Magnetic Resonance Imaging (MRI) is a widely used medical diagnosis technique. However, the quality of MR image is affected by the noise which is caused by mechanical and environmental reasons during the MR image acquisition process. For decades, various kinds of methods including filtering approaches, transform domain approaches and statistical approaches have been applied to the MR image denoising problem, while there are also some drawbacks exiting in these methods such as arising undesirable change of texture and long computation time needed. In this paper, we proposed a novel MR image denoising method called DESN which is a neural network method and has a novel network architecture with well-designed loss function. In DESN, the convolutional neural networks itself is considered as a regularizer or image prior information for the inverse problems such as denoising. The network architecture of DESN is designed from the auto-encoder architecture, it has three main parts: encoder network for extracting low-resolution image features, decoder network for restoring high-resolution features and skip connections for transmitting abstract information from encoder network to decoder network. Besides, we also design a novel loss function which contains two main parts: data fidelity loss ($L_{fidelity}$), image quality penalty ($L_q$) and three loss terms: mean squared error term ($L_{MSE}$), image structure similarity term ($L_S$), image information entropy term ($L_{IE}$). We compare the performance of DESN with DIP and some state of the art denoising methods, and the performance of our network with different loss terms are also compared in three MR modalities. The comparative results show that DESN have the superior performance in generating high-quality MR image with enough edge and texture information.

磁共振成像 (MRI) 是一种广泛使用的医学诊断技术。然而，在MR图像采集过程中，由于机械和环境原因引起的噪声会影响MR图像的质量。几十年来，包括滤波方法、变换域方法和统计方法在内的各种方法已应用于MR图像去噪问题，但这些方法也存在一些缺点，例如会引起不希望的纹理变化和需要较长的计算时间。在本文中，我们提出了一种称为 DESN 的新型 MR 图像去噪方法，它是一种神经网络方法，具有新颖的网络架构和精心设计的损失函数。在DESN中，卷积神经网络本身被认为是去噪等逆问题的正则化器或图像先验信息。DESN 的网络架构是从自动编码器架构设计的，它包含三个主要部分：用于提取低分辨率图像特征的编码器网络、用于恢复高分辨率特征的解码器网络和用于从编码器网络向解码器传输抽象信息的跳过连接网络。此外，我们还设计了一个新颖的损失函数，它包含两个主要部分：数据保真度损失（用于恢复高分辨率特征的解码器网络和用于将抽象信息从编码器网络传输到解码器网络的跳过连接。此外，我们还设计了一个新颖的损失函数，它包含两个主要部分：数据保真度损失（用于恢复高分辨率特征的解码器网络和用于将抽象信息从编码器网络传输到解码器网络的跳过连接。此外，我们还设计了一个新颖的损失函数，它包含两个主要部分：数据保真度损失（${升}_{F\mathrm{一世}d\mathrm{电子}升\mathrm{一世}吨是}$), 图像质量惩罚 (${升}_q$) 和三个损失项：均方误差项 (${升}_{米秒乙}$)、图像结构相似项(${升}_{秒}$)、图像信息熵项(${升}_{\mathrm{一世}乙}$）。我们将 DESN 与 DIP 和一些最先进的去噪方法的性能进行了比较，并且我们还在三种 MR 模式下比较了具有不同损失项的网络的性能。对比结果表明DESN在生成具有足够边缘和纹理信息的高质量MR图像方面具有优越的性能。