Learning ECOC Code Matrix for Multiclass Classification with Application to Glaucoma Diagnosis | JoVE Visualize

Area of Science:

Ophthalmology
Machine Learning
Computer Vision

Background:

Accurate classification of angle closure glaucoma (ACG) mechanisms is crucial for effective medical diagnosis.
Error-Correcting Output Codes (ECOC) offer a robust framework for multiclass classification tasks.
Existing ECOC methods may require optimization for improved performance in complex medical image analysis.

Purpose of the Study:

To develop and evaluate a novel ensemble learning method based on Error-Correcting Output Codes (ECOC) for classifying four distinct angle closure glaucoma (ACG) mechanisms.
To enhance the accuracy and diversity of dichotomizers within the ECOC framework.
To improve the separability of ECOC codes through a new selection criterion incorporating regularization.

Main Methods:

Proposed a new ensemble learning method utilizing optimized dichotomizers within an ECOC framework.
Employed a wrapper approach with cross-validation to determine optimal feature sets for individual dichotomizers.
Introduced a novel criterion with a regularization term for selecting competitive dichotomizers to maximize ECOC code separability.
Applied the method to classify four ACG mechanisms using 84 features extracted from anterior segment optical coherence tomography (AS-OCT) images of 152 patients.

Main Results:

The proposed ECOC-based ensemble learning method achieved a weighted average classification accuracy of 87.65% using leave-one-out cross-validation (LOOCV).
The method demonstrated significantly superior performance compared to other existing ECOC approaches.
Individual dichotomizers were optimized for accuracy and diversity, enhancing the overall ECOC framework's effectiveness.

Conclusions:

The developed ensemble learning method provides accurate classification of four angle closure glaucoma (ACG) mechanisms.
The optimized ECOC approach shows significant promise for application in clinical glaucoma diagnosis.
Individual dichotomizer optimization and a novel selection criterion are key to the method's success.