Figure 3. Top Row: Susceptibility maps with 0.65 mm isotropic resolution for the 3D-GRE acquisition (left column) and 3D-EPI acquisitions with different acceleration factors (right three columns). Enlarged regions for each map are shown in the bottom two rows.

Figure 1. Axial, sagittal and coronal 3D-EPI images, taken with no parallel imaging (PAT) and at 0.65 mm isotropic resolution. Yellow lines are the outline of the corresponding 0.65 mm isotropic 3D-GRE sequence, indicating low levels of signal loss and minimal distortion.

Figure 4. In-vivo results of (a) ChEST, (b) MMSR-STI, and (c) STI reconstructions. The two subject results show similar contrasts in all maps. In particular, the CE maps show a much smaller contrast range compared to that of MMS and MSA (see display ranges), confirming our assumption in QSM that the contribution of CE is much smaller than that of susceptibility.

Figure 1. (a) An illustration of the iterative algorithm for ChEST. In the first step, STI is conducted from the CE-removed frequency shift map, producing PEV. Then, the least-squares algorithm is performed while fixing PEV to derive CE, MMS, and MSA. Through iterations, MMS, MSA, PEV, and CE maps are reconstructed. (b) The plot of RMSE between the source frequency shift maps and the reconstructed frequency shift maps in the numerical simulation. The algorithm converged after 200 iterations.

Fig. 1. Physical model for the training data synthesis, examples of the training data, and the neural network architecture for mapping the total frequency contrast f to either of the two source distributions χ or f_ρ.

Fig. 4. Three axial slices from a volunteer (f, 33y.). A) f_ρ has a stronger positive non-susceptibility frequency shift in gray matter than in white matter, and cerebrospinal fluid (CSF) has the smallest shift. CSF was reconstructed homogenously (orange arrow). The iron-rich basal ganglia (B) and dentate nuclei (C) were isointense in f_ρ (blue arrows). Compared to the HEIDI solution, DEEPOLE QUASAR’s χ of the inner capsule was more similar to surrounding white matter (purple arrow).

Figure 4. DECOMPOSE-QSM results of a susceptibility-mixture phantom showing the parameters and composite susceptibility maps in comparison with thresholding original QSM. Note that the subplots relate to the diamagnetic component are displayed with inverted dynamic range to have a better visual contrast.

Figure 1. (A) A cartoon illustration of the 3-pool signal model. (B) A flowchart of the proposed algorithm. The algorithm takes inputs of echo dependent QSM and Magnitude to compose the local signal. The proposed alternating direction solver processes the local signal and outputs the estimated unknowns. Note that $$$C_0$$$ is not shown due to the simple relation $$$C_0=1-C_+-C_-$$$. With the estimated parameters, maps of paramagnetic component susceptibility (PCS) and diamagnetic component susceptibility (DCS) are constructed respectively.

Figure 1 Manhattan plots comparing GWAS associations with QSM and T2* (reversed) IDPs in thalamus, pallidum and putamen. Here, the x-axis shows the SNPs (sorted by chromosome), with the y-axis revealing the significance level. Hits passing significance threshold (two grey horizontal lines) are in black. Selected lead SNPs in each genetic locus were labelled with names of the related gene and colour-coded for different biological functions involved. Common and distinct genetic loci are identified in these regions and between QSM and T2* IDPs (e.g. HFE – T2* and GFAP – QSM in Thalamus).

Figure 4 Voxel-wise correlation maps of 3 selected SNPs in genes related to iron homeostasis with susceptibility (left) and T2* (right) maps in MNI152 space. Colour overlays represent the strength of each voxel’s correlation. Susceptibility and T2* maps generally have similar spatial patterns of associations with these SNPs. These iron-related SNPs have very distinct spatial patterns of associations (purple arrows). Additional associated areas in the cerebellum (green arrows) were identified.

Figure 3: Water–fat imaging and QSM results in a subject with mainly osteoblastic bone metastasis. All 4 QSM methods agree with CT. However, the BFR+LFI methods show BFR artifacts especially in the spinal process region (top arrow) and the fat region in the height of the L5 vertebra (bottom arrow). The TFI method shows significant noise amplification in the anterior to the spinal chord (bottom arrow) and artifactual paramagnetic susceptibility values in all intervertebral discs. The wfTFI shows no noise amplification, no BFR artifacts and no artificial paramagnetic values in the IVDs.

Figure 4: Water–fat imaging (first row), clinical T1w TSE, T2 IP, T2 water, CT and QSM (second row) results in a female patient diagnosed with kidney cancer and mainly osteolytic bone metastases of the vertebrae T4 and T5 according to CT. Yet, the metastatic lesion appears T1- and T2-hypointense, thus suggesting an osteoblastic mass. The results of the wfTFI QSM method, however, suggest an osteolytic pattern, thus being in agreement with CT as the reference standard.

Figure 3. Statistical results

ANOVA results from comparisons of quantitative susceptibility mapping (QSM) within the segmented lesion based on the staging groups. The linear regression between the region of interest standard deviation (red line) and estimated days post-admission is shown (top corner). (a) Morphological Enabled Dipole Inversion (MEDI) method (P_ANOVA = 0.043), (b) Total Field Inversion (TFI) method (P_ANOVA = 0.0004). * = denotes statistical significance for post-hoc comparisons.

Figure 4. Comparison of MEDI and TFI measurements within the hemorrhagic lesion

Linear regressions between the Total Field Inversion (TFI) Quantitative Susceptibility Mapping (QSM) against the Morphological Enabled Dipole Inversion (MEDI) reconstruction (a), the MEDI-QSM against the CT_index (b), and the TFI-QSM against the CT_index (c) within the lesion. In (d), the shadowing artifact is shown around the hemorrhaged region using the MEDI reconstruction (top) , which is improved in TFI (bottom).

Figure 3: $$$\chi$$$ maps and difference images illustrating the effects of all tilt correction schemes in the numerical phantom. An axial and a coronal slice are shown for a volume tilted at 25° and a reference 0° volume with all $$$\chi$$$ maps calculated using the iterative Tikhonov method. The ROIs analysed are also shown (bottom left). Qualitatively, RotPrior performs the best while NoRot results in substantial $$$\chi$$$ errors across the whole brain. The results from TKD and weighted linear TV (not shown) are very similar.

Figure 1: Schematic illustration of oblique acquisition and proposed tilt correction methods for QSM. The scanner frame (x,y,z) and the image frame (u,v,w) are shown with respect to the main magnetic field B₀=B₀z (left). Proposed tilt correction methods are shown with the k-space dipole (right). RotPrior involves rotation of the tilted image frame into alignment (u',v',w') with the scanner frame. NoRot represents incorrectly misaligning the dipole kernel with B₀z simulating a common error.

Figure 5: Illustration of the effect of individual chemical shift and relaxation corrections. The susceptibility maps corrected for Type 1 and Type 2 CSA (column 4) clearly depict the paramagnetic fatty areas. The susceptibility maps generated from the conventional images (column 1) and from the SMURF images recombined without corrections for chemical shift effects (column 2 and 3) are more blurred and the values are visibly overestimated. Corrections for T₁ and T₂^* relaxation differences (columns 5 and 6) remove the dominant influence of fat signal in mixed voxels.

Figure 2: Comparison of QSM maps generated from individual echo images and echo-combined images for the head-and-neck (top) and brain-only (bottom). While in the neck, the first-echo QSM map correctly depicts the paramagnetic fatty areas and provides good CNR (top left), in the brain – with long T₂^* – the CNR is low. This is improved for later echoes at the expense of short T₂^* areas, such as the neck, sinuses and veins, where errors occur (centre). In the QSM maps generated from the echo-combined images (right) these errors are reduced and high CNR is achieved over the entire head-and-neck.

Single Orientation QSM results from QSMART, RESHARP/iLSQR and V-SHARP/iLSQR, and the difference maps between the susceptibility values calculated by QSMART and the two other methods. Dark cortical surface artifacts (red arrows) and x-shaped artifacts caused by veins and high susceptibility sources (green arrows) are successfully suppressed by QSMART.

Components of the SDF algorithm. Volume rendered A: Brain mask, B: Vasculature mask, and C: Indentation mask. D: Exemplar sagittal, coronal and axial slices of a spatially dependent filter size map ($$$\alpha$$$ map) derived in the SDF procedure to remove the background field.

(a) Fiber tracking results of the minor eigenvector, where fibers shorter than 5mm are not shown (left: STI, right: aSTI). The aSTI reconstruction recovers much more of the white matter fiber tracks. (b) The minor eigenvector direction of the mouse brain, weighted by SI (left: STI, right: aSTI). The aSTI color maps show more coherent fiber directions, and the SI also serves as a better indicator of white matter.

The χ₁₂ (top left) and χ₂₁ (top right) terms in the aSTI reconstruction, along with the symmetric (bottom left) and antisymmetric (bottom right) decompositions. The antisymmetric part contains mostly noise and streaking artifacts. As a result, after discarding the antisymmetric part, the symmetric part shows reduced noise and streaking compared to the top two images.

Figure 1. (A) Schematic of the QSM and STI pipelines. The brain used in for the data acquisitions of this abstract comes from a wild, female Chimpanzee (Pan troglodytes verus) named “Emma”, 6 years old, from Tai National Forest, Ivory Coast, West Africa. Cause of death was an attack by another group of Chimpanzees. (B) The general pipeline used for training of the machine learning model.

Figure 2. The three main tensor components (diagonal) from STI (A) and DTI (B). Differentiation of GM/WM is clear in both tensors. (C) Eigenvalues derived by susceptibility (sign is inverted) and diffusion tensors. (D) Maps of mean diffusivity (MD) and mean magnetic susceptibility (MMS). (E) Maps of fractional anisotropy (FA) and magnetic susceptibility anisotropy (MSA). MSA exhibits some common spatial characteristics with FA in white matter but with noticeably more variations.

Figure 3: Model fits of $$$\chi^B$$$. To the left, $$$\chi^B$$$ is fitted without the mesoscopic contribution corresponding to classical QSM. In the middle, $$$\chi^B$$$ is fitted with the mesoscopic contribution using $$$p_{2m}$$$ estimated by diffusion. To the right, the susceptibility difference $$$\Delta\chi^B$$$ is plotted. Visible changes are numbered: 1,4. Corpus callosum 2. Striatum 3. Optical tract 5. Cerebellum 6. Anterior Commisure

Figure 1: Model overview of fiber bundles. A bundle consists of long cylinders of different orientations. Each fiber may be constructed by several concentric layers. $$$\langle\Omega\rangle_k$$$ is averaged over all the interstitial bi-layers including the intra- and extra-axonal space. For simulations, the bundles are packed by randomly packing cylinders. Cylinder orientations are randomly selected up to a chosen cut-off in polar angle. The position is randomly chosen until a non-overlapping placement is found. A total volume fraction around 15% can be achieved in this way.

Figure 3. Comparison between ex-vivo 𝜒-separation results vs. iron and myelin histology. (a-b) Frequency shift and R2' maps. (c-d) Positive and negative susceptibility maps from 𝜒-separation. (e) Iron distribution from LA-ICP-MS. (f) Myelin images from LFB staining. Similar features are observed between positive susceptibility map vs. iron image as well as negative susceptibility map vs. myelin image. 𝜒-separation maps also delineate brain anatomy observed in iron and myelin histology (see color markers).

Figure 5. Comparison between in-vivo 𝜒-separation results (a-f) vs. iron and myelin histology from literatures (g-l). The positive and negative susceptibility maps demonstrate similar distributions with histologically stained iron and myelin images, respectively. The 𝜒-separation maps render exquisite details of the brain anatomy that agree with the iron and myelin histology (white arrows).

Figure 3. Parameter maps, paramagnetic component susceptibility (PCS) and diamagnetic component susceptibility (DCS) compared to thresholding QSM. The line artifact in the third column is due to a dicom file error of two slices in the magnitude images. The increasing trend of PCS is visible, while the temperature related change in DCS is minimal.

Figure 2. DECOMPOSE QSM results for 16 echoes data set. Mean value of each parameters of all 10 scans are shown. Temperatures range from 37 °C to 21 °C. Mean value is calculated from the non-zero mean of one representative slice. Note the paramagnetic component susceptibility is increasing with decreasing temperature as expected.

Fig. 2: Transverse and coronal slices of quantitative R1, MD, R₂*, χ and f_ρ maps after fixation. Unlike in vivo, R1 maps show enhanced values in iron-rich DGM, similar to R₂* and χ, and greatly reduced GM & WM contrast and overall increased rates. MD in WM is much lower than typical in-vivo values (5-7x10^-4mm²/s). QUASAR χ shows fixation artefact at the brain surface (bright ring, orange arrows). The non-susceptibility source map contrast is dominated by fixation artefact (gradient change from brain surface toward the centre), which are also visible in R1 and MD maps (purple arrows).

Fig. 3: Top 2 rows show ROIs of excised samples. 3^rd row shows χ_i and χ_a of excised specimens measured by external field. Bottom 2 rows show residual field inside the sample fitted with Eq.4. WM at the genu and splenium of the corpus callosum (CC7-9) showed both the strongest diamagnetic isotropic susceptibility but very poorly described residual fields. The CST sample had residual DGM tissue in the corner of the sample, explaining the strong positive susceptibility compared to other WM. These 4 samples were excluded from the correlation analysis in Fig.4.

QSM results using the QSM challenge dataset. Our two-pass QSM technique reduces streaking artefacts near the brain lesion and includes more brain detail.

QSM results using our gel phantom dataset. On the left is a photo of the gel phantom containing a gold fiducial marker with 1.5% iron content (yellow square), a pure gold fiducial marker (blue square), and a tooth piece (green square). The latter three images depict the same slice from the magnitude, single pass, and two-pass QSM. The two-pass QSM image has a clear reduction of streaking artefacts surrounding the strong sources.

Figure 4: In vivo healthy liver results show a reduced contrast between high and low PDFF regions in the field- and χ-map. The susceptibility map within the liver (box) shows a significant underestimation when using in-phase echoes, although the liver fat content is negligible. Further, when using in-phase echo times the overall sharpness and contrast is reduced in the χ-map, the depiction of fine structures such as vessels is limited (left arrow), the susceptibility values are heterogeneously distributed within the liver and more BFR artifacts are apparent (bottom right arrow).

Figure 2: Numerical simulation results of the lumbar spine. In-phase echo times induce a 24% higher error in the fieldmap and 44% higher error in the susceptibility map when compared to standard water–fat separation echo times. The overall contrast of the QSM map is reduced using in-phase echo times and the intervertebral disc region shows an artefactual paramagnetic increase (arrow). When Laplacian unwrapping is additionally performed on the field-map based on the water–fat separation echo times, the estimated susceptibility-map does not change significantly.

Figure 3: Boxplots of subject mean whole-brain a) venous susceptibility and b) venous density from five different acquisition sequences (different colors) and six QSM methods (columns) and c) statistical analysis. Subject mean values were calculated across all voxels obtained from multiscale vessel filtering (MVF) on individual susceptibility maps within a common minimum-size brain mask. Differences in subject mean venous susceptibility are greater for different QSM reconstruction methods (effect size η_p²=0.861) than for different acquisition settings (η_p²=0.016).

Figure 1: First-echo magnitude images and susceptibility maps from different QSM reconstruction methods (columns) in a representative healthy subject. The same axial and sagittal slices are shown for a) Full FC, b) No FC, and c) TE1 FC CS sequences (rows). Differences between the six QSM methods are clearly visible in the extent of brain erosion (orange arrows), the delineation of the straight sinus (blue arrows), and the contrast between various brain tissues. First-echo magnitude data (first column) are shown in arbitrary units and scaled within the same intensity range.

Figure 1: Sample images from the 1st echo magnitude, R2* and QSM of one subject for scan and rescan. Although magnitude contrast varies across sites (due to TR and flip angle variation), quantitative maps are similar except at large susceptibility-gradient regions (arrows).

Figure 3: Correlation analysis of R2* and QSM reproducibility. Measurements were highly correlated with root mean square error (RMSE) ≤1.1 s^-1 and ≤3.8 ppb, and slope of 1.04±0.04 and 0.99±0.02 for R2* and QSM, respectively.

Figure 5: Axial knee/ upper thigh susceptibility maps of a conventional QSM pipeline without fat water separation (A), of a fat water separation pipeline and subsequent conventional QSM pipeline (B) and the UNET prediction (C) of two subjects. Susceptibility maps of both subjects are proximal to the knee joint, while the susceptibility maps of Subject 2 are further distal at the level of the patella. Black arrows highlight artifacts at the border between subcutaneous fat and muscle, which are improved in the UNET prediction.

Figure 2: Illustration of the training process. Randomly cropped input and target data of the above data set, which served as training data for a UNET. The order of the input phase patches was selected randomly for each data set to avoid overfitting and to ensure that the network is not focusing on specific echo times recognized solely by the input order and not due to the image contrast.

Figure 1. Performance comparison of the networks, QSMnet_1x1x1 and QSMnet_1x1x3, trained by the datasets of different resolutions (1x1x1 mm³ vs. 1x1x3 mm³). The network trained with the same resolution data (e.g., QSMnet_1x1x1 with the test data of 1x1x1 mm³) outperformed the other network.

Figure 3. Performance comparison of the networks of two different hyperparameter sets. The hyperparameter of the original QSMnet was empirically determined for the 1x1x1 mm³ training dataset. When the same network was trained with a new training dataset of 1×1×3 mm³ resolution, it resulted in NRMSE of 57.8 ± 7.1% (QSMnet_1x1x3-ref). Since the training dataset characteristic has changed, we can further improve the performance by hyperparameter tuning, resulting in NRMSE of 56.6 ± 7.0% (QSMnet_1x1x3-hyper). The calculated p-value was 4.7e-4.

Figure 1. Reconstruction results of QSMnet⁺ in various test data types. Underestimations in susceptibility values are observed (red arrows) in the higher resolution (0.5 mm) in-vivo brain, numerical brain, and geometric shape images. To overcome this limitation, we propose a data augmentation method to expand the diversity of spatial gradient of training images. The underestimation is reduced in the proposed method (blue arrows).

Figure 4. QSM maps of the five inputs from Figure 3 reconstructed by the three networks (Unet_High, Unet_Low, and Unet_High+Low). The ground-truth susceptibility maps and corresponding spatial gradient maps are shown in the first two rows. The best performance is highlighted by green boxes. As highlighted by red arrows, the network with high spatial gradient (Unet_High) underestimates susceptibility values in low spatial gradient regions, revealing training data dependency. The augmented Unet_High+Low shows high performance in all inputs (yellow and green boxes).

Illustration of the NeXtQSM pipeline including the training data generation process (blue) and the two deep learning models trained jointly in one optimization (red). In the data generation, we apply the QSM forward operation to the synthetic brain with and without external sources to have the inputs for the two learning steps. In the learning part, the two architectures can be seen as a unique one because of the end-to-end training fashion.

Illustration of the training dataset. The left column shows the starting synthetic brain, the center shows the data after application of the QSM dipole model and on the right the data after applying the forward model including the effect of external sources.

Figure 3: Optimal reconstructions of the RC2 numerical phantom for all regularization methods (TDV, TV and TGV). TDV results show better depiction of the veins and streaking artifact suppression relative to TV and TGV (relevant areas highlighted with red arrows).

Figure 1: Optimal reconstructions and error maps of COSMOS-based forward simulations (RC1) for all regularization methods (TDV, TV and TGV). A sagittal, coronal and axial slice are shown. All methods were terminated after 500 iterations. TDV shows better depiction of cortical areas and less staircasing and streaking artifacts (highlighted with red arrows) than TV and TGV and achieves better RMSE and XSIM scores. Detailed comparisons between TV and TDV are shown for each labeled region (a-d).

Fig 1. Neural network architecture of QSMInvNet+. It has an encoder-decoder structure with 9 convolutional layers (kernel size 3x3x3, same padding), 9x2 domain-specific batch normalization layers, 9 ReLU layers, 4 max pooling layers (pooling size 2x2x2, strides 2x2x2), 4 nearest-neighbor upsampling layers (size 2x2x2), 4 feature concatenations, and 1 convolutional layer (kernel size 1x1x1, linear activation). To address the unsupervised domain adaption on real data, domain-specific batch normalization layers are utilized while sharing all other model parameters.

Fig 3. Comparison of QSM of a multi-orientation dataset. TDK (a), MEDI (b) and uQSM (e) results show black shading artifacts in the axial plane and streaking artifacts in the sagittal plane. QSMInvNet (c) displays better quality then TKD and MEDI, but shows subtle artifacts. QSMInvNet+(d) results have high-quality with clear details and invisible artifacts.

Table 1. Convergence study. We computed the number of iterations of each iterative scheme which resulted in the lowest NRMSE when compared to COSMOS, multi-orientation NDI, and multi-orientation HANDI for five different datasets.

Table 3. Comparison of NRMSE for HANDINet, variational network NDI (VaNDI), and variational network with standard dipole deconvolution (VarNet).

Figure 5: Comparison of Auto NDI_CG and Automatically stopped NDI with FANSI in vivo. High quality NDI reconstructions are possible even with extremely noisy voxels. Auto NDI_CG and Auto NDI produce very similar susceptibility maps (with substancial control of streaking artifacts and sharp fine details) but NDI_CG is much faster and showed no signs of susceptibility underestimation.

Figure 1: Spatial frequency components of the RC1 ground truth and NDI_CG reconstructions for different numbers of iterations. With more iterations, frequencies close to the magic angle are amplified. To assess QSM optimality (RMSE: 18 iterations), the mean absolute values of the frequency coefficients in regions M1-M3 have been used previously8. Here, regions M4 and M5 near the cone were used. All regions were defined as a function of the dipole kernel coefficients and radial frequencies 0.60-0.95 mm-1.

Figure 2: COSMOS forward simulation results. 500 iterations were performed on all reconstruction methods. In the results with snr = 40, it is observed that the Huber loss has a better noise reduction capacity than the L1 and L2 norms. In the results with snr = 100, it is observed that nlHu, unlike FANSI manages to reconstruct the veins in the cortical area and, unlike nlL1, it has a less noisy appearance.

Figure 1: The graph on the left shows the cost functions, you can see how the huber loss penalizes less the large values. The graph on the right shows the penalty functions, you can see that as the parameter $$$\delta$$$ decreases the Huber loss converges to a soft threshold(L1-norm).

Figure 1: Optimal (RMSE-based) reconstructions of the COSMOS forward simulation.

Figure 3: Results in in-vivo data. The optimal number of iterations was selected by visually inspecting the result of each iteration.

Figure 1 Associations of the corrected volume in hippocampus and temporal cortex, QSM in ENT, CN, PT and GP, to the annual change in global cognitive functions in all participants (n = 150). Annual change of cognitive function was calculated for each participant based on his or her trajectory of decline in the follow-up years after baseline MRI (up to 5 years), and it was plotted against baseline predictor, i.e. corrected volume or QSM. CI (Max-Min) = confidence interval of the difference between the annual change in global cognitive score at the minimum and maximum values of the predictor.

Table 1 Associations of structure volume, tissue iron load (QSM) and β-amyloid load (PET DVR) with rate of change in global cognitive composite scores (variable×time interaction in mixed-effects models) in all participants and participants in the PET group. d: effect size.

Figure 1. Mean images of SUVR and QSM for Aβ positive and negative patients.

Figure 2. Evaluation results.

Figure 4: In each plot the diagnostic accuracy for each group pair (HC vs MSA-p, HC vs MSA-c and MSAp vs MSAc) is reported. The colors represent the values of the area under the ROC curve (AUC) for each TE, for each ROI and for each histogram feature. It can be noticed that AUC values are higher for short TEs.

Figure 1: The top row shows the ROIs used in the analysis overlaid onto the study-specific template, which was computed from the across-TEs average of the T2*-weighted image of each subject. In the bottom row the susceptibility maps computed from the first TE averaged over all the subjects is displayed.

Figure 1. Quantitative susceptibility maps within thalamus (purple), caudate (blue), pallidum (red) and putamen (yellow) of a 38 years-old female RRMS vs. 40years-old female HCs.

Table 3. Spearman’s correlation between iron-rich regions in deep grey matter in MS cohort only.

Figure 2: Positive, negative and conventional susceptibility maps, as well as a T₁-weighted image for lesions from three different patients. For the enhancing lesion, additionally a contrast-enhanced T₁-weighted image is shown. Lesion age is indicated on the left and lesions are ordered according to their age. Green indicates a contrast enhancing lesion, yellow arrows point to non-enhancing lesions.

Figure1: The cancellation of the susceptibility effects within the same voxel in QSM are illustrated.

Figure 1. Regional magnetic susceptibilities measured in the brain of healthy control infants and those diagnosed with CHD in the first two months of life. Asterisks denote significant differences between control and CHD infants when controlling for PMA and accounting for repeated measures of the same infant.

Table 2. Associations between magnetic susceptibilities measured in pre/post-operative MRI and Bayley-III receptive/expressive language scores. P values in bold indicate significant associations.

Figure 2. Boxplot of CSVO2 percentages by group. Grey circles are the ROI measurements from each subject.

Figure 1. Sample internal veins selected after thresholding out the lower 99.75% χ values. As can be seen in these sagittal, coronal, and axial views from a sample healthy term neonate, the major veins that were left over include the straight sinus, inferior sagittal sinus and the internal cerebral vein. Note the weak contrast between grey and white matter and the basal ganglia due to the low myelin and iron content of the newborn brain.

a) Sagittal slice from the QSM in a representative sickle cell anaemia (SCA) subject with the superior sagittal sinus (SSS) region of interest overlaid in red. b) Comparison of QSM venous oxygenation (Y_v) in the SSS in subjects with sickle cell anaemia (SCA) and healthy controls (HC). Significantly lower Y_v was observed in the SCA group which also showed a wider range of Y_v values than the HC group.

Comparison of QSM venous oxygenation (Y_v) measured in the superior sagittal sinus of subjects with sickle cell anaemia (SCA) and healthy controls (HC) with (SCI+) and without (SCI-) silent cerebral infarcts (SCI). In SCA, significantly lower Y_v was measured in subjects with SCI relative to those without. No significant differences in Y_v were observed between the SCI+ and SCI- groups in HC subjects.

Figure 3. Relationship between venous oxygen saturation and hemoglobin measured by QSM and CISSCO.

Figure 1. Acquired image and the internel cerebral vein. Magnitude in axial view (top left), magnitude in coronal view(bottom left), phase in coronal view (bottom right) and processed susceptibility map (top right).

a) Axial slice of a T2-weighted FLAIR image showing two silent cerebral infarcts (SCI) in a representative sickle cell anaemia subject (right: SCI ROI overlaid in red). b) Axial slice of the T1-weighted MP-RAGE image showing the same lesions appearing hypointense. c) Similar axial slice of the Std-GRE magnitude at TE = 27ms in the same subject. d) A different axial slice from the MP-RAGE showing the segmented juxtacortical, deep and periventricular white matter regions. e) QSM in the same axial slice as in c) showing a slight χ difference between the lesions and normal appearing WM.

Mean SCI lesion χ within each brain lobe in the combined sickle cell anaemia and healthy control groups. Mean SCI χ are significantly lower for lesions in the parietal lobe relative to the frontal lobe (diff = -0.017ppm, p<0.001). In white matter, myelin is the key diamagnetic (negative) susceptibility source. Therefore, these results suggest reduced myelin concentrations within lesions in the frontal lobe relative to the parietal lobe. No significant χ differences were observed between SCI any other regions.

Figure 5. Compared to the original liver QSM method (a: test, b: re-test), the proposed method (c: test, d: re-test) produces more consistent (repeatable) susceptibility maps. The improved repeatability may be explained by the fact that the proposed method is less sensitive to errors in the field map induced by air/tissue interface due to adaptive data weighting, which, together with fat constraint, reduces shading artifacts.

Figure 3. Correlation of R₂* values in the liver with susceptibility values and the corresponding linear regression lines for test (red) and re-test (blue) exams obtained with (a) the original method (slopes 0.002/0.002, y-intercepts -0.334/-0.289, R²=0.858/0.939); (b) the proposed method (slopes 0.0029/0.003, y-intercepts -0.577/-0.566, R²=0.982/0.985). These slopes/intercepts are consistent with previously reported⁵ correlation results (slope 0.0028, y-intercept -0.54).

Figure 1: Training of DL QSM: (1) Digital torso 3D phantom with various random-susceptibility regions; (2) Augmented with rotation and local elastic deformation; (3) Convolved with a dipole kernel (voxel:1.56x1.56x0.9mm³, matrix:256³) to create a field map; (4) A random background field was added; (5) The field map was down-sampled to 8mm slices (in-vivo resolution), then interpolated to 2.25mm to enable a deeper network; (6) 64³ patches (n=100) were randomly pulled for training.

Figure 2: Representative susceptibility maps from a previously proposed L2-regularized liver QSM method (top), and the proposed CobraChi DL-based method (bottom), in patients with various levels of liver iron. Compared to the L2-regularized method, CobraChi may provide better robustness at high iron levels, although some artifactual shading remains at low iron levels. For both QSM methods, Δχ of the liver is measured relative to subcutaneous fat.

Figure 1: Magnitude, SWI, tSWI, QSM and mIP from SWI and tSWI in a patient with an acute Day 2 hemorrhage. It demonstrates the blooming effect in SWI of hemorrhage.

Figure 2: A comparison between phase, SWI and tSWI in a patient with Day 12 hemorrhage to show phase wrap artifacts are present in SWI for large hemorrhage which gets removed in tSWI as shown by the red arrow.

Figure 1. Distributions of the most important variables of the random forest by diagnosis group.

Figure 2. Example of QSM images of the primary motor cortex region obtained from ALS and mimics patients.

Figure 2. Representative axial QSM images of a healthy control, NPSLE patient and non-NPSLE patient. The regions of interest are indicated in the healthy control. TH = Thalamus; CN = Caudate nucleus; PT = Putamen; GP = Globus pallidus.

Figure 5. Iron in control and SLE brain. Globus pallidus showed higher staining intensity compared to putamen. Within the putamen small areas of increased iron were found originating from myelinated fiber bundles. Iron was predominantly found in oligodendrocytes and myelin, to a lesser extent, in neurons, microglia and astrocytes (arrows) in both control and SLE.

Figure 2: Maps of the relative signal changes (A-B)/B∙100% between the different scan settings. The colour bar is scaled from -30 to 30%. Temporal and cerebellar regions benefit when increasing FA, but signal is lost in the central brain region (top). Turning the RF spoiling OFF increased the signal in the overall brain and particularly for the CSF (middle). Increasing the FA for RF spoiling OFF, however, shows more signal loss in the overall brain (bottom).

Figure 1: WM signal magnitude (A) for different TE, TR and FA settings following equations 1 and 2, phase SNR (B), scanning efficiency (C) and the corresponding FAs and imposed SAR (D).

Figure 2. Comparison of the MS lesion susceptibility trajectories estimated using CSF_vent mask (left) and the full CSF_brain (right) masks. Note synchronous jumps in measured susceptibility values occurring with CSF_vent

Figure 3. Comparison of the CSF_vent for different datasets in longitudinal study

Figure 1. PDM is the absolute value of the difference between the full mask reconstruction, and the perturbed, 5mm eroded mask reconstruction. Artifacts shown with yellow arrows.

Figure 4 a) QSM and PDM in a hemorrhage case. Shown are MEDI, MEDI SMV, TFI, TFIR, and the field uncertainty. b) shadow score and PDM average for each method for 11 cases. Both scores decrease together as visual image artifacts decrease in a).

Figure 4– (left) Representative hemorrhage case with reconstructions, arrows show artifacts, (right) Shadow scores for 12 cases.

Figure 1 –Algorithmic flow chart for POSSUM

Table1. The susceptibility of tumorous and peritumoral from MVI+ and MVI− HCC.

Figure1. Hepatectomy pathological confirmed hepatic carcinoma (Yellow arrow) with MVI− (Top row) and MVI+ (Bottom row). ROIs were placed on the maximal slice of tumor (red zone) and peritumor (green zone). The tumorous and peritumoral susceptibility of HCC with MVI− and MVI+ was -0.12 ppm, -0.09 ppm, 0.13 ppm, 0.16 ppm, respectively.