Deep Learning-Based Facial and Skeletal Transformations for Surgical Planning.

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Biomedical Engineering

Background:

• Accurate prediction of facial and skeletal shapes is crucial for virtual surgical planning (VSP) in orthognathic surgery.
• Understanding craniofacial relationships and achieving accurate soft tissue-bone transformations remains challenging due to complex anatomy and nonlinearities.

Purpose of the Study:

• To develop and validate a novel bidirectional 3D deep learning framework (P2P-ConvGC) for accurate subject-specific transformations between facial and skeletal shapes.
• To enhance the precision of virtual surgical planning in orthognathic procedures.

Main Methods:

• A bidirectional 3D deep learning framework, P2P-ConvGC, was developed using a large-scale dataset.
• A 2-stage point-sampling strategy generated high-resolution facial and skeletal point subsets for input.
• Separate subnetworks predicted corresponding skeletal and facial point subsets.

Main Results:

• P2P-ConvGC outperformed existing state-of-the-art algorithms in shape and landmark error calculations.
• Total landmark errors were 1.964 ± 0.904 mm (upper skull), 2.398 ± 1.174 mm (mandible), and 2.226 ± 0.774 mm (facial soft tissues).
• Clinical validation showed average surface deviation errors of 0.895 ± 0.175 mm (facial) and 0.906 ± 0.082 mm (skeletal).

Conclusions:

• The P2P-ConvGC model demonstrates high performance in subject-specific facial and skeletal shape prediction.
• The framework shows significant clinical application potential for postoperative facial prediction and VSP in orthognathic surgery.