Evaluating the Effect of Annotation Detail in AI-based Caries Detection

Sammanfattning

Caries is the most common disease worldwide, with more than a third of the adult population around the world suffering from untreated caries. Early detection is crucial for optimal treatment, and in recent years, artificial intelligence has played an important role in caries detection within the field of dentistry. This study investigates how the level of annotation detail impacts the performance of a pre-trained convolutional neural network model in caries detection. Roboflow was used to train several RF-DETR models on a subset of the Dental X-Ray Panoramic Dataset with augmented annotations. The results suggest that less exact but more consistent annotations may be more effective when training caries detection models, potentially saving valuable time and reducing the resources necessary for preparing training data.