基於支持向量機辨識肺音之病徵

Recognition of lung sound classification based on support vector machines

曾安慈(An-Tzu Tseng)

指導教授 ： 王昭男

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201801527

摘要

肺部呼吸音是重要的醫療診斷訊號源，相關呼吸疾病的診斷平時都必須依賴專業醫師的經驗進行聽診，才能夠得知肺部的情況是否良好，本研究希望藉由聽診器輸出訊號即可自動辨別出肺部呼吸音為正常、哮喘或粗囉音等問題。
在實際的聽診器輸出訊號中，心跳聲音明顯大於肺部呼吸音許多，本文採用基於小波轉換的穩態－非穩態濾波器(wavelet transform-based stationary-non stationary filter, WTST-NST)方法處理肺音受心音干擾的問題，此濾波器可分離穩態與非穩態的訊號，符合心音與肺音兩者波形間的差異性，明顯降低肺音訊號受心音成份所影響的雜訊。此外為了讓每段訊號有辨別的標準，使用梅爾頻率倒頻譜得到的一組係數計算出作為該段肺音訊號的特徵值，並由支持向量機作為分類器，在資料點屬於高維度特徵值的情況下難以區分類型，利用映射函數轉換資料點到更高的維度後，找到一個可以區分兩類別的超平面來進行分類，由此訓練出分類模型後即可得知輸入訊號屬於何種類別，本研究在辨識的準確度上有不錯的效果，可實現肺音辨識的目的。

關鍵字

肺音 ； 辨識 ； 基於小波轉換的穩態－非穩態濾波器 ； 梅爾倒頻譜係數 ； 支持向量機

並列摘要

Respiratory sound is an important source of medical diagnosis signal. It must rely on the experience of professional doctor to know whether the lung is in well condition. We hope to automatically identify lung sound.
In the output signal of lung sound, the heartbeat sound is much larger than the lung sound. The lung sound is disturbed by the heartbeat sound. In this thesis, the wavelet transform-based stationary-nonstationary filter is used to separate the heartbeat sound. This filter can separate the steady-state and non-steady-state signals, which is consistent with the difference between the heart sound and the lung sound. Significantly, it reduces the noise of the lung sound signal affected by the heart sound component. In addition, in order to make the identification of each lung sound, a set of coefficients obtained by the mel-frequency cepstral coefficients is used to calculate the eigenvalues of the lung sound signal, and the support vector machine is used as the classifier. In the case of high dimension eigenvalues, it is difficult to distinguish the types. After using the mapping function to convert the data points to a higher dimension, we find a hyperplane that can distinguish between the two categories for classification. And then training the classification model, we can find what the input signal classification. This study has outstanding accuracy of identification, and it attained the purpose of lung sound recognition.

並列關鍵字

lung sound ； recognition ； the wavelet transform-based stationary-nonstationary filter ； mel-frequency cepstral coefficients ； support vector machine

參考文獻

[1] G. Hedenstierna and L. Edmark, “The effects of anesthesia and muscle paralysis on the respiratory system”, Intensive Care Medicine, 31, 1327–1335, 2005.

[2] C. Gu and C. Li, “Assessment of human respiration patterns via noncontact sensing using Doppler multi-radar system”, Sensors, 15, 6383-6398, 2015.

[3] H. Pasterkamp, S.S. Kraman and G.R. Wodicka, “Respiratory Sounds. Advances Beyond the Stethoscope,” American Journal of Respiratory and Critical Care Medicine, 156, 974-987, 1997.

[4] B. Erickson, “heart sounds and murmurs across the lifespan”, Fourth Edition, Mosby, 2003.

[5] I. Hossain and Z. Moussavi, “Finding the Lung Sound-Flow Relationship in Normal and Asthmatic Subjects,” IEEE Engineering in Medicine and Biology Society Annual Conference, 3852-3855, 2004.