毕业生知识库

   磁粒子成像（Magnetic particle imaging, MPI）是一种可以对磁性纳米粒子（Magnetic nanoparticles, MNPs）示踪剂在动物体内的浓度分布进行定量的实时高分辨率三维成像技术，MPI已在灌注成像、动脉粥样硬化斑块成像、肿瘤检测和细胞示踪等研究领域中得到应用。虽然MPI具有高时空分辨率，但该技术对硬件系统中的随机噪声和背景噪声较为敏感，导致重建出的图像中包含较大的噪声。此外，当视场中磁粒子样本相距较近（≤ 10mm）且浓度相差达到8倍以上时，MPI成像系统的动态范围受限于模数转换器转换精度，使得MPI难以准确定量同一视场中低浓度的磁粒子分布。

本文针对MPI成像过程中复杂动态噪声信号影响定量精度的问题，开展了基于特征融合的磁粒子成像图像去噪算法研究，从模型对MPI图像的去噪精度，去噪后对图像细节的保持能力，和对不同水平噪声的鲁棒性三方面对所提出的算法模型进行构建，实现对不同浓度磁粒子成像中的图像噪声进行有效去除，提升MPI图像的定量精度，并且在磁粒子成像设备获取的图像上进行验证。

   本文的主要研究工作如下：

（1） 针对磁粒子成像去噪任务中，现有方法在去除噪声的同时，难以兼顾对MNPs样本形状细节保持的问题，本文提出了一种基于多层次特征融合的内容学习网络（Multi-level Feature Fusion Neural Network, MFFNet），通过反卷积和注意力机制实现同一尺度空间内不同层次特征的高效融合，并使用融合特征对真值图像进行回归预测。实验结果表明，MFFNet能够有效抑制MPI图像中的噪声，提升MPI图像质量。

（2） 针对MPI成像系统的动态范围有限，噪声影响MPI图像定量精度的问题，本文在MFFNet的基础上，提出一种基于内容与噪声特征融合学习的磁粒子成像图像去噪网络（Content-Noise Feature Fusion Neural Network，CNFFNet），CNFFNet同时学习输入MPI图像的MNPs分布特征与噪声分布特征，并对其进行高效融合，使用融合后的特征对真值图像进行回归预测。实验结果表明，CNFFNet在去噪精度、去噪后MNPs样本形状的保持、对不同水平噪声的鲁棒性三方面优于其他对比方法。同时，CNFFNet能够有效提升同一视场中低浓度MNPs样本的定量精度。

Magnetic Particle Imaging (MPI) is a tomography imaging technique that enables real-time, high-resolution, and quantitative in vivo imaging of the concentration distribution of magnetic Nanoparticles (MNPs) tracer in animals. MPI has been applied in various biomedical applications including perfusion imaging, atherosclerotic plaque imaging, tumor detection and cell tracking. Although MPI is an imaging technique with high temporal and spatial resolution, it is sensitive to random noise and background noise due to the imperfection of hardware, resulting in low signal-to-noise ratio in the reconstructed images. Moreover, when the MNPs samples are relatively close (≤ 10mm) in the field of view and the concentration difference reaches more than 8 times, the dynamic range of the MPI imaging system is limited by the conversion accuracy of the analog-to-digital converter, which makes it difficult to accurately quantify the distribution of low-concentration MNPs in the same field of view.

By addressing the aforementioned problem in MPI, a feature fusion-based image denoising algorithm has been developed for MPI image denoising. In order to effectively denoise MPI images of MNPs samples with different concentrations in the field of view, we developed a deep learning model with high denoising accuracy, the ability to preserve image details after denoising, and the robustness to different levels of noise. The proposed method has been verified both on simulated phantoms and images acquired by a commercial MPI scanner.

The main contributions of this thesis are summarized as follows:

1. To simultaneously achieve noise removal and shape details preserving for MPI imaging of MNPs samples, we proposed a content learning network based on multi-level feature fusion (Multi-level Feature Fusion Network, MFFNet), which achieved efficient fusion of features of different levels in the same scale space through deconvolution and attention mechanisms. Fused features were used to perform regression prediction on the ground truth image. Experimental results showed that MFFNet can effectively suppress the noise in MPI images.
2. To address the limited dynamic range of MPI imaging systems and the impact of noise on MPI image quantitative accuracy, we further developed a Content-Noise Feature Fusion Neural Network (CNFFNet) based on MFFNet for MPI image denoising. CNFFNet simultaneously learns MNPs distribution features and noise distribution features of raw MPI images and fuse them efficiently. CNFFNet use the fused features to perform regression prediction on the ground truth image. Experimental results showed that CNFFNet outperforms other comparative methods in denoising accuracy, structures preservation and noise reduction at different noise levels. In addition, CNFFNet can effectively improve the quantitative accuracy of low-concentration MNPs samples in the same field of view.