Investigation On Deep Learning-based Reconstruction Method For Overlapping-echo Quantitative MRI

Magnetic resonance imaging(MRI)has been widely concerned and rapidly developed because of its harmlessness to human body and unique advantages in tissue imaging.However,most of the current magnetic resonance imaging techniques are qualitative parameter-weighted imaging.For qualitative imaging,the images obtained from different instruments and different imaging sequences may be different.Quantitative magnetic resonance parametric imaging can effectively-solve this problem.However,conventional quantitative imaging requires long scan time to complete multiple data acquisitions,which makes it sensitive to motion artifacts.Many methods have been proposed to accelerate quantitative magnetic resonance parametric imaging,but the scan time is still not short enough for dynamic imaging.Recently,our group has proposed an overlapping-echo quantitative T2 mapping method which can synchronously acquire multiple echo signals with different T2 weighting in a single scan.This method reduces the scan time to millisecond level.Quantitative T2 maps can be reconstructed from the acquisition data via deep learning.In this thesis,the reconstruction method is further studied in order to achieve more accurate and higherquality reconstruction,and provide the basis for the follow-up reconstruction of other quantitative magnetic resonance parametric maps.This thesis is mainly composed of the following four parts:1.The basic principle of MRI is briefly introduced.The principle of MR T2 imaging and the conventional quantitative T2 mapping methods are described,including spinecho,spin-echo echo planar imaging(EPI)and multi-echo EPI.The applications and the significance of developing new methods for fast quantitative T2 mapping are set forth.2.The principle of overlapping-echo quantitative T2 mapping,the reconstruction methods for quantitative T2 maps based on optimal separation algorithm and deep learning are elaborated.The deep learning and convolution neural network are briefly introduced,where the residual network used in overlapping-echo quantitative T2 imaging reconstruction is focused.3.A new method based on U-Net is proposed for reconstructing quantitative T2 maps.Compared with the residual network-based method,the new method makes use of the down sampling process of U-Net to enlarge the receptive field of the convolution kernel.At the same time,jump connection is used in the network to enhance the utilization of input information.The results of experiments on numerical brain phantom,water phantom and human brain show that the quantitative T2 maps reconstructed by the new method has smaller errors,and the texture details are closer to the reference images.Moreover,B1 field maps can be obtained4.A detailed comparative analysis is carried out on the factors that influence the quality of reconstructed T2 maps,such as signal-to-noise ratio,types and size of network training samples,loss function.It is found that the noise level of training samples,the matching degree between the features of the samples and the measured data,and the loss function have significant impacts on the reconstruction results.To achieve good reconstruction results,they should be set reasonably according to the reconstruction target.A sufficient number of training samples is conducive for the network to learning mapping relationship,and the scattered echo-focusing position helps to improve the texture quality of reconstructed images.