毕业生知识库

磁粒子成像是一种新兴的分子影像技术。该技术利用磁纳米粒子（Magnetic Nanoparticles，MNPs）作为示踪剂，通过直接检测和量化MNPs在交变场激励下产生的磁化信号，利用重建算法实现对MNPs浓度的高灵敏和定量成像。在磁粒子成像中，x-space是一种常用的重建算法。x-space算法通过将接收到的MNPs响应信号从时域投影到空间域进行重建，重建得到的图像为MNPs浓度分布图和点扩散函数（Point Spread Function，PSF）的卷积。低信噪比、弱梯度场强以及小MNPs粒径等多种因素会影响PSF的质量和准确性，导致重建的图像出现模糊，从而使得成像分辨率降低。解卷积作为一种常用的后处理方式，难以有效去除PSF造成的模糊。此外，由于各方向梯度场强的不同和扫描轨迹的不均匀性，还会造成成像分辨率的各向异性的问题。

针对上述问题，本文提出了基于对比学习的深度学习方法，通过图像特征的对比获取更具判别性的特征表示，从MNPs浓度分布图中提取更多的细节特征，并保持原始图像结构信息的完整，从而实现磁粒子成像分辨率的提升。本文的主要研究内容和创新点归纳如下：

（1）在基于x-space的重建方法中，针对传统解卷积方法难以有效去除二维磁粒子成像中PSF带来的模糊的问题，本文提出了一种基于多层图像块特征对比学习的双向对抗生成网络模型（Dual Adversarial Network，DANet）。DANet中基于图像块特征的对比损失帮助模型保留了磁粒子成像图像的空间结构；双向对抗生成网络架构提升了特征提取的效果。在仿真数据及成像设备真实采集的数据上的实验表明，DANet在不同水平的噪声环境中有效地去除了模糊和噪声，提高了二维磁粒子成像分辨率。同时，不需要配对的数据集来训练模型。在仿真测试集上，DANet将磁粒子成像图像的峰值信噪比指标从8.12 dB提升至31.95 dB，结构相似性指数测量指标从0.38提升至0.88。

（2）针对扫描轨迹稀疏导致的三维磁粒子成像分辨率的各向异性的问题，本文提出了一种基于结构蒸馏对比学习的网络模型（Contrastive Structural Distillation Network，CSDNet）。CSDNet充分利用了二维各向同性教师网络提取的空间特征，并将其蒸馏至三维分辨率提升学生网络，有效地改善了扫描轨迹稀疏导致的各向异性问题。同时，CSDNet中基于图像特征的对比损失提升了空间特征的迁移效果；引入的空间注意力模块自适应地调整了二维网络在不同空间位置的关注度，提取了图像中轨迹稀疏区域的特征。实验结果表明，CSDNet在分辨率提升效果、对图像细节特征的保持能力、对不同程度噪声的鲁棒性三方面均优于RRDB、RCAN、EDSR和Conv-FNO三维图像分辨率提升模型。相较于在对比模型中表现最佳的Conv-FNO，CSDNet在峰值信噪比和结构相似性指数测量指标上分别提升了3.09 dB和0.32，均方根误差指标上降低了2.32。

关键词：磁粒子成像，分辨率提升，对比学习，结构蒸馏，空间注意力

Magnetic particle imaging (MPI) is an emerging medical imaging technique. This technology utilizes magnetic nanoparticles (MNPs) as tracers to directly detect and quantify magnetization signals generated by MNPs when subjected to alternating field excitation. By employing a reconstruction algorithm, MPI enables highly sensitive and quantitative imaging of MNPs concentration. In MPI, x-space is one of the commonly used reconstruction methods which projects the received magnetization signals from time domain to space domain to reconstruct image of MNPs concentration. The reconstructed native image can be modeled as a convolution of the magnetic particle concentration with a point-spread function (PSF). However, the PSF may be deteriorated by the low signal-to-noise ratio, the weak gradient field intensity, and the small size of MNPs etc., which can introduce blurring in the reconstructed image, leading to imaging resolution reduction. It is challenging to effectively remove the blur caused by PSF via the conventional deconvolution methods in MPI. In addition, the anisotropy of imaging resolution arises due to variations in the gradient field strength along different directions and the non-uniformity of the scanning trajectory.

To address the abovementioned challenges, this thesis proposes a novel deep learning approach based on contrast learning. By leveraging the power of contrast learning, this thesis aims to derive a more discriminative feature representation through the comparison of image features. It effectively captures the features in the reconstructed MPI images while preserving the structure, thereby enhancing the resolution of MPI. The main research contents and innovative points of this thesis are summarized as follows:

(1) To address the issue of ineffective deblurring in 2D MPI by traditional deconvolution, the thesis proposes a Dual Adversarial Network (DANet) based on contrast learning of multi-layer image patch features. The contrast loss based on image patch features in DANet enables to preserve the spatial structure of MPI images. The bidirectional adversarial generation network architecture can improve the feature extraction efficiency. The experiments on the simulation images and images acquired by a commercial MPI scanner show that DANet can effectively reduce blur and noise and improve the resolution of 2D MPI. Notably, DANet can improve resolution for 2D MPI without using paired datasets and achieve robust performances at different noise levels. Specifically, on the simulated test dataset, DANet improves the peak signal-to-noise ratio of MPI images from 8.12 dB to 31.95 dB and enhances the structural similarity index measurement from 0.38 to 0.88.

(2) To address the issue of anisotropic resolution in 3D MPI caused by sparse scanning trajectories, this thesis proposes a Contrastive Structural Distillation Network (CSDNet) based on structural distillation contrastive learning. CSDNet can effectively remove blurring in 3D networks while leveraging the spatial features extracted by a 2D isotropic teacher network to tackle the anisotropy issue arising from sparse scanning trajectories. Moreover, the contrastive loss based on image features in CSDNet can enhance the transfer effects of spatial features. The introduced spatial attention module can adaptively adjust the focus of the 2D network at different spatial positions, allowing for the extraction of features in sparsely scanned regions of the image. Experimental results demonstrate that CSDNet outperforms other 3D image resolution enhancement models such as RRDB, RCAN, EDSR, and Conv-FNO in terms of resolution improvement, preservation of original structures, and robustness to different levels of noise. CSDNet achieves a 3.09 dB improvement in peak signal-to-noise ratio, a 0.32 improvement in structural similarity index measurement, and a 2.32 reduction in root mean square error than Conv-FNO, which shows the best performance among the other 3D image resolution enhancement models for comparison.

Key Words: Magnetic particle imaging, Resolution improvement, Contrastive learning, Structural distillation, Spatial attention