68/100

Moderate evidence

Cite
More
EmailLinkedInX

A study of 210 patients found AI detected more tooth decay than dentists' exams

Study title: DIAGNOSTIC ACCURACY OF ARTIFICIAL INTELLIGENCEBASED SYSTEMS FOR DETECTION OF DENTAL CARIES COMPARED WITH CONVENTIONAL DIAGNOSTIC METHODS

Evidence type
Controlled or validated comparison
Sample
n = 210
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: DIAGNOSTIC ACCURACY OF ARTIFICIAL INTELLIGENCEBASED SYSTEMS FOR DETECTION OF DENTAL CARIES COMPARED WITH CONVENTIONAL DIAGNOSTIC METHODS

A comparative diagnostic accuracy study of 210 adults at Shahida Islam Dental College in Lodhran, Pakistan, tested a deep-learning caries detector on bitewing radiographs. It reached 90.9% overall accuracy, with 91.5% sensitivity and 90.2% specificity, against a consensus reference diagnosis by two clinicians.

Read the study abstract

Objective: To evaluate the diagnostic accuracy of an artificial intelligence (AI)-based dental caries detection system compared with conventional clinical and radiographic diagnostic methods in patients presenting for routine dental examination. Methods: This comparative diagnostic accuracy study included 210 patients undergoing dental examination. Each participant received a standardized clinical examination followed by digital intraoral bitewing radiography when indicated. Dental caries were assessed using con…

The gain

In a 210-patient study at a Pakistani dental college, an AI caries detector had 91.5% sensitivity, versus 82.2% for dentists' clinical examination and 87.3% for their radiograph reading.

In brief

Dentists' visual-tactile examination reached 84.3% accuracy (82.2% sensitivity) and their reading of the radiographs 88.1% (87.3% sensitivity). The AI's area under the curve was 0.934, versus 0.846 for clinical examination and 0.882 for radiographic interpretation.

The gap narrows against combined conventional assessment, which scored 89.3% accuracy and an AUC of 0.908. The AI was weakest on early enamel lesions, where sensitivity was 78.4%, and the authors say it should support dentists rather than replace them until tested in larger multicentre groups.

Main points

  1. The study enrolled 210 consecutive adults (126 men, 84 women, mean age 34.8) attending routine dental assessment; 118 (56.2%) had caries by the reference standard.
  2. The AI system scored 91.5% sensitivity, 90.2% specificity, 91.5% PPV, 90.2% NPV and 90.9% accuracy.
  3. Clinical examination scored 82.2% sensitivity, 86.9% specificity and 84.3% accuracy; radiographic interpretation scored 87.3%, 89.1% and 88.1%.
  4. Combined conventional assessment reached 89.0% sensitivity, 89.6% specificity and 89.3% accuracy, close to the AI result.
  5. AUC was 0.934 (95% CI 0.901–0.967) for AI, 0.908 for combined conventional assessment, 0.882 for radiographs and 0.846 for clinical examination.
  6. AI sensitivity rose with lesion depth: 78.4% for early enamel, 89.7% for established enamel/proximal, 93.5% for early dentinal and 97.1% for advanced dentinal lesions.
  7. Specificity stayed between 90.6% and 92.4% across lesion categories.

The full story

Each participant had a standardized mirror-and-probe examination under adequate illumination, followed by digital intraoral bitewing radiography when indicated. Experienced dentists read the radiographs while blinded to the AI output, and the same images were run through the deep-learning system, which labelled each surface caries-positive or caries-negative.

The reference standard was a consensus diagnosis by two experienced dental clinicians, with a senior oral and maxillofacial radiologist adjudicating disagreements. Patients with extensive restorations obscuring surfaces, severe developmental abnormalities, non-diagnostic radiographs or incomplete records were excluded.

Clinical examination missed the most cases. The AI's sensitivity was 91.5% against 82.2% for clinical examination and 87.3% for radiographic reading. Its advantage over the combined conventional approach was smaller, at 90.9% versus 89.3% accuracy.

Performance depended on lesion depth. For early enamel lesions the AI had 78.4% sensitivity and 86.7% accuracy; for advanced dentinal lesions it reached 97.1% sensitivity and 94.8% accuracy. The authors say subtle early demineralization remains a diagnostic challenge.

Sources

1 on file
  1. Peer-reviewed2026-11-17

Talk it over

Three questions to start on, two more if you have time
The AI caught 91.5% of cases against 82.2% for clinical examination, but only 78.4% of early enamel lesions. How would you weigh its overall edge against that weaker spot?

Start with a question

A study of 210 patients found AI detected more tooth decay than dentists' examshttps://truvace.com/item/diagnostic-accuracy-of-artificial-intelligencebased-systems-for-detection-of-den

  1. The AI caught 91.5% of cases against 82.2% for clinical examination, but only 78.4% of early enamel lesions. How would you weigh its overall edge against that weaker spot?

    Answer in the thread
  2. The 210 adults all came through routine assessment at Shahida Islam Dental College in Lodhran, Pakistan. What might change in the results at a clinic with different patients, equipment or dentists?

    Answer in the thread
  3. The AI scored 90.9% accuracy and combined conventional assessment scored 89.3%. What would you want to see before deciding the AI should be used alongside dentists in a clinic?

    Answer in the thread
If you have more time
  • How does a dental practice you know decide when to take bitewing radiographs, and how might an AI reader change who gets one?

    Answer in the thread
  • The reference diagnosis came from two clinicians, with a radiologist settling disagreements. What other kind of evidence could show whether the AI's extra detections helped patients over time?

    Answer in the thread

What readers say