Browsing by Author "Kallio, Kari Hannu"

Now showing 1 - 1 of 1

Hengitysäänien luokittelu itseorganisoituvilla piirrekartoilla
(2010) Kallio, Kari Hannu
School of Science | Master's thesis
Lung and bronchial diseases, especially asthma and allergies, have been increasing in frequency for a long time because of environmental factors and the changed living circumstances. Returning to nature is not possible and smoking has not decreased remarkably therefore lung problems have come to stay. Lung sound research has rapidly become intensified with the possibilities granted by new technology. Traditionally respiratory sound analysis is based on the auscultation made by the doctor with the stethoscope and the sounds heard and empirically recognized. Signal processing and pattern recognition offer new tools for the separation of breath sounds. Self-organizing feature map (SOM) is a neural network that imitates the biological neural system. It has successfully been utilized in speech recognition already during three decades. In this thesis, we have for the first time in the world applied the method of self-organizing feature map created by academician Teuvo Kohonen for the classification of lung sounds. The feature vector is assembled according to the model used in speech recognition with signal rms-value and FFT frequency components. The program composed according to Kohonen's theory comprises pattern computation, teaching of the self-organizing map, recognition of lung sounds and visualization of classification results. The method organizes the teaching sound patterns on the map forming clusters of the lung sound groups according to their features. The organized map can be used for the recognition of normal and adventitious breath sounds and for the separation of diseases. The results of the experiments indicate that the method is suit-able for classification of lung sounds. However, the findings of this study suggest that the feature vector modified from speech recognition is not the best possible. A fairly extensive literature review has been done in this study. Tentatively, some alternative feature vectors have been evaluated referring to publications concerning classification of lung sounds. Preliminary suggestions are given to plan a new feature vector. Based on the findings of the present study, future research should focus on the improvements of the method and the classification results. New possibilities to significantly improve the process of classification have been opened by European lung sound standards and the growing international networks. These will also facilitate the set up of a cross-border research database and the collection of respiratory sounds. In addition this thesis proposes a model to build up an intelligent automated classifier for lung sound quality control and the actual classification.