Reconstruction of 3D Models of the Mandible from 2D Lateral Cephalometric Radiographs by Deep Learning

Background: The aim of this proof-of concept study was to develop and evaluate the accuracy of a novel deep learning framework capable of reconstructing accurate three-dimensional (3D) models of the mandible from two-dimensional (2D) lateral cephalometric radiographs (LCRs) and assess regional accuracy in different segments of the mandible across anatomical axes. Methodology: Using paired synthetic LCRs and Cone Beam Computerized Tomography (CBCT) derived segmented mandibular models, a convolutional neural network (CNN) was trained to deform a template ellipsoid into predicted mandible models. The final testing set consisted of 178 CBCTs from 147 patients to test the accuracy of the model. Chamfer Distance (CD), Earth Mover’s Distance (EMD), and Hausdorff Distance (HD) were used to measure the geometric similarity between the predicted shape reconstruction model of the mandible and the ground truth mandible segmentation from the CBCT. After initial validation of the model, each mandible was evaluated along the anatomical axes to assess for regional differences in accuracy. Results: The model achieved an overall CD value of 3.11 mm, EMD value of 12.15 mm, and HD value of 14.31 mm. The mixed-model ANOVA test showed that the posterior segment showed lower accuracy compared to the middle and anterior segments (P<0.001), the transverse halves were not significantly different an accuracy from one another (P=1.000), and the superior segment showed lower accuracy compared to the middle and inferior segments (P<0.001). Conclusions: The proposed framework reliably reconstructs coarse 3D mandible models from 2D synthetic LCRs. Regional accuracies varied, with superior and posterior areas showing reduced accuracy in their respective axes.