Application of a deep learning method based on CBCT panoramic reconstruction images for differentiating operationally defined cystic and neoplastic jaw lesions from osteomyelitis: a single-center retrospective study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Oral and Maxillofacial Surgery II, The First Affiliated Hospital of Zhengzhou University, No. 1 Jianshe East Road, Erqi District, Zhengzhou, 450052, Henan, China.
- Henan University of Science and Technology, No. 263 Kaiyuan Avenue, Luolong District, Luoyang, 471023, Henan, China.
- Zhengzhou University, No. 100 Science Avenue, Zhongyuan District, Zhengzhou, 450001, Henan, China.
- The First Affiliated Hospital of Zhengzhou University, No. 1 Jianshe East Road, Erqi District, Zhengzhou, 450052, Henan, China.
- Department of Oral and Maxillofacial Surgery II, The First Affiliated Hospital of Zhengzhou University, No. 1 Jianshe East Road, Erqi District, Zhengzhou, 450052, Henan, China. [email protected].
Abstract
This single-center retrospective study developed and internally validated a two-dimensional deep learning model based on cone-beam computed tomography (CBCT) panoramic reconstruction images for the preliminary differentiation of operationally defined cystic and neoplastic jaw lesions from osteomyelitis of the jaw. A total of 400 patients were included, comprising 200 patients with jaw mass lesions and 200 patients with osteomyelitis. The jaw mass lesion cohort included cystic, benign neoplastic, and malignant neoplastic lesions and was analyzed as a single positive class for preliminary screening rather than lesion subtype classification. In this study, the term "jaw mass lesions" was used as an operational imaging-based umbrella term for jaw lesions presenting as clinically relevant expansile, destructive, or mass-like osseous abnormalities on CBCT panoramic reconstruction images. This term does not imply that all included lesions were tumors or soft-tissue masses; in particular, odontogenic keratocyst was classified as a cystic lesion. Images were randomly divided at the patient level into training, validation, and independent test sets in a 70:10:20 ratio. ResNet-50 was used as the baseline model and was compared with EfficientNet-B4, Swin Transformer, and ConvNeXt-Tiny under the same evaluation protocol. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), accuracy, sensitivity, specificity, precision, and F1 score. Gradient-weighted Class Activation Mapping (Grad-CAM) was used for post hoc visualization, and an exploratory reader study was conducted to preliminarily assess the potential influence of artificial intelligence (AI) assistance on image interpretation. In the independent test set, ResNet-50 achieved an AUC of 0.9781 and an AUPRC of 0.9714. The proposed model is intended as a preliminary screening and triage-support tool rather than a replacement for contrast-enhanced MDCT, magnetic resonance imaging (MRI), histopathological examination, or multidisciplinary assessment.