Students versus artificial intelligence in assessing dental plaque remotely : who outperforms?

Abstract

Aims: Our community health project aimed to assess and compare the accuracy of dental plaque assessments by dental students versus an Artificial Intelligence (AI) model. Methods: Healthy volunteers were recruited, and digital images of their anterior teeth were obtained following plaque disclosure. Cropped images of individual teeth were created in Photoshop®. Plaque assessments from the individually cropped images were conducted employing three standardized measures: 1) O’Leary (1972), 2) Greene-Vermillion (1964), and 3) Turesky Modified Quigley Hein (TMQH) (1970). Students were trained and calibrated in the use of the three indices. With assistance from a computer scientist and dentist, a deep learning model (AI) was developed to assess plaque according to the criteria of the three plaque indices. The accuracy of students' assessments of plaque was determined with respect to the gold standard (group facilitator’s assessments). Likewise, the accuracy of the AI models' assessment of plaque was determined with respect to the gold standard. Students’ accuracy of plaque assessments were compared with that of the accuracy of the AI models. Results: Of the 546 cropped images obtained, 97.8% (n=534) were amenable for plaque assessment analysis. Students' inter- and intra-examiner reliability across the three plaque indices was good/excellent (Kappa ≥0.70). For the O’Leary index, the accuracy of students’ was high (Team A 0.99, Team B 0.97), and for the AI model accuracy was 0.94. For the GV index, the accuracy of students’ assessment ranged from 0.94-0.99 and for the AI model it ranged from 0.89-0.99. For the TMQH index, the accuracy of students’ assessment ranged from 0.93 -0.99 and for the AI model, it ranged from 0.82- 0.98. Conclusions: Dental students’ accuracy in assessing plaque levels remotely from photographic images was consistently high, irrespective of the plaque assessment index criteria. The artificial intelligence (AI) models (deep learning) accuracy in assessing plaque levels was consistently high and comparable to dental students’ assessments. The findings of our community health project have implications for using AI in assessing dental plaque remotely.