YoLeTooth: A Unified Framework for Joint Tooth Segmentation and Periapical Lesion Detection in Panoramic Radiographs

Background and Objective: Periodontitis is a chronic inflammatory disease of teeth-supporting tissues, and one of the major causes of tooth loss worldwide, affecting about 3.5 billion individuals globally. The conventional diagnosis, based on subjective manual clinical inspection and radiographic interpretation, is time-consuming and liable to possible mistakes. This paper presents a unified framework for joint tooth segmentation and automated periapical lesion detection in panoramic radiographs that simultaneously identifies individual teeth and detects any associated lesions with fast processing speeds suitable for clinical use.Methods: Our approach employs a joint process: first, a deep learning model identifies and segments individual teeth according to standard dental numbering systems, while a second one detects periapical lesions within the tooth regions obtained from the segmentation outputs in the first stage. The framework incorporates an advanced loss function (Powerful IoU v2) to improve bounding box regression accuracy and a spatial association mechanism to map detected lesions to specific teeth based on geometric overlap analysis. A variety of experiments have been conducted, optimizing various aspects including loss functions, model architectures, learning rates, optimizers, and training procedure.Results: Our proposed tooth segmentation model achieved a mean Average Precision of 97.7% at the standard 0.5 threshold and 79.0% when averaged across thresholds from 0.5-0.95, with a mean Dice coefficient of 93.5% and a mean Intersection over Union of 87.8%, while the periapical lesion detection model achieved a mean Average Precision of 91.9% at 0.5 threshold and 77.9% at thresholds from 0.5-0.95. The region-of-interest-based approach provides a 3.49x computational speedup compared to full-image processing, with an average inference time of only 0.159 seconds per radiograph. The explicit tooth-to-lesion mapping enables precise association of detected lesions to individual teeth based on spatial overlap.Conclusions: To the best of our knowledge, this is the first work that achieves lesion classification with explicit tooth mapping in a unified framework using exclusively open-source datasets, making it reproducible and suitable for clinical deployment in resource-constrained environments.