多種肺音自動辨識之研究

Abstract

本研究提出一種可辨識多種肺音的演算法，可自動辨識包括正常與異常肺音種類達到15種。本系統的取樣頻率為11025 Hz，並整合了小波轉換與自我迴歸模型為特徵值擷取方法。其中在小波轉換方面，分析了三種不同系列共12種的小波基底，包括了Coiflet-2、Coiflet-3、Coiflet-4、Coiflet-5、Symlet-5、Symlet-6、Symlet-7、Symlet-8、Daubechies-5、Daubechies-6、Daubechies-7及Daubechies-8。分析的結果為Daubechies-8有最好的辨識正確率。本系統亦加入了自我迴歸模型(Autoregressive model, AR model)做為特徵值選取的工具，並以Akaike模型選取資訊準則(Akaike information criterion, AIC)為模型階數評估標準。整合了小波轉換與自我迴歸模型為特徵值擷取方法改善了傳統上只使用小波轉換為特徵值選取的工具，在辨識效果上有顯著的提升。 在分類器方面，本研究除了使用傳統的倒傳遞網路做為型態辨識分類器外，並且調整隱藏層神經元數目予以最佳化，使得辨識正確率達到85.3%。同時亦比較了另外兩種不同架構的圖形分類器：學習向量量化網路及徑向基網路，其正確率分別可達到90.4%及93.3%。最後將這三種不同架構的分類氣整合成一大系統，達到類似投票系統的功能，提高系統的可靠度，其辨識正確率達到94.4%。未來可望朝居家看護儀器發展或進行肺音資料庫之建立。

An algorithm of automatic recognition 15 kinds of lung sounds, including normal and adventitious lung sounds, is presented in this thesis. The proposed system utilizes 11025 Hz sampling rate, and integrates wavelet transform and autoregressive model (AR model) for extracting features of lung sounds. As the aspect of wavelet transform, 12 kinds of wavelet basis, including Coiflet-2, Coiflet-3, Coiflet-4, Coiflet-5, Symlet-5, Symlet-6, Symlet-7, Symlet-8, Daubechies-5, Daubechies-6, Daubechies-7 and Daubechies-8, are comprehensively compared in this study. Among 12 wavelet basis, Daubechies-8 has been confirmed to yields best recognition result through comprehensive comparisons. The system also combines autoregressive model as another feature extraction method, and Akaike information criterion is utilized as principle of order determination. The experimental result clearly shows that by combining AR model and wavelet transform, the proposed method performs better than traditional approaches that only used wavelet transform as primary feature extraction method in recognizing lung sounds. As the aspect of classification, back propagation neural network is been utilized to perform the task of recognizing pattern in lung sounds. It also has been confirmed that the number of neurons in the hidden layer is optimized. The experimental result shows that average recognition rate yielded by the proposed algorithm reaches as high as 85.3%. In addition, other two neural networks, learning vector quantization network and radio basis function network, are compared along with the previous one, which the recognition accuracy yielded by those methods are 90.4 and 93.3%, respectively. Finally, the three neural networks are integrated together as a vote system that can increase the reliability of the proposed system, and the recognition accuracy yielded by the integrated system further improved to 94.4%. Thus, the system can be developed for home care equipment, or an assist for establishing lung sound databases.