Figure 3: Quantitative error for selected submissions (we show only the submission with the lowest errors). First row: MSE for isotropic upsampling task of the whole brain, white matter and gray matter. It can be observed that the competing top 5 submissions indicate quite similar errors. Second row: Mean squared error for anisotropic task of the whole brain, white matter and gray matter.

Figure 4: Error heat maps for the single TE of 80ms across b-values of 1000, 2000 and 3000 s/mm^2 per row. The columns depict different submissions from the best performing five ones. Submission 12 is cubic spline interpolation and the other 4 are deep learning based approaches

Figure.2 Comparison between DTI of the thoracic spinal cord using (upper row) full FOV PSF-EPI (8 shots), (lower row) rFOV ZOOM-EPI (1/2.6 FOV, single shot), and full FOV SS-EPI (R_PE=4). Detailed imaging parameters see Table 1, scan 3-5. Color-encoded FA map (cFA), non-DWI image (b0) and mean-DWI (mDWI) image are shown. For better comparison, the major structural boundaries extracted from the T2W reference were overlaid on the cFA using PSF-EPI. Additionally, the diffusivity maps using PSF-EPI are also demonstrated.

Figure.4 Comparison of EPI-based methods for distortion correction in thoracic DTI. From left to right: T2W TSE, 4-shot MUSE with full FOV (Table 1, scan 6), 4-shot MUSE+FOCUS with 1/2 FOV (scan 7), 8-shot PSF-EPI (scan 8), SS-EPI with full FOV and R_PE=2 (scan 9), SS-EPI+FOCUS with 1/2 FOV (scan 10). Upper: T2W TSE reference and non-DWI images. Lower: mDWI images.

Statistical descriptors derived from the 6D $$$\bf D$$$-$$$R_1$$$-$$$R_2$$$ distributions shown for a representative axial slice. Displayed are the median values of 100 per-bootstrap means $$$\mathrm{E}$$$, variances $$$\mathrm{V}$$$, and covariances $$$\mathrm{C}$$$. Bin-resolved fractions and means obtained by binning the parameter space to isolate WM ("thin"), GM ("thick"), and CSF ("big"). $$$\bf D$$$ is characterized in terms of the isotropic diffusivity $$$D_{iso} = (D_{||} + 2 D_\bot)/3$$$ and normalized anisotropy $$$D_\Delta = (D_{||} - D_\bot)/D_{iso}$$$.

Experimental data in three representative voxels. (A) $$$S_0$$$ map with labeled voxels: WM in the corpus callosum (red), cortical GM (green), and CSF in the frontal ventricle (blue). (B) Normalized signal $$$S$$$ versus acquisition index $$$n_{acq}$$$ (measured: black, fitted: colors correspond to the selected voxels). (C) Selected 2D projections of the 6D $$$\bf{D}$$$-$$$R_1$$$-$$$R_2$$$ nonparametric distributions obtained by the Monte Carlo inversion.

Figure 5 Mismatch of perfusion and diffusion after acute spinal cord injury. Maps of filtered axial diffusivity (fADC_||) and spinal cord blood flow (SCBF) are shown for three acutely injured animals with the labeling plane (red) indicated. Areas of decreased fADC_|| (green) or SCBF (red) were manually outlined and along with maps of lesion overlap mean lesion values reflect the extent of each contrast or overlap. Across three animals, the perfusion abnormality was clearly and consistently smaller than the diffusion abnormality.

Figure 2 Image quality of sagittal DW_prep-RARE in the spinal cord. Prominent artifacts exist in the EPI image (A) without diffusion weighting (b₀) that are not apparent in the RARE (B), noting these are from different animals. With b(perpendicular) =2000 s/mm², subtle artifacts were evident in both respiratory (C) and dual (D) gating conditions with both m1 (E) and m2 (F) compensation eliminating the artifacts. Data shown for n=4 animals obtained from ROIs within the spinal cord as shown (B).

Fig. 3. Demonstrate of noise removal pipeline (USD) on dMRI data in an ex vivo mouse brain (4b=0 images + 60 DWIs). The noise level in DWIs of b = [1, 2] ms/µm² was reduced dramatically after denoising, and the image residual maps have no anatomical structures. Further, the noise maps before and after re-normalization ($$$\hat{\sigma}$$$ and σ) are both smooth, and the number P of signal components in PCA domain is low. Finally, the histogram of image residual r is normally distributed and below the reference line of slop -1/2 in semi-log scale, indicating that USD only removes the noise.

Fig. 1. The pipeline of Universal Sampling Denoising (USD) for non-Cartesian k-space data, as detailed in Theory section. Here we take the radial trajectory as an example. The key idea in USD is to de-correlate the noise statistics in gridded/Cartesian k-space before applying any denoising algorithm in the image space. The noise covariance matrix in Cartesian k-space is determined by coefficients of density compensation and convolution with interpolation kernel. After denoising using MP-PCA algorithm, re-normalization of the denoised image recovers the original image contrast.

Explains the flow of Patch2Self Framework. Also depicts the self-supervised loss used in the J-invariant training.

Regressor Comparisons

Figure 4 Diffusion weighted imaging with spherical encoding at b = 4 ms/μm² in coronal view. A vastly higher contrast is observed between the cerebellar cortex and white matter using super-resolution reconstruction (right, contrast ratio 1.82) compared to direct high-resolution sampling (left, contrast ratio 1.06).

Figure 5 Signal retention using diffusion-weighted imaging with spherical encoding at b = 4 ms/μm² (upper left) and estimation of $$$f_{dot}$$$ (upper right) show agreement with cell-stained histology (lower right plot shows human brain histology from the BigBrain atlas [15]). As expected, regions of high signal correspond to the cerebellar cortex where granule cells are densely packed, whereas the white matter is suppressed by the spherical diffusion encoding (lower left for morphological reference).

B₁+ inhomogeneity and Gibbs ringing (A) Flip angle map (in degrees; top) and deviation from the nominal 90 degrees (bottom) (B) tSNR for MESMERISED with ESxMB=3x3 (top) for both STE and SE b0 volumes and PGSE with MB=2 (bottom). (C) Average b0 image MESMERISED STE volumes (top) and PGSE (bottom). Magnification of red box and line-plot at the red line (D) The Fraction of Sticks (FS) map for MESMERISED (ESxMB=3x3) STE volumes (top) and PGSE (bottom). Magnification of blue box and line-plot at the blue line.

Full coronal, sagittal and axial views of (A) absolute flip angle (B) deviation from the nominal 90 degrees, and for for PGSE (MB=2) volumes: (C) average b0 with line-plot at the red line, (D) b0 tSNR (E) FS map with line-plot at the blue line, (F) FS map uncertainty with line-plot at the blue line.

The incoherent k-q under-sampling illustrated. Here, the full k-space sampling involves using all the 4 shots (denoted using 4 colors) for every q-space point. In (a), an acceleration of R=4 is achieved by using under-sampled multi-shot trajectory. Specifically, only 1 shot (illustrated using bold lines) is used to sample a given q-space point. Different q-space points are sampled using differnt shots (denoted by the different colors). In (b), an acceleration of R=8 is achieved by skipping half of the points in q-space.

The orientation distribution (row 1) and the neurite density (row 2) computed using the NODDI model for the fully sampled case and the accelerated cases. The average value in the difference map computed with respect to R=1 is less than .06 and .04 and for ODI and NDI respectively for R=4,8.

Figure 2: Example of slice-to-volume reconstruction (SVR) in a fetus (28 wPMA) with severe head motion. The images show the mean of each shell before and after SVR, in the spin echo and gradient echo contrast. Grayscale values are min-max normalised per row.

Figure 4: Age trends in the mean tissue response function (RF) in the parenchyma. Tissue RFs for each subject were averaged across bi-weekly age intervals. The spin echo RF shows increasing signal anisotropy and stronger attenuation at young gestational age, as expected. On the other hand, the gradient echo shows stronger T₂*-weighting at later gestational ages, most prominently in the b=0 signal.

Figure 4: Axial, coronal and sagittal planes of whole-brain diffusion weighted images with 1.3mm isotropic resolution using single-shot spiral sampling. (a) b0 (T2W) images. (b) single diffusion weighted images (DWI1). (c) mean DWI and (d) colored FA maps.

Figure 3: Colored FA maps of 10 representative slices covering the whole brain. The results indicate that single-shot spiral diffusion imaging can provide accurate DTI metrics.

Figure 2: Direct comparison of the b=0 s/mm² images between CP and pTx acquisition in three orientations. The right most column shows the percentage signal change between the pTx and CP acquisition. The signal change was calculated as 200% × [signal(pTx) – signal(CP)] / [signal(pTx) + signal(CP)] .

Figure 4: (a) Colour FA maps acquired with CP pulses (left) and pTx spokes pulses (right). The maps show FA (in the range of [0.25 1]) with colours representing the orientation of the first eigenvector (red: left-right; green: anterior-posterior; blue: inferior-superior). (b) Zoomed region of interest as denoted by the red box in (a). Diffusion fibres are more clearly seen in the pTx acquisition especially in the brainstem.

Figure 5. (Upper column) An example of noise-like artifacts in SENSE. The nerve root region is noisy, which may contribute to the reduced coefficients of variation in this region. (Lower column) There is a lot of noise in the area outside the nerve root. In the region of noise, ADC value is low and FA value is high. When the ROI is placed outside the nerve, the values are greatly shifted. It is important to draw ROIs on noise-free images.

Figure 3.(Upper column) Fractional anisotropy (FA) and coefficient of variance (CVFA) of EPICS, SENSE, and SENSE×2. FA was significantly different among the three protocols, and CVFA was significantly smaller in EPICS and SENSE than in SENSE×2. (Lower column) Focusing on the nerve roots, FA in EPICS was significantly smaller and CVFA was smaller than SENSE.

Figure 4: Mean ADC values from the 3D IR-OGprep-GRASE sequence measured in several GM and WM regions at different frequencies (60Hz, 40Hz, 20Hz, and 0Hz). (A) ADCs in the hippocampal head were lower than that in the hippocampal body or tail, and its t_d–dependent change was faster (from PG to OG-20Hz) than the other two parts. (B) The t_d–dependency of cortical GM was distinct compared to other GM and WM regions, with a slower t_d–dependent change from PG to OG-20Hz.The error bar represents the standard deviation.

Figure 3: Diffusion-time (t_d) dependent ADC measured using the 3D IR-OGprep-GRASE and OGprep-GRASE sequences at different frequencies (60Hz, 40Hz, 20Hz, 0Hz). For the 3D IR-OGprep-GRASE sequence, all ROIs showed t_d-dependent changes of ADC. In comparison, for the 3D OGprep-GRASE sequence without IR module, the t_d-dependent effect was not observable for regions close to the ventricle and sulci, including the hippocampus, cortical gray matter, and splenium of corpus callosum. (*p < .05 and **p < .01 by post-hoc t-test)

Figure 1. Illustration of the echo-train shifted acquisition, 3-shot spatiotemporal encoding and subspace reconstruction of ACE-EPTI.

Figure 4. Reconstructed images and SNR maps of ss-EPI and 3-shot ACE-EPTI from the phantom and in-vivo experiments. For ACE-EPTI, the echo-combined image is shown and is used for the SNR analysis. The images acquired by ss-EPI show obvious distortion in the front part of phantom and brain (yellow arrows), while ACE-EPTI eliminates all these distortion artifacts. ACE-EPTI also shows a 30-40% increase in SNR when compared to a time-matched ss-EPI with 3 average in both phantom and in-vivo experiments.

Fig.5 Testing of the final model on a new subject randomly chosen from the 7T HCP database. Shown are mean diffusion-weighted images with b1000 (averaged across all directions) and color-coded FA maps in sagittal, coronal and axial views. The final model with tuned hyperparameters was trained on our entire dataset of 5 subjects. Note that the final model (ResNet) substantially enhanced the image quality by restoring signal dropout observed in the lower brain regions (as marked by yellow arrowheads), producing color-coded FA maps with reduced noise levels in those challenging regions.

Fig.3 Example b0 and b1000 diffusion images (of one diffusion direction) for single transmission (sTx) vs deep-learned pTx (ResNet) in reference to acquired pTx images. Shown is a representative axial slice in lower brain from one subject, in which case the model with tuned hyperparameters was trained on data of the other 4 subjects. Note that the use of ResNet substantially improved the image quality, effectively recovering the signal dropout observed in the lower temporal lobe (as marked by the yellow arrowheads) and producing images that were comparable to those obtained with pTx.

Figure 5: (a)-(c) The b0 images, MD and FA maps of the same slice as in Fig. 4 from 6 out of 48 echo times.

Figure 3: A comparison of b0 and diffusion images acquired with conventional 4-shot EPI and cEPTI, with a T2 FLAIR image as a distortion-free reference.

Figure 5: In-vivo results, acquired at 2 mm³ isotropic resolution, different transverse slices; a) b=0 s/mm², b) b=500 s/mm² (iso), c) b=1000 s/mm² (iso), d) ADC map (10^-3 mm²/s).

Figure 3: Low resolution reconstructions for a 19-shot acquisition for a single direction of b=500 s/mm². Top left: magnitude images, top right: phase images, bottom: nRMSE and weights for each shot. Shot 9 was determined to be the reference shot. Shot 6 (red square) showed most corruption in both magnitude and phase, resulting in a high error.

Figure 2. Multi-shot DWI with 16-fold segmentation reconstructed using the proposed joint estimation approach. The shot-to-shot phase variations ($$$\phi_{vp}$$$) were initialized with the low-pass filtered phase images of individual shots (bottom left). The object (O) was initialized with the MUSE-reconstructed image (top left). The diffusion contrast similar to the DWI with lower segmentations (top row of Fig. 1) was recovered in the final image (right column).

Figure 1. DWIs (b=1000s/mm²) with varying number of shots (n_s) reconstructed using the MUSE method. Overall, the MUSE approach provided high-quality images. However, less diffusion contrast was observed for higher segmentations (blue vs. red arrows).

MD-dMRI maps with and without EBS. Mean Diffusivity (MD), Mean Isotropic Kurtosis (MK_I), Mean Anisotropic Kurtosis (MK_A)

Artifact vs. b-value. Images are shown with EBS on (top row) and visible ghosting artifacts, EBS off (middle row) with no visible artifacts, as well as a difference image (bottom row)

Figure 3: The reconstruction results of the 0.86-mm isotropic MUSIUM data. (a) The diffusion images at b=500, 1000 and 2000. (b) The FA map in three orthogonal planes. The data acquisition time is ~12.5 minutes.

Figure 1: The RF excitation and k-space sampling trajectories of the MUSIUM sequences. (a) The RF excitation of a MUSIUM sequence. n_sms × n_shot slices are excited simultaneously. (b) The sampling trajectories of MUSIUM sequence.

Figure 3: k-space shifts versus rotational velocities at TE=68ms. (a,b) the same as in Fig.2. (a) Without PMC, the signal of ~75% of the repetitions (158/216) is outside the sampling window; blue a=14.5, b=2.5, r²=0.93. (b) less than 4% (4/216) of repetitions show signal loss. Almost all acquisitions show 2 motion updates; blue a=-0.5, b=3.5, r²=0.1. (c) A linear dependency of the residual echo-shift from the k-space center upon rotational acceleration is now visible; blue a=0.5, b=-0.3, r²=0.4.

Figure 2: k-space shifts versus rotational velocities at TE=80ms. (a) Without PMC, the signal of less than 50% of repetitions (93/216) is in the sampling window. This occurs for rotational velocities $$$|\omega_{r}|<6^{\,\circ}/\mathrm{s}$$$, in agreement with the simulations (grayed band). From theory: $$$\Delta k_{x}^{max}=a\omega_{r}+b$$$; blue a=17.5, b=3.4, r²=0.97. (b) With PMC on, less than 4% (2/216) of repetitions are lost. Two main subpopulations are present, depending upon whether 2 (blue) or 3 (red) motion updates have been applied; blue a=1.8, b=6.3, r²=0.3.

Figure 4. Time series of DW three-shot images across dynamics for all gradient shapes (left), and the complex average across dynamics (right). All images capture the same slice of one subject. The diffusion-sensitizing gradient was aligned with the z-direction. Significant artifacts confound each dynamic and the subsequent average for PGSE acquisitions; no such issue occurs for either OGSE acquisition, for which fine anatomical detail can be seen.

Figure 2. Phase differences with respect to the first dynamic (different columns) across subsequent single-shot dynamics for each diffusion sensitization scheme (different rows). Phase is observed to have more pronounced fluctuations (i.e., of greater magnitude) among PGSE dynamics than among dynamics of either OGSE acquisition. Between both forms of OGSE, phase variations are similar.

Figure 3: Proposed joint recovery: (a) shows the images from the proposed sampling scheme with SMS & joint k-q acceleration. The goal is to recover all DWIs jointly & simultaneously unfold the slices from the MB excitation. (b) During the iterative recovery, the 1st term in Eq [1] performs unaliasing arising from the multi-band excitation & k-space under-sampling. The q-space prior performs voxel-wise projection of the diffusion signals to the learned manifold (illustrated in (c)). The TV prior imposes DWI smoothness. The joint recovery maximally exploits the redundancy in the data.

Figure 4: Proposed joint reconstruction from two samples DWIs. (a)-(c) shows the reconstruction from the fully sampled data from all 4 shots, which were accelerated using multi-band imaging only at MB=3. (b) shows the individual SMS MUSSELS reconstruction and (c) shows the joint reconstruction using the proposed method. (d)-(e) corresponds to the acquisition with additional 4-fold acceleration from joint k-q under-sampling. (e) shows two sample DWIs from the joint recovery of all the 20 DWIs from all the 3 slices simultaneously.

Figure 5:Comparison of HA, E2A, MD and FA for a single midventricular slice in a porcine ex vivo heart using a spiral readout and an EPI readout. Both using the same STEAM approach, multiple averages (both using complex averaging) and the same total acquisition time of 2hours.

Figure 4:A) DT-CMR maps reconstructed from data acquired with similar total acquisition duration, but increasing numbers of spiral interleaves/decreasing numbers of interleaves. Total acquisition time was 2h for rows 1-4. B) Plot of mean transmural HA-R² of the data shown in A. C) Plot of Transverse Angle standard deviation of the data shown in A.

Figure 4. The 1 mm isotropic mean DWI images (a), MD (b) and FA maps (c) of SMSlab DTI with 20 diffusion directions, from three orthogonal views. The images are reconstructed by k-space-based reconstruction with a synthesized 3D navigator.

Figure 3. The reconstructed images from two volunteers, by k-space-based reconstruction with (a) the acquired (for SMSlab) or extracted (for multi-slab) 2D navigator, (b) the synthesized 3D navigator and (c) the acquired 3D navigator (only for multi-slab).

Figure 2: T2-weighted images (b=0) and diffusion-weighted images (b=1000 s/mm²) acquired by different spiral acquisition schemes. Single-shot EPI DWI were also shown as a reference. Based on the proposed acquisition strategy, for 4-shot spiral acquisition, the readout duration was reduced from 41.0 ms to approximately 30.0 ms. The blurring effects were significantly reduced, and the image quality was improved.

Figure 3: Colored FA maps of six slices. The results show that 4-shot spiral variants can provide fine DTI metrics, which indicates that POCS-ICE algorithm can successfully correct motion-induced phase errors for these 4-shot spiral variants.

Figure 3. Representative T2WI, DWI, MUSE, and SRGAN reconstructed images.

Figure 4. Detailed comparison among DWI, MUSE, SRResNet, and SRGAN reconstructed images.

Figure 5: Data from a high-resolution, trace-weighted acquisition with and without SMC. Three out of 20 slices are displayed. The potential protocol for clinical use shows minimized SNR loss and contrast changes in the DW images. The SNR loss in the T₂*W case is higher, but this is more than offset by four-fold averaging (from b-value of zero and three diffusion-gradient directions with b = 1000 s/mm²). Contrast change in T₂*W case is mainly due to high CSF signal saturation.

Figure 2: Measured and simulated saturation for both contrast types in the phantom as a function of the T₂*W excitation flip angle with and without the use of split-slice GRAPPA during image reconstruction. The trend of the three curves is almost identical. The simulation overestimates saturation. Additionally, standard deviations of experimental data are higher than the simulation would have suggested. Reasons could be the uncorrected frequency drift and normal differences between two separated acquisition as can be seen in the table below.

Figure 2 Comparisons of both non-diffusion and diffusion weighted images of mouse brain acquired using Diffusion-prepared 3D VF-RARE and single-shot EPI. a: T2-RARE was acquired as anatomy reference. b, c : b=0 s/mm² images of sequence DP VF-RARE and SS-EPI. d, e : b=500 s/mm² images of sequence DP VF-RARE and SS-EPI.

Figure 3 Comparison of in vivo FA maps (a, b) and color-encoded FA maps (c,d) of an adult mouse brain acquired using DP 3D VF-RARE and SS-EPI. The areas of cortex and hippocampus were used in the following MD’s and FA’s calculation.

Fig. 5. Colored FA maps of different methods. The proposed method can recover fine fiber structures. With the introduction of FA loss function, the contrast contamination among diffusion directions can be reduced.

Fig. 3. Selected comparison results and zoomed-in images of in-vivo DWI data. b0 and mean DWI results from 2 representative slices are shown. In the zoomed-in images, the arrows point to the structures that SRCNN and SRResNet fail to recover.

Figure1. Briefly, in the signal extract layer, signals were extracted from each voxel as the input for dictionary A. Each unit in the red dashed box was comprised of a nonlinear operator layer and a dictionary layer B. We chose ReLU as the Nonlinear layer. This process was repeated 10 times. Then the output was transferred to the Normalization. [f, D, D*] can be obtained using the linear combination layer (Eq [4-6]). Lastly, the output parameters are fed into the signal reform layer according to IVIM model to reconstruct the signals which are compared with the input as part of the loss function.

Figure2. (A) Estimated f, D, and D* parameter maps using the least square or Baysian fitting of the bi-exponential model, multilayer perceptron, and the proposed SC-DNN. The ground truth was obtained by adding noise after recovering from the previous parameters according to the IVIM model. (B) Residual error maps between the fitted parameters and the ground truth via different methods.

Fig. 1. Diagram of the proposed joint denoising method. Within each iteration: (1) extracting reference patches using a sliding window and searching for similar patches through block matching; (2) for each reference patch, stretching its similar patches to vectors and stacking them into a matrix to form a low-rank patch matrix; (3) for each patch matrix, estimating a noise-free patch matrix through a weighted nuclear norm minimization (WNNM) model; (4) converting estimated patch matrices back to images.

Fig. 3. Denoising results with in vivo DW brain images (A) and resulting diffusion metric maps computed from denoised DW images (B). The image set contains one b0 image and 6 DW images with b =2000 s/mm². Only DW image along one direction is shown. The image set of NEX=1 was used for denoising, while the image set of NEX=4/10 was used as high SNR reference. At very low SNR, the proposed method was still robust and more effective than MPPCA in reducing noise while recovering structural details when compared to reference. It achieved image quality and FA map comparable to those using 4 averages.

FIGURE 1. Schematic comparison of the conventional DTI model fitting and deep learning methods SuperDTI for generating various diffusion quantification maps.

FIGURE 4. Comparison of fiber tractography generated from 6 DWIs using different methods. Corpus callosum, internal capsule/corticospinal tract, and superior longitudinal fasciculus generated by MF (a), MLP (b), BM4D (c), proposed SuperDTI (d), proposed SuperDTI+ (e) with additional k-space reduction (l), respectively, and the corresponding difference map (g-k) with 6 DWIs. The model-fitted tractography from 90 DWIs (d, k, r) is also shown as a reference.

Fig.1 Violin plots of distributions of the quantitative metrics between M0 and Mcyc.

Fig.2 T2WI, ADC, a-DWIb1000, a-DWIb1500, cal-DWIb1500, non-optimized syn-DWIb1500 vs. optimized syn-DWIb1500 of four different patients. Patient A: A 76-year-old man with chronic prostatitis. Patient B: A 63-year-old man with prostatic hyperplasia. Patient C: An 87-year-old man with prostatic cancer in the right transition zone. (Gleason score, 4+3). Patient D: A 69-year-old man with prostatic cancer in the peripheral zone invading the rectum (Gleason score 4+5).

FIGURE 1. A nodular lesion (arrow) in the right upper lobe. A, T2WI-SPIR. B, GRASE-IVIM (b = 800 s/mm2 ). C, TSE-IVIM (b = 800 s/mm2 ). D, EPI-IVIM (b = 800 s/mm2 ).E,T2WI fused with GRASE-IVIM. F, T2WI fused with TSE-IVIM.G, T2WI fused with EPI-IVIM. The distortion and displacement of the lesion are shown on EPI-IVIM (arrow). On TSE and GRASE-IVIM ,the lesion, along with its air bronchus, can be clearly shown. Fused images showed that GRASE and TSE-IVIM is perfectly matched with T2WI-SPIR.

FIGURE 2. Comparison of SNR,CNR and Distortion between GRASE-IVIM,TSE-IVIM and EPI-IVIM.

Brain images for the reference scheme (constant FA of 120°) and optimized FA scheme, respectively. (a-b): Data sampled with linear phase-encoding order before and after applying calculated filters. (c-d): Data sampled with center-out phase-encoding order before and after applying calculated filters.

The Hanning target function (a) and corresponding PSF (b) having a low FWHM and limited ripples. The optimized FAs (c) together with the reference scheme of 120° flips. The first four echoes are marked as “transients” and discarded for improved robustness. The resulting filters for both echo families, E1 and E2, normalized with the coefficient used for the center of k-space.

Figure 1: (A) Coronal DWI-EPI (NEX = 1) and (B) Coronal DWI-MAVRIC (bins=3) images of 13 PVP vials contained in phantom. (A)DWI-EPI contains low EPI distortion while (B) DWI-MAVRIC contains no EPI distortion.

Figure 2: ADC mean values for eight acquired scans (six coronal, two axial) at all six polymer concentrations. The dashed lines show the references values from Palacios et al.(10) for all six concentrations.

Fig.3. (a) Representative T2W(b=0) images and associated ADC maps: ADC's are shown for each of the four evaluated methods. Significant distortion seen in the SS-EPI, both with and without the presence of the metal high susceptibility source. Without the metal, image distortions were relatively mitigated with MUSE EPI. Qualitatively, DW-Prop and MSI-Prop were less susceptible to susceptibility artifacts. Vial 6 was largely non-viable in the SS-EPI and MUSE-EPI, but showed perfect geometry in DW & MSI-Prop. Vial 6 were compromised in DW-Prop, but were fully recovered in MSI-Prop.

Fig.4. (a) Representative EPI and MSI-Prop T2W (b=0) images and spinal cord ROIs. Significant distortions can be seen for the EPI. MSI-Prop provided more anatomically accurate imaging of the cord. (b) ADC values computed with MSI-Prop were roughly 20% larger on average than the FOCUS EPI images. The computed bias was 293·10-6mm2/s among 13 subjects’ mean cord ADC values.

Figure 1: The proposed model for multi-coil high-resolution DWI reconstruction. The network combines losses from four channels to back propagate. Each channel consists of a de-aliasing layer and a data consistency layer to recover from aliasing artifacts.

Figure 2: Comparison of reconstructed images for multiple b-values in healthy volunteers. The calculated ADC maps are shown leftmost. The b-values are listed on each column. SSIMs are listed at the bottom-right of each image. The rightmost four columns represent the corresponding error maps. The error was x5 amplified for better visualization. The aliasing artifacts are shown by arrows.

Figure 2. The MIP of direct coronal whole body DWIBS at 3.0T were shown. Image based B0 shimming provided improving depiction of axillary lymph nodes (red circle). Blip-up blip down distortion correction also decrease distortion effect due to the presence of air (blue circle).

Figure 4. The MIP of transverse DWIBS at 1.5T were shown. Similar findings to 3.0T were obtained at 1.5T.

`FA maps new' denotes the phase correction results obtained by the renewed procedures. Artifacts are significantly eliminated for all three methods, while the FA values in corpus callosum is increased properly.

`TV new', `CF new', and `MPPCA new' denote the phase correction results obtained by the renewed procedures. The renewed procedures yield more accurate FA values, especially for MPPCA. Black color of white matter means high accuracy achieved. MAG denotes the magnitude images.

Figure 5. fODFs over two slices highlight exemplary insets as # directions is uniformly decreased. Interestingly, at half the sample size of the superset, the principal component is retained. Even with the sample size scaled to a 1/4^th of the original dataset, the principal component is retained, though, a slight uptick in noise is evident in the fiber crossing and interfacial regions. (“noisy” lobes: yellow circles).

Figure 4. Normalized root mean square error (NRMSE) over 76 WM bundles for FA and Orthogonal Kurtosis highlighting bias developed as # directions sampled is decreased. As is evident, the NRMSE for both FA and kurtosis grows more slowly, indicating stability compared to the standard, however, past 90 directions, the sqrt(N) scaling factor breaks down for both measures albeit the bias is not at the same scale. Interestingly, with uniform sampling over a 3-shell configuration, the bias in metrics can be reduced. Note the y-axis scale for FA and Kurtosis is not the same

Figure 1 (a) The proposed design of interleaved DW-EPI with navigator echo. The acceleration factor R of imaging can be 4, 8, or 16, while R for the navigator echo is fixed at 4. (b) The reconstruction pipeline based on proposed k-d SVD with contrast compensation at final step.

Figure 5 (a)The comparison of colour FA map derived from gold standard images and the images reconstructed with the different strategies in Simulation 1. (b) The comparison of colour FA map derived from gold standard images and our proposed k-d SVD method at different acceleration factors (R = 4/8/16).

Two representative sagittal slices reconstructed with linear phase correction, SENSE, and LORAKS-based reconstruction. Note the marked ghosting reduction obtained with LORAKS-based k-space reconstruction.

Figure 2 Representative section from Figure 1 normalized to highest intensity of each acquisition. The noise amplification in the midbrain and brain stem is easily visualized here (right-most sections in each row).

Figure 3 Boxplots representing quality control metrics (described on y-axis labels) over different acceleration schemes (x-axis). Blue bars indicate young subjects, orange indicates healthy, old subjects, and green indicates MCI subjects. p-values calculated using a Generalized Estimating Equation (GEE) with Gaussian assumption and exchangeable correlation accounting for repeated subject scans. p-values reported after false discovery rate (FDR) correction; significance: *p<0.05, **p<0.01, ***p<0.001, ****p<0.0001.

Figure 4. Super-resolution of human brain diffusion weighted images. From top to the bottom row: A. initial Low resolution data; B. cubic interpolation; C. Zero-filling interpolation; D. ResNet output by zero-filling as pre-step interpolation; E. High resolution Reference).

Figure 1. The architecture of ResNet architecture (input patch size could be three dimensional 21x21x21 (or m×n×k) or two dimensional m×n, and convolution kernel should be corresponded as three dimensional and two dimensional).

Figure 1. (a) A generalized imaging sequence along one coordinate direction, and (b) a general framework for an automated and accurate calculation of b matrix.

Figure 2. 2D SE imaging sequence

Figure 3. Scattered plots of HHI values for all slices for an example b1000 (a) and b2000 (b) cases. Motion detected by PITA-MDD are all dots fall below the HHI threshold, defined as 0.7 for b1000 and 0.6 for b2000 dMRI (red lines). Motion detected by APITA-MDD are denoted in red. Notice slices missed by PITA-MDD but detected with adaptive thresholding in APITA-MDD (circle). Arrows indicate time when motion instruction was given.

Figure 4. Heat maps of HHI (a) and the deviation score M (b) for representative b1000 and b2000 acquisitions with no motion, head rotation, and leg crossing motions. In each map, the horizontal axis represents volume index, and vertical axis represent slices index. Range for HHI is [0,1]. Range for deviation score is from 0 to maximum value 25.

Figure 4 iShim DWI for patient with primary rectal cancer of Grade 1

DW images of b800 (A), b1600 (B), ADC map (C), T2WI (D), infusion images of both T2WI and DWI (E) and dynamic T1WI (F) showed the same location of lesion. Average ADC value measured was 1025.39 mm2/s (C).

Figure 2 Comparison between iShim- and SS-EPI-DWI in patients after CRT.

Images on the left side were on the same location in patients with iShim-DWI. Images on the right side were on the same location in patients with SS-EPI-DWI. iShim-DWI of b800 (A) and b1600 (B) showed higher SNR and CNR with lower signal noise compared with SS-EPI-DWI of b1000 (C). ADC map (E) of iShim-DWI showed better image quality compared with ADC map (F) of SS-EPI-DWI. T2W images (F and G) and dynamic T1W images (H and I) showed the same location of lesion (arrows) for iShim- and SS-EPI- cohort, respectively.