Deep Learning with Generative Adversarial Networks Predicts Postoperative Imaging of Single Level ACDFs

Introduction: Artificial intelligence has contributed to innovations in spine surgery by providing tools for patient education and surgical planning. Recent deep learning innovations yet to see application in spine surgery are generative adversarial networks (GANs), unsupervised models that use competing neural networks to generate photorealistic transformations from input images. The purpose of this study is to evaluate a custom-trained GAN in generating predictive postoperative anterior cervical fusion Xrays from preoperative Xrays.

Methods: A custom GAN adapted from the open source pix2pix image translation network was constructed. The model was designed to output digitally generated radiographs (DGRs) predicting the postoperative instrumented fusion from preoperative XR inputs. A blinded team of neurosurgeons compared these DGR outputs to ground truth postoperative imaging for level accuracy in patients who had undergone single level anterior cervical discectomy and fusion (ACDF).

Results: Preoperative and postoperative XRays were included from 72 single-level ACDF patients. As a proof of concept, 50 paired AP and lateral XRays of C5-C6 ACDFs were used to train the model and 22 XRs were used to test the model in generating predicted postoperative fusions. Neurosurgeons were able to identify the presence of a single level ACDF fusion in 100% of the DGRs and were able to identify the correct level of fusion at C5-C6 in 54.5% (27.3%-81.8%) of the DGRs.

Conclusion : Deep learning with GANs is capable of accurately generating realistic postoperative imaging for single level cervical spinal fusions from pre-operative Xray inputs despite the low volume and low-resolution training images in this early preliminary analysis. These promising results justify further investigation into modeling additional regions and surgeries as this model is refined with more training data and higher resolution training images. Downstream implications of predicting postoperative imaging using GANs include patient education and decision-making, surgical training, and prediction of spinal parameter changes resulting from surgery.