Reconstruction of 3D Mandibular Models from 2D Lateral Cephalometric Radiographs Using Deep Learning

Abstract

Background: This proof-of-concept study aimed to develop and evaluate the accuracy of a novel deep learning framework designed to reconstruct three-dimensional (3D) mandibular models from two-dimensional (2D) lateral cephalometric radiographs (LCRs), and to assess regional accuracy across different mandibular segments along anatomical axes. Methodology: A convolutional neural network (CNN) was trained using paired synthetic LCRs and Cone Beam Computed Tomography (CBCT)-derived segmented mandibular models to deform a template ellipsoid into predicted mandible shapes. The final testing dataset included 178 CBCT scans from 147 patients to evaluate model accuracy. Geometric similarity between the predicted reconstructions and ground truth CBCT segmentations was quantified using Chamfer Distance (CD), Earth Mover’s Distance (EMD), and Hausdorff Distance (HD). Following initial validation, each reconstructed mandible was analyzed along anatomical axes to identify regional variations in accuracy. Results: The model achieved overall values of 3.11 mm for CD, 12.15 mm for EMD, and 14.31 mm for HD. Mixed-model ANOVA revealed that the posterior segment demonstrated lower accuracy than the middle and anterior segments (P < 0.001). Accuracy between the transverse halves did not differ significantly (P = 1.000), while the superior segment exhibited lower accuracy compared with the middle and inferior segments (P < 0.001). Conclusions: The proposed deep learning framework can reliably reconstruct coarse 3D mandibular models from 2D synthetic LCRs. However, regional performance varied, with reduced accuracy observed in the superior and posterior regions.