Toward Semantically-Consistent Deformable 2D-3D Registration for 3D Craniofacial Structure Estimation From a...

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Deep neural networks and statistical shape models aid 2D-3D registration but struggle with fine anatomical details.
• Recovered volumetric images often lack fidelity and cross-dimensional semantic consistency.

Purpose of the Study:

• To develop a semantically-consistent deformable 2D-3D registration method for detailed volumetric image recovery.
• To improve the accuracy and fidelity of 3D anatomical structures reconstructed from single radiographs.

Main Methods:

• Inferred a voxel-wise registration field between cone-beam computed tomography and lateral cephalometric radiographs.
• Refined initial registration fields using craniofacial structural details and semantic consistency.
• Employed a self-supervised scheme for a voxel-level refiner and a weakly supervised semantic consistency measure.

Main Results:

• Achieved significant performance gains in deformable 2D-3D registration over state-of-the-art methods.
• Demonstrated improved volumetric fidelity and fine-grained craniofacial structural detail recovery.
• Validated the effectiveness of the weakly supervised semantic consistency measure.

Conclusions:

• The proposed method offers a simple yet effective solution for semantically-consistent deformable 2D-3D registration.
• This approach enhances volumetric image recovery from single radiographs, improving anatomical detail.
• The framework advances radiograph-based volumetric reconstruction and registration accuracy.