Research And Application On Synthesizing Magnetic Resonance Imaging Method Based On Deep Learning

Magnetic resonance imaging(MRI)is a widely used neuroimaging technique that can provide images of different contrasts(i.e.,modalities),providing doctors with diverse diagnostic information.Yet,acquiring multi-modal MR images is very time-consuming.Some modalities may be corrupted by a lot of noise or artifact,which could seriously affect the image quality.Therefore,it is of great significance to synthesize images of missing modalities or to restore modalities of poor quality,which could improve the diagnostic utility.For multi-modal synthesis of MR images,current methods learn a nonlinear intensity transformation between the source and target images,either via nonlinear regression or via deep learning.However,most of the existing algorithms only rely on mono-modality image for synthesizing MRI,suffering from the loss of structural details in synthesized images.Most of the algorithms based on deep learning are large-scale networks with many parameters,facing the problem of overfitting on small datasets.In view of the above problems,this article takes the task of synthesizing T2-weighted images as an example and proposes the following solutions:1.Based on deep learning algorithms,this paper proposes a small network—Dual-Channel Multi-Feature Fusion Network(DCMFF)that based on multi-modal MR information including T1-weighted image and down-sampling T2-weighted image.The dual-channel network and the dilated residual dense blocks are added to DCMFF,which could learn characteristics for each individual modality better and obtain multi-scale image information without losing image resolution and increasing the parameters of the network.In addition,a multi-feature fusion block(MFFB)is innovatively proposed in order to effectively utilize the correlation between multi-modal.On the IXI dataset and MSSEG dataset,for the case of 4x down sampling rate and 8x down sampling rate,our network DCMFF performs best by comparing with Unet and Dense_Unet.The parameters of our network is less than 1/5 of Unet,and there is no overfitting when it runs on small datasets.For visualizing results,our proposed network can synthesize high quality T2-weighted image with more details,and it can remove the artifact exists in the original down-sampling T2-weighted image.2.For the task of synthesizing T2-weighted images in this article,we test the datasets scanned by different devices to verify the generalization ability of the DCMFF network under different scanning devices and different magnetic field strengths.We conclude that:when the testing dataset and the training dataset come from different scanning devices and magnetic field strengths,the T2-weighted image synthesized by the DCMFF network can still maintain the contrast and structure information similar to the original T2-weighted images.The DCMFF network basically removes the artifact exists in the down sampling T2-weighted image,confirming that the DCMFF can synthesize T2-weighted image of good performance on other MRI datasets.