Accelerating Quantitative Magnetic Resonance Imaging Based On Deep Learning

After decades of development,magnetic resonance imaging(MRI)technology has been widely used in the world.However,MRI is often used for qualitative or weighted analysis of diseases.For example,when a tissue is compared with its surrounding environment,the image intensity is usually used to distinguish it,rather than specific quantitative analysis.Therefore,MRI may not be able to accurately show the severity of the disease.In recent years,quantitative magnetic resonance imaging(qMRI)methods that can visualize the ultrastructure of human tissues,such as T2* Mapping,have been widely used in clinical and neuroscience fields.The T2* Mapping technique not only visualizes the microstructure and the biochemical changes in the cartilage matrix,but also provides an early warning of morphologic damage to the knee tissue.In addition,the T2* Mapping technique is particularly sensitive to changes in the water content,collagen composition and tissue anisotropy of human cartilage,and can achieve accurate measurement of biochemical parameters that have important diagnostic basis for osteoarthritis(OA).Therefore,the T2* Mapping technique can be used as a meaningful assessment tool for the diagnosis and follow-up of cartilage abnormalities.T2* Mapping technique traditionally detects T2* decay in multiple TE intervals through gradient echo(GRE)sequences,that is,a series of images with different echo times are fitted according to exponential decay,and T2* relaxation time is calculated.This acquisition method of time is very long,so people put forward a variety of ways to accelerate the acquisition of T2 * Mapping technology:(1),such as the SENSE and GRAPPA algorithm,through multi-channel coil undersampling signal collection,and then the collected signal data redundancy information as prior knowledge is used to restore the whole k-space signal,finally reconstruct MR image and T2 * Map.Method(1)usually increases the cost and complexity of the imaging system.(2)the k-space undersampling technique is adopted to reconstruct MR images and T2* Map from the undersampled data by compressed sensing,sparse or low-rank conditions.Method(2)reduce the time required by T2* Mapping technique by reducing the collection process of k-space data.However,the introduction of undersampling technology will lead to the inevitable artifacts and noise in the image,which will finally affect the accurate calculation of T2*.The purpose of this study is to remove artifacts and noise in under-sampled images through deep learning,and achieve the goal of the T2* Mapping process based on undersampled data.The main work of this study is divided into four steps:In the first step,full and undersampled k-space data of knee joint were collected through 3D UTE-Cones sequence,and then MR images were reconstructed using Regridding and ESPIRiT reconstruction algorithms.The second step is to input the MR image into a three-layer residual convolution network,to learn the mapping relationship between the multi-under-sampled MR image and the full-sampled MR image,and to train a model that can remove the artifacts and noise of the under-sampled image.In the third step,this model is used to optimize the undersampled MR image,and then the single exponential fitting method is used to perform the T2* fitting calculation of the optimized MR image,and finally a complete T2* Map image is obtained.The fourth step is to use qualitative and quantitative analysis methods to analyze whether the proposed method can achieve the goal of the T2* Mapping process based on undersampled data.In this paper,we realize the importance of quantitative magnetic resonance imaging in assisting doctors' decision making based on objective evidence,among which T2* Mapping technique can be used as a reliable tool to evaluate cartilage status at different stages in the process of OA disease.Aiming at the problem that undersampling technology can speed up T2* Mapping technology but will affect the accuracy of T2*Map,this paper proposes a deep learn-based image post-processing method.Through analysis,this method is superior to other conventional reconstruction algorithms in image quality and T2*Map accuracy,and there is no significant difference between the measured value of T2* obtained by this method and the measured value of T2* obtained by the full-sample MR data.Therefore,the proposed method can improve the accuracy of T2*Map based on undersampled data,and even realize the T2* Mapping process based on undersampled data.