Transformer and Attention Enhanced Deep Learning Approach for CBCT-Based Mental Foramen Classification and Segmentation

Determining the location and morphological characteristics of the mental foramen is of critical clinical importance in dental surgical procedures. In this study, the mental foramen was classified and segmented using deep learning methods on CBCT images. The study utilized a unique and manually annotated CBCT dataset. An ensemble deep learning model incorporating Transformer and Attention mechanisms was developed for classification. A UNet structure, also incorporating Transformer and Attention mechanisms, was proposed for segmentation. The models were trained and tested on images containing variations in the anatomical location, shape, and size of the mental foramen. The proposed DAFNet model achieved 95.20% validation accuracy and 91.25% test accuracy with an AUC of 0.9648, while the proposed TA-UNet segmentation model achieved a Dice score of 0.8804 and IoU of 0.8386, outperforming conventional deep learning architectures. The findings demonstrate that the mental foramen can be automatically detected with high accuracy in CBCT images and its segmentation can be reliably performed.