文章摘要

文章信息/Info

Title:: Deep magnetic resonance imaging reconstruction based on K-space data

Author(s):: HUANG Min; GUAN Zhihui; ZHOU Dao; CHEN Junbo; College of Biomedical Engineering, South-Central University for Nationalities, Wuhan 430074,China

Keywords:: Deep magnetic resonance imaging; Image reconstruction; Convolutional neural network; Transfer learning; K-space

摘要:: 传统的核磁共振(magnetic resonance imaging,MRI)成像技术中，图像重建算法与脉冲序列和K空间采样轨迹等因素密切相关。深度MRI成像采用了全新的重建方法。本研究采用深度卷积神经网络W-net对数据样本进行学习，从欠采集的K空间数据快速重建出高质量的图像。采用迁移学习方法，优化原模型参数，提升模型对各方向扫描、含病灶（如肿瘤）的大脑，以及结构较简单的膝盖等MRI数据的泛化能力。对比不同欠采样率的K空间输入数据，分析模型性能；并添加数据更新层，改进模型结构。测试结果表明，改进后的模型重建质量更优，对病灶和小脑纹理细节的恢复更好。

Abstract:: In traditional MRI technology, image reconstruction algorithms are closely related to factors such as pulse sequence and K-space sample trajectory. Deep MRI imaging adopts a new reconstruction method. We adopted deep convolutional neural network W-net to learn data samples. High-quality images were rapidly reconstructed from under-sampled MRI K-space data. The transfer learning method was adopted to optimize the parameters of the original model, and enhanced the generalization ability of the model for MRI data from different orientation scan, brain with lesions (such as tumors), and knee with relatively simple structure. We analyzed the performance of the model by comparing the input data of different sampling rates. The model structure was also improved by adding the data update layer. The test results show that the improved model has better reconstruction quality. It can restore lesions and texture details of cerebellum better.