Multiclass Detection of Third Molar Impaction on Panoramic Radiograph Using Convolutional Neural Network

Impacted molars, which fail to erupt properly, represent a common dental pathology often causing pain and complications. Panoramic radiography remains the primary diagnostic tool, but its effectiveness is limited by image quality issues such as noise artifacts and reduced sharpness, which may lead to diagnostic errors. This study develops an automated classification system using convolutional neural networks (CNNs) to enhance the diagnosis of impacted third molars. We compared ResNet-50 and ResNeXt-50 architectures on a dataset of ${3, 0 0 0}$ manually cropped panoramic radiographs, applying data augmentation techniques including flipping, color jitter, and Gaussian blur to improve model robustness. The system evaluates three key diagnostic parameters: angulation, position, and impaction class. Results show that ResNeXt-50 outperforms ResNet-50. In angulation classification (84 % accuracy), position classification has (61 % accuracy), and class classification has (62 %), while ResNet-50 achieves lower accuracy. The comparative analysis indicates that ResNeXt-50's residual blocks enable more effective feature extraction from limited dental imaging data, making it the preferred architecture for clinical applications. This CNN-based approach significantly enhances diagnostic consistency for impacted third molars.