Fig. 3: Overall procedure.

a Comparison of surgical outcome between real post-cephs and spost-cephs. b Evaluation of spost-cephs and assessment of their clinical utility. c Generative prediction for orthognathic surgery using ceph network (GPOSC-Net) model architecture, which utilizes a convolutional neural network (CNN)-based image embedding module (IEM) and a GCNN-based landmark topology embedding module (LTEM) to vectorize lateral cephalograms and landmark data, respectively. These vectors are concatenated and fed into a multi-layer perceptron (MLP) to predict the landmark movements caused by surgery. To generate spost-cephs, a latent diffusion model is employed with a few conditions, such as surgical movement value predicted by IEM and LTEM, pre-cephs, pre-operational landmarks, profile lines, and intended amount of surgical movement (IASM), which can control the virtual setback amounts of spost-cephs.