Enhancing wisdom teeth detection in panoramic radiographs using multi-channel convolutional neural network with clinical knowledge

Abstract

This study presents a novel artificial intelligence approach for detecting wisdom teeth in panoramic radiographs using a multi-channel convolutional neural network (CNN). First, a curated dataset of annotated panoramic dental images was collected, with bounding box annotations provided by a senior oral and maxillofacial surgeon. Each image was then preprocessed and split into three input channels-full, left-side, and right-side views-to replicate the diagnostic workflow of dental professionals. These channels were simultaneously fed into a classification-based CNN model designed to predict the presence or absence of wisdom teeth in each of the four quadrants. Unlike traditional segmentation or object detection approaches, our method avoids pixel-level labeling and offers a simpler, faster pipeline with reduced annotation overhead. The proposed model achieved an accuracy of 82.46 %, with an AUROC of 0.8866 and an AUPRC of 0.8542, demonstrating reliable detection performance across diverse image conditions. This system supports consistent and objective diagnosis, particularly benefiting less experienced practitioners and enabling efficient screening in clinical settings.

Keywords: Convolution neural network; Deep learning; Image augmentation; Panoramic radiographs; Wisdom teeth detection.