Classification of Normal and Crackles Respiratory Sounds into Healthy and Lung Cancer Groups

Abstract

Lung cancer is the most common cancer worldwide and the third most common cancer in Malaysia. Due to its high prevalence worldwide and in Malaysia, it is an utmost importance to have the disease detected at an early stage which would result in a higher chance of cure and possibly better survival. The current methods used for lung cancer screening might not be simple, inexpensive and safe and not readily accessible in outpatient clinics. In this paper, we present the classification of normal and crackles sounds acquired from 20 healthy and 23 lung cancer patients, respectively using Artificial Neural Network. Firstly, the sounds signals were decomposed into seven different frequency bands using Discrete Wavelet Transform (DWT) based on two different mother wavelets namely Daubechies 7 (db7) and Haar. Secondly, mean, standard deviation and maximum PSD of the detail coefficients for five frequency bands (D3, D4, D5, D6, and D7) were calculated as features. Fifteen features were used as input to the ANN classifier. The results of classification show that db7 based performed better than Haar with perfect 100% sensitivity, specificity and accuracy for testing and validation stages when using 15 nodes at the hidden layer. While for Haar, only testing stage shows the perfect 100% for sensitivity, specificity, and accuracy when using 10 nodes at the hidden layer.

Deep Lung Auscultation Using Acoustic Biomarkers for Abnormal Respiratory Sound Event Detection

“…Lung diseases are among the leading causes of death globally [9,10]. Due to its high prevalence, lung sound analysis becomes utmost important in early detection of the health related problems.…”

“…The use of OST features was shown to give improvements over the constrained features like STFT. Although, there have been several studies using wavelets for respiratory sounds detection [23,10,24], the features are limited to entropy, energy and their high-level statistics. In contrast to these, in this paper we extend the wavelet based feature set.…”

“…This step is repeated iteratively, pro-ducing a set of approximation signals at different detailed levels and a final approximation of the signal. Here, we use db7 wavelet due to their efficient time frequency localization properties in respiratory sound analysis [24,10]. The choice for level of decomposition is based on prominent frequency components of the signal.…”

Distinctive features for normal and crackles respiratory sounds using cepstral coefficients

“…In a study by Abdul Malik et al [12], fifteen different features are extracted from each segment of the respiratory sounds and Artificial Neural Network (ANN) was used as the classifier. DWT was used to decompose the respiratory sounds into seven different frequency band based on Daubechies (db7) and Haar mother wavelet.…”

“…Result of the study showed that db7 outperform Haar with perfect 100% sensitivity, accuracy and specificity in both testing and validation stage by using 15 nodes at the hidden layer. Meanwhile, using 10 nodes at the hidden layer, Haar was able to obtain perfect 100% sensitivity, specificity and accuracy for testing stage only [12].…”