Vision Transformers for Accurate Inferior Alveolar Nerve Classification in Cone Beam Computed Tomography

The precise localization of the Inferior Alveolar Nerve (IAN) is crucial for safe dental and maxillofacial surgical procedures. This study compares and contrasts five Transformer architectures - Vision Transformer (ViT), Shifted Window Transformer (SwinT), Class-Attention in Image Transformers (CaiT), Gradient Focal Transformer (GFT), and the Data-Efficient Image Transformer (DeiT) - for automated classification of IAN position in Cone Beam Computed Tomography (CBCT) images. SwinT outperformed the other models and achieved the highest accuracy of 80.00% due to its hierarchical, multi-scale window attention mechanism, which effectively captures sensitive anatomical information and generalizes well on limited medical imaging data. In contrast, GFT performed lowest, mainly because of extreme overfitting caused by the gating operations, which enhanced training characteristics but were unable to generalize to new data. Overall, the results reveal that transformer-based designs have great potential for precise and reliable IAN position classification in CBCT for universal healthcare systems, especially those with hierarchical attention like SwinT.