Propostas de técnicas para caracterização e classificação automática de sons pulmonares adventícios

Abstract

In this thesis, the investigation of methods to characterize and classify adventitious lung sounds by spectral analysis is described. To accomplish this task, two novel techniques were developed, through Multiressolution Analysis, based on the Discrte Wavelet Transform. The first technique aims to detect abnormal sounds and classity them info four groups: normal, continuous and discontinuous adventitions lung sounds, also notifying their simultaneous occurence. During its processing, the respiratory cycle signal is decomposed up to its tenth level, and the energy present in the detail and approximation coefficients for each decomposition level is calculated, resulting on a curve of energy versus decomposition level. The resulting curves show different signatures for each kind of adventitious sound. These signatures are used as data source for a classifier system based on Radial Basis Function Artificial Neural Networks. This technique was tested for ten different wavelets, training a hundred neural networks for each wavelet, totalizing a thousand neural networks trained. The best performance rates for each wavelet reach values from 88% to 92.36% for the test group, in a set of 275 respiratory cycles. In the second technique, named Filtering by Selective Spectral Analysis, the lung sound is decomposed until its fourth level, the approximation coefficients spectra are calculatedand, based on the highest frequency component found on those coefficients, a multiband FIR filter is determined. This filter is used to eliminate all frequency components in the approximation coefficients except the highest one. After the filtering procedure, the signal is recomposed by wavelet reconstruction. In order to evaluate the proposed technique, ten wavelets were used in the decomposition and reconstruction stages. The wavelet which presented the best performance attenuated heart sounds 6 dB more than the adventitious sounds that occur in the same spectral band. For measuring this attenuation, the Power Spectral Density was used. This procedure showed satisfactory results, elimination the normal airflow noise and cardiac sounds, leaving only the adventitious sounds in the recorded lung sounds.