Conditional generative modeling of postoperative radiographs in adolescent idiopathic scoliosis

Patient-specific visualization of immediate postoperative alignment may ultimately support preoperative planning for adolescent idiopathic scoliosis, but the accuracy of simulations under alternative fusion segments has not been established. We developed SVV-Net, a conditional generative neural network that simulates postoperative anteroposterior radiographs from preoperative radiographs and surgeon-defined fusion segments. The model was trained and internally tested in 720 retrospectively collected patients using a patient-level split and prospectively evaluated in 79 additional patients. SVV-Net demonstrated favorable overall image-generation performance compared with the comparator generative models and showed high agreement with real postoperative radiographs for Cobb angle and correction magnitude, with minimal discrepancies in the representation of the predefined upper and lower instrumented vertebrae. Expert reviewers rated the generated images favorably for clarity, authenticity, accuracy, reasonableness, and usability. SVV-Net provides a visual simulation under predefined fusion segments rather than independently selecting fusion levels. In the present study, validation was limited to the fusion range actually used for each patient; its potential use for simulating and comparing alternative fusion plans requires dedicated future evaluation. Multicenter validation and incorporation of lateral radiographs and additional clinical parameters into the model are also needed before clinical use.