NR-GAN: Noise Reduction GAN for Mice Electroencephalogram Signals
NR-GAN: Noise Reduction GAN for Mice Electroencephalogram Signals
复制标题
DOI：
10.1145/3366174.3366186
复制
发表时间：
2019-10
期刊：
影响因子：
--
通讯作者：
Yuki Sumiya;Kazumasa Horie;Hiroaki Shiokawa;H. Kitagawa
中科院分区：
文献类型：
--
作者：
Yuki Sumiya;Kazumasa Horie;Hiroaki Shiokawa;H. Kitagawa
为了支持基础睡眠研究，已经提出了几种用于小鼠的自动睡眠阶段评分方法。虽然这些方法可以根据小鼠的脑电图（EEG）和肌电图（EMG）信号准确地对小鼠的睡眠阶段进行评分，但它们对噪声，特别是对EEG信号的干扰很脆弱。最简单的解决方法是在评分前减少或消除噪音。然而，不存在用于减少生物信号中的噪声的方法。由于EEG信号包含多种类型的噪声，因此很难预测所有类型的噪声，这抑制了诸如频率滤波器之类的手工设计方法的使用。此外，使用深度学习模型的降噪方法不适用，因为它们需要记录噪声，并且这里考虑的噪声不能与生物信号分开测量。在这项研究中，我们使用对抗训练来解决这个问题，这是一种不需要噪声记录作为训练样本的深度学习模型方法。我们提出了一种新的降噪模型，称为“NR-GAN。“它的训练过程需要一组嘈杂的信号和一组清晰的信号。由于这些集合可以独立测量，NR-GAN可以减少小鼠EEG信号中的噪声。
To support basic sleep research, several automated sleep stage scoring methods for mice have been proposed. Although these methods can score mice sleep stages accurately based on their electroencephalogram (EEG) and electromyogram (EMG) signals, they are fragile against noise, especially in EEG signals. The simplest solution is to reduce or eliminate noise before scoring. However, a method for reducing noise in biological signals does not exist. Because EEG signals contain many types of noise, predicting all of them is difficult, which inhibits the use of hand-engineered methods such as frequency filters. Additionally, noise reduction methods with deep learning models are not applicable as they require records of noise, and the noise considered here cannot be measured separately from biological signals. In this study, we address this problem using adversarial training, which is a method for deep learning models that does not require noise records as training samples. We propose a new noise-reduction model called "NR-GAN." Its training process requires a set of noisy signals and a set of clear signals. Since these sets can be measured independently, NR-GAN can reduce noise in mice EEG signals.