Deep-learning-based mandibular third molar recognition and classification system

Background: Traditional impacted-tooth classification methods are limited, as they rely heavily on clinicians' subjective judgment and often exhibit poor consistency. To enhance diagnostic accuracy, provide clinical utility, and develop a scalable, intelligent decision-support system for oral and maxillofacial surgery, this study proposes an objective and quantitative approach that integrates deep learning-based image segmentation with geometric feature quantification for classifying mandibular third molar impaction types. Methods: A decoupled segmentation-classification workflow was implemented, utilizing a fusion model that integrates U-Net, U-Net++, and DeepLabV3+ for tooth contour segmentation. Results: The fusion model achieved superior performance, with an accuracy of 0.937 and sensitivity of 0.889 for third molar segmentation. Classification based on the α-angle, defined as the angle between the long axis of the third molar and the reference line of adjacent teeth, reached an overall accuracy of 96.2%. Distoangular impaction demonstrated the highest specificity (0.983), highlighting the reliability of the system. Conclusions: The combination of multimodel segmentation techniques and a newly defined angle criterion significantly improves classification accuracy, offering an efficient and quantifiable tool for clinical practice. This approach reduces diagnostic subjectivity, leverages the advantages of panoramic radiography, and demonstrates strong potential for advancing intelligent dentistry. Future studies will focus on model optimization to accommodate diverse imaging data and further advance clinical applications in oral and maxillofacial surgery.