Cross-Examination for Angle-Closure Glaucoma Feature Detection.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Anterior segment imaging is crucial for diagnosing angle-closure glaucoma (ACG).
• Identifying the specific ACG mechanism is essential for effective treatment.
• Feature selection methods are vital for accurate diagnosis from complex imaging data.

Purpose of the Study:

• To evaluate the efficacy of redundant features in diagnosing ACG mechanisms using anterior segment optical coherence tomography (AS-OCT) images.
• To compare the performance of supervised (MRMR) and unsupervised (L-score) feature selection algorithms in ACG diagnosis.
• To assess the utility of an AdaBoost classifier for categorizing various ACG mechanisms.

Main Methods:

• Anterior segment optical coherence tomography (AS-OCT) images were analyzed.
• Two feature selection algorithms were employed: Minimum Redundancy Maximum Relevance (MRMR) and Laplacian score (L-score).
• An AdaBoost machine learning classifier was used to classify five ACG mechanisms: iris roll, lens, pupil block, plateau iris, and no mechanism.

Main Results:

• The unsupervised L-score method, utilizing redundant features, demonstrated improved accuracy in ACG diagnosis.
• The MRMR method, focusing on minimum redundancy, showed comparatively lower diagnostic accuracy.
• The AdaBoost classifier effectively categorized the different ACG mechanisms when combined with the selected features.

Conclusions:

• Redundant features selected by the L-score algorithm are beneficial for enhancing ACG diagnosis accuracy.
• Unsupervised feature selection may offer advantages over supervised methods for complex disease diagnosis in AS-OCT imaging.
• The findings support the use of advanced feature selection techniques for improved clinical diagnosis of angle-closure glaucoma.