Diagnostic Performance Analysis using CBCT Assisted Artificial Intelligence for Detection of Supernumerary Teeth and Odontomes: A Retrospective Study

Introduction: Artificial Intelligence (AI) is revolutionising dentistry by enhancing diagnostic precision, particularly for complex conditions like supernumerary teeth and odontomes. Accurate detection of these anomalies is essential for effective management and improved patient outcomes. Aim: To evaluate the diagnostic performance of the machine learning models- Logistic Regression (LR), Random Forest (RF), and k-Nearest Neighbours (kNN), Gradient Boosting (GB), Artificial Neural Network (ANN), Stochastic Gradient Descent (SGD)- in detecting supernumerary teeth and odontomes using Cone Beam Computed Tomographies (CBCTs). Materials and Methods: The present retrospective study analysed CBCT images of the maxillary and mandibular arches from patients who visited Saveetha Dental College and Hospital between February and September 2024 in Chennai, Tamil Nadu, India. A total of 240 high-resolution CBCT scans were included, categorised into three groups: supernumerary teeth (n=80), odontomes (n=80), and normal dentition (n=80), and patients exhibiting other dental anomalies were excluded. The dataset was divided into training (70%) and testing (30%) sets. Following preprocessing, multiple machine learning modelsLR, RF, kNN, GB, ANN, and SGD- were employed for analysis using Orange data mining software. Results: The present study evaluated the diagnostic performance of six machine learning models- LR, kNN, RF, ANN, SGD, and GB- for classifying CBCT images into supernumerary teeth, odontomes, and normal dentition. LR achieved the highest overall performance with an Area Under Curve (AUC) of 0.960, accuracy of 86.3%, and Mathews Correlation Coefficient (MCC) of 0.829, demonstrating strong generalisation across all classes. NN and GB also performed well, with comparable AUCs (0.960 and 0.944, respectively) and balanced classification metrics. RF showed good performance (AUC=0.951) but with a higher LogLoss, indicating less confident predictions. SGD and kNN underperformed, with kNN yielding the lowest accuracy (53.7%) and high misclassification rates. Confusion matrix analysis further supported these findings, highlighting LR, ANN, and RF as the most reliable models for distinguishing among the three diagnostic categories on CBCT images. Conclusion: The LR proved to be the most effective model for detecting supernumerary teeth and odontomes. Future research should address dataset diversity and class imbalances while exploring ensemble modelling approaches to enhance diagnostic capabilities further.