Artificial intelligence (AI) has been successfully employed in dentistry from managing workflow in the clinic, e.g., booking and coordinating appointments, to assisting with clinical diagnosis and treatment planning [1,2]. Its efficacy in clinical diagnosis primarily stems from its performance in analyzing diverse types of dental imagery, including photographs, radiographs, transillumination images, and 3-D computed tomography scans. AI methods have been effectively employed for a wide range of tasks on dental imagery, such as tooth classification and segmentation, cephalometric landmark detection, caries identification or predicting the risk of dental complications following third molar extraction [3,4]. Integrating AI into clinical practices may increase the chances of early detection of pathologies and appropriate treatments, resulting in improved patient outcomes and reduced dental care expenditures [5].

While accuracy gains by using AI have been demonstrated by a wealth of studies [4], further aspects like the impact on care processes or costs have been investigated to a much lesser degree [6,7]. One question of interest, for example, how AI leads impacts the diagnostic process, for example by directing attention towards relevant areas of interest (AOI) and reducing attention towards irrelevant features and areas devoid of pertinent information, or vice versa. A deeper understanding in this direction could help to improve AI systems for diagnostic support, but also to safeguard users from diagnostic biases when using AI or addressing these biases during dental education. To assess the effects AI has on the diagnostic process, mere diagnostic accuracy studies comparing AI against dentists are not sufficient: Instead, prospective clinical studies are needed.

In a recent randomized controlled trial, we compared AI-assisted detection of proximal caries on bitewing radiographs with that of non-assisted dentists and demonstrated a significant increase in sensitivity when using AI [6]. In parallel to recording accuracy estimates, we also employed eye tracking to precisely determine where dentists focus on during image analysis and to record their eye movements while detecting caries [8]. Previous work using eye tracking has shown that dentists employ a task-dependent gaze known as scanpath which comprises of ‘fixations’ (attentional information) and ‘saccades’ (transitions to attentional areas) [9,10]. Moreover, different types of radiographs are assessed differently. For instance, when examining panoramic radiographs, a holistic representation of the content is formed at a glance [11] and then a systematic spiraling scan pattern [12,13] or a circular scan pattern has been observed [14], [15], [16]. For intraoral periapical radiographs, dentists commonly adopt a tooth-by-tooth viewing approach [17]. Using data from the control group of the randomized trial (i.e., dentists not assisted by AI), we showed that the dentists employed a heightened focus on certain image areas, with respect to their task [8]. Also, they generally examined the entire image in a systematic tooth-by-tooth pattern for caries detection [8].

The present study aimed to compare gaze pattern and scanpaths of dentists detecting caries on bitewing radiographs when they are assisted or not assisted by an AI software in the aforementioned randomized controlled trial. Our results may offer valuable insights into the interaction between dentists and AI, and how it may impact their diagnostic performance in a real-world setting. We did not aim to compare the vectorial differences in the gaze patterns. We hypothesized that dentists using AI would exhibit shorter viewing durations, heightened focus on relevant AOI, especially inconspicuous carious lesions, and, in general, more efficient gaze patterns compared to those without AI. The assumption underlying this hypothesis was that the dentists would incorporate the findings of the AI software into their caries detection strategies thus resulting in more efficient gaze patterns.