Distinct Feature Labeling Methods for SVM-Based AMD Automated Detector on 3D OCT Volumes

Abstract

Today's automated detectors of Age-related macular degeneration (AMD) on optical coherence tomography (OCT) volumes using the support vector machine (SVM) are widely researched in the field of ophthalmology. Additionally, an OCT volume is three-dimensional (3D) data composed of several OCT images. Therefore, two feature labeling methods, the slice-chain labeling method and the slice-threshold labeling method, are investigated for the 3D OCT volume in this paper. The two labeling methods are evaluated in this paper because they influence detection accuracy for the SVM-based AMD automated detector and the number of features stored in the memory of SVM hardware. According to the quantization analysis, we can easily compare several types of feature extraction in the local binary patterns (LBP) and linear configuration patterns (LCP) in the data that have to be stored in the RAM. From the experiment results, the slice-threshold labeling method achieves a high detection accuracy of 96.36% with 35.34% features saved in the memory of SVM hardware compared with the slice-threshold labeling method.