Fig 1.
The schematic diagram of the new complex-domain DM-PCA based two-channel denoising procedure.
Fig 2.
Comparison of DWI denoising through filtering signals across nearest neighboring voxels and diffusion-matched voxels: The red dot in (a) shows a target voxel (displayed on top of mean DWI map), whose signals in 6-direction DWI scans are to be denoised. In many existing denoising methods, signals of nearest neighboring voxels in a patch (see b) are the input of a filtering procedure. In contrast, we identify a group of voxels that demonstrate very similar signal variation patterns along the diffusion dimension but are not necessarily neighboring (see c) for subsequent filtering procedures. Panels d, e and f show an input image, nearest-neighboring PCA produced image, and DM-PCA produced image, respectively. Residual maps obtained with nearest-neighboring PCA and DM-PCA methods are shown in panels g and h, respectively.
Fig 3.
A simulation study for comparing magnitude-domain DM-PCA and two-channel complex-domain DM-PCA in terms of the accuracy in ADC fitting: (a) Noise-free DWI data corresponding to b = 0, 200, 400 … 2200 (s/mm2). (b) DWI data affected by Rician noise. (c) DWI data denoised by magnitude-domain DM-PCA. (d) DWI data denoised by a two-channel complex-domain DM-PCA procedure. (e) Signal intensities of noisy DWI data (solid curve in orange) and the ground truth (dashed curve in blue). (f) Signal intensities of magnitude-domain DM-PCA produced data (solid curve in orange) and the ground truth (dashed curve in blue). (g) Signal intensities of complex-domain DM-PCA produced data (solid curve in orange) and the ground truth (dashed curve in blue). (h) Errors in ADC fitting for data with different SNR levels.
Fig 4.
A simulation study for comparing magnitude-domain DM-PCA and two-channel complex-domain DM-PCA in terms of the accuracy in fitting ADC values from parallel DWI data: (a) Images reconstructed with the 2xSENSE algorithm from noise-free under-sampled k-space data corresponding to b = 0, 200, 400 … 2200 (s/mm2). (b) Images reconstructed with the 2xSENSE algorithm from noisy under-sampled k-space. (c) SENSE-produced data denoised by magnitude-domain DM-PCA. (d) SENSE-produced DWI data denoised by a two-channel complex-domain DM-PCA procedure. (e) Signal intensities of noisy parallel DWI data (solid curve in orange) and the ground truth (dashed curve in blue). (f) Signal intensities of magnitude-domain DM-PCA produced data (solid curve in orange) and the ground truth (dashed curve in blue). (g) Signal intensities of complex-domain DM-PCA produced data (solid curve in orange) and the ground truth (dashed curve in blue). (h) Errors in ADC fitting for data with different SNR levels.
Fig 5.
FA maps obtained from high-resolution DWI images (0.85 mm3 voxel size), before and after DM-PCA based denoising, corresponding to different SNR levels in input data.
Fig 6.
(a) and (b) show coronal-plane mean DWI and FA maps, respectively, derived from high-resolution data after DM-PCA based denoising, with arrows indicating the left hippocampus. The corresponding zoom-in images shown in (c) and (d), respectively. The FA map derived from images without DM-PCA denoising is shown in (e). The coarse hippocampal structures revealed by the mean DWI map are highlighted in (f). Anatomic structures that can be identified from color-coded FA map are shown in (g) and (h). Region 1 in (g) corresponds to the dentate gyrus; Region 2 shows fibers that connect hippocampus and entorhinal cortex to other brain areas; Region 3 contains hippocampal CA1, CA2, and CA3.
Fig 7.
Application of DM-PCA denoising to human brain DWI data at conventional resolution (1.8 mm3 voxel size): Panels a and b compare one of the DWI images before and after DM-PCA denoising, respectively, for 4 of the participants. Panels c and d show the corresponding FA maps obtained from data before and after DM-PCA denoising, respectively.