Performance of a deep learning system for simultaneously diagnosing radiolucent and radiopaque lesions in the anterior maxilla on panoramic radiographs.

Aim: To validate the performance of a deep learning system with detection and classification functions for a mix of radiolucent and radiopaque lesions in the anterior maxilla on panoramic radiographs. Methods: Patients with radiolucent or radiopaque lesions in the anterior maxilla on panoramic radiographs were selected retroactively from May 2022 until Feb 2002 to obtain 100 preoperative radiographs each for nasopalatine duct cysts (NDCs), radicular cysts (RCs), and impacted supernumerary teeth (ISTs). An additional 100 patients with no lesions in the anterior maxilla were selected. Two deep learning systems (Systems 1 and 2) were created and tested. For System 1, the models were created and tested using datasets of radiolucent lesions (NDCs and RCs) and No lesions. For developing System 2, the data of radiopaque lesions (ISTs) were added to those used in System 1. The neural network used was You Only Look Once ver. 7 (YOLOv7). The recall, precision, F1 score, and accuracy calculated from the confusion matrix were used to evaluate diagnostic performance. Results: The performance of System 2, which included the IST data, was worse than that of System 1. Even when NDCs and RCs were addressed as a joint category of radiolucent lesions, the addition of IST data resulted in a worse performance than that of System 1. Conclusion: Our results indicate that combined use of radiopaque lesions (ISTs) with radiolucent lesions (NDCs and RCs) reduces the deep learning performance for radiolucent lesions with the volume of data used in the present study. [ABSTRACT FROM AUTHOR]