Semi-Supervised Tooth Instance Segmentation from Panoramic X-ray Images

Abstract

Tooth image segmentation is a critical task in medical image analysis, where precise segmentation outcomes are essential for oral diagnosis and treatment planning. Dental X-ray images often present challenges such as boundary blurring, tooth misalignment, and complex anatomical variations, which hinder accurate tooth edge delineation, especially for wisdom tooth assessment. This study integrates semi-supervised learning with the nnU-Net deep learning framework and proposes a novel tooth image segmentation method that incorporates a dual-attention mechanism. By utilizing a limited set of labeled data alongside abundant unlabeled samples, the proposed approach effectively handles complex backgrounds and diverse tooth morphologies. Experimental evaluations show superior performance, with an average Dice coefficient of 92.05%, mean Intersection over Union (mIoU) of 86.27%, Normalized Surface Distance (NSD) of 94.72%, and Identification Accuracy (IA) of 85.48%. This method enhances diagnostic precision for oral clinicians, streamlines dental image analysis workflows, and holds significant potential for advancing digital dentistry in diagnosis and treatment.