Context-guided fully convolutional networks for joint craniomaxillofacial bone segmentation and landmark digitization

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Cone-beam computed tomography (CBCT) is crucial for diagnosing and planning treatments for craniomaxillofacial (CMF) deformities.
• Accurate 3D modeling and anatomical landmark digitization from CBCT scans are essential clinical tasks.
• Existing methods often treat bone segmentation and landmark digitization as separate tasks, neglecting their inherent relationship and spatial context.

Purpose of the Study:

• To propose a novel framework for joint bone segmentation and landmark digitization from CBCT images.
• To leverage spatial context information, modeled via displacement maps, to improve accuracy in both tasks.
• To address the limitations of standalone approaches in CMF analysis.

Main Methods:

• Developed a Joint bone Segmentation and landmark Digitization (JSD) framework using context-guided fully convolutional networks (FCNs).
• Utilized displacement maps to represent spatial context, indicating voxel-to-landmark relationships.
• Employed a multi-task FCN to perform bone segmentation and landmark digitization concurrently, guided by displacement maps.

Main Results:

• The proposed JSD method demonstrated superior performance compared to state-of-the-art approaches.
• Experimental validation on 107 subjects confirmed the effectiveness of the joint approach.
• The framework successfully improved accuracy in both bone segmentation and landmark digitization tasks.

Conclusions:

• The proposed JSD framework effectively integrates bone segmentation and landmark digitization by utilizing spatial context.
• This joint approach offers significant improvements for CMF analysis using CBCT data.
• The method provides a more accurate and comprehensive solution for clinical diagnosis and treatment planning.