Lateral Cephalometric Landmark Annotation Using Histogram Oriented Gradients Extracted from Region of Interest...

Area of Science:

• Medical Imaging
• Orthodontics
• Machine Learning

Background:

• Two-dimensional cephalometric image analysis is vital for orthodontic diagnosis and treatment planning.
• Deep learning models for landmark annotation face challenges due to limited clinical data acquisition and labeling.
• Automating anatomical landmark annotation is crucial but hindered by data limitations.

Purpose of the Study:

• To propose a novel model leveraging conventional machine learning for enhanced cephalometric landmark detection accuracy.
• To address the challenge of limited datasets in clinical cephalometric image analysis.
• To improve the accuracy of anatomical landmark annotation in orthodontic diagnosis.

Main Methods:

• The model employs region of interest (ROI) extraction for coarse localization and Histogram-Oriented Gradient (HOG) features for fine localization.
• The Light Gradient Boosting Machine (LGBM) algorithm classifies image patches containing landmark pixels.
• Performance was evaluated on the ISBI Cephalometric and Dental Cepha datasets using cross-validation, targeting 2 mm radial precision.

Main Results:

• The model achieved 77.11% accuracy within a 2 mm radial precision range on the ISBI Cephalometric dataset.
• Cross-validation yielded a mean accuracy of 78.17%, confirming the model's robustness.
• Remarkable landmark detection accuracy of 84% was achieved on the Dental Cepha dataset.

Conclusions:

• Traditional machine learning techniques are effective for accurate cephalometric landmark detection, even with limited data.
• The proposed model demonstrates potential for clinical applications where large labeled datasets are unavailable.
• This approach offers a viable solution for improving orthodontic diagnostic tools.