Classification of Lung Sounds with Convolutional Neural Network

In this work, respiratory sounds are classified as normal and pathological sounds using convolutional neural network (CNN) and effect of the depth (number of layers) of CNN on classification performance is investigated. Respiratory sound records which were obtained from various sources are down sampled to a common sampling frequency and normalized to ensure that all the data lies in the same data space. Training and test data is chosen randomly from the set of normal and pathological sounds. Architecture of the CNN has 1, 2, and 3 convolution layers for 3 different experiments and each model has one fully connected layer after convolution layers. Classification is completed with logistic regression which is the last step of the architecture. When the results of the experiments are evaluated, it is understood that convolutional neural networks for classification of respiratory sounds is an active research area.