A Point Cloud Generation Network for Automatic Prediction of Postoperative Maxillofacial Soft Tissue.

Abstract

Purpose: In the planning of maxillofacial surgery, accurately evaluating the postoperative soft tissue area is crucial. This allows doctors to provide patients with better morphological recovery while ensuring the restoration of normal functional areas. This study aims to develop an advanced automatic algorithm for the completion of soft tissue defects, enhancing the accuracy and effectiveness of surgical planning.

Methods: We introduce a point cloud completion method based on Generative Adversarial Networks. Firstly, we remove the defective regions, reserving the healthy tissues and converting them into point clouds. Then, using the soft tissue completion network, we reconstruct the defective areas and generate the corresponding point cloud images. Finally, we combine the healthy point cloud with the generated defective regions to produce a complete soft tissue image and model of the patient's face.

Results: To validate our approach, we conduct qualitative and quantitative experiments on 20 normal individuals (10 males and 10 females). Compared with several existing algorithms, our method shows superior significance in soft tissue prediction. The Root Mean Squared Error between the generated model and ground truth is 1.45 ± 0.25 mm, and the Surface Distance error is 0.69 ± 0.13 mm. The visualization results show that using our network to generate soft tissue at the defect site is more structurally consistent.

Conclusion: This study introduces a novel point cloud generation network for reconstructing facial soft tissue images, which provides accurate and structurally consistent morphological outcomes. The proposed method shows great potential in improving the quality and accuracy of surgical planning in maxillofacial surgery.