Virtual reconstruction of midfacial bone defect based on generative adversarial network.

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Artificial Intelligence in Medicine

Background:

• Bony midfacial defects pose reconstruction challenges.
• Accurate reconstruction is crucial for restoring form and function.
• Existing methods may lack precision or require specific references.

Purpose of the Study:

• To evaluate the accuracy of generative adversarial networks (GAN) for reconstructing bony midfacial defects.
• To assess the feasibility of GAN-based virtual reconstruction for clinical use.
• To determine if GANs can reconstruct defects without a direct reference.

Main Methods:

• The bony midface was divided into five anatomical subunits.
• Artificial defects were created on CT images and reconstructed using GAN.
• Clinical defects (midspan and unilateral) were reconstructed and evaluated quantitatively and qualitatively.
• Cosine similarity and mean error were used as accuracy metrics.

Main Results:

• GAN achieved high accuracy in reconstructing artificial and unilateral clinical defects (cosine similarity ~0.97, mean error ~0.5 mm).
• Reconstruction of midspan clinical defects was deemed acceptable by experts.
• Statistical analysis showed no significant difference in reconstruction errors between artificial and real defects (P > 0.05).

Conclusions:

• GAN-based virtual reconstruction technology demonstrates high accuracy for bony midfacial defects.
• The approach shows potential for similar performance on real patient data.
• This study presents a novel solution for reconstructing bony midfacial defects without a reference.