Lung disease detection using EasyNet

Abstract

One of the major causes of deaths worldwide is the pulmonary diseases. There is an increasing need for an efficient technique that can automatically diagnose these diseases with high accuracy. In this paper, we propose a deep learning architecture to automatically detect pulmonary diseases. The raw pulmonary sound signals are taken from two popular datasets: ICBHI and KAUH datasets. These signals have diverse sampling frequencies of 4 kHz, 10 kHz or 44.1 kHz. The signals from KAUH dataset have a duration of minimum 5 s, while ICBHI signals have a duration of 10 to 90 s. These signals undergo pre-processing, which involves re-sampling them to a common 4 kHz frequency, and segmenting them into frames lasting 3 s. The frames are then normalized and passed to the proposed EasyNet model for training and classification. The EasyNet architecture contains only two convolution layers, which reduces the model complexity. The model’s performance is analyzed for both binary detection as well as multi-class detection. Our method performs well in all the considered evaluation scenarios, and yields an accuracy, sensitivity, and specificity of 1.0 for the KAUH dataset, while for the ICBHI dataset, an accuracy of 0.997, sensitivity of 0.999, and specificity of 0.997 is achieved. For the combined dataset, we have achieved an accuracy of 0.998, with a sensitivity and specificity of 0.999. These values are better than the existing state-of-the-art methods. The proposed architecture is quite simple yet effective in detecting lung diseases.