Automatic Detection Of Osteonecrosis Of The Jaws On Panoramic Radiographs

This study aimed to propose an automatic detection method for osteonecrotic lesions (ON) on panoramic radiographs using a deep learning model. We conducted two experiments using a total of 300 panoramic radiographs. Experiment 1 developed a deep learning-based EfficientDet model for the automatic detection of ON, including osteomyelitis (OM) and medication-related osteonecrosis of the jaw (MRONJ). Experiment 2 aimed to classify OM and MRONJ using the same model as in Experiment 1. Both experiments included 20 normal images in the test set to evaluate the lesion detection ability of the model. The labeling process was performed by an oral radiologist with 20 years of experience, who identified all common radiological features of OM and MRONJ. The performance of the model was analyzed using accuracy, sensitivity, and specificity. In Experiment 1, which focused on detecting ON, the model had an accuracy of 75%, a sensitivity of 68.33%, and a specificity of 95%. In Experiment 2, the model achieved an accuracy of 55% in classifying OM and MRONJ. The proposed model demonstrated its potential to detect ON on panoramic radiographs. However, the accuracy of Experiment 2 in classifying OM and MRONJ was not high. These errors are understandable, as OM and MRONJ are typically based on a history of antiresorptive or antiangiogenic medication. The developed model can assist dentists with diagnosing ON and provide a second opinion. The ability to differentiate OM from MRONJ may be improved by collecting more detailed patient histories.