Localized regions of inhomogeneity observed on the same patient in both 3D FLAIR and T1-w diagnostic images. In this case the localized region (red arrows) was thought to be suspicious of potential herpes encephalitis and follow-up imaging was conducted.

Two experienced radiologist’s asymmetry scores (assessment localized to the hippocampus), are plotted against the percent difference in intensity measured in the hippocampus (left side > right side) with automated atlas-based segmentation. The same 30 image volumes from 30 different patients were scored. The color of the data points corresponds to the response of each radiologist to the question: “Is clinical follow-up required?” where: definitely not (black +), uncertain (blue +), definitely yes (red +).

Figure 3. Self-correction of images acquired with single-shot spiral and EPI (R=2) trajectories: blurring and geometrical distortions could be effectively mitigated. The algorithm was initialized with no information on B₀ given. Overlays of the self-corrected and the uncorrected images show vast differences. The corrected images show only minor differences to the reference images and in some areas even more details in the spiral images. The estimated field maps are similar for spiral and EPI trajectories in both slices.

Figure 5. Single-shot self-correction shown for spiral-out and EPI read-outs for varying undersampling factors. In the fully sampled cases field maps with 15x15 coefficients were fitted; in the undersampled scenarios (R=2-4) maps modelled with 13x13 coefficients were fitted to increase the regularization. B₀-introduced blurring and geometrical distortions in spirals and EPIs (lines added for assistance), respectively, could effectively be reduced without requiring any prior knowledge on B₀.

a) TSE Imaging parameters: TR/TE: Repetition/Echo Time, FoV: Field of View, Resol: Spatial Resolution, Aver./Conc.: Averages/Concatenations, iPat: Grappa Acceleration Factor, ROBW: Readout Bandwidth, SAR: Specific Absorption Rate, b) evaluation criteria for qualitative analysis of through-slice chemical-shift artifacts (CSA).

The chemical-shift artifact (*) width (green double arrow) was measured at two locations (black arrows): (i) at 50 % condyle height and (ii) separated by 10 mm along the posterior direction (a). b) Stronger artifacts due to unmatched and moderate RF-bandwidths (UMB), compared to smaller artifacts obtained with matched and increased RF-bandwidths (MIB) (c). The phase encoding direction is along the head-feet direction (dashed white arrow). d) Artifact size for both locations and UMB/MIB. e) Artifact rating for bone and muscle tissues and UMB/MIB (*: significant difference).

Figure 1: Schematic representation of the proposed method in a spin-echo double diffusion encoding sequence. a) Bipolar diffusion-weighting gradients (blue) are used for double diffusion encoding, here for instance along the physical x and y axes. These gradients induce additional concomitant fields which cause severe artifacts. b) Oscillating gradients (green) are implemented to null the concomitant fields induced by the bipolar gradients.

Figure 4: Comparison of acquired magnitude images for the long phantom cylinder orientated along the z-axis. The images were corrected for diffusion attenuation. a) Without concomitant field compensation, the image exhibits strong artifacts, especially signal voids (q = 46 mm^-1). b) No concomitant field related artifacts can be seen when the concomitant phase is removed by the oscillating gradients (q = 46 mm^-1). c) Image acquired with q = 0 mm^-1 as reference.

Figure 4. Comparison of 2D SE, MAVRIC-SL, and SEMSI. Images on the same slice location by 2D spin-echo (A), Product 3D MSI (MAVRIC-SL) (B), and the proposed SEMSI sequence (C). The proposed SEMSI sequence achieves comparable off-resonance correction with improved sharpness compared to the MAVRIC-SL sequence.

Figure 2. SEMSI Image Reconstruction. The reconstructed slices images are first FFTed along the echo time dimension after windowing to generate spectral images. Then based on the center frequency of each spectral bin, the spectral images are re-registered to its original location and combined to form the final image.

Figure 1: Illustration of CACTUS algorithm.

Figure 4: (a) T2 maps of two representative subjects using locally low rank (LLR) only, LLR with once-pruning ACS, LLR with twice-pruning ACS, and LLR with CACTUS; (b) ROIs on the LLR T2 maps and (c) mean and standard deviation (in parenthesis) of T2 values in millisecond within ROI. Rectangular bounding boxes are placed on both arms for the estimation of interference array correlation statistics.

Fig 5: The effect of shim change on the accuracy of a conventional (static) GE-based fieldmap and the dynamic EPI-based REFILL fieldmap. An imposed shim change, mimicking the effect of motion or physiology in fMRI, leads to additional distortion in frontal regions at 7T (top row, left). The change in field is not captured in the static fieldmap, leading to an incomplete correction (second column). The dynamic REFILL fieldmap for this volume is accurate, and gives a complete correction (third column) - see outline transferred from the distortion-free reference (top row, right).

Fig.2: Dependence of the readout gradient, φ_G, on field strength, sequence, receiver bandwidth and echo time. φ_G was close to linear and quite consistent over phase-encode lines (left) and slices (not shown). It was larger at 3T (red symbols) than 7T (black symbols) and followed opposing trends with TE and bandwidth at the two field strengths.

FIG. 5 Representative zoomed-in temperature images of the epilepsy patients obtained during MRgLITT treatment. The temperature maps were superimposed onto the post-ablation T1w+Gd images. (a) 38 years old, female, hippocampal sclerosis. (b) 34 years old, male, focal cortical dysplasia (FCD). Results without susceptibility correction (top row, traditional algorithm, TE=19ms) and with correction (bottom row, multi-echo based algorithm, multiple TEs) are both demonstrated.

FIG. 4 Selected thermal images of a Doberman. From left to right: (1) single-echo TE=4ms and (2) TE=19ms temperature images calculated via the traditional PRF algorithm; (3) the multi-echo (TE=4/9/14/19ms) combined temperature image using the proposed multi-echo algorithm. Note that the proposed method outperforms its single-echo counterparts, providing robust correction of the susceptibility induced artifacts at the heating center, and effective suppression of the CSF flow-induced artifacts near the ventricles (artifacts pointed by the yellow arrows).

Figure 3. Maximum intensity projection (MIP) maps of all possible pathways for (A) the optimized crusher scheme and (B) the same scheme with reverse polarities of the gradients along FH direction. The MIP maps depict the number of occurrence among all pathways per B₀ map (PB-2^nd) voxel where φ_crush matches -φ_{inhom.gradient}. The red arrows point to the locations with more than 3 occurrences through pathways.

Figure 4. Spectra with PB-2^nd VOI shim setting. (A) Time and frequency domain signal from a voxel of 2 × 2 × 2 cm³. (B) Comparison of spectra (NSA=1) acquired from CSI (spatial resolution = 27 cm³, matrix size = 7 × 7 × 6) and SV MRS (NSA=16) with the same localization (semi-LASER, TE = 31 ms). The frequency domain signals from the central slice, contain the VOI (red box), of a matrix size of 7 × 7 × 6 (left). All the spectra in the grid are scaled by a factor of 10 compared to the VOI. Resulting spectrum from SV MRS which was placed on the VOI can be seen in the right-bottom of the figure.

Fig. 1. Optical tracking system designed for the 64ch head coil. The four MR-compatible cameras record the patient’s face in the head coil through the mirror.

Fig. 2. Face images obtained from the four MR-compatible cameras. Green squares indicate the facial features for tracking.

Qualitative results on a representative axial slice of an abdominal MRI scan reconstructed (a) without any streak suppression algorithm (reference), and with (b) SW, (c) CR+OP, (d) B-STAR (i.e., manual ROI), (e) RGB-STAR (Auto Center), and (f) RGB-STAR (Auto Arms).

Bland-Altman plots showing agreement between B-STAR and the two automated RGB-STAR interference covariance matrix generation methods.

Figure 2. Illustration of spiral aliasing artifacts in a numeric phantom and in-vivo. The numeric phantom is based on XCAT with SPGR contrast (a), using simulated coil sensitivities (b), and gradient non-linearity maps of the GE Zoom coil. Simulated image shown in (c) illustrates the distortion and the hotspot outside the FOV (green arrow). Simulated spiral reconstructions without (d) and with (e) readout LPF both contain spiral aliasing artifacts within the FOV (blue arrows). This behavior closely matches in-vivo data, with a representative example shown in (f).

Figure 1. Flowcharts of artifact reduction methods. The data was temporally combined and reconstructed to 2.5×FOV to obtain coil images. Root sum-of-squares coil combination was performed, and the resulted image was smoothed. A mask M₁ was automatically detected around the pixel with the highest intensity outside the FOV by thresholding. The LF method can then be performed by applying masks M₁ and M₂ during reconstruction. For the ES method, M₁ is applied on the coil images, and k-space of the aliasing source was estimated and subtracted from the raw k-space for further processing.

Figure 4. Example image quality of the single-echo test results with 6.3x to 3.1x accelerator factors. The input shows the undersampled images, the output shows the images after our network destreaking, and the target shows the fully-sampled ground truth images.

Figure 5. Example image quality of the multi-echo test results of a 4.2x acceleration factor. The input shows the undersampled images, the output shows the images after our network destreaking, and the target shows the fully-sampled ground truth images.

Illustration of the estimation procedure for the trajectory shifts.

Blue arrows represent the first blade.

Red arrows represent the second blade.

Reconstructed images obtained from the stack-of-stars volunteer scan, under free-breathing

(a) Image reconstructed without correction

(b) Image reconstructed with the linear correction

(c) Image reconstructed with the non-linear correction.

Comparing (b) to (c), streak artifacts (arrow 1) and image shading (arrow 2) are suppressed in (c).

The scan parameters: TE=1.3ms, TR=3.6ms, FOV=35cm*35cm, matrix-size=256*256, sampling-pitch=6$$$\mu$$$s, number-of-slices=50, slice-thickness=4mm, number-of-trajectories=800.

Figure 4: Two reconstructed slices of the brain dataset. Reconstructions were performed using TS with $$$L=10$$$ (second column), TS with $$$L=20$$$ (third column) and the proposed SVD-based method (fourth column).

Figure 5: The upper part of the figure shows the normalized root mean squared deviation (NRMSD) of the reconstructed slices. The middle part displays the number of basis functions $$$L$$$ employed by the SVD based approximation. The bottom plot shows the NRMSD in dependence of the iterations of the PDHG solver.

Figure 4. Axial slice through 3D EPI image in a volunteer acquired in the reference position (a), combined image from reference position and nose touching (b), and augmented reconstruction using ground-truth MGRE field maps (c) and FIDnav field estimates (d). Difference images relative to the reference image (e–g). FID-navigated augmented reconstruction successfully compensates for artifacts in the combined image with B₀ variations.

Figure 2. Comparison of field difference maps (in Hz) measured from multi-gradient-echo (MGRE) data and FIDnavs for changing shim settings in the ADNI phantom. FID-navigated field estimates are in excellent agreement with ground-truth pixel-wise field maps.

Figure 3: Experimental results for three different MR-systems. Top row: single channel measured values of line-by-line δk_RO- δk_RO,nav. For the 3T system there are no oscillations or offset observed. For the two 7T systems, the temporal development is quite similar, with both initial oscillations and a significant offset between the navigator-measured δk_RO and the equilibrium line-by-line value. Middle row: reconstruction based on standard navigator-method. Bottom row: reconstruction using line-by-line δk_RO correction. W/L = [0, 0.1*median(phantom)].

Figure 2: Appearance of Nyquist ghost for different temporal development of δk_RO. Top row: plot of δkRO in units of Δk_RO as function of k_PE. Middle row: resulting images with window/level (W/L) = [0, max], bottom row: same images but with W/L = [0,0.1*max]. Increasing the linear phase offset (columns 1-2) produces the expected Nyquist ghosting; an oscillating offset introduces more complex ghosting dependent on the frequency (columns 3-4), where a decaying oscillation will have a reduced effect according to the decay time (column 5).

Figure 3. Representative images with different level of artifact intensity and corresponding Dice Index values as a function of posterior-anterior template shift. From top to bottom: high, medium, and low level. The prominent peak in each graph corresponds to the best-matching offset of template and artifact.

Figure 2. A brief overview of the tool processing steps from input image to artifact detection. a) input image; b) skull mask and template; straightened image; d,e) image gradient; f, g, h) image gradient merged and shifted for edge enhancement; i) base image for template matching; j) overlay of base image and template. A detailed description of each step is given in the text.

Figure 1. Population-level standard deviation maps for TR for all four phase-encoding directions. As evident in the temporal lobes of AP and PA data, ghost artifacts originating from the eyes significantly increased the variability in these regions. Additionally, in the Pons region, the variability was again significantly higher in AP and PA data compared to RL and LR. The reproducibility at mid-brain level was mostly similar for all phase-encoding directions except a small region of high variability near the caudate for the PA data.

Number of regions each PED produced the lowest and highest variabilities. The columns labeled “# lowest" and "# highest" variability indicate the number of ROIs where a given PED yielded the lowest or highest raw standard deviation values. The columns “# clearly best” and "worst” show the number of times the difference between the automatically computed "more reproducible" and "less reproducible" cluster centroids was more than 20%. These values indicate an overall trend where LR and RL PED generally produced more reproducible results and PA PED was the least reproducible.

Fig. 3. Comparison between state-of-the-art methods with the proposed method using the test subject. 2 representative slices of b0 and mean DWI images are selected. The red arrowheads point at regions where residual distortion and blurring artifacts still exist after distortion correction.

Fig. 2. Validation results of the proposed method. Input LR SSh-EPI, T2w-TSE, reference PSF-EPI and the reconstructed results are shown in the figure. b0 and mean DWI results show consistency with the reference image.

Figure 3. The correction results of the images with 1.3mm (a) and 1mm (b) isotropic resolutions. The uncorrected images (1st column), SMSlab NPEN (2nd column), CPEN (3rd column) and the reference images (4th column) are shown. In each sub-figure, the upper row shows the b0 images, and the bottom row shows the b=800 s/mm² images. The acquisition time is marked on top of each column. The nominal SNRs of the reference and the SMSlab images are comparable.

Figure 4. MD (a) and FA (b) maps of the 32-direction dataset. 1 mm isotropic resolution and b=1000 s/mm² are used. The acquisition time is 33 mins. The uncorrected images (1st column), and the results of CPEN (2nd column) are shown.

FIG. 3. Quantitative assessment of eddy current correction quality between FSL eddy only (FSL), our time-constant model (TCEDDY), our time-varying model (TVEDDY), and field monitored eddy current correction (FM). For both PGSE and OGSE acquisitions, the MSE was measured between all volumes with inverse distortions of both subjects, normalized by the mean MSE of each subject without correction, and averaged over all directions. Statistically significant changes in MSE between adjacent correction techniques is indicated by *(p < 0.01).

FIG. 2. Qualitative comparison between eddy current corrupted OGSE slice without correction, and after correction with FSL eddy, TCEDDY, TVEDDY, and field monitored eddy current modelling. Yellow arrow: blurring artefact corrected by modelling eddy current decay.

Figure 4. Prospectively acquired data with optimal view ordering but apparent eddy current artifacts and prediction following application of 3D U-Net. (Left) A 2D cardiac cine dataset is prospectively acquired with 12X acceleration, then reconstructed using a model-based deep learning approach. Eddy current artifacts manifest as residual aliasing across the image. (Right) The images on the right are corrected using our 3D U-Net approach, resulting in reduction of eddy current artifacts.

Figure 2. Data processing pipeline. A) Eddy current effects of random phase encoding orderings were simulated by scaling each k-space line by a random amplitude in a randomly chosen range. B) 3D-Unet Architecture with an additive skip connection to enable residual learning. The input and output sizes are Nx224x224x32.

Figure 3 Volunteer T1 weighted image

(a): conventional phase control

(b): proposed phase modulation

(c): difference between (a) and (b), Window width is 1/5 of (a) and (b)

(d): line profile on FID artifact

Figure 1. RF phase of proposed method (a), and signal intensity of FID artifact

overview of the 3D subvoxel shifts approach. Each panel shows three orthogonal cuts through the 3D image. The approach is identical to the original 2D version, but applies the filtering and subvoxels shifts along all three spatial axes, with modified filters appropriate for the 3D case. For each axis, the image is Fourier filtered to attenuate high frequencies in the directions orthogonal to the axis of interest. Gibbs ringing removal is then applied using 1D subvoxel shifts along that axis. The resulting three images are then summed together to give the final output.

demonstration of the proposed method in practice, using a T2-weighted 3D acquisition (details). The top two rows show 3 orthogonal cuts through the original image with evident Gibbs ringing artefacts and the output of the proposed method. The bottom two rows show magnified sections of the same images for closer inspection.

Figure 4 Comparison of results. A mask computed based on the Laplacian of Δf0 was used to combine the results of the standard deblurring (a) and the proposed method (c). The combined results are shown in (b). Water and fat in (a) were also low-pass filtered (d). Note that the residual ringing inside the yellow boxes is significantly less in (c) (d) compared to (d).

Figure 5 Results of 2D T2-weighted images with spiral readout length of 45 ms. (a) is the blurred TE image. the field map and and its gradient and laplacian are shown in (b)-(d), respectively. A mask computed based on the Laplacian of Δf0 (d) was used to combine the results of the standard deblurring (e) and the deblurring with low-pass filtered kernels (g). The combined results are shown in (f). The ringing artifacts (pointed by the arrows) in (e) are mitigated in (f) and (g).

Fig 3. Numerical simulations of Gibbs-ringing removal in 2D magnitude images of the Shepp-Logan phantom with phase. The first two rows show images of the signal magnitude before and after ringing removal, and the third row shows residual maps. For pf = 7/8 and 6/8, the RPG pipeline robustly removes PF-induced ringings, whereas for pf = 5/8 the corrected image still has non-trivial amount of ringings after applying RPG, potentially caused by the interaction of the images phase in high resolution ground truth and the convolution kernel in Eq. (1).

Fig 4. Mean diffusivity (MD) map baed on diffusion-weighted-images (DWIs) before and after Gibbs-ringing removal using local subvoxel-shifts (SuShi) and RPG, with the residual of MD before and after corrections. The unit of MD is $$$\mu$$$m$$$^2$$$/ms.

(A) Simulated spin echo and gradient echo images of a titanium rod acquired on a 3 T MRI system. (B) Measured spin echo and gradient echo image of the titanium rod. (C) Overlay of the artifacts from image (A) and (B). Red color: both artifacts are congruent; yellow: measured artifact only; blue: simulated artifact only.

This figure compares the artifact size of the titanium (A-C) and stainless steel (D-F) rod of the simulations (light grey) and the measurements (black) in mm2 separated by the different field strengths, (A,D) 1.5 T, (B,E) 3 T, (C,F) 7 T. Artifact sizes are shown for different implant orientations in regard to B0 (para: parallel and perp: perpendicular), sequences (GRE and SE), slice orientations in regard to implant orientation (cor, sag, tra), and phase-encoding direction (FH, AP).

Figure 1: Simulation Pipeline. 3D XCAT body and implant masks are utilized to create tissue parameter maps. The susceptibility map is used to calculate the field shift. K-space is simulated using the field shift, proton density (PD) and MSI sequence parameters with phase encoding in both left-right and anterior-posterior directions. Complex Gaussian noise is added in k-space. Spectral bin images are reconstructed and combined with quadrature summation.

Figure 2: MAVRIC-SL simulation of total hip arthroplasty. Central coronal slices from 3D volumes are shown for B₀ = 0.2T, 0.55T, 1.5T, 3T, and N_bins = 12, 18, 24. Other parameters: BW/pixel = 625Hz; RF bandwidth = 2.25kHz and bin spacing 1kHz. Implant contours are shown with yellow outlines. Artifacts around the implant are indicated with red arrows. A small artifact adjacent to implant surface at 0.55T is shown with green arrow. Total scan times are 3:53, 5:49, and 7:45 for N_bins of 12, 18, and 24, respectively.

Figure 3: Simulated phantom study of a ball with magnetic susceptibility 900 ppm at 3T. Input images exhibit readout blur and ripple artifact due to VAT. Ripple artifacts are reduced in the initial image (red arrow) and eliminated in the final estimation of $$$m$$$. Grid lines are sharper in the final image (yellow arrow), indicating mitigation of the VAT blur effect. Final $$$f$$$ agrees well with the true field within the region of excited spins.

Figure 5: Results in shoulder phantom shown for an off-center coronal slice. Ripple-like intensity artifacts are apparent in the input and initial images (red arrow), but significantly reduced in the final image. Differences in readout resolution are less apparent than in Figure 4 due to the absence of grid lines.

Figure 1. Image processing flow chart and the uneven echo spacing. The image processing pipeline is presented in panel A. Phase images were first combined with a graph-cut algorithm to conduct high SNR phase maps. The central frequency and unwrapping artifact were further calibrated from the derivative of the phase difference maps between the uneven DTEs. The phase accumulation from the uneven echo spacing is presented in panel B.

Figure 5. In-vivo images from a healthy human volunteer with an ICD on the chest. In the presents of ICD, the compromised image SNR and strong phase accumulation from the off-resonance source lead to obvious phase unwrapping artifacts and central frequency shifts in the peer algorithms (arrows). The proposed AUTO-HDR shows a much smoother field map around the ICD and higher SNR at a region further away from the ICD. The improved field map demonstrates the capability of AUTO-HDR in resolving fast varying phase change while preserving high SNR in human subjects with metallic implants.

Fig.1 (a-b) Under the default window width and window level, the images show the artifacts, including the low (long arrow) and high (arrowhead) signal intensities. The screw without protons displays a low signal intensity (short arrow).

Fig.1 (a-b) Under the default window width and window level, the images show the artifacts, including the low (long arrow) and high (arrowhead) signal intensities. The screw without protons displays a low signal intensity (short arrow).

Figure 2: Results from the head axial FLAIR experiments. The denoising CNN reduced image noise at the cost of significant smoothing. Blended images demonstrate the use of proposed method to generate images at different target SNR values. The target and measured SNR values closely match each other in all the blended images.

Figure 3: Results from the lumbar spine sagittal STIR and knee sagittal IW experiments. Images processed by the denoising CNN appear artificially smooth. Images blended using the proposed method have a natural appearance. The target and measured SNR values closely match each other.

Figure 5: Comparison of the high SNR spectrum obtained with NSA of 96 (left) and a jSPaMP denoised spectrum with NSA of 16 (right).

Figure 2: Description of the proposed jSPaMP algorithm.

Figure 1. Denoising of a single identical Lorentzian peak with independent noise. The black line shows the noiseless signal, the red line shows a single voxel’s data after denoising (no denoising applied in “Noisy”, all spectra mean shown in “Avg”). The Shaded grey region shows ± two standard deviations of all 128 spectra. All denoising approaches show high apparent noise level reduction, but LP denoising does not show a reduction in peak amplitude variance.

Figure 5. Voxel-wise maps of mean metabolite concentration (left, red/yellow) and standard deviation of metabolite concentration (blue) across the ten repeated acquisitions for each denoising case. Data is shown for three metabolites and is overlaid on a T1w structural image. Colour scales are fixed for each metabolite.

Figure 1 - Faraday cage attenuation measurements of eleven clinical MRI systems: measurements were acquired at 65MHz, 128MHz and 297MHz at the magnet room door, console room window and penetration panel. Box and whisker plots adjacent to the data show the median in red, interquartile range within the box, whilst the whiskers show the range of data excluding outliers, which are indicated with a red cross.

Figure 2 - Summary of RF attenuation measurements in eleven clinical MRI systems

Representative case. (upper row) Conventional wavelet based denoising method, (lower row) DLR based denoising method. The DLR based denoising method offered decrease in image noise, artifact, and clear depiction of pituitary grand and small objects as compared with the conventional wavelet based denoising method.

A box plot shows the SNR of CSF. There was a progressive increase in SNR of CSF with DLR based denoising method with increase of the denoising level in any location. On the other hand, the SNR of conventional wavelet-based method was not increased at high denoising levels (4-5).

Signal-to-Noise ratios on the images of 2 mm, 4 mm, and 8 mm of slice thickness with various parameters of Deep Learning Reconstruction (DLR) were shown. DLR had a powerful performance to increase SNR in all SNR ranges.

Structural similarity (SSIM) of the images were shown. Decreasing of SSIM higher in thinner slice images is compared to that in thicker slice images. Edge enhancement recovered SSIM higher, especially in thinner slice images.

Five echoes after QGRG spatial unwrapping are shown across the columns. From left to right, the first echo is at 5.0 ms, followed by 10.4 ms, 15.8 ms, 21.2 ms and then 26.6 ms. The top row is without temporal unwrapping (TU), the middle row is with temporal unwrapping and the bottom row is a plot of the line profiles for each column. Dataset C2 was used.

Six different techniques (three Laplacian, top row, and three path-following, bottom row) were used to spatially unwrap phase data from dataset C2. The output from each unwrapping technique is shown with an axial cross-section of the brain, and a line profile beneath it. Panel A represents the Laplacian from the STI suite, B is the Laplacian from the MEDI toolbox, C is the Laplacian from a collaborator, D represents QGRG, E represents SEGUE and F represents region growing from the MEDI toolbox.

Figure 4 : Comparison of the SWI with mIP without correction (A), with correction using the acquired field map (B) and with correction using the estimated field map (D) with the same slices as Fig3. (C) and (E) are respectively the difference between (A) and (B), and between (A) and (D).

Figure 1 : Pipeline of the field map estimation followed by an off-resonance correction. The magnitude and associated mask estimated with the adjoint are used to weight and stabilize the phase unwrapping algorithm.

Figure 1. The figure shows the variation of fieldmaps with TR and plots the associated statistics of the bias.

Figure 5. Slice select gradient reversal along with PE and FE gradient reversal (a) cancels most of the bias in the fieldmap (b) and reduces to a mean of 0.5 Hz and a 97th percentile of ~2 Hz (c).

Representative T1 weighted image and T2* maps from the in vivo experiment, a) before correction and b) after macroscopic field inhomogeneity correction. Average increase in the T2* values in both GM and WM is evident. GM/WM T2* contrast is also increased.

Proton density (a) and examples of T2* maps calculated without (b) and with F-term correction (c). Regions substantially affected by the B0 field inhomogeneity are indicated by red arrows. F-function correction significantly reduces magnetic field inhomogeneity artifacts on T2* maps and sharpens T2* contrast in gray and white matter.

Figure 4:
Simulated versus measured dependency of MP-FLAIR image intensity to B₁⁺. Dots: Measured MP-FLAIR intensities as a function of measured DREAM B1+. Colors depict different RF gains. Other sources of signal loss (e.g. B₁^-/B₀, fluid-content) cause the measured intensities to “fill out” the area under the curve. Black line: Simulated FLAIR intensities. Increasing discrepancy between simulation and measurements is seen below 60% and above 140% of B₁⁺, which is ascribed to inherent bias in the measured B₁⁺. Red line: The MP-FLAIR dependency to B₁⁺ used for bias-field correction.

Figure 3:
5 repetitions of an MP-FLAIR scan with different RF gains (0.6 – 1.4). For low RF gains, only the mid brain (where B₁⁺ is relatively large) experience close to nominal flip-angles. For RF gains higher than 0.8, the mid brain shows as hypo-intense due to flip-angles being larger than the nominal flip-angle. The lower occipital lobe (where B₁⁺ is relatively small), show as hypo-intense for all but the highest RF gain, where it experiences close to nominal flip angles and shows as hyper-intense, due to the surrounding tissue experiencing higher than nominal flip-angles.

Fig.1 Cartoon depicting the proposed acquisitions and processing pipeline. The images contained in the purple square (top left) represent the four SE-EPI scans that need to be acquired (each in 15 s).

Fig.4 Sample of the B1 map produced using the proposed method compared to the B1 map produced using a typical SE-DAM protocol in our Centre (as per ref.2). The horizontal and vertical profiles are plotted on the right to show the very good agreement between these two B1 maps.

Figure 1. Example of bias correction applied to native T1-WI. Maps were generated by dividing the corrected image by the native image. Bias correction methods N3 (top left), N4 (top right), N4_W(bottom left) and N4_ITER(bottom right).

Figure 2. Summary of iterative WM N4 bias correction technique. 1) Initial WM mask is generated with FSL-FAST on native T1-WI and used for N4Wcorrection. 2) WM mask is segmented from N4 corrected image with FSL-FAST. 3) Volume of WM mask from step 2, is compared to the volume of the WM mask from the previous iteration, or initial input WM mask if it is the first iteration. 4) If the change in volume is greater than 0.05%, the updated mask is used to perform N4Won the native T1-WI. 5) If the change in volume is less than 0.05 percent, the process is terminated.