departed from a right-angle projection, the more the X-ray underestimated bone loss. That has a direct consequence for AI "progression tracking." A 0.5 mm change between visits can come from a different film holder angle as easily as from disease.

Can AI Detect Periodontitis Better Than Dentists?

Not on current evidence. The 2026 meta-analysis measured how well AI models perform against a reference standard. It did not test whether AI beats dentists in routine practice.

Head-to-head studies give mixed results. In the earlier systematic review of 30 studies, one multicenter model matched three periodontists and outperformed three general dentists on panoramic images. Another model performed close to five experienced dental hygienists. In a third study, six dentists caught more disease than the model (92% versus 81% sensitivity), but with far more false alarms (63% versus 81% specificity).

The most useful real-world test so far is an independent study in BMC Oral Health. Researchers ran one commercial tool, Overjet, on full-mouth radiographs from 103 patients. They compared it with a general dentist's manual reading and a periodontist's full clinical and radiographic diagnosis. For moderate to severe disease, the AI reached 82% sensitivity and 89% specificity, close to the general dentist's 90% and 90%. It finished in one to two minutes, compared with about 10 minutes by hand.

The larger story in these studies is the variation among humans. In one multicenter study from the 2024 review, periodontists reached 81% accuracy on panoramic images and general dentists 69%. The model reached 80% in 0.03 seconds per image, compared with about 6 seconds for periodontists and 13 for general dentists. The spread between specialists and generalists was wider than the spread between AI and specialists. The same independent study notes that only about 27% of US periodontal disease cases receive treatment. If AI has a clear role, it is narrowing the detection difference between a busy general practice and a periodontal office.

The better question is whether dentists using AI make better decisions than dentists without it. That requires randomized studies with consistent reference standards, and few periodontal studies have been designed that way.

Can AI Replace Periodontal Probing?

No, not on the current evidence. A radiograph shows bone, while periodontitis is defined by loss of attachment and active inflammation in the soft tissues.

An X-ray cannot measure probing depths, bleeding on probing, clinical attachment levels, recession, mobility, or furcation involvement on the buccal and lingual surfaces. It also cannot tell active disease from a stable, treated case with the same amount of old bone loss. Radiographic bone loss also lags behind attachment loss, so the earliest disease may not show on film at all.

The 2017 classification uses both kinds of evidence. Stage relies on clinical attachment loss or radiographic bone loss, tooth loss, and complexity factors. Grade adds the rate of progression and risk factors such as smoking and diabetes. Our guide to periodontitis staging and grading walks through the framework.

The vendors draw the same line. Pearl's support documentation states that its bone level feature does not replace periodontal charting or probing depth assessment.

Disease activity is the blind spot that matters most in daily practice. A patient treated for periodontitis 10 years ago may carry 40% bone loss on every molar yet be stable, with shallow pockets and no bleeding. Another patient with the same X-ray may have active 6 mm bleeding pockets. AI sees the same image for both, while the treatment plans differ completely: routine maintenance for one, active therapy for the other. The reverse case also occurs. Gingival recession and early attachment loss can be clinically significant before the crest moves enough to show on a bitewing.

What Did Earlier Studies Find About AI Periodontal Diagnosis?

The 2026 numbers look stronger than earlier pooled results, but the difference needs careful reading. A 2024 systematic review in Dentomaxillofacial Radiology included 30 studies and pooled 10 of them. It reported 87% sensitivity, 76% specificity, and 84% accuracy for AI-based periodontal bone-loss assessment.

That review also graded study quality with the APPRAISE-AI tool. None of the 30 papers reached the "very high quality" band, and the weakest areas were robustness of results and reproducibility. Many papers did not share code or data. The authors concluded the models were not yet good enough to serve as automated screening tools.

The difference between 76% and 88% specificity does not prove the technology improved by 12 points. The two reviews used different inclusion criteria, reference standards, diagnostic tasks, and statistical methods. The earlier review treated each measure as a simple pooled proportion, while diagnostic meta-analyses often model sensitivity and specificity together.

Other 2025 and 2026 reviews tell a similar story. A panoramic-only review found that studies using a clinical periodontal diagnosis as the reference reported lower performance than those using expert X-ray annotation. Only four of its nine studies were externally validated.

What Are The Biggest Limitations Of AI Periodontal Detection?

The 2026 review flagged a high risk of bias in patient selection and limited external validation. Those two problems run through most of this field, and they explain why strong pooled numbers do not translate directly into practice.

The main weaknesses include:

• Patient-selection bias, since many datasets come from one university clinic with clear-cut cases

• Little external validation, so many models were never tested on images from another institution

• Inconsistent reference standards, with expert X-ray annotation often used instead of a clinical diagnosis

• Narrow datasets that underrepresent some ages, populations, sensors, and restoration types

• Variable image quality, including cone cuts, overlap, and processing differences

• Almost no prospective trials testing AI in day-to-day clinical care

A model trained at one institution can stumble at another. New sensors, different exposure settings, or a population with more crowns and implants can all lower performance.

Disease prevalence also changes what a positive result means. The table below applies the 2026 patient-level estimates to 1,000 hypothetical patients at three prevalence levels:

In a low-risk population, more than half of the AI's positive flags would be false alarms at this level of accuracy. The independent Overjet study shows the same effect in real data. For "any periodontal disease," only 15% of the AI's positive calls were correct, against 46% for the general dentist.

False positives lead to unnecessary deep cleanings and eroded patient trust. False negatives delay treatment. Both risks grow when clinicians accept AI output without checking it, a concern our guide to AI governance in dentistry addresses in detail.

Which Dental AI Software Can Detect Periodontal Bone Loss?

Several US companies now hold FDA 510(k) clearance for AI tools that measure radiographic bone levels. These products measure bone and flag findings for a clinician to review. None of them is cleared to diagnose periodontitis independently. The table summarizes what each company has published:

Clearances, age ranges, and modalities change as companies file new submissions. Confirm each product's current indications in the FDA 510(k) database before purchase. Our roundup of FDA-cleared dental AI tools for CBCT review and our analysis of whether Pearl AI is worth the investment cover adjacent capabilities and costs.

There is also a payer side that practice owners should know about. Overjet's software is used by large dental insurers to review claims, according to the company. Claims for scaling and root planing typically need documented bone loss and pocket depths. A practice submitting a claim may therefore have its radiographs read by an algorithm on the insurer's side, whether or not it uses AI itself. Complete charting alongside the images is the best protection against a denied claim.

How Could AI Improve Periodontal Diagnosis In General Practice?

The clearest near-term value is consistency. Manual bone-level reading varies between clinicians and even for the same clinician on a busy day. AI applies the same measurement rules to every image in seconds.

That consistency supports several practical uses. AI can flag suspicious bone loss on routine bitewings that a rushed review might miss. Standardized millimeter readings make it easier to compare findings from one visit to the next. Color-coded overlays help patients see why a deep cleaning is recommended. They can also support referral to a periodontist when bone loss is advanced.

A realistic workflow keeps the clinician in charge:

1. Take a clinically indicated radiograph.

2. Review the AI's findings and measurements.

3. Compare them with the full periodontal charting and examination.

4. Confirm or reject each AI suggestion.

5. Document the diagnosis, stage, grade, and treatment plan.

A few habits make AI output more trustworthy at the chair. Treat a single flagged interproximal site with suspicion when the neighboring sites are normal, since overlapping contacts and cervical burnout can mimic crestal loss. Compare bone levels only between images taken with similar projections, ideally with the same holder. When the AI and the probe disagree, the probe usually wins for diagnosis, but the disagreement deserves a second look. A deep pocket over normal-looking bone may be a pseudopocket from gingival enlargement. Bone loss without pockets may reflect a stable, previously treated site.

The last step is where findings become decisions. A structured dental treatment plan records the clinician's confirmed diagnosis and phased treatment. Treatment planning software and AI periodontal detection are different capabilities. The plan should rest on the clinician's confirmed findings. Patients facing treatment can review typical deep cleaning costs before their appointment.

What Should Dentists Check Before Choosing AI Periodontal Software?

Marketing claims rarely answer the questions that matter for a purchase. This checklist covers what to ask:

• Which specific periodontal functions are FDA-cleared, and for which ages and image types

• Whether the tool supports your sensors and your mix of bitewings, periapicals, and panoramics

• Whether independent, externally validated studies exist beyond the vendor's own submission data

• Sensitivity and specificity for the exact task you need, such as measurement or classification

• Integration with your imaging software and practice management system

• How clinicians review, override, and document disagreement with AI findings

• Data security, storage location, and how the vendor monitors model performance after updates

• Pricing model, contract length, training and implementation time

Ask vendors for performance data from practices similar to yours. A tool validated in a periodontal referral clinic may produce more false positives in a general practice with a low-risk population. Our 2026 survey of 300 dental professionals shows how practices are approaching these decisions.

What Comes Next For AI In Periodontology?

The research priorities are clear, and most of them involve proving value in real clinics. A recent systematic review in the Journal of Periodontal Research found staging accuracy lower than simple disease-versus-health classification. On panoramic images, staging accuracy ranged from 64% to 91%.

The next wave of studies needs to address several questions:

• Multicenter prospective trials in general practices

• External validation on new sensors and populations

• Longitudinal tracking of bone levels across years of recall images

• Integration of radiographic AI with digital periodontal charting

• Outcome studies showing whether AI assistance changes treatment and tooth retention

Some of this is starting. Vendors already market bone-level tracking over time, and AI-assisted charting tools are emerging. Neither has yet been shown to improve long-term periodontal outcomes in published trials. For now, those remain research goals rather than demonstrated benefits.

Bottom Line

AI periodontal disease detection performs well in research. The September 2026 meta-analysis pooled 93% sensitivity and 88% specificity at the patient level across 50,080 radiographs.

Those figures come from studies with a high risk of patient-selection bias and limited external validation. They describe research models rather than the software in any given practice. An X-ray also shows bone, while periodontitis is diagnosed from attachment loss, inflammation and risk factors as well.

AI is most useful today as a consistent second reader for radiographic bone loss. It can speed up review, standardize measurements and help patients understand their findings. Periodontal probing and clinical judgment remain the foundation of diagnosis.