Landmark annotation through feature combinations: a comparative study on cephalometric images with in-depth analysis...

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Automated anatomical landmark localization is crucial for cephalometric analysis.
• Current methods often lack interpretability and robust feature extraction.

Purpose of the Study:

• To evaluate machine learning techniques for automating landmark localization in cephalometric images.
• To explore feature extraction, combinations, contextual analysis, and model interpretability using Shapley Additive exPlanations (SHAP) values.

Main Methods:

• Experimentation on 300 lateral cephalograms using pixel feature descriptors (raw pixels, gradient magnitude, gradient direction, Histogram of Oriented Gradients - HOG).
• Evaluation of features in local, pyramid, and global contexts.
• Classification for landmark/non-landmark pixel discernment.
• Application of SHAP values to interpret Light Gradient Boosting Machine (LGBM) models.

Main Results:

• Histogram of Oriented Gradients (HOG) and gradient direction features showed significant performance across contexts.
• Global texture context offered superior performance but increased test time.
• HOG in the local context achieved the highest Success Detection Rate (SDR) of 75.84%.

Conclusions:

• Feature combinations significantly impact landmark annotation accuracy.
• The study highlights the importance of feature selection and context in automated cephalometric analysis.
• Explainability methods like SHAP values facilitate understanding and further development of landmark-specific feature combinations.