Back to all papers

[Diagnostic performance of a dual-stage deep learning model based on cone beam CT-derived multi-view two-dimensional slices for assessing ankylosed impacted teeth].

July 30, 2026pubmed logopapers

Authors

Wu GL,Lu Z,Du TJ,Qiao H,Zhang YYR,Ba H,Wang S

Affiliations (1)

  • Key Laboratory of Shaanxi Province for Craniofacial Precision Medicine Research, College of Stomatology, Xi'an Jiaotong University, Xi'an 710004, China.

Abstract

<b>Objective:</b> To develop and validate a dual-layer deep learning model based on cone-beam CT (CBCT)-derived multi-view two-dimensional (2D) slices for the assisted assessment of ankylosed impacted teeth. <b>Methods:</b> The CBCT images of 432 patients with impacted teeth were collected at the Hospital of Stomatology, Xi'an Jiaotong University from January 2023 to June 2025, comprising 600 target teeth and approximately 22 000 derived multi-view 2D slices. Imaging reference standards were established by two experienced oral and maxillofacial residents, with adjudication by a professor in cases of disagreement. Data were split into training and validation sets at a 4∶1 ratio by target tooth. A two-stage deep learning model was developed, including a slice-level classifier for predicting dentoalveolar ankylosis and a lesion-level detector for localizing suspicious ankylotic regions. Model performance was evaluated using area under the receiver operating characteristic curve (ROC-AUC), (PR-AUC), sensitivity, specificity, Brier score, average precision (AP) at intersection over union (IoU) 0.5, and ablation analysis. <b>Results:</b> In the validation set, the slice-level classification model achieved an area under the receiver operating characteristic curve of 0.850, an area under the precision-recall curve of 0.439, and a Brier score of 0.071. At a threshold of 0.5, the sensitivity and specificity were 36.8% and 96.4%, respectively. The lesion-level detection model achieved an average precision of 0.582 at an intersection-over-union threshold of 0.5. Ablation experiments showed that eliminating hard example mining significantly decreased positive recall and overall model performance with an increase in the area under the precision-recall curve from 0.439 to 0.373. <b>Conclusions:</b> The proposed model can provide probabilistic cues and lesion localization evidence for ankylosed impacted teeth, showing potential as an auxiliary tool for radiographic interpretation.

Topics

English AbstractJournal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAI Slice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.