Deep Learning-Assisted Localization of Cystic Lesions and Benign Tumors in the Maxillofacial Region Using Panoramic Radiographs: A Preliminary Feasibility Study

Background/Objectives: Automated localization of cystic lesions and benign tumors on panoramic radiographs may support lesion recognition in the maxillofacial region. This preliminary feasibility study aimed to develop and evaluate a deep learning model based on Mask R-CNN for the localization of dentigerous cysts (DCs), radicular cysts (RCs), odontogenic keratocysts (OKCs), and ameloblastomas using panoramic radiographs. Methods: A total of 215 panoramic radiographs were retrospectively collected from Taichung Veterans General Hospital (2018–2023). After excluding postoperative, recurrent, or low-quality images, 184 lesions were allocated to the training set and 47 lesions to the testing set. Lesions were annotated based on pathology-confirmed diagnoses. The Mask R-CNN model was trained to localize and classify four lesion types. Model performance was evaluated using precision, sensitivity (recall), and F1 score at an Intersection over Union (IoU) threshold of 0.1. Results: In the testing set (n = 47), 26 lesions were correctly localized, yielding an overall sensitivity of 55.3% and a precision of 83.9%. The corresponding F1 score was 66.7%. Lesion-specific sensitivities were 40.0% for ameloblastomas, 37.5% for OKCs, 36.8% for RCs, and 93.3% for DCs. Conclusions: This study suggests the preliminary feasibility of a deep learning-assisted approach for lesion localization on panoramic radiographs. However, the absence of lesion-free control images and the limited dataset size restrict the generalizability and clinical applicability of the findings. Further validation using larger and more balanced datasets is required.