Enhancing Dental Radiographic Interpretation by Collaborating with AI Systems to Minimise Interpretive Errors

Abstract

Interpretive errors in dental radiology pose risks to patient care, often resulting from limitations in human capabilities, including visual detection, pattern recognition and clinical reasoning. Despite the critical role of radiographic assessments, there is a lack of evidence on the prevalence and cause of dental radiographic interpretation, their consequences and the solutions to mitigate them. This PhD thesis addressed this gap by investigating the factors contributing to interpretive errors and evaluating the effectiveness of machine learning (ML) algorithms as cognitive aids in improving diagnostic accuracy.

The research involved a systematic review, surveys of Australian dental practitioners and students, and a comparative study assessing cognitive aids (ML algorithms and checklists) for diagnosing caries on bitewing radiographs. Errors of omission were most frequent, leading to undertreatment (72%), increased costs (62%), legal issues (82%) and reputational damage (75.6%).

ML algorithms significantly enhanced diagnostic performance, achieving a higher sensitivity (79%) and diagnostic odds ratio (20.3) than the other methods. Participants in the ML group also reported greater confidence in diagnosis. However, concerns regarding accuracy, trust and job displacement remain barriers to AI adoption in dentistry. Beyond caries detection, the potential applications of AI span dentomaxillofacial radiology, implantology and prosthodontics. Additionally, this thesis developed a novel explainability method, enabling clinicians and computer scientists to interpret ML-generated outputs better.

In conclusion, ML algorithms serve as valuable assistive cognitive aids, reducing interpretive errors and enhancing clinician confidence in dental radiology. The findings support AI integration as a means to improve diagnostic accuracy and clinical decision making in dentistry.