Variable Augmented Network For Invertible Magnetic Resonance Coil Compression

As an important imaging technology for clinical medical diagnosis,magnetic resonance imaging is widely used for various disease diagnosis and physical examination screening.However,due to its own physical characteristics,clinical data acquisition usually requires relatively long scanning time.Parallel imaging technology uses multi-channel coils to acquire MRI data simultaneously,which improves the speed of MRI.But the increasing number of coils multiplies the amount of data storage and makes the reconstruction phase time-consuming.Coil compression provides an effective way to alleviate these problems.Coil compression refers to compressing the data acquired by multiple receiving coils into fewer virtual coils by some suitable transformation.The classical coil compression methods need to consider the sampling method and the boundary threshold of effective information and noise,which leads to the lack of flexibility and the loss of more amount of information.To address the above problems,this study proposes an invertible magnetic resonance coil compression method based on variable augmented network in combination with deep learning.The main contributions are as follows:(1)For coil compression method on image domain,this study calculates the square root of the sum-of-square for the network input and network output separately,and uses the difference between them as the loss function.This study directly uses the calculation result of the network input as the label,and does not need to collect additional data corresponding to a small number of coils as label data,avoiding the difficulty of pairwise data set collection.Also,this loss function does not need to consider the sampling pattern and the threshold boundary delineation of effective information and noise,which has higher flexibility and less information loss.Secondly,this study uses a variable augmentation strategy by viewing the network output results as several sets of virtual coil data with a small number of coils.After that,averaging is performed to obtain the final compression results,and the correlation between multi-channel coils is used to make the compression results more excellent while implicitly reducing the dimensionality.Finally,this study utilizes the dual loss function of the invertible network,which not only achieves the recovery of the compression results to the original data,but also constrains the positive output of the network in this way to make the compression effect more stable.The experimental results show that this study has greater advantages over the classical coil compression methods,both in terms of visual effects and quantitative metrics.(2)To make the present method compatible with clinical medical applications,the coil compression method on the K-space domain is further explored in this study.Due to the large dynamic range of the K-space value domain,the model performance will be very poor if the model is trained directly on the K-space domain.Therefore,this study converts the K-space data into image-domain data for the calculation of the loss function,which enables the network to better extract image features while retaining the advantages of coil compression on image domain.The experimental results show that this study still has a good compression effect with different acceleration factors and sampling methods,and consumes less time.(3)For each different data type and coil configuration,this study requires separate training of the model for the corresponding configuration,which consumes a lot of time.To address this drawback,this study also investigates in model generalizability.First,this study trains the network using different anatomical structures,and the trained model can be compressed for different anatomical structures;secondly,this study supplements the all-zero matrix during data preprocessing to keep the number of channels consistent for all training data,and the trained model can be compressed for different number of coils;finally,this study combines different loss functions,and the trained model can compress the data to results with different number of coils.The experimental results prove that this study still has a good compression effect after model generalization.