Abstract
Purpose: :
To exploit machine learning algorithm for automated classification of different angle closure mechanisms based on the quantitative assessment of Anterior Segment Optical Coherence Tomography (AS-OCT; Visante OCT Model 1000; Carl Zeiss Meditec, Inc., Dublin, CA) imaging.
Methods: :
148 subjects were recruited from glaucoma clinics in Singapore with diagnoses of primary angle closure and angle closure glaucoma. One eye from each subject (only nasal and temporal quadrants) was imaged with AS-OCT under dark conditions. The images were classified by 4 glaucoma experts into 4 classes based on anatomical configuation of anterior chamber angle (ACA). These 4 classes are pupil blick (PB, n=51), plateau iris (PI, n=23), peripheral iris roll (PIR, n=21), and large antero-posterior lens diameter (LAPLD, n=53)) . A customized software was used to analyze the AS-OCT images and to measure parameters. Four machine learning algorithms were tested: Support Vector Machines (SVM), Naive Bayes (NB), Random Forest (RF), and Classification Trees (CT). The heuristic similar backward elimination techniques were used to generate the reduced parameter sets. The area under the receiver operating characteristic curve (AUC) was used to measure diagnostic performance of these algorithms.
Results: :
A total of 90 parameters were evaluated by the software. From the initial run, the number of parameters was reduced to 15. Among the tested algorithms, the SVM attain the best performance with the 0.909 AUC. In contrast, NB, RF, and CT obtained 0.750, 0.825 and 0.889 AUC respectively.
Conclusions: :
Automated machine learning classifiers of AS_OCT image parameters might be useful for enhancing the utility of AS_OCT for detecting different angle closure mechanisms.
Keywords: imaging/image analysis: clinical • computational modeling