Automatic landmark identification in cone-beam computed tomography.

Area of Science:

• Medical Imaging
• Computer Vision
• Dental Technology

Background:

• Accurate landmark identification is crucial for quantitative analysis in cone-beam computed tomography (CBCT).
• Manual landmark placement is time-consuming and prone to inter-observer variability.
• Existing automated methods may lack robustness or accessibility for widespread use.

Purpose of the Study:

• To present and validate ALICBCT, an open-source, fully automated tool for landmark placement in CBCT scans.
• To reformulate landmark detection as a classification problem using a virtual agent within volumetric images.
• To enable precise quantification of bone morphology and tooth position changes in clinical studies.

Main Methods:

• A novel approach, ALICBCT, was developed and trained on 143 large and medium field-of-view CBCT scans.
• A virtual agent navigated volumetric space, guided by DenseNet features and fully connected layers, to identify landmarks.
• 32 ground truth landmarks were identified by clinicians, with subsequent training for 119 common clinical landmarks.

Main Results:

• ALICBCT achieved high accuracy with an average error of 1.54 ± 0.87 mm for 32 landmark positions.
• The tool demonstrated rare failures and a computation time of approximately 4.2 seconds per landmark on a conventional GPU.
• The algorithm was validated for identifying 119 clinically relevant landmarks.

Conclusions:

• ALICBCT is a robust and accurate tool for automatic landmark identification in CBCT.
• The open-source algorithm has been deployed on the 3D Slicer platform for clinical and research use.
• Continuous updates are planned to further enhance the precision of the ALICBCT tool.