MCVoxel2MeshNet: Voxel-To-Mesh Generation for Mandibular Canal Segmentation in CBCT

Accurate segmentation of the mandibular canal (MC) is crucial in maxillofacial surgical planning to prevent nerve damage and improve treatment outcomes. However, manual MC segmentation from cone-beam computed tomography (CBCT) images is time-consuming and prone to observer variability. While CNN-based segmentation models have shown promise in automating this process, they often suffer from spatial discontinuities and low surface quality due to their reliance on voxel-wise classification. To address these limitations, we propose MCVoxel2MeshNet, a neural implicit representation (NIR)-based voxel-to-mesh generation network for MC segmentation in CBCT images. Unlike conventional CNN-based methods, our approach models the MC surface as a continuous function, enabling seamless segmentation and high-fidelity 3D mesh generation. Additionally, we introduce a modality-aware loss function that enforces multi-level supervision at the mask, grid, point, and mesh levels, improving segmentation accuracy and mesh smoothness. Quantitative and qualitative evaluations demonstrate that MCVoxel2MeshNet outperforms existing segmentation networks, achieving the highest Dice Score (DS) of 0.85 ± 0.06 % and Chamfer Distance (CD) of 1.97 ± 1.85 mm. Furthermore, our method produces continuous, noise-free MC structures, effectively eliminating false positives and artifacts. These results highlight the potential of MCVoxel2MeshNet for accurate and robust MC segmentation, paving the way for enhanced computer-assisted maxillofacial surgical planning.