Automated Classification of Alveolar Bone Defects for Preoperative Augmentation Planning Using Deep Learning

Determining alveolar bone deficiencies prior to dental implant surgeries is critical for surgical planning. This study aimed to develop a deep learning-based framework that can automatically detect and classify alveolar ridge deficiencies (healthy, horizontal, vertical, and combined defects) using Cone Beam Computed Tomography (CBCT) slices and to evaluate the diagnostic performance of four different Convolutional Neural Network (CNN) architectures.

A novel dataset consisting of 1305 anonymous CBCT cross-sectional images, classified and labeled by experts into four categories, was created: Healthy (n=325), horizontal Defect (n=359), vertical Defect (n=310) and combined Defect (n=311). Four different CNN models (RegNetY-008, EfficientNetV2-S, ResNet50, and MobileNetV3-Large) were trained and tested. Model performance was evaluated using accuracy, weighted precision, recall, F1-score, and epoch duration.

The RegNetY-008 model demonstrated the highest performance with a 93.87% accuracy rate and a 93.88% weighted F1-score, processing the data with the fastest epoch time (8.04 sec.). This was followed by EfficientNetV2-S with a 93.10% accuracy rate. ResNet50 achieved 90.80% success, while MobileNetV3-Large showed the lowest performance (85.82%). Confusion matrices revealed that RegNetY-008 is particularly effective at distinguishing complex combined defects with minimal misclassification.

Deep learning models, particularly the RegNetY architecture, can effectively classify alveolar bone defects from CBCT slices.

The proposed automated system provides clinicians with a quick and objective second opinion in preoperative planning. By accurately classifying bone defects, it can assist dentists, particularly those with limited experience, in deciding between standard implant placement and augmentation procedures, thereby potentially reducing surgical complications and planning time.