Research On Self-supervised Diffusion Weighted Image Denoising Algorith

Diffusion weighted(DW)magnetic resonance imaging is currently the unique non-invasive technique for detecting diffusion information of water molecules in in vivo tissues.However,due to the limitation of imaging nature,DW images are severely affected by noise during the acquisition process,resulting in low signal-to-noise ratios,which influences the estimation accuracy of the diffusion metrics and makes it difficult to accurately reflect the fiber structure.Therefore,investigating the effective denoising algorithms for DW images is of great significance for their subsequent applications.Although the conventional and deep learning-based denoising methods have shown the potential to deal with effectively the noise in DW images,most of them are redundant information dependent or require the noise-free images as golden standard.To deal with these problems and considering the intrinsic properties of DW images,this work proposes two self-supervised learning models that can effectively remove noise from single b-value and multiple b-value DW images.The detailed research contents are as follows:(1)For denoising single b-value DW images,a self-supervised structural similarity-based edge-weighted convolutional neural network(SSECNN)is designed in this work.Considering that the DW images acquired along different diffusion directions have structural similarity,and the noise in these DW images is independent and identically distributed,the structural similarity-based matching algorithm is proposed to search for the most similar DW images.Such similar noisy DW image pairs are then used as the input and target of the denoising network SSECNN,which consists of several convolutional and residual blocks.Through the self-supervised training with these image pairs,the network can restore the clean DW images.In addition,to avoid the over-smoothing problem,this paper design a novel edge-weighted loss which enables the network to adaptively adjust the loss weight of each pixel with iterations and therefore to improve the detail preserve ability of SSECNN.Through comparisons and analyses with various denoising methods on both simulated and acquired datasets,it is demonstrated that SSECNN is effective in the DW image denoising task,which can remove DW image noise while preserving detail information as much as possible.(2)For denoising the multiple b-value DW images,this work proposes a novel convolution self-adapting denoising network(CSDNet).It can adaptively remove the noise of DW images with different b-values using the same model.CSDNet first processes the data using the optimal diffusion gradient direction matching algorithm and diffusion tensor model to generate multiple subsets and pseudo-labeled data with consistent contrast for network training.The noise level of DW images with different b-values is estimated using the noise level estimation sub-network(CNN_E).After generating the estimated noise level features,they are input to the denoising sub-network(CNN_D)together with the original DW images for noise removal.In this paper,a compound loss function is designed to constrain the training of CSDNet.Finally,the averaging method is applied to the outputs of multiple subsets to obtain the final denoising result.The experimental results show that the proposed method can effectively remove noise from multiple b-value DW images and accurately reconstruct fiber structure from the denoised DW images.Compared with other methods,the proposed method achieves the best qualitative and quantitative evaluation.