A Transfer Learning Approach for Detecting Buds of Posterior Teeth

Artificial Intelligence-based medical image analysis represents a critical step towards more sustainable and prevention-oriented diagnosis and healthcare. Early and accurate identification of dental anomalies in pediatric patients is an underexplored field of study, despite dental pathologies being one of the most costly medical conditions in the world. This study proposes a transfer learning approach for automatic detection of first and second premolar tooth buds on pediatric panoramic radiographs, through a YOLOv8 architecture, optimized on a labeled dataset of 619 images. Evaluation metrics such as mean Average Precision (mAP@50 = 0.984), Dice coefficient (0.847), and Structural Similarity Index Metric (SSIM = 0.947 demonstrate the effectiveness of the proposed method. Howeve the model performance decreases with more stringent localizatio criteria (mAP@50-95 = 0.610) and is sensitive to image quality with degraded radiographs negatively impacting the detectio accuracy. These limitations highlight the importance of high quality image acquisition and suggest the need for further studie to improve generalization across different clinical acquisitio conditions. Despite these challenges, the results suggest that th model could represent a first step in an automated early diagnosi and treatment planning pipeline.