Accuracy of AI in Detecting Bifid Mandibular Canal on CBCT: A Diagnostic Accuracy Study

Overview

The goal of this observational study is to evaluate how accurately a deep learning-based artificial intelligence (AI) model can detect and segment bifid mandibular canals (BMCs) on cone-beam computed tomography (CBCT) scans in Egyptian patients. This condition is a key anatomical variation that, if missed, may cause surgical complications such as nerve injury. The study uses previously collected CBCT scans of individuals aged 15 and older from the Oral and Maxillofacial Radiology Department at Cairo University. The scans will be analyzed retrospectively. The main questions it aims to answer are: How closely does the AI model's segmentation of the mandibular canal match the expert manual segmentation? How accurate is the AI model in identifying the presence or absence of bifid mandibular canals? Participants are not actively involved. Instead, anonymized CBCT data will be analyzed using the AI model and compared to expert annotations to measure diagnostic performance.

This retrospective diagnostic accuracy study aims to evaluate the performance of a novel artificial intelligence (AI) model in the segmentation and detection of bifid mandibular canals (BMC) on cone-beam computed tomography (CBCT) scans. The study will be conducted on anonymized DICOM data collected from a dental radiology archive, with no direct patient contact or intervention. Study Rationale and Background Accurate identification of bifid mandibular canals is crucial for avoiding complications during oral surgical procedures. Manual identification on CBCT by experts is the current standard, yet it is time-consuming and subject to interobserver variability. Deep learning-based models have the potential to offer automated, reproducible, and efficient alternatives. AI Model and Methodology The AI model used in this study will be developed using a 3D U-Net architecture implemented in the MONAI framework. Manual segmentation of mandibular canals, including bifid variants, will be performed using 3D Slicer software to serve as ground truth data. The AI model will be trained and validated using this ground truth. Technical Workflow Data Collection: Retrospective DICOM CBCT images will be extracted and anonymized. Manual Segmentation: Gold standard segmentation will be performed manually by an expert using 3D Slicer. Model Development: The 3D U-Net model will be developed and trained using the MONAI framework in Python. Evaluation: The trained model will be evaluated on unseen test data and compared to manual segmentation using standard performance metrics (Dice Similarity Coefficient, sensitivity, specificity, and accuracy). Statistical Analysis: The Dice score will be used as the primary metric of agreement. Sample size was calculated based on expected performance values. Confidence intervals were computed, and statistical significance was assessed. Sample Size Assessment A power calculation was conducted based on the sensitivity and specificity of the model using normal approximation methods. The total sample size was determined to ensure reliable performance evaluation with adequate power and precision. Data Quality Assurance and Handling A structured process will be employed to ensure that CBCT scans used are of sufficient diagnostic quality. All manual segmentations will be reviewed by a second radiologist to reduce bias. All model outputs will be verified against the manual ground truth by a blinded expert. Data inconsistencies, if present, will be flagged, documented, and excluded from the analysis. Missing Data Management Given the retrospective nature, no missing data will be filled in or replaced because only complete and usable scans will be included in the study. Statistical Plan The primary analysis will involve evaluating the overlap between the AI model and expert segmentations using the Dice coefficient. Secondary measures will include the sensitivity and specificity of the AI model in detecting bifid branches. All statistical analyses will be performed using Python libraries such as SciPy and NumPy.