Diffusion MRI is an important technology for detecting human brain nerve pathways,aiding in neuroscience and clinical diagnosis.However,the Multi-ShellMulti-TissueConstrainedSphericalDeconvolution(M-CSD)method,which i...Diffusion MRI is an important technology for detecting human brain nerve pathways,aiding in neuroscience and clinical diagnosis.However,the Multi-ShellMulti-TissueConstrainedSphericalDeconvolution(M-CSD)method,which is a significant technique for reconstructing thefibre orientation distribution func-tion(fODF),requires multishell data with a considerable number of gradient direc-tions to achieve high accuracy.As multishell data are not easy to acquire,the Single-Shell Single-Tissue CSD(S-CSD)suffers from the Partial Volume Effect(PVE).It would be more convenient if we could use single-shell data to reconstruct better fODFs.We propose a novel method that utilizes the spatial structure and anisotropy of dMRI data through a spherical convolution network.We reduce the need for high-quality data by utilizing b=1000 s/mm2 with 60 gradient directions or even less.Our results show that our method outperforms the traditional S-CSD when compared to the M-CSD results as our gold standard.展开更多
基金This work was supported by the Natural Science Foundation of Heilongjiang Province(LH2021F046).
文摘Diffusion MRI is an important technology for detecting human brain nerve pathways,aiding in neuroscience and clinical diagnosis.However,the Multi-ShellMulti-TissueConstrainedSphericalDeconvolution(M-CSD)method,which is a significant technique for reconstructing thefibre orientation distribution func-tion(fODF),requires multishell data with a considerable number of gradient direc-tions to achieve high accuracy.As multishell data are not easy to acquire,the Single-Shell Single-Tissue CSD(S-CSD)suffers from the Partial Volume Effect(PVE).It would be more convenient if we could use single-shell data to reconstruct better fODFs.We propose a novel method that utilizes the spatial structure and anisotropy of dMRI data through a spherical convolution network.We reduce the need for high-quality data by utilizing b=1000 s/mm2 with 60 gradient directions or even less.Our results show that our method outperforms the traditional S-CSD when compared to the M-CSD results as our gold standard.
