Multimodal Synthesis Of MRI Knee Image Based On Hi-Net

Multiple imaging modalities of magnetic resonance imaging(MRI)can provide doctors with diagnostic information from multiple perspectives,but it is usually difficult to obtain a complete MRI multi-modal sequence due to the acquisition problem,so it is meaningful to synthesize other missing modal images from the existing partial modal images.Based on MRI images of human knee joint,MRI images of different modalities have different gray ranges and distributions,which makes the gray of the same tissue in different modalities vary greatly;And the texture information characteristics of the internal structure of the MRI knee image is very important,some tissues with the same pixel gray level can only be judged by texture and regional location.These differences in gray and texture information have a significant impact on the performance of multimodal synthesis tasks.Based on this,this thesis proposes an improved hybrid fusion network model for the multi-modal synthesis task of MRI knee images.Firstly,aiming at the lack of structural information in the synthesized image,a gradient detector is proposed to integrate the gradient information of the image on the basis of the original Hi-Net.The gradient information is respectively fused into the generator and the discriminator,so that the gradient information can also be reversely learned,the problem that the image structure content information is lost in the multimodal synthesis of the MRI knee joint image is effectively solved,and more details of the synthesis image are reserved.Secondly,in order to solve the problem of inconsistent gray distribution of the image,this thesis proposes a convolution neural network to learn the gray distribution information of the target modal image based on the idea of depth-separable convolution.Considering the complexity of the network and the training time of the model,two convolution layers are replaced by the depth-separable convolution used in the lightweight network model.The network is used to monitor the brightness,contrast and gray distribution of the image,and solve the problem that the gray difference between different modes affects the result of the synthetic image.Thirdly,in order to make use of the correlation between multi-modality more effectively,this thesis improves the feature fusion strategy in Hi-Net,and proposes the most suitable feature fusion strategy for the data set of MRI knee images.MFB module is used to adaptively weight different feature fusion strategies to improve the performance of the algorithm.Finally,the ablation experiments are carried out on the MRI human knee dataset.The experimental results show that each improved module can improve the image synthesis effect to a certain extent,which proves the effectiveness of the improved module.In addition,this thesis compares the use of two kinds of gradient information through experiments,and verifies the effectiveness of the proposed method.In addition,this thesis is based on MRI knee data sets from different devices,and compared with other mainstream medical image multi-modal synthesis algorithm,the proposed algorithm has obvious advantages in various indicators,which proves the effectiveness of the proposed algorithm.