TCFNet: Bidirectional face-bone transformation via a Transformer-based coarse-to-fine point movement network

Area of Science:

• Computer Science
• Medical Imaging
• Biomedical Engineering

Background:

• Computer-aided surgical simulation is vital for orthognathic surgery planning.
• Traditional methods face limitations in accuracy, computational time, and data processing.
• Existing deep learning approaches struggle with large point clouds and complex registration.

Purpose of the Study:

• To develop an accurate and efficient deep learning framework for face-bone shape transformation in orthognathic surgery.
• To overcome the limitations of current simulation methods, including computational cost and accuracy.
• To introduce a coarse-to-fine network that learns complex correspondences for dense point cloud transformations.

Main Methods:

• Proposed a Transformer-based coarse-to-fine point movement network (TCFNet).
• Integrated a local information aggregation network (LIA-Net) to enhance precision by modeling local geometric structures.
• Employed a gated recurrent unit for global feature guidance and an optional auxiliary loss inspired by medical image registration.

Main Results:

• TCFNet demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.
• Achieved outstanding evaluation metrics and visual results in face-bone point cloud transformations.
• The framework is unsupervised and offers an optional expert-guided loss for critical organ reconstruction.

Conclusions:

• TCFNet provides a robust and accurate solution for computer-aided orthognathic surgical simulation.
• The proposed method effectively addresses the limitations of previous simulation techniques.
• TCFNet offers a promising advancement for improving the precision and applicability of surgical planning tools.