Incorporating support vector machine to the classification of respiratory sounds by Convolutional Neural Network

• Published in: Biomedical signal processing and control Vol. 79; p. 104093
• Main Authors: Cinyol, Funda, Baysal, Uğur, Köksal, Deniz, Babaoğlu, Elif, Ulaşlı, Sevinç Sarınç
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.01.2023
• Subjects: Convolutional Neural Networks; Deep learning; Lung diseases; Respiratory sounds; Softmax; SVM; Transfer learning; Deep learning; Transfer learning; Lung diseases; Respiratory sounds; Softmax; Convolutional Neural Networks; SVM
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2022.104093

Classification of respiratory sounds (RS) by artificial intelligence (AI) methods has been studied by many groups, and a preferred method belonging to Deep Neural Networks (DNN) is Convolutional Neural Networks (CNN), where the Softmax function is one of the most popular classifiers used in the last layer of the network. However, there have also been studies examining the use of linear support vector machine (SVM) instead of Softmax in an artificial neural network architecture. This work focuses on incorporating SVM to CNN in multi-class RSs classification. Moreover, two-layer CNN model was added to the VGG16 model using transfer learning and similarly, the classification successes were compared with Softmax and SVM. The dataset used in this work has been obtained by clinical experts from a total of 294 subjects (105 diagnosed as normal and the rest of the patients have adventitious sound namely, 116 crackle and 73 rhonchi). During the classification phase, two-layer CNN architecture and SVM were added to this architecture, then various classifiers are implemented, namely CNN-Softmax, CNN-SVM, VGG16-CNN-Softmax, VGG16-CNN-SVM with 10-fold cross validation. In addition, state-of-art models (DenseNet 201, VGG16, InceptionV3, ResNet 101) and VGG16-CNN-KNN have also been applied to the dataset. It is found that the best classification accuracy figures have been found as 83% with VGG16-CNN-SVM model.