# Deep learning-based automated detection of supernumerary teeth in pediatric panoramic radiographs

## TL;DR

This study evaluates a deep learning model for detecting extra teeth in children's dental X-rays, showing strong classification accuracy and potential for clinical use.

## Contribution

A novel YOLOv8-based deep learning pipeline for automated detection and classification of supernumerary teeth in pediatric radiographs.

## Key findings

- The classification model achieved 100% accuracy, precision, recall, and F1-score on both internal and external datasets.
- The segmentation model showed 100% precision but only 38% recall, indicating strong localization but limited sensitivity.
- Model predictions were not significantly different from expert decisions (McNemar’s test, p > 0.05).

## Abstract

Supernumerary teeth are a common developmental anomaly in pediatric patients, potentially leading to complications such as impaction, crowding, and delayed eruption. Accurate and early detection is critical to prevent these sequelae and guide appropriate intervention strategies. This study aims to evaluate the diagnostic accuracy and clinical applicability of a convolutional neural networks-based deep learning model (YOLOv8) for the automated localization and binary classification of supernumerary teeth on pediatric panoramic radiographs.

A retrospective analysis was conducted on 2000 pediatric panoramic radiographs following ethical approval. Three calibrated pediatric dentists independently examined the dataset and annotated a representative subset of 140 radiographs (71 positive, 69 negative), achieving substantial inter-rater agreement (Cohen’s κ = 0.92). Performance was assessed in two stages: (1) segmentation of supernumerary teeth and (2) binary classification of radiographs. An independent validation set of 20 radiographs was used for secondary evaluation. Evaluation metrics included precision, recall, F1-score, and McNemar’s test to compare model predictions with expert labelling.

The mean age of the patients was 9.6 ± 2.3 years; 52% were male, 48% were female. The segmentation model yielded 100% precision, 38% recall, and an F1-score of 55%, indicating strong localization when detections were made but limited sensitivity. The classification model achieved 100% accuracy, precision, recall, and F1-score on both internal and external datasets. McNemar’s test revealed no statistically significant discrepancy between the model and expert decisions (p > 0.05). The segmentation model demonstrated high precision in localizing supernumerary teeth; however, recall performance was more modest, indicating occasional under-detection. Due to the limited validation sample size, these findings should be interpreted with caution.

The YOLOv8-based pipeline demonstrated robust diagnostic accuracy in classifying panoramic radiographs for supernumerary teeth and promising but preliminary results in lesion-level segmentation. These findings highlight the potential utility of advanced deep learning systems in augmenting early diagnosis and streamlining pediatric dental radiology workflows.