Abstract:

Objective To develop an end-to-end lightweight deep learning network to automatically predict impacted maxillary anterior teeth in young children using panoramic X-rays. Methods Panoramic X-rays of 216 patients aged 5-12 years were collected, and 6 anterior teeth of each sample were selected, with a total of 1 296 tooth samples as the research objects. By constructing a lightweight deep network model(FA-UNet) based on the fusion attention mechanism, taking panoramic X-rays as input and the clinician’s diagnosis as the gold standard, the potential mapping relationship between image features and clinical diagnosis results was learned to achieve automatic prediction and segmentation of impacted teeth in the maxillary anterior teeth of young children. The performance was evaluated using both pixel-level metrics (accuracy, sensitivity, specificity, PPV, mIoU, F1-score) and tooth-level classification accuracy. Results Taking manual diagnosis as the standard, FA-UNet model demonstrated the following performance metrics for pixel-level prediction of maxillary anterior impacted teeth in pediatric patients: accuracy of 0.866, sensitivity of 0.551, specificity of 0.995, PPV of 0.977, mIoU of 0.544, and F1-score of 0.705. Additionally, it attained a tooth-level classification accuracy of 87.5%. Conclusion FA-UNet model effectively improves the prediction accuracy of impacted teeth in the maxillary anterior teeth of children.

Key words: impacted tooth prediction, maxillary anterior teeth, panoramic X-ray, fusion attention mechanism, Mamba