Predicting orthognathic surgery results as postoperative lateral cephalograms using graph neural networks and diffusion models.

Summary

The authors introduce GPOSC-Net, combining a landmark prediction module with a latent diffusion model to synthesize realistic post-operative lateral cephalograms from pre-operative inputs. Multi-dataset validation, a visual Turing test, and simulation studies show accurate landmark prediction and high-fidelity image generation, supporting use in planning and patient communication.

Key Findings

• Introduces GPOSC-Net that fuses cephalometric landmark prediction with a latent diffusion image generator.
• Accurately predicts post-surgical cephalometric landmarks across diverse datasets.
• Generates high-fidelity synthesized post-operative lateral cephalograms validated by a visual Turing test and simulation.

Clinical Implications

May improve shared decision-making, expectation management, and procedural planning by previewing plausible post-op morphology; could reduce revisions by aligning plans with predicted outcomes.

Limitations

• External, multi-center clinical validation and outcome correlation are not reported.
• Impact on surgical decisions and real-world revision rates remains untested.

Future Directions

Prospective clinical trials to assess decision impact, integration with 3D CBCT, and open benchmarking across centers with code/data sharing.