Received May 26, 2020, accepted May 31, 2020, date of publication June 4, 2020, date of current version June 16, 2020. Digital Object Identifier 10.1109/ACCESS.2020.3000111 Classification of Lung Sounds With CNN Model Using Parallel Pooling Structure FATIH DEMIR 1 , ARAS MASOOD ISMAEL 2 , AND ABDULKADIR SENGUR 1 1 Electrical-Electronics Engineering Department, Technology Faculty, Firat University, 23119 Elazig, Turkey 2 Information Technology Department, Technical College of Informatic, Sulaimani Polytechnic University, Sulaimani, Iraq Corresponding author: Aras Masood Ismael (aras.masood@spu.edu.iq) ABSTRACT The recognition of various lung sounds recorded using electronic stethoscopes plays a signif- icant role in the early diagnoses of respiratory diseases. To increase the accuracy of specialist evaluations, machine learning techniques have been intensely employed during the past 30 years. In the current study, a new pretrained Convolutional Neural Network (CNN) model is proposed for the extraction of deep features. In the CNN architecture, an average-pooling layer and a max-pooling layer are connected in parallel in order to boost classification performance. The deep features are utilized as the input of the Linear Discriminant Analysis (LDA) classifier using the Random Subspace Ensembles (RSE) method. The proposed method was evaluated against a challenge dataset known as ICBHI 2017. The deep features and the LDA with RSE method provided the best accuracy score when compared to other existing methods using the same dataset, improving the classification accuracy by 5.75%. INDEX TERMS Lung sound, CNN model, parallel pooling, deep features, RSE method. I. INTRODUCTION Lung disease ranks third among fatality causes worldwide. According to the World Health Organization (WHO), more than 3 million people die each year due to respiratory dis- eases [1]. Lung sound attributes and their diagnosis play a significant role in the pulmonary pathology. Lung sounds can generally be grouped as ‘‘normal lung sounds’’ or ‘‘abnormal lung sounds.’’ Normal lung sounds are when no pulmonary disease exists, whilst abnormal lung sounds are heard when a pulmonary disease is present [2], [3]. An abnormal lung sound is a supplementary respiratory sound that is heard in addition to the normal lung sound. Abnormal lung sounds are known as continuous if they con- tain wheezes, and discontinuous if they contain crackles. The presence of such sounds mostly indicates the presence of a lung disease [4]. Auscultation is a method by which doctors evaluate and diagnose lung diseases using a stethoscope. It is known as a low-cost, easy to apply, and reliable test that requires minimal diagnosis duration [5]. The test is able to provide considerable information about lung diseases and their symptoms [6]; how- ever, the classical auscultation process using a stethoscope is The associate editor coordinating the review of this manuscript and approving it for publication was Zahid Akhtar . not infallible as it depends on the skill of the physician and their hearing sensitivity. Because of the inclusion of non-stationary signals, lung sounds can be difficult to analyze and separate using conven- tional auscultation techniques. Hence, the use of an electronic stethoscope combined with an artificial intelligence system can be used as a means to overcoming the limitations of con- ventional auscultation, and thereby providing a more reliable and efficient method through automated diagnosis [7]. From the outset of machine learning and pattern recog- nition, numerous studies have put forwards proposed meth- ods for the automatic classification of lung sounds. In the literature, conventional methods have generally been used, consisting of classifiers and hand-crafted features for the cat- egorization of lung sounds. In [6], features are extracted with the frequency ratio of Power Spectral Density (PSD) values and the Hilbert-Huang Transform (HHT) method, and then evaluated using Support Vector Machine (SVM) algorithm. In [8], the features extracted from time-frequency and time- scale analysis methods are utilized for the detection of normal lung sounds and crackles, with k-Nearest Neighbors (k-NN), Multilayer Perceptron (MLP) and SVM used for the classifi- cation stage. The best accuracy was achieved with the SVM. In [9], the feature set is constituted by instantaneous kurto- sis, discriminating function, and entropy in order to classify 105376 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 8, 2020