Research On Adaptive Cross-domain Translation Of Multi-center Diffusion Magnetic Resonance Imaging

With the development of medical image processing technology,multi-center research becomes more and more popular.The acquisition of medical image data is a basic problem in medical image processing,multi-center research can effectively expand the data amount of medical image data by collecting image data jointly in several medical centers,which can dramatically increasing the statistical power,expecially in the study of some rare diseases.Diffusion magnetic resonance imaging usually have significant inter-site difference due to its sensitivity to collecting configurations.Thus,it is necessary to harmonize the diffusion magnetic resonance imaging data before further jointly analysis.In this work,we utilized the spheric harmonics representation and proposed a deep learning based harmonization method,the main contribution of this work can be summarized as follows:Firstly,we proposed a diffusion magnetic resonance image harmonization model based on three-dimensional convolution.A local unsupervised constraint is also implemented by introducing a Markovian discriminator to achieve an adversarial learning.Such constraint can induce the generator to pay more attention to local details,which is a suitable complement with the supervised constraint which pays more attention to the global information.Also,the unsupervised constraint can effectively alleviated the fuzzy problem in harmonization results.Since the medical image contains additional spatial information,the three-dimensional convolution used in our model can make use of this spatial information of the input image and effectively improve the inter-frame continuity of the harmonization results.Secondly,we propsed a Markovian discriminator weight allocation strategy based on the spatial distribution characteristics of brain regions in diffusion MRI.Traditional markovian discriminator uses the mean value of each patch's discriminative probability as the final probability of the whole image.However,since the brain regions always display in the center of image,the mean-value strategy is no longer reasonable.In this work,we used the brain region mask as additional information to introduce the spatial distribution characteristics of input image,and proposed a weight allocation strategy suitable for brain imaging.The proposed strategy can effectively utilize the prior spatial distribution information of the input image.Thus can further induce the generator to pay more attention to the harmonization of brain region.Thirdly,we proposed an deep supervised constraint based on the auxiliary brain region segmentation task.The classic segmentation model,3D U-Net,is introduced as an auxiliary loss network to segment the brain region in R0 feature,which has the clearest organizational structure among all rotational invariant spherical harmonic features.During the harmonization process,the pretrained 3D U-Net is fixed and introduced to the harmonization process as an auxiliary loss network.The segmentation map of the harmonization results is then constrained utilizing such loss network,to be specific,the segmentation map of the harmonization results should be closed to the segmentation map of target output.To achieve this,the extracted deep features of the harmonization results should be consistent to target output's deep features.Thus,this constraint is a supplement to original loss functions,which can further improve the harmonization quality of diffusion data.Experimental results in this work showed that the proposed deep learning harmonization method based on spherical harmonics representation can effectively reduce the inter-site difference of diffusion magnetic resonance imaging,and provide an effective solution for the integration of diffusion data in multi-center study.