AI-Assisted Classification of Mandibular First Molars According to the Presence of an Additional Distal Canal on CBCT Images: Automated CBCT Mandibular Canal Detection.
Authors
Affiliations (4)
Affiliations (4)
- Faculty of Dentistry, Department of Endodontics, Cankiri Karatekin University, Cankiri, Turkey. [email protected].
- Faculty of Dentistry, Department of Maxillofacial Radiology, Cankiri Karatekin University, Cankiri, Turkey.
- Faculty of Engineering, Electrical-Electronics Engineering, Cankiri Karatekin University, Cankiri, Turkey.
- Faculty of Engineering, Computer Engineering, Cankiri Karatekin University, Cankiri, Turkey.
Abstract
A hybrid fusion architecture based on deep convolutional neural networks (CNNs) was implemented for AI-assisted classification of mandibular first molars (MFMs) according to the presence or absence of an additional distal canal on cone-beam computed tomography (CBCT) images. The hybrid deep fusion architecture was trained using an augmented dataset derived from 253 CBCT scans, including 175 three-canalled and 78 four-canalled MFMs. The original CBCT scans were first divided at the scan/tooth level into training, validation, and test subsets, and data augmentation was subsequently applied separately within each subset to prevent cross-subset leakage of augmented derivatives. The reference standard was established through consensus between an independently calibrated dentomaxillofacial radiologist and an endodontist. Primary diagnostic performance was evaluated on the 38 original, independent, non-augmented test cases at the patient/tooth level. When the presence of an additional distal canal/four-canal morphology was defined as the clinically positive condition, the proposed Fusion IncepDenseCBAM model achieved an accuracy of 97.4% (95% CI: 92.3-100.0%), sensitivity of 91.7% (95% CI: 76.0-100.0%), specificity of 100.0% (95% CI: 100.0-100.0%), positive predictive value of 100.0% (95% CI: 100.0-100.0%), negative predictive value of 96.3% (95% CI: 89.2-100.0%), F1-score of 95.7% (95% CI: 83.5-100.0%), and AUC of 0.958 (95% CI: 0.880-1.000). The model also demonstrated high agreement and balanced classification performance, with an MCC of 0.940 (95% CI: 0.812-1.000) and Cohen's kappa of 0.938 (95% CI: 0.817-1.000). These findings indicate that the proposed hybrid fusion architecture shows promising preliminary diagnostic performance for detecting an additional distal canal on CBCT images. However, further external validation using larger independent datasets is required before clinical implementation as a decision-support tool in endodontic diagnostic workflows.