Automatic voice based disease detection method using one dimensional local binary pattern feature extraction network. (1st December 2019)

• Record Type: Journal Article
• Authors: Tuncer, Turker; Dogan, Sengul; Ertam, Fatih
• Is Part Of: Applied acoustics. Volume 155(2019)
• Page Start: 500
• Page End: 506
• Publication Date: 2019-12-01
• Subjects: 1D local binary pattern based network -- Neighborhood component analysis -- Pathological voice detection -- Pattern recognition; Acoustical engineering -- Periodicals; Periodicals; 620.2
• DOI: 10.1016/j.apacoust.2019.05.023
• Languages: English
• ISSNs: 0003-682X

Abstract

Abstract: Voices have been widely used for disease detection in the literature but these methods are non-invasive. In this article a 1D local binary pattern (LBP) based feature extraction network (1D-LBPNet) is proposed to extract stable features from voices. The proposed 1D-LBPNet is inspired by convolutional neural networks (CNN) for instance AlexNet, GoogleNet, ResNet. Then, a voice based disease recognition method is presented in this paper. The presented voice based disease recognition method consists of feature extraction using 1D-LBPNet, feature concatenation, feature reduction using neighborhood component analysis (NCA) and classification phases. In the feature extraction phase, 1D-LBPNet extracts 256 × 8 = 2048 features because it has 7 layers. The extracted features are concatenated in the feature concatenation phase. To reduce the concatenated features, a NCA based feature reduction method is used. 1 nearest neighbor (1NN) classifier is utilized as classifier to demonstrate distinctive of the extracted features. To test performance of the proposed method, Saarbruecken Voice Database (SVD) is used in this article. /a/ vowels of the Cordectomy and frontolateral resection diseases are chosen to test the proposed 1D-LBPNet based recognition method. 10 cases are defined using single and concatenated voices for each disease. The results and comparisons clearly shown that the proposed 1D-LBPNet achieved high success rates and these results clearly proved success of the proposed method.