基于深度神经网络的新型肠鸣音检测分析方法

A New Method for Detection and Analysis of Bowel Sounds Based on Deep Neural Networks

摘要: 肠鸣音是人体重要的生理信号，对肠鸣音的检测和分析具有重要的临床价值。传统的听诊器检测法主观性强，且无法做到连续、动态的监测，导致数据的时效性及精确性差。另一种方法是借助声音传感器采集和数字化肠鸣音，然后利用计算机对肠鸣音进行处理和特征分析，以实现对肠鸣音客观，定量检测。但由于肠鸣音具有环境噪声干扰大、周期性差、随机性强等特点，人工提取普适稳定的特征极为困难，导致这种识别方法检测准确率较低。项目组提出一种有效的肠鸣音检测方法，即先提取肠鸣音的MFCC特征，然后采用深度神经网络提取更加稳定抽象的特征，最后采用softmax识别出肠鸣音出现的位置和肠鸣音具体的类别。实验表明，这种方法能够准确检测出肠鸣音出现的时刻，同时也有较高准确率识别肠鸣音类别，值得在临床中推广应用。

Abstract: Bowel sounds is an important physiological signal of the human body, and it has important clinical value for the detection and analysis of bowel sounds. The traditional stethoscope detection method is highly subjective and cannot be continuously and dynamically monitored, resulting in poor timeliness and accuracy of data. Another method is to use a sound collector to collect and digitize bowel sounds, and to use a computer to reduce noise, amplify and manually extract features, and identify them through a classifier. However, due to the characteristics of large environmental noise interference, poor periodicity, and strong randomness, artificial extraction of universal features is extremely difficult, resulting in lower classifier recognition accuracy. In this paper, an effective method is proposed for the detection of bowel sounds, namely, the MFCC feature extraction of bowel sounds, and then the deep neural networks is used to detect the category of bowel sounds. The confirmative study of large samples based on the bowel sounds database shows that this method has high recognition accuracy and can avoid the high complexity of human extraction features, so it is worth popularizing in clinical application.