Virtual Coil Augmentation Technology For Magnetic Resonance Imaging Based On Invertible Network

Magnetic resonance imaging is a medical imaging modality that creates images with adjustable soft tissue contrast for anatomical and functional evaluations and plays a pivotal role in clinical applications.However,one of the major technical bottlenecks of MRI is that it takes a relatively long time to capture all the data needed for imaging,which cannot be met for rapid imaging requirements such as cardiac imaging,dynamic imaging and functional brain imaging.With the continuous development of MRI scientific research and clinical applications,MRI methods with higher image quality and shorter scanning time have become the frontier of research.In this thesis,a deep learning-based virtual coil augmentation technology（VCA）is proposed to address the problem of how to improve the quality of magnetic resonance.This algorithm utilizes the powerful nonlinear mapping capability of the generation models in image processing to generate multi-channel MR images.Without modifying the parallel imaging reconstruction algorithm,the high dimensional information formed by coil augmentation is used as priori information for parallel imaging to improve the reconstruction performance of parallel imaging.The main contributions of this article are as follows:（1）An invertible network-based method for coil augmentation of image domains（VCA-I）is investigated.The data preprocessing process employs variable enhancement techniques that can satisfy the requirement of invertible networks to make the output dimension consistent with the input dimension,and the use of variable enhancement techniques can enhance the correlation between images.The real part and the imaginary part of the multi-channel complex MR images are stacked into two channels as the input of the network to reconstruct the multi-channel complex MR images,thus preserving the phase information features of the images.To improve the robustness and stability of the network,thesmooth_L1loss function is used for bidirectional training.In the forward process the sum-of-squares is performed separately for the labeled image and the network forward output,and the difference is solved for the loss function,and the same operation is performed for the labeled image and the network inverse output in the inverse process.（2）An invertible network-based coil augmentation method in the k-space domain（VCA-K）is investigated.Based on the idea of VCA-I variable augmentation and channel stack replication,the input and output of VCA-K come from the measured data in k-space.Since magnetic resonance k-space data is different from natural images with a large dynamic range of amplitudes,the model cannot be trained well by calculating the objective function directly in the k-space domain,hence the algorithm performs the calculation of the loss function in the image domain to enable better feature extraction during the network training.（3）The experimental results show that VCA-I and VCA-K can achieve different coil numbers of augmentation.In addition,the reconstruction performance of parallel imaging with coil augmentation is better than that without coil augmentation under different acceleration factors and under-sampling patterns.