Deep Learning-based Research On The Structure Of Human Myocardial Fibers In Vivo

The myocardial fiber structure is closely related to cardiac function.The investigation of myocardial fiber structure has important theoretical significance and clinical value in explaining the causes and early diagnosis of various cardiovascular diseases.At present,diffusion magnetic resonance imaging(d MRI)is the unique technique that can detect the structure of in-vivo myocardial fibers without injury and radiation,but this imaging technology is particularly sensitive to motion.During the acquisition process,the free-breathing,the periodic heart beating and the physiological motion of organs will reduce the signal-to-noise ratio of the diffusion weight image(DWI)and cause signal loss.Therefore,from which,diffusion tensor image(DTI)cannot be accurately reconstructed due to signal loss,which seriously impact on the measurement of microstructural characteristics of the in-vivo human cardiac fiber.Considering the superiority of deep learning in image restoration and reconstruction,this work intends to use deep learning models to compensate the signal loss in DWI caused by motion and noise,and then from the restored DWI,the DTI can be reconstructed accurately and therefore the cardiac fiber structure can be correctly estimated.The detailed works are as follows:(1)Based on the end-diastolic in-vivo cardiac DWI data acquired at multiple TDs,this paper designs an unsupervised dense encoder-decoder network(DEFD-net)to compensate the signal loss in DWI.This network uses the encoder to construct the feature map of DWI,then uses the fusion strategy to fuse the DWI features at multiple TDs,and finally reconstructs the fused DWI features through the decoder.Experimental results show that DEFD-net can effectively reduce the signal loss caused by freebreathing motion in in-vivo cardiac DWI,and can describe the cardiac fiber structure accurately.(2)Based on in-vivo cardiac DWI of end diastole,systole,and affected by organ motion,this paper proposes a novel motion compensation method WSCNN(Wavelet Scattering CNN).This method uses wavelet scattering to extract multi-directional and multiple-scales invariant features,such features are robust to the motion and noise.Then mapping the fused scattering features to DWI images by designing CNN to achieve motion compensation.Compared with the existing methods,the proposed WSCNN method can not only compensate for the signal loss caused by free-breathing,but also compensate for the signal loss caused by organ motion,from which,one can obtain in-vivo cardiac fiber structure with high quality.