Lung sound disease detection using attention over pre-trained efficientnet architecture

作     者：Nair, Anuja Vadher, Himanshu Patel, Pal Vyas, Tarjni Bhatt, Chintan Bruno, Alessandro 

作者机构：Department of Computer Science & ampEngineering School of Technology Nirma University Gujarat Ahmedabad382481 India Department of Computer Science and Engineering School of Technology Pandit Deendayal Energy University Gujarat Gandhinagar382007 India Department of Business Law Economics Consumer Behaviour ’Carlo A. Ricciardi’ Faculty of Communication IULM University Milan20143 Italy 

出 版 物：《Multimedia Tools and Applications》 (Multimedia Tools Appl)

年 卷 期：2024年

页      面：1-35页

核心收录：

学科分类：0808[工学-电气工程] 0809[工学-电子科学与技术（可授工学、理学学位）] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

主　　题：Butterworth filters 

摘      要：Air pollution has been on the rise for quite a while now and with it is the increasing number of cases involving respiratory diseases. These respiratory diseases range from the mild ones to the most severe ones. Therefore, we need to detect these diseases. The most efficient way of checking the health of the lung is through the use of lung sounds. But the lung sounds carry with them the minute variations that a doctor or any human sometimes misses. To tackle the same, this paper proposes a deep learning model that detects a person’s lung disease by processing lung sounds. The ICBHI 2017 dataset comprising the lung sound audios and their corresponding labels is used to train the proposed model. The dataset consists of lung sounds and some additional background noise has been removed using the bandpass Butterworth filter, whose frequency range has been set to 250Hz - 2000Hz. Later, the amount of data is multiplied through data augmentation for audio signals. The augmented data is later converted to mel spectrograms, which are, in turn, fed into the deep learning model. The model that gives the best accuracy is EfficientNet-B0 + attention with a training accuracy of 99.72%, validation accuracy of 99.82%, precision, recall and f1-scores of 99.82% as well. Furthermore, a comparison of the pre-trained model’s training and inference times has been performed. Although the proposed model takes a significant amount of training time, it takes the least amount of inference time compared to the other pre-trained models that have been tested. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.