Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (12): 2150-2160.DOI: 10.3778/j.issn.1673-9418.2004071

Previous Articles    

Fast Brain MRI Reconstruction Using Deep Iterative Convolutional Neural Network

DU Nianmao, SONG Wei   

  1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2020-12-01 Published:2020-12-11

深度迭代卷积神经网络的快速脑部MRI重建算法

杜年茂宋威   

  1. 江南大学 人工智能与计算机学院,江苏 无锡 214122

Abstract:

Human brain MRI is usually multi-slice, and there is data redundancy between adjacent slices. Deep learning has become a powerful tool in the field of undersampled MRI reconstruction. However, the current reconstruction algorithms based on deep learning are mainly for a single MRI image. In order to make full use of the data redun-dancy in brain MRI data and obtain higher reconstruction quality and acceleration factor, a deep iterative convolu-tional neural network (DICNN) is proposed. In each iteration, a bi-directional convolution module (BDC) is used to explore the data redundancy between adjacent slices, and then a 2D convolution module (refine net, RNET) is used to further explore the data redundancy within a single MRI slice. Simulation experiments on a single-coil brain MRI dataset show that the proposed algorithm is better than the algorithm based on a single MRI image under different undersampling factors. This method can not only effectively make use of the data redundancy between brain MRI slices and recover more tissue structure details, but also meet real-time MRI reconstruction at a speed of 49 slices per second.

Key words: brain magnetic resonance imaging (MRI), deep learning, image reconstruction, convolutional neural network (CNN)

摘要:

人体脑部MRI通常是多切片的,并且相邻切片间存在数据冗余。深度学习已经成为欠采样MRI重建领域的有力工具,然而目前基于深度学习的重建算法主要是针对单幅MRI图像。为了充分利用脑部MRI数据中的数据冗余,以获取更高的重建质量与加速因子,提出了一种深度迭代卷积神经网络(DICNN)。在每次迭代中,首先使用双向卷积模块(BDC)探索相邻切片间的数据冗余,然后用2D卷积模块(RNET)进一步探索单幅MRI切片内部的数据冗余。在单线圈的脑部MRI数据集上的仿真实验表明,提出的重建算法在不同欠采样因子下的重建效果优于基于单幅MRI图像的重建算法。该方法不仅能够有效地利用脑部MRI切片间的数据冗余,恢复更多的组织结构细节,还能进行实时的MRI重建,速度可达每秒49张。

关键词: 脑部核磁共振成像(MRI), 深度学习, 图像重建, 卷积神经网络(CNN)