Scholarlink

Objectives <p>To evaluate the diagnostic accuracy of the Diagnocat™ artificial intelligence (AI) system for caries detection on bitewing radiographs compared with expert human examiners, with emphasis on differences between enamel and dentin lesions.</p> Materials and methods <p>A sample of 100 digital bitewing radiographs (1540 surfaces) was retrospectively selected from the European University Cyprus dental clinic database using a systematic backward screening method. Radiographs were obtained with a standardized phosphor plate system and anonymized before analysis. Two independent experts (operative dentistry and oral radiology) established the reference standard. AI and human assessments were binarized (caries/no caries; enamel/dentin) and compared. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated, and statistical significance was tested across detection categories.</p> Results <p>Diagnocat™ showed high specificity (94.3%, 95% CI: 92.4%–96.0%) and NPV (96.1%, 95% CI: 94.7%–97.3%), with an overall accuracy of 91.6% (95% CI: 89.7%–93.4%). Sensitivity was moderate (73.1%, 95% CI: 65.9%–79.9%), and PPV was 64.7% (95% CI: 57.7%–71.5%). Agreement with the expert consensus was substantial (Cohen’s κ = 0.638). For enamel lesions, sensitivity and specificity were 73.3% (95% CI: 62.8%–82.7%) and 92.9% (95% CI: 91.0%–94.7%) with moderate agreement with the consensus (Cohen’s κ = 0.492) and for dentin lesions they were 72.8% (95% CI: 61.8%–83.8%) and 92.8% (95% CI 90.9%–94.6%) with moderate agreement with the consensus (Cohen’s κ = 0.468). NPV remained high (≥ 98.0%), while PPV was low (42.0% and 39.2%), across lesion types. Detection patterns differed significantly between AI and the reference standard (<i>p</i> &lt; 0.001).</p> Conclusions <p>Diagnocat™ demonstrated good diagnostic performance in ruling out caries. However, its overall lower sensitivity emphasizes the need for clinician oversight, especially in detecting early-stage disease.</p> Clinical relevance <p>This study offers an independent validation of Diagnocat™ using bitewing radiographs. It demonstrates lesion-depth–specific insights, showing that while AI is highly reliable for excluding disease, its predictive value remains limited.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating AI diagnostic accuracy in approximal dental caries detection on bitewing radiographs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月29日

查看全文
  • Kostis Giannakopoulos,
  • Argyro Kavadella,
  • Dimitrios Paraskevis,
  • Aristidis Arhakis,
  • Miltiadis A. Makrygiannakis,
  • Eleftherios G. Kaklamanos

摘要

Objectives

To evaluate the diagnostic accuracy of the Diagnocat™ artificial intelligence (AI) system for caries detection on bitewing radiographs compared with expert human examiners, with emphasis on differences between enamel and dentin lesions.

Materials and methods

A sample of 100 digital bitewing radiographs (1540 surfaces) was retrospectively selected from the European University Cyprus dental clinic database using a systematic backward screening method. Radiographs were obtained with a standardized phosphor plate system and anonymized before analysis. Two independent experts (operative dentistry and oral radiology) established the reference standard. AI and human assessments were binarized (caries/no caries; enamel/dentin) and compared. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated, and statistical significance was tested across detection categories.

Results

Diagnocat™ showed high specificity (94.3%, 95% CI: 92.4%–96.0%) and NPV (96.1%, 95% CI: 94.7%–97.3%), with an overall accuracy of 91.6% (95% CI: 89.7%–93.4%). Sensitivity was moderate (73.1%, 95% CI: 65.9%–79.9%), and PPV was 64.7% (95% CI: 57.7%–71.5%). Agreement with the expert consensus was substantial (Cohen’s κ = 0.638). For enamel lesions, sensitivity and specificity were 73.3% (95% CI: 62.8%–82.7%) and 92.9% (95% CI: 91.0%–94.7%) with moderate agreement with the consensus (Cohen’s κ = 0.492) and for dentin lesions they were 72.8% (95% CI: 61.8%–83.8%) and 92.8% (95% CI 90.9%–94.6%) with moderate agreement with the consensus (Cohen’s κ = 0.468). NPV remained high (≥ 98.0%), while PPV was low (42.0% and 39.2%), across lesion types. Detection patterns differed significantly between AI and the reference standard (p < 0.001).

Conclusions

Diagnocat™ demonstrated good diagnostic performance in ruling out caries. However, its overall lower sensitivity emphasizes the need for clinician oversight, especially in detecting early-stage disease.

Clinical relevance

This study offers an independent validation of Diagnocat™ using bitewing radiographs. It demonstrates lesion-depth–specific insights, showing that while AI is highly reliable for excluding disease, its predictive value remains limited.