毕业生知识库

磁粒子成像技术是一种基于磁纳米粒子在交变磁场中的非线性磁化响应原理的成像技术。该技术能够对磁纳米粒子的浓度实现高灵敏度的在体定量成像， 被广泛应用于多种生物医学研究中。然而，由于硬件带来的各种噪声影响磁粒子成像响应信号的信噪比，降低了磁粒子成像的成像质量。目前，主要通过在硬件改进和算法设计两方面提升磁粒子成像信号信噪比。然而，现有的硬件改进和算法设计无法有效去除动态噪声，对信号信噪比的增强有限。

针对上述问题，本论文提出基于多尺度网络的深度学习方法，通过将信号先验知识融入多尺度网络，提升信号增强算法性能。本论文的主要工作内容与创新点归纳如下:

1. 提出了一种频域-时域有监督多尺度信号增强算法

针对磁粒子成像响应信号受到多种动态噪声影响导致信噪比下降的问题，本论文提出了一种频域-时域多尺度网络，结合磁粒子成像信号的频域数据和时域数据，提高多尺度网络对响应信号的预测结果精度。另外，磁粒子成像响应信号作为一维数据，易产生特征提取不充分的问题，因此，在网络中设计了多通道特征提取模块，通过串联两个并行多尺度卷积，提高网络噪声滤除能力。通过仿真信号和实测数据的一维磁粒子成像响应信号数据集上的实验结果表明，设计的双域多尺度网络能够充分提取信号和噪声特征，实现对噪声的抑制和响应信号 的增强。在实际测量噪声的数据集上，提出的算法实现了将 MPI 信号的信噪比从 6.88 dB 提升至 29.11 dB，同时信号的均方根差百分比从 51.26 降至 3.96，均方根误差从 65.07 × 10−4 降至 5.29 × 10−4。

2. 提出了一种融合磁粒子成像响应信号先验知识的自监督多尺度信号增强算法

针对磁粒子成像响应信号无公开数据集，并且低噪声的磁粒子成像信号获取难度大，采集时间长的问题，在利用多尺度网络实现磁粒子成像响应信号特征提取的基础上，本论文进一步提出了一种基于自监督的多尺度信号信噪比增强 算法。另外，本文在自监督算法中融合了信号周期内磁粒子成像信号呈中心对称的形状先验知识，利用不同周期的磁粒子成像信号数据抑制磁粒子成像信号噪声，通过最小化不同周期信号间的均方误差实现信号信噪比的增强。在一维和二维自监督多尺度网络数据集上的实验结果证明，本论文提出的自监督方法优于其他对比算法，具有更好的噪声抑制能力，实现更好的信噪比增强结果和图像重建结果。

Magnetic particle imaging technology is an imaging technology based on the principle of nonlinear magnetization response of magnetic nanoparticles in an alternating magnetic field. This technology can achieve high-sensitivity in vivo quantitative imaging of the concentration of magnetic nanoparticles and is widely used in a variety of biomedical research. However, due to various noises brought by hardware, the signal-to-noise ratio of magnetic particle imaging response signals is affected, which reduces the imaging quality of magnetic particle imaging. At present, the signal-to-noise ratio of magnetic particle imaging signals is mainly improved through hardware improvements and algorithm design. However, existing hardware improvements and algorithm designs cannot effectively remove dynamic noise, and the enhancement of the signal-to-noise ratio is limited.

In response to the above issues, this thesis proposes a deep learning method based on multi-scale networks, which improves the performance of signal enhancement algo- rithms by integrating signal prior knowledge into multi-scale networks. The main work content and innovative points of this thesis are summarized as follows:

1. A frequency-time domain supervised multiscale signal enhancement algorithm is proposed

This thesis proposes a frequency-domain multi-scale network to address the issue of the decrease in signal-to-noise ratio caused by various dynamic noises in the response signal of magnetic particle imaging. By combining the frequency-domain and time-domain data of magnetic particle imaging signals, the accuracy of the prediction results of the multi-scale network for the response signal is improved. In addition, as one-dimensional data, the magnetic particle imaging response signal is prone to insufficient feature extraction. Therefore, a multi-channel feature extraction module was designed in the network to improve the network’s noise-filtering ability by concatenating two parallel multi-scale convolutions. The experimental results on the one-dimensional magnetic particle imaging response signal dataset of simulated signals and measured data show that the designed dual domain multi-scale network can fully extract signal and noise features, and achieve noise suppression and response signal enhancement. On the dataset 

containing measurement noise, the proposed algorithm improves the signal-to-noise ra- tio of the MPI signal from 6.88 dB to 29.11dB, while reducing the root mean square error percentage of the signal from 51.26 to 3.96, and the root mean square error from 65.07 × 10−4 to 5.29 × 10−4.

2. A self-supervised multiscale signal enhancement algorithm incorporating prior knowledge of magnetic particle imaging response signals is proposed

In response to the lack of a publicly available dataset for magnetic particle imaging response signals, as well as the difficulty and long acquisition time of obtaining low noise magnetic particle imaging signals, this thesis further proposes a self-supervised multi-scale signal to noise ratio enhancement algorithm on the basis of utilizing multi- scale networks to extract the characteristics of magnetic particle imaging response signals. In addition, this article integrates the prior knowledge of the centrosymmetric shape of magnetic particle imaging signals within the signal period in the self-supervised algorithm and uses magnetic particle imaging signal data from different periods to suppress the noise of magnetic particle imaging signals. By minimizing the mean square error between signals from different periods, the signal-to-noise ratio is enhanced. The experimental results on one-dimensional and two-dimensional self-supervised multi-scale network datasets demonstrate that the self-supervised method proposed in this thesis outperforms other comparison algorithms, has better noise suppression ability, and achieves better signal-to-noise ratio enhancement results and image reconstruction results.