Automatic multi-class classification of 3D relationships between the mandibular third molar and canal on cone-beam computed tomography using a geometry-aware network

OBJECTIVES: Accurate 3D evaluation of the spatial relationship between the mandibular third molar (M3) and the mandibular canal (MC) is critical for assessing the risk of nerve injury during M3 extraction. The purpose of this study was to automatically classify the diverse spatial relationships between M3 and MC into 5 distinct categories based on the degree of anatomical contact or involvement as well as spatial proximity on cone-beam computed tomography (CBCT) images using a geometry-aware deep learning framework. METHODS: The 3D spatial relationships between M3 and MC were categorized into 5 distinct types. The proposed framework consisted of 2 stages: first, the modified mAttUNet, an improved 3D U-Net architecture augmented with an attention mechanism was used for segmentation of M3 and MC; second, the proposed DenseAttNet was developed for multi-class classification. By incorporating dense attention mechanisms and signed distance map (SDM) inputs, the proposed network effectively captured both geometric and anatomical features, leading to more accurate and reliable multi-class classification. RESULTS: The mAttUNet outperformed other conventional models, achieving the highest segmentation performance with average precision scores of 0.96 for MC and 0.84 for M3. The proposed DenseAttNet demonstrated superior and consistent performance, achieving an overall AUC of 0.97 and maintaining high accuracy across all 5 relationship types, effectively and reliably distinguishing the various spatial relationships between MC and M3. CONCLUSIONS: This automated and accurate classification of M3-MC spatial relationships offers valuable clinical utility, supporting enhanced risk evaluation and optimized surgical planning, and ultimately helping to reduce complications such as inferior alveolar nerve injury.