基於類神經網絡之異常音與呼吸週期偵測系統

Abstract

本研究旨在研究「應用類神經網路於自動化呼吸週期及異常肺音的偵測」及分析多聲道聽診器對於異常音分類的影響。肺部疾病是世界上第三大死因，因此，肺部疾病的早期介入及治療一直都是醫療領域著重的一環，當中聽診扮演了重要的角色。醫療人員藉由聽診器進行聽診，進行於肺部疾病的診斷。相較於傳統聽診器，電子聽診器解決了傳統聽診器無法收集、判斷只能仰賴資深醫療人員的問題。因此，如何利用電子聽診器收集的肺音及建構高精確度的自動化肺音分類系統成為重要的指標。 本研究使用的肺音資料是藉由多聲道聽診器於實際醫院收集，收集的部位為左右上胸，左右側胸，及左右下胸，資料包含環境噪音，人聲等其他噪音。因此，如何在高噪音資料中擷取肺音特徵成為自動化肺音分類重要的一環。本研究建立呼吸週期偵測系統及異常音偵測系統以達到建立自動化肺音系統的目的，其中異常音偵測系統採用兩階段判斷，第一階段為判斷音檔是否含有異常音，第二階段為幀的判斷。本研究採用短時距傅立葉轉換（STFT) 作為資料的特徵擷取，並以卷積神經網絡（CNN）、深度殘差網絡（Deep Residual Network,ResNet）及CNN-BLSTM作為分類器。結果顯示，呼吸週期偵測系統準確率達87%、異常音第一階段準確率達95%，及第二階段準確率達90%。儘管如此，為了確保在真正的工作環境裡可有效的被應用，後期實驗加入了外部麥克風，希望藉由同步的環境噪音可達到除噪的效果，以提升深度學習網絡的準確率。本實驗提出了早期融合和晚期融合兩大策略。以CNN-BLSTM作為分類器，晚期融合策略將準確率提高了8%，證明了此策略可有效地提高分類的性能。

Treatment of lung diseases, which are the third most common cause of death in the world, is one of great importance in the medical field. Many studies using lung sounds recorded with stethoscope have been conducted in the literature in order to diagnose the lung diseases with artificial intelligence-compatible devices and to assist the experts in their diagnosis. In this paper, a database which includes noise and background sounds were collected by using a novel multi-channel data acquisition system from six different positions over the anterior chest, then was used for the classification of lung sounds. The purpose is to build an automated lung sound detection system which consist of breathing phase detection system and adventitious lung sound detection system. The adventitious lung sound detection system is a two-stage classifier which includes adventitious respiratory sound detection and frame-based adventitious respiratory detection. Short-time Fourier transform (STFT) was adopted as statistical feature extraction for further analysis. Network model such as CNN, ResNet, and CNN-BLSTM. The results of the breathing phase detection system achieved 87% of accuracy. Besides, the first stage of adventitious respiratory sound system reached an accuracy of 95% while the second stage which adopted only sound file from lateral chest obtained an accuracy of 90%. However, to apply in real working environment more efficiently, this research used the simultaneous environmental noise for noise reduction which collected by an external microphone during the subsequent stage of this research. The environmental noise spectrograms and noisy lung sound spectrograms are adopted to estimate the corresponding Ideal Ratio Mask (IRM) for further classification. The early-fusion strategy (EF) and late-fusion strategy (LF) are proposed. The evaluated performance of CNN-BLSTM, CNN-BLSTM+EF, and CNN-BLSTM+LF, which showed the precision of 65%, 72%, and 73%, respectively, demonstrating the proposed method could effectively improve classification performance.