The Research On Magnetic Resonance Imaging Modal Transfer Technology Based On Generative Adversarial Network

Recently,more and more attention is drawn to brain imaging technology in medical field.Among them,magnetic resonance imaging(MRI)plays a vital role in clinical diagnosis and lesion analysis of brain diseases.Different sequences of MR images provide more comprehensive information and jointly help doctors to make accurate clinical diagnoses,however,its costs are particular.In order to overcome these limitations,MRI modal transfer technology has been widely applied.By learning the mapping relationship between different modal MR images,it is possible to reconstruct MRI images from one modality to another,providing more auxiliary information for clinical diagnosis by doctors.The commonly used MRI modal transfer technology involves the conversion of one modality image to another in a particular domain.In clinical diagnosis,MR images are obtained by scanning with various pulse sequences.It is more conducive to synthesize the missing modal data by utilizing the feature information from these modalities and the target modalities.Furthermore,image-to-image synthesis in the medical field is generally divided into supervised and unsupervised learning methods.Supervised learning methods require labeled datasets,which are typically difficult to obtain.To address the problem of multimodal image transfer,we introduce a self-supervised learning cycle-consistent generative adversarial network(BSL-GAN)for brain imaging modality transfer.The framework constructs multi-branch input,which enables the framework to learn the diversity characteristics of multimodal data.In addition,their supervision information is mined from large-scale unsupervised data by establishing auxiliary tasks,and the network is trained by constructing supervision information,which not only ensures the similarity between the input and output of modal images,but can also learn valuable features for downstream tasks.The experimental results show that BSL-GAN has clearer organizational details and contours in multimodal image synthesis compared to other multimodal transfer models.Furthermore,to address the problem of data pairing,we propose an unsupervised learningbased Generative Adversarial Network with Adaptive Normalization(AN-GAN)for synthesizing T2-Weighted MR images from rapidly scanned DWI MR images.Different from the existing methods,the deep semantic information is extracted from the high-frequency information of the original sequence images,then added to the feature map in deconvolution layers as a modality mask vector.This image fusion operation results in better feature maps and guides the training of generative adversarial networks.Then we introduce adaptive normalization,a conditional normalization layer that modulates the activations using the fused feature map.Experimental results show that our method in synthesizing T2 images has better perceptual quality and greater details than other state-of-the-art methods.