應用深度學習模型於異常肺音訊號分析之自動化聽診辨識

Automatic auscultation classification of abnormal lung sounds through deep learning models

吳昱陞(Yu-Sheng Wu)

指導教授 ： 袁賢銘

國立交通大學/資訊學院/資訊科學與工程研究所/碩士(2020年)

摘要

肺部呼吸音在醫療診斷中是非常重要的指標，尤其在當今新冠肺炎全球肆虐的情況下，如何將聽診設備結合人工智慧來搭建遠程醫療系統顯得相當重要。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。本研究希望藉由聽診器輸出訊號並透過深度學習的方法自動辨識出肺部呼吸音為正常、喘鳴或濕囉音等問題。在臨床聽診器輸出訊號中，心音訊號強度會大於肺部呼吸音，並且與臨床人員大量觀察頻譜圖後，我們擷取肺部訊號主要頻段100Hz到1600Hz來進行建模，以降低心音與其它環境音的干擾。為了有效辨識每段訊號，我們將短時快速傅立葉轉換產生的頻譜圖以及梅爾頻率倒頻譜係數作為該段肺音訊號的特徵值。並且發現將兩種特徵合併後會使辨識準確度有明顯的提升，接著我們使用此合併特徵來挑選模型。本研究實際比較了以影像和時序數據兩大類的深度學習模型在聽診音辨識的表現，並發現影像辨識為主的模型表現較佳。此外我們也採用了深度可分離卷積的技術，並參考MobileNet的結構建模，以達到高準確率、低參數量的目的。

關鍵字

深度學習 ； 異常肺音 ； 特徵提取 ； 聽診 ； 卷積神經網路

並列摘要

Pulmonary breath sounds are representative indicators in medical diagnosis, especially in the current situation of the coronavirus COVID-19 pandemic in the world, how to combine auscultation equipment with artificial intelligence to build a remote medical system becomes very important.The diagnosis of related respiratory diseases usually depends on the auscultation of experienced doctors. This research aims to use the output signals of stethoscope and classify them through deep learning model automatically. The dataset used in this work consists of four classes, normal, wheezing, crackles, and unknown.During clinical auscultation, the intensity of the heart sound signal will be greater than the lung breathing sound.After extensive observation of the spectrogram with clinical staff, we filter the main frequency band of the lung signal from 100Hz to 1600Hz to reduce the noise caused by heart sounds and other environmental sounds. In order to effectively classify each signal, we use the spectrogram generated by the short-time fast Fourier transform and the Mel frequency cepstrum coefficient as the feature value of each lung sound signal. And found that combining the two features will significantly improve the classification accuracy, and then we use this combined feature to do model selection. This research compares the performance of the image-based models and the time-series models in auscultation sound classification and the result shows that image-based models outperformed the time-series models. In addition, we also adopt Depthwise separable convolution technic, and refer to the architecture of MobileNet, in order to achieve the purpose of high accuracy and low model parameter.

並列關鍵字

Adventitious lung sound ； Convolutional neural network ； Feature extraction ； Deep learning ； Auscultation

參考文獻

[1] A. Bohadana, G. Izbicki, and S. S. Kraman, “Fundamentals of lung auscultation,” New

England Journal of Medicine, vol. 370, no. 8, pp. 744–751, 2014.

[2] P. Forgacs, “The functional basis of pulmonary sounds,” Chest, vol. 73, no. 3, pp. 399–405,

1978.

[3] A. Marques and A. Oliveira, “Normal versus adventitious respiratory sounds,” in Breath