Research On Super-resolution Reconstruction Algorithm Of Diffusion Weighted Imagin

Diffusion Weighted Imaging(DWI)is currently the unique non-invasive technology that can detect diffusion information of water molecules in in vivo tissues,which has been widely used in many applications.However,due to the limitations of hardware equipment and acquisition time,the spatial resolution of DWI is not high enough,making it difficult to meet the requirements of clinical precise diagnosis.With the emergence of deep learning,using deep learning models to realize the superresolution reconstruction of DW images has achieved promising results,however,most deep learning models are based on the supervised learning,requiring a large amount of low-resolution(LR)and high-resolution(HR)data pairs to train the model,such dataset is difficult to obtain in practice.To address these issues,this work fully utilizes the properties of DW images and proposes two DWI super-resolution reconstruction algorithms based on semi-supervised learning,detailed as follows:(1)A semi-supervised DWI super-resolution reconstruction algorithm based on multiple references(Multiple References Super Resolution,MRSR)is designed.In MRSR,the prior information of multiple high-resolution reference images is migrated into a residual-like network to assist SR reconstruction of DW images to improve the texture quality of SR images.At the same time,a Cycle GAN-based semi-supervised strategy is used to train the network with 30% matched and 70% unmatched LR-HR image pairs.To evaluate the performance of the MRSR,we compare it against several state-of-the-art methods on HCP dataset in terms of the quality of reconstructed DW images and diffusion metrics of DT images.MRSR achieves the best performance,with the mean PSNR/SSIM of DW images being improved at least by 14.3% / 28.8% and1% / 1.4% respectively relative to SOTA methods,and with the fiber orientations(FO)deviating from the ground-truth about 6.28° on average,the RMSEs of fractional anisotropy(FA),mean diffusivity(MD),axial diffusivity(AD)and radial diffusivity(RD)being 3.0%,4.6%,5.7% and 4.5% respectively relative to the ground truth.This work validates the effectiveness of multiple references,Cycle GAN-based semisupervised learning strategy and the proposed network structure through the ablation studies.(2)A semi-supervised DWI super-resolution reconstruction algorithm based on multiple directions(Multiple Directions Super Resolution,MDSR)is designed.The MDSR model integrates information from low-resolution DW images along multiple diffusion gradient directions to achieve better texture quality.Furthermore,a semisupervised learning strategy based on reconstruction consistency is used to train the model with 50% labeled data.MDSR achieved better performance than the supervised state-of-the-art models on HCP dataset.The PSNR/SSIM of DW images has increased by at least 1.1%/0.4%,and the results of fiber tracking(FT),fiber orientation,fractional anisotropy,mean diffusivity,axial diffusivity,and radial diffusivity are closest to the HR reference.The ablation study illustrates the effectiveness of MDSR and the corresponding training strategies.