Assessment of the Impact of Missing Teeth on Deep Learning-Based Tooth Numbering Method From Panoramic X-Rays

Automatically analyzing panoramic dental X-rays through AI-based approaches has attracted attention and concern among dentists with the recent advancement of AI techniques. Recent studies have shown that these approaches have achieved high accuracy in recognizing and numbering teeth. However, the datasets utilized in such studies contain only a small proportion of panoramic X-rays that include missing teeth, despite such X-rays often being observed in actual diagnosis. Accordingly, the suitability of these approaches for practical panoramic X-rays remains unconfirmed. This study verifies the efficacy of three powerful and efficient object detection models for tooth numbering on panoramic X-rays with varying numbers of missing teeth. The experimental results indicate that the performance of AI models is affected by the existing tooth count appearing in panoramic X-ray dataset images, and images with a lower tooth count tend to exhibit a lower tooth-numbering accuracy that reaches a precision of 62.2%. Additionally, our study highlights the features in low tooth-count images that influence this accuracy. Our results thus also verify the unsuitability of AI models in numbering on panoramic X-ray images with a low tooth count. Importantly, this study suggests that disclosing the tooth count distribution of datasets is indispensable for proving the effectiveness of AI-based approaches in the tooth numbering task.