| In the current medical imaging tasks,because the acquisition of magnetic resonance(MR)images is expensive and time-consuming,it is necessary to obtain multi contrast MR images with the help of technical means.Nowadays,it has become a hot topic to study the synthesis of magnetic resonance images through deep learning algorithm to improve the diagnostic efficiency.With the generation countermeasure network(GAN)playing an important role in the field of image processing in recent years,this thesis will mainly study the synthesis of different contrast MRI images combined with the method of generation countermeasure.In the existing one-to-one MR image synthesis methods,most of them only focus on the whole image,while the local details are ignored.Various derived generation antagonistic depth learning methods can basically obtain seemingly complete images in the field of MR image synthesis,but in fact,there are problems of blurred edges or reduced resolution,and the training of antagonistic objective function is prone to collapse.Based on the method of conditional generation countermeasure network(CGAN)and Cycle GAN,aiming at the problems of fuzzy synthetic image and unstable training,this thesis studies the synthesis of multi contrast MR image.The main research work is as follows:(1)A multi-contrast MR images synthesis method based on edge information retention is proposed,which is improved and innovated in many aspects under the guidance of the method proposed by CGAN.Firstly,the generator of this method is redefined as a structure composed of multiple codecs in series,which can explore more dimensional features by strengthening the depth of the network;Secondly,the method uses the edge feature information of the image to further promote the integrity of the edge of the tissue or structure in the synthetic image;Finally,the discriminator of Patch GAN structure is combined with spectral normalization constraint to stabilize the training process.Combined with the above methods,the generalization error is reduced by adding a regularizer to improve the synthesis effect of multi contrast images and obtain high-definition images.(2)A multi-contrast MR images synthesis method based on bidirectional generation is proposed.The method uses paired images to operate end-to-end at the image level,and the circular generation method is used to convert a given contrast MR image into another contrast MR image.Learning nonlinear mapping through bidirectional generation,while acquiring image features,the convolution module(ECM)of extended receptive field is used to enhance the acquisition of neighborhood information;In addition,the content enhancement network is proposed to further capture the image details.The experimental results show that the bidirectional generation network has good performance in SSIM and PSNR indexes. |
