AbstractPurpose: This study aimed to design and evaluate a novel method for the registration of 2D lateral cephalograms and 3D craniofacial cone‐beam computed tomography (CBCT) images, providing patient‐specific 3D structures from a 2D lateral cephalogram without additional radiation exposure.Methods: We developed a cross‐modal deformable registration model based on a deep convolutional neural network. Our approach took advantage of a low‐dimensional deformation field encoding and an iterative feedback scheme to infer coarse‐to‐fine volumetric deformations. In particular, we constructed a statistical subspace of deformation fields and parameterized the nonlinear mapping function from an image pair, consisting of the target 2D lateral cephalogram and the reference volumetric CBCT, to a latent encoding of the deformation field. Instead of the one‐shot registration by the learned mapping function, a feedback scheme was introduced to progressively update the reference volumetric image and to infer coarse‐to‐fine deformations fields, accounting for the shape variations of anatomical structures. A total of 220 clinically obtained CBCTs were used to train and validate the proposed model, among which 120 CBCTs were used to generate a training dataset with 24k paired synthetic lateral cephalograms and CBCTs. The proposed approach was evaluated on the deformable 2D–3D registration of clinically obtained lateral cephalograms and CBCTs from growing and adult orthodontic patients.Results: Strong structural consistencies were observed between the deformed CBCT and the target lateral cephalogram in all criteria. The proposed method achieved state‐of‐the‐art performances with the mean contour deviation of 0.410.12 mm on the anterior cranial base, 0.480.17 mm on the mandible, and 0.350.08 mm on the maxilla, respectively. The mean surface mesh ranged from 0.78 to 0.97 mm on various craniofacial structures, and the LREs ranged from 0.83 to 1.24 mm on the growing datasets regarding 14 landmarks. The proposed iterative feedback scheme handled the structural details and improved the registration. The resultant deformed volumetric image was consistent with the target lateral cephalogram in both 2D projective planes and 3D volumetric space regarding the multicategory craniofacial structures.Conclusions: The results suggest that the deep learning‐based 2D–3D registration model enables the deformable alignment of 2D lateral cephalograms and CBCTs and estimates patient‐specific 3D craniofacial structures.