Study Of Fast Multi-channel Magnetic Resonance Imaging Method Based On Complex Convolutional Neural Network

Magnetic resonance imaging is a non-invasive medical imaging method without ionizing radiation,which has now been widely used in clinical auxiliary diagnosis.However,in practical applications,the application of this technology is limited due to the slow speed of magnetic resonance imaging and the prone to motion artifacts.The parallel magnetic resonance imaging method uses multiple receiving coils to collect magnetic resonance signals at the same time,which speeds up the magnetic resonance imaging speed.However,in the case of under-sampling,the obtained images will have aliasing artifacts.Parallel magnetic resonance reconstruction algorithms can use spatial sensitivity information to calculate the missing phase encoding information for image reconstruction,but when the acceleration factor is large,the quality of the image reconstructed by these algorithms will be significantly reduced.Deep learning is an automatic learning process that uses deep neural networks to perform feature representation,and it has been gradually applied in image recognition,image segmentation and other fields.In recent years,researchers have begun to apply deep learning to fast magnetic resonance imaging,and the most representative of these is the fast magnetic resonance imaging method based on convolutional neural networks.Since the originally collected magnetic resonance image data is complex number data in k-space,the image after Fourier reconstruction is also a complex number image.In the early stage of the research,researchers mainly used real-number convolutional neural networks for magnetic resonance imaging research,but this approach would lose phase information.From the phase information,important information including blood flow velocity,blood flow rate,and quantitative magnetization map can be obtained.Therefore,the use of complex convolutional neural networks for rapid magnetic resonance imaging research is of great significance.This article mainly studies the fast multi-channel magnetic resonance imaging method based on complex convolutional neural network.The research content mainly includes parallel magnetic resonance imaging algorithms SENSE and GRAPPA,and PCU-Net and C-RAKI methods based on complex convolutional neural networks.The specific research content is as follows:(1)The principles of SENSE and GRAPPA algorithms are studied.The SENSE algorithm uses the sensitivity distribution map to expand the aliased image in the image space,and the GRAPPA algorithm uses the automatic calibration signal line and the collected data to reconstruct the missing k-space data and reconstruct the image in the k-space.Experimental results show that the image quality reconstructed by these two algorithms is better in the case of a lower acceleration factor,but the image quality reconstructed by the two algorithms will be significantly reduced in the case of a high acceleration factor.(2)The fast multi-channel magnetic resonance imaging method based on PCU-Net network is studied.This method is mainly based on U-Net network and multi-channel complex module to study fast multi-channel magnetic resonance imaging method.On the basis of the PCU-Net network,the network structure is improved,and a PCAU-Net network with an asymmetric network structure is proposed.This network reduces the network scale in the decoding part.Due to the large scale of these two network parameters and large memory consumption during training,network training can be carried out by cyclically importing network parameters.The experimental results show that the reconstruction quality of the PCU-Net network is significantly better than that of the SENSE and GRAPPA parallel magnetic resonance imaging algorithms in the case of higher acceleration factors.The PCAU-Net network has achieved similar reconstruction quality as that of the PCU-Net network,and reduces the model complexity and training time.(3)The fast multi-channel magnetic resonance imaging method based on C-RAKI network is studied.RAKI is a convolutional neural network used for k-space interpolation and reconstruction,which uses the real part and the imaginary part as two real channels for feature extraction.The main advantage of this method is that it does not need any prior information,it uses its own ACS data for training without any additional training samples.It can non-linearly estimates missing k-space data from the collected k-space data,which improves noise immunity ability.This paper further improved the RAKI method,and proposed a C-RAKI method based on the multi-channel complex convolution module.The calculation of the multi-channel complex convolution module can retain the mapping relationship between the real part and the imaginary part,thereby improving the ability to extract phase information features.The experimental results show that the reconstruction quality of the RAKI method is significantly better than that of the SENSE and GRAPPA parallel magnetic resonance imaging algorithms in the case of higher acceleration factors.Compared with RAKI,C-RAKI has advantages in reconstructed images and image quantification indicators.