Relational reasoning network for anatomical landmarking.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate anatomical landmarking is essential for craniomaxillofacial (CMF) surgery planning and deformation analysis.
• Traditional methods often rely on explicit bone segmentation, which can fail with severe pathology or bone deformation.
• This limitation necessitates novel approaches for robust landmark identification.

Purpose of the Study:

• To develop and evaluate a deep learning network for anatomical landmarking of CMF bones (mandible, maxilla, nasal) without explicit segmentation.
• To accurately learn local and global landmark relationships within CMF bones.

Main Methods:

• Proposed a novel, efficient deep network architecture: Relational Reasoning Network (RRN).
• RRN operates end-to-end, utilizing learned landmark relations via dense-block units.
• Treated landmarking as a data imputation problem, predicting missing landmarks from given inputs.

Main Results:

• Applied RRN to 250 patient cone-beam computed tomography scans.
• Achieved an average root mean squared error of $$ per landmark using fourfold cross-validation.
• Demonstrated accurate identification of missing landmarks, even with severe pathology or bone deformations.

Conclusions:

• The RRN system successfully identifies anatomical landmark relations using deep learning, a novel approach.
• This method overcomes limitations of segmentation-based techniques, offering improved accuracy in challenging cases.
• The findings contribute to enhanced deformation analysis and surgical planning in CMF procedures.