计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (13): 228-236.DOI: 10.3778/j.issn.1002-8331.2304-0099

• 图形图像处理 • 上一篇    下一篇

融合注意力和空洞编码解码的3D-MRI超分辨率算法

张金迪,贾媛媛,祝华正,李洪碧,杜井龙   

  1. 1.重庆医科大学 医学信息学院,重庆 400016
    2.重庆科技学院 智能技术与工程学院,重庆 401331
  • 出版日期:2024-07-01 发布日期:2024-07-01

3D-MRI Super-Resolution Algorithm Fusing Attention and Dilated Encoder-Decoder

ZHANG Jindi, JIA Yuanyuan, ZHU Huazheng, LI Hongbi, DU Jinglong   

  1. 1.College of Medical Informatics, Chongqing Medical University, Chongqing 400016, China
    2.School of Intelligent Technology and Engineering, Chongqing University of Science and Technology, Chongqing 401331, China
  • Online:2024-07-01 Published:2024-07-01

摘要: 为了提高脑部磁共振成像(MRI)图像的空间分辨率,目前基于深度卷积神经网络(CNN)的超分辨率(SR)重建方法已经取得了显著的效果。为了提高重建质量,可以通过增加深度或宽度扩大网络的感受野,但会因参数量大、消耗资源多等问题导致模型难以训练;同时,网络对图像的中高频特征关注度不足,影响了重建质量。针对以上问题,设计了一种基于注意力和空洞编码解码的3D-MRI图像超分辨率重建网络。在低分辨率空间进行特征提取,通过对称连接的空洞编码解码器提取图像特征,并缓解棋盘效应;构建串联的通道-空间注意力模块,利用通道注意力获取各特征通道间的相互依赖关系,根据空间注意力自动学习不同位置的特征的权重,有效增强网络对中高频信息的学习;使用亚像素卷积将特征图上采样到目标图像尺寸,降低模型的内存消耗。基于公开数据集Kirby21和BraTS的实验结果表明,相较于传统SR算法和基于CNN的主流SR算法,所提出的算法在峰值信噪比(PSNR)、结构相似度(SSIM)和主观视觉效果上均优于对比算法。

关键词: 超分辨率重建, 注意力机制, 亚像素卷积, 空洞卷积, 3D-MRI

Abstract: In order to improve the spatial resolution of brain magnetic resonance imaging (MRI) images, the super resolution (SR) reconstruction method based on deep convolutional neural network (CNN) has achieved remarkable results. Increasing network depth or width usually can expand the receptive field of the network to improve the quality of reconstruction, but it is difficult to train the network due to the large parameter quantity and huge computing requirements. Meanwhile, the network does not pay enough attention to the medium and high frequency features of the image, which affects the quality of reconstruction. To solve the above problems, this paper proposes a cascaded channels-space attention and encode-decode network for 3D-MRI image SR reconstruction. Firstly, feature extraction is carried out in the low resolution space, and the image features are extracted by the symmetric connected encoder-decoder with dilated convolution to alleviate the checkerboard artifacts. Secondly, a serial connected channel-space attention module is constructed to capture the interdependence between the feature channels by channel attention, and the spatial attention is increased to assign weight to information of different positions, effectively enhancing the network’s learning of medium and high frequency information. Finally, subpixel convolution is used to sample the feature map to the target image size and meanwhile reduce memory consumption. Experimental results on public Kirby21 and BraTS datasets show that the proposed method is superior in both quantitative peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and subjective visual quality when compared with traditional SR algorithms and mainstream CNN-based SR algorithms.

Key words: super-resolution reconstruction, attention mechanism, subpixel convolution, dilated convolution, 3D-MRI