Abstract: High-resolution Magnetic Resonance（MR） images can provide a clearer anatomical image to facilitate early diagnosis of the disease. However, the inherent defects of medical imaging system make the acquisition of high-resolution medical images face many problems, and one of the methods to solve these problems is to use the super-resolution reconstruction technique. In view of the problem of medical image super-resolution reconstruction, a feedforward full connection convolution neural network is designed, which includes five-layer convolution layer and five residual blocks, and the loss function based on feature is also used to solve the problem caused by mean square error loss function which cannot meet the human visual sense. This method realizes the image 4 times magnification reconstruction in the network and avoids the checkerboard artifacts, which are often occurred when using deconvolution layers to up-sample images. The effectiveness of the method is verified by experiments, and both the visual and numerical results are improved.

Key words: medical image, super-resolution reconstruction, convolution neural network, feature-based loss

摘要： 高分辨率的磁共振图像可以提供更加清晰的解剖图像，从而促进疾病的早期诊断。但是医疗成像系统的固有缺陷，使得高分辨率医学图像的获取面临许多问题，解决这类问题的方法之一就是使用超分辨率重建技术。针对医学图像超分辨率重建问题，设计一个前馈全连接卷积神经网络，网络包括五层卷积层和五个残差块，并且使用基于特征的损失函数，解决了使用均方误差损失函数不符合人视觉感的问题。该方法在网络内部实现图像4倍放大重建，避免了使用反卷积层上采样时出现的棋盘伪影。通过实验验证了方法的有效性，在视觉和数值结果上都有所提高。

关键词: 医学图像, 超分辨重建, 卷积神经网络, 特征损失