Figure 4: (A) Diagnostic performance of different microstructural markers to differeciate clinically significant (GS>1) and insignificant (GS≤1) cancers in terms of area-under-the curve (AUC), accuracy, sensitivity and specificity, based on a five-fold cross-validation using the linear discriminator. (B) Receiver-operating-curves (ROCs) of the different markers. (C) Classifying clinically significant (red) and insignificant (blue) cancers using the combined marker of the cellularity and D_PGSE(30ms). The mis-classified cases are marked as “x”.

Figure 1: Microstructural maps of the prostate tissues of Gleason score (GS) from 0-4. Intracellular fraction (f_in), cell diameter (d), cellularity, and extracellular diffusivity (D_ex) fitted from the IMPULSED model and the diffusivity maps from PGSE, OGSE (17Hz), and OGSE (33Hz) data were shown. The white contours and red arrows indicated the cancerous regions based on manual delineation.

Figure 3: RSI cancer-likelihood map of a patient with prostate-cancer metastases in the pelvis and femur (cyan arrows), compared against conventional MR images. Bone lesions show a very high likelihood value [probability of being cancerous; P(cancer)] compared to surrounding normal tissue. Normal tissue is generally less pronounced on the likelihood map than on conventional MR images. False positive signal remains, however, in organs with dense cellular arrangement like the kidneys and brain.

Figure 2: RSI signal distributions for normal tissue and bone lesions. The joint C₁,C₂ probability density functions (PDFs) are shown for normal control tissue (left) and bone lesions (middle). Both PDFs are shown after log transformation to better show less frequent combinations of C₁ and C₂. The posterior probability distribution on the right is derived from the PDFs and shows the likelihood of cancer [P(cancer)] given particular C₁ and C₂ values. High C₁ signal in particular is indicative of cancer.

Figure 1. Schematic of biomimetic lipid nanoparticle phantom to reflect multi-compartment contribution to DWI voxel signal in PCa (GS7 histology) and example fit parametric maps for diffusion-kurtosis model. Sizes of the lipid nanoparticles can be controlled to be between 200 and 2,000 nm, depending on the composition of the nanoparticle shell.

Figure 3. Scatter plot of apparent kurtosis, K_a, versus apparent diffusion, D_a, for in vivio PCa and selected diffusion-kurtosis phantom materials. The range of values found in PCa GS6 and GS7 lesions for D_a and K_a (estimated by fitting DWI decay curves to the isotropic diffusion kurtosis model) can be spanned by fabrication of lipid nano-materials with appropriate physical properties.

Figure 2. Representative FROC-derived parametric images of one patient with micro-vascular invasion.

Fig. 4: Diagnostic performance of FROC-derived parameters for predicting the MVI in patients with HCC.

Figure 1: overview of the two-step DR-HIGADOS framework. In the first step, an extension of the IVIM model is fitted to diffusion-relaxation measurements. In the second step, model parameters are mapped to biophysical properties (e.g. average intrinsic cell diffusivity and size; cellularity), using information derived from Monte Carlo simulations.

Figure 2: parametric maps provided by DR-HIGADOS in vivo on two MRI vendors. Left to right: non-DW image, vascular signal fraction f , vascular diffusivity and T2 (D_v and T_2V), tissue diffusivity, kurtosis excess and T2 (D_T, K_T and T_2T), average cell diffusivity and size (D₀ and L), cellularity C. Top to bottom: healthy volunteer on Ingenia system; healthy volunteer on Avanto system; patient on Avanto system.

MR Microscopy, Pdx1-Cre;Kras^G12D mouse pancreas. A – PanIN in ADM background. B – Area of PDAC. As observed in the previous example, PanIN (yellow arrows) are most conspicuous with high b values and longer echoes, with high peripheral anisotropy values. PDAC (green arrows) presents with similar contrasts, but with lower anisotropy values in its solid central portions, possibly due to a different cellular and fibrotic content.

In vivo dMRI. (A) The healthy pancreas has low signal intensity in DWI and low anisotropy values. (B) A large abdominal mass is seen, compatible with PDAC (green arrows), with high DWI signal intensity and high anisotropy values in its solid areas. (C) Much smaller lesions, compatible with PanIN (yellow arrows), are depicted with similar contrasts. (D) Similar lesions are observed (yellow arrows), along with diffuse pancreatic changes compatible with PanIN/ADM (blue arrows).

Figure 1: Workflow of the study.

Figure 3: T1 image in a), CTRW maps, D_m (in mm²/msec) in b), α in c), β in d), and co-registered histology-based blended probability maps in e) for three representative slices from the NAs. The blended probability maps in e) were generated by determining the “assigned” heterogeneity level (AHL) as the one with the highest probability; and displaying AHL’s probability with a customized color bar. For example, a yellow pixel with a value of 0.90 was predicted to have a heterogeneity level of L3 with a 90% probability. The L2 and L3 ROIs given in f) were drawn with the guidance of AHLs.

Figure 1: Example b=0 images of a non-deescalated and deescalated patients at pre-treatment and week 4 into the treatment. The metastatic lymph nodes are noted by arrows.

Figure 3: Box-whisker plots of diffusivity and diffusion kurtosis of metastatic lymph nodes measured at long diffusion times. Black square boxes are for non-deescalated patients (n=6) and red boxes with notches are for deescalated patients (n=12).

Figure 5: Scatter plots of post-NAC (pre-operative) ADC_median vs. tumor cell fraction (left panel) and change in ADC_median after three or four cycles of neoadjuvant chemotherapy and percentage necrosis.

Figure 4: Normalised probability density functions for ADC estimates from ovarian, omental, and peritoneal lesions, and lymph nodes at baseline (pre-NAC) and after three or four cycles of neoadjuvant chemotherapy (post-NAC, pre-operative).

Figure.1. Images in a 49-year-old woman with type I, low-grade (grade 1) EA (arrowheads, Ki-67 = 30%). (a) DWI original maps (b = 1000 s/mm²), (b) Pseudo colored maps of K_app, (c) Pseudo colored maps of D_app, (d) APTWI original maps, (e) Pseudo colored maps of MTRasym (3.5ppm), (f) Pathological images (original magnification, ×100).

Figure.1. Images in a 49-year-old woman with type I, low-grade (grade 1) EA (arrowheads, Ki-67 = 30%). (a) DWI original maps (b = 1000 s/mm²), (b) Pseudo colored maps of K_app, (c) Pseudo colored maps of D_app, (d) APTWI original maps, (e) Pseudo colored maps of MTRasym (3.5ppm), (f) Pathological images (original magnification, ×100).

Fig. 3 Simulations for two compartments separated by a membrane of permeability $$${w}=0.6\,\mu m/ms$$$. Compartment sizes: $$$L_L=4.5\,\mu m$$$ (left), $$$L_R=5.5\,\mu m$$$ (right), diffusivities: $$$D_L=3\,\mu m^2/ms$$$, $$$D_R=2\,\mu m^2/ms$$$. Time-scales: $$$\tau_L=L_L^2/\pi^2D_L$$$, $$$\tau_R=L_R^2/\pi^2D_R$$$, $$$\tau_{ex}=\sqrt{D_L D_R}/{w}^2$$$. Top to bottom: true propagator, estimated propagator, EAP. Left three columns: near-ideal parameters. Last column: $$$\delta_1=200\,ms,\delta_{2,3}=1.2\,ms$$$, and $$$G_{max}=10\, T/m$$$.

Fig. 1 (a) Effective gradient waveform of the Stejskal-Tanner sequence [1]. The signal can be transformed into the ensemble average propagator [2]. (b) Effective gradient waveform of the sequence by Laun et al [3], which can be utilized to obtain the long-time form of the diffusion propagator. (c) Effective waveform for one realization of the experiment considered. Here, $$$-\mathbf q-\mathbf q’$$$, $$$\mathbf q$$$, and $$$\mathbf q’$$$ are the signed areas under the first, second, and third gradient pulses, respectively. The signal can be transformed into the actual propagator.

Figure 2. Decay curves for the diffusion signals parallel (light shade) and perpendicular (dark shade) to the fiber axis plotted as a function of q-value, for ROIs in the SLF (green) and CST (blue). Curves are shown for diffusion times of 30 ms (solid lines, filled markers) and 50 ms (dashed lines, open markers).

Figure 1. Left: Sample extracted from a coronal slab of a human hemisphere. Right: Locations of SLF (left) and CST (right) ROIs from one representative coronal slice, overlaid on the primary fiber orientation vectors and corresponding FA map from GQI reconstruction.

Figure 2. Segmented parcels #1-#13 of the corpus callosum; and parallel diffusivity measurements at different OGSE frequencies of the corpus callosum. Different colors represent different subjects. L: left; R: right.

Figure 1. OGSE with different frequencies and PGSE with flattened waveforms used for diffusion imaging (top), and the corresponding diffusion spectrum (bottom).

Figure 4: Differential ADC (A, C) and differential kurtosis (B, D) maps for both subjects produced from the images shown in Figure 3. ΔD and ΔK maps are generated from the subtraction of PGSE images from OGSE images using ADC maps and apparent kurtosis maps respectively. The distortions in the frontal lobe of Subject 1 are due to B₀ inhomogeneity.

Figure 1: Gradient waveforms (A, C) and spectral densities (B, D) of a conventional N = 2 cosine OGSE waveform (A) and our abbreviated oscillating gradient waveform (C). Here the second diffusion gradient is inverted to reflect the effect of the refocusing RF pulse (not shown). A reduction of the total diffusion gradient duration is apparent for the new waveform (C) compared to OGSE (A) at the same oscillation frequency of 25 Hz. Comparison of the respective spectra in (B) and (D) shows similar spectral selectivity and demonstrates zero DC spectral component for both waveforms.

FIGURE 3 - The 'classical' view of impact of SNR on eigenvalues, mean diffusivity, radial diffusivity and fractional anisotropy. In this simulation, the number of averages was not increased to compensate for the loss of SNR. This result provides useful insights into the impact of noise. E.g. we see both over-estimation of the largest eigenvalue (when FA^true is low) and under-estimation (FA^true is high), in line with previous literature^1,2 on repulsion and squashing peanuts

FIGURE 2 - Results of pair-wise t-tests, between all possible [SNR, N] couples. Blue entries show where the hypothesis that the samples come from the same distribution cannot be rejected at the p < 0.05 level. Yellow entries are where we have to reject this hypothesis and deem that the means are significantly different. Note the pattern depends both on FA^true and on the metric itself.

Fig. 1. The results from 488 samples for both shelled (61×8) and non-shelled point sets ¹⁸ in the presence of Gaussian noise. (a) the mean and std of the estimated signal versus b-value using MAP-MRI method with N_max = 6 for five different noise floors, σ_g, and three different dispersion values, κ. The thickness of the blue band is twice the standard deviation of the signal estimates and its center is the mean. The dashed black line shows the ground truth and the red dots and bars show the results of the SH (L = 6). (b) the mean and std of the d₁ and d₂ measures for different methods.

Fig. 5. (a) the estimated d₁ and d₂ for three different κ values, 344 (43×8) point sets in the presence of five different Gaussian noise levels. (b) the results of MAP-based interpolation of the orientationally-averaged data from Knutsson method (`MAP, Knutsson, s8' and `MAP, Knutsson, s43') on 43×8 shelled Lebedev ²⁵ and the interpolation of original data (before averaging) on shelled (`MAP, s8' and `MAP, s43') and non-shelled point sets (`MAP, ns8' and `MAP, ns43') ¹⁸.

Figure 1 – Graphical overview of proposed CED assessment pipeline

Figure 2 – Simulated (top row) vs. clinical (middle row) ROC curves for Zhao’s dataset⁷, for Tasks 1-3. The bottom row compares clinical AUC values to the distribution of simulated AUC values obtained by sub-sampling the pipeline dataset to match Zhao’s sample sizes. All ROC curves are qualitatively similar; the relative performance (AUC values) of different IVIM parameters are equal; all AUC values are in numerical agreement once clinical sample size is considered

Figure 2: Example of the parametric maps obtained using the whole available dataset (REFERENCE) and the minimum acquisition scheme (MAS), and the percentage error maps (ERR [%]) for the main parameters for each model.

Figure 1: Frequency of appearance for each shell in the minimum acquisition schemes computed in the population. Red shells represent the population minimum acquisition scheme (MAS) for each model.

Figure 3. Contour plots of ln (1- a) for different b_max (from 2,000 to 7,000) and different b-spacing power (r) values for four different SNRs (single real or imaginary channel) of 50 (A), 100 (B), 200 (C), 300 (D). The log spacing option (gray arrows) was appended to the far right to the r values in each panel with white shading for differentiation. For normal pancreatic tail tissue, a b_max under 3000 s/mm² is sufficient for practical achievable SNR and even b-spacing pattern is often among the optimal choices. See Fig. 4 for more details.

Figure 4. The numerically determined radius (R) of curvature reciprocal for the “true” DWI curve (Fig. 1 solid curve) is plotted against the b-values in panel A. When 1/R approaches zero, the b-space decay approaches a straight line. Panel B plots the derivative of (1/R) with respect to b. It quantifies how quickly the curvature in panel A changes with respective to b. For this case, the most dramatic change occurs when b < 3,000 s/mm², matching the Fig. 3 simulation results. (The ordinate units are not given.)

Figure 2. Repeatability test of the new mcT₂ algorithm on in vivo brain data. Parametric maps of white matter (WM) segments from 3 consecutive scans of the same subject using the suggested algorithm. (a-d) mask and myelin water fraction (MWF) maps of genu of corpus callosum (GCC). (e-h) mask and MWF maps of splenium of corpus callosum (GCC). (i-k) mask and MWF maps of the cortical WM segment (CSEG). MWF maps are presented with the same color scale and on top of a T₂ map presented in gray scale.

Figure 1. Flowchart describing how to identify spatially global microscopic features for a given white matter (WM) segment. (a) A correlation-based probability score is computed between WM segment data and dictionary elements. (b-c) All scores are normalized and powered by β to prioritize the signals according to their probability to be found within the segment. (d) Summation of all signals scores to assign a global score for each dictionary element. (e) Selection of a L elements with the highest score.

Maximum achievable b-value as a function of TE, for the gradient head insert (solid lines) and body gradients (dashed lines). RF pulse durations of 4 and 6 ms were assumed. The curves are shown for varying readout times [0, 5, 10, 15] ms e.g. depending on resolution and FOV, with a readout time of 0 ms representing a readout strategy starting at the centre of k-space (e.g. spiral imaging^5,14).

a) Signal decay as a function of TE at 3T (blue) and 7T (red), with t_TE the threshold where the signals are equal. A linear increase in signal was assumed as a function of field strength and the apparent T₂ was set to 50 and 77 ms at 7T and 3T respectively. b) t_TE as a function of other settings for the relative signal gain at 7T - which can depend on e.g. body noise¹¹ - and T₂.

Standard Model. We show parametric maps for Standard Model parameters: axonal volume fraction $$$f$$$, free water fraction $$$f_w$$$, angular dispersion $$$\theta_{disp} = \cos^{-1}\sqrt{(2p_2+1)/3}$$$, axonal (neurite) diffusivity $$$D_a$$$, longitudinal extra-axonal diffusivity $$$D_e^\parallel$$$, and radial extra-axonal diffusivity $$$D_e^{\perp}$$$. Streamline tractograms are are colored by their correspsonding SM parameters.

RMT Reconstruction. (A) 5/32 coils and their corresponding residuals are shown before and after noise removal via RMT. Properties of the removed normalized residuals $$$r$$$ are evaluated via (B,C) histograms $$$p(r)$$$, and (D,E) power spectrum analysis. (F) Standard and RMT reconstructions following parallel imaging and coil combination of a $$$b=4000\,$$$s/mm$$$^2$$$ image are displayed for several slices.

Figure 4: The k_bo map from DCE-MRI and AXR map from FEXI on the same slice. k_bo’s underlay is enhanced image, and AXR’s underlay is T2-weighted image. The k_bo and AXR shows similar spatial patterns as pointed with red arrows.

Figure 2: (a) The illustration of FEXI sequence. (b) At equilibrium, MR-visible water molecules (white dots) are located in blood with fast diffusivity and interstitium with show diffusivity. Signal from fast diffusing intravascular water is suppressed after filter applied, indicated by black dots, leading to a reduction in the ADC′. After t_m, water molecular exchange leads to a recovery of MR-visible water molecules in the blood regions and the ADC’ approaches the equilibrium ADC. Relaxation back to the equilibrium ADC can be described by the apparent exchange rate constant, AXR.

Figure 1. Parameter maps of a representative subject obtained for the different combination of gradient orientations. The rows correspond to the results for all acquired gradient directions (reference: 30 dir. x2), 30x1, 15x2 and 15x1 directions. The maps obtained with at least 30 samples resulted were similar. On the other hand, the DTI fits on only 15 directions resulted in overestimation of FA and disrupted HA maps.

Figure 2. DTI parameters projected on the AHA segments and averaged for all subjects. MD and FA are most robust to different gradient direction selection approaches. On the other, FA and HAT are sensitive to a reduction in the number of samples, as can be observed in the differences of the results obtained with 15 Dir. x1.

Figure 1: Effective gradient waveforms g(t), dephasing waveforms q(t), normalized spectral trace s(ω) and diagonal components of the dephasing cross power spectral density. Individual gradient components (XYZ shown in red, green, blue) from the motion compensated STE waveforms (M0-, M1-, M2-STE) were used as M0-, M1-, M2-LTE. with all moments nulled up to zeroth, first and second order (M0, M1, M2). Their power spectra is shown below s(ω). Note the shift of spectra from zero frequency to higher frequencies, which alters sensitivity to time-dependent diffusion.

Figure 2: ROI-average mean diffusivity in the left ventricle myocardium in hearts 1 (top) and 2 (bottom) and theoretical prediction. The ROIs are shown the right. Left column: measured MD (markers) and prediction (lines) vs cylinder radius, R. Right column: measured MD vs prediction with error bars corresponding to the distance between measurement and prediction (dotted line for unity relation). The n=1,2,3 labels for Mn-LTE correspond to the X,Y,Z channels of Mn-STE.

Fig. 4.　Results of MR brain study Comparison of DTI images, quantitative maps, and fiber tractographies, for standard (std) and for our water suppression (T2wsup). The T2wsup provided better tissue-specific values for the tissues close to the boundary regions of the ventricle, as the body of fornix (orange arrows), or genu of corpus callosum (yellow arrows). For the tractographies of fornix, the fibers at the central portions of two seed ROIs (yellow arrow) for T2wsup was thicker and better connected than for the standard, and artifactual fibers were drawn for the standard (blue arrow).

Fig. 3. Simulation results with 2-point method. A: MD and FA as a function of V_w (%) for standard SE-DWI (std) and T2wsupSE-DWI (T2wsup) obtained by the two-point method with denoted combinations of b = (b₀, b_n) [s/mm2]. B: SIs as a function of b-value (b) for three DWI methods (a–c) each as a parameter of water volume ratio (V_w), and overlapped version a and b (d). C: Theoretical SNRs of ADC values as a function of the second b-value, b_n of pure tissue (V_w = 0, ADC = 0.8×10⁻³ mm²/s) for three DWI techniques with two-point method when the SNR of SE b0 image (S(b₀)/σ) = 50.

Figure 4- Mean diffusivity (MD), fractional anisotropy (FA), and microscopic anisotropy (µFA) maps for 3 transverse slices.

Figure 3- Reconstructed images of 3 slices. The first column on the left shows images acquired by the Cartesian GRE sequence for comparison. Other columns show reconstructed STE images for 4 b-values with the same diffusion direction. Images of each b-value are presented with a different scale for better visibility. The yellow arrow shows remaining artifacts caused by static field inhomogeneities in the frontal lobe.

Figure 1: Top: In-vivo DWI images and white matter regions of interest (ROI) identified with the Jülich fiber atlas¹⁰, a) optic radiation (or), b) cortico spinal tract (ct), c) superior longitudinal fasciculs (slf) and d) callosum body (cb). In every one of the four ROIs, three voxels were chosen (FA images bottom, for slf only one is shown here) and DKI parameters were estimated with the standard DKI model without RBC. These 12 sets of DKI parameters were then used to generate ground truth signals for the simulation.

Figure 2: Bias and standard deviation for the estimated AxDKI model parameters, averaged over 2500 noise samples per SNR across all ROIs. Note the differently scaled y axes for the diffusion (top) and kurtosis parameters (bottom). The axial-symmetric fit with RBC (magenta line) produces the overall least biased estimators.

Figure 3. Example in vivo mean diffusivity (top) and microscopic fractional anisotropy (bottom) slices estimated using the DTI model (left), DKI model (center-left), shifted DKI model (center-right), and FWE-DTI model (right) with data from STE acquisitions.

Figure 1. Summary of the models used to estimate diffusivity from STE signal data in voxels containing CSF and tissue: (a) DTI, (b) DKI, (c) FWE-DTI (with separate compartments shown by dashed lines), and (d) shifted DKI. The C terms represent constants. In FWE-DTI, signal due to CSF-PVEs is removed by separating the overall signal into diffusion components for free water and brain tissue. In shifted DKI, signal due to CSF contamination is attenuated via the use of higher b-values.

Figure 3. Frequency-dependent stiffness maps of a central slice from both DTI-MRE and conventional MRE.

Figure 4. Mean diffusivity and fractional anisotropy maps of the same central slice from DTI-MRE and conventional DTI excluding fluid-filled ventricles.

Figure 1. Example bone metastasis illustrating B0 inhomogeneity induced distortion on DWI. (A) Pre-DisCo DWI b=2000 s/mm² (B) Lesion annotation (pink) overlaid on post-DisCo DWI b=2000 s/mm² (C) Distortion map; voxel values represent extent of displacement at each voxel. Red and blue values denote displacement in the posterior and anterior direction, respectively. (D) Pre-DisCo DWI b=0 s/mm² (E) Post-DisCo DWI b=0 s/mm². F. T2-weighted image. In each subfigure, bounding box was drawn 10 voxels from the lateral edge of the lesion. Abbreviations: DisCo=distortion correction.

Figure 3. Distribution of RMS distortion and MI between b=0 s/mm2 and T2-weighted images. (A) RMS distortion for all lesions (black) as well as the distribution within specific imaging stations. (B) Change in MI values between the b=0 s/mm² and T2 images after DisCo. A value larger than 0 indicates improved agreement between b=0 and T2 images. (C) RMS distortion for lesions broken down by anatomic group. (D) Change in MI values after DisCo by anatomic group. Abbreviations: RMS=root mean square, MI=mutual information.

Table 1. Baseline characteristics for all the 41 patients. The blood-based biomarkers investigated were the Prostate-Specific Antigen (PSA), Hemoglobin level (HGB), Lactate dehydrogenase level (LDH) and Alkaline phosphatase level (AlkpH).

Figure 1. A) Probability of response derived from the trained logistic regression model which expects as features the relative change of Median gADC and tDV. B) Probability of response and 95% confidence interval (dashed lines) derived from fitting the logistic regression model input only the relative change of Median gADC. Estimates of the probability of response higher than 60% and lower than 40% showed low uncertainty. However, samples with a probability of response from 40 to 60% (and around the decision boundaries as showed in Figure 1A) shown higher uncertainty.

Figure 3. Assess model performance on 3 patients of the test data. Coronal and sagittal Maximum-Intensity-Projection (MIP) of SNRmap. The manual delineation of bone lesions performed by an experienced radiologist was overlayed on the MIP of SNRmap (in red). The first colormap shows the ground truth, annotated axial images with (in red) and without (in blue) metastatic bone disease. The second colormap shows the prediction performed by the trained transfer learning model.

Figure 5. Axial Fat fraction, b900, RGB image and CAM from test patient 2 in Figure 3. Fat fraction combined with b900 images are typically used for manual detection of bone disease. However, using the DWI data alone, our model was able to predict presence of disease in one slice (top-row, red outline), and absence of disease in two different slice locations. CAM emphasizes where the model is ‘looking’ to derive the relevant regions within the images that lead to a particular decision. From the heatmap, red = important and blue = not important.

Figure 3: Coronal whole-body images of a patient with metastatic bone lesions in the pelvis and femur (red arrows). Conventional MR images are shown in the top row. The bottom row shows the signal-contribution (C_i) maps for the optimized 4-compartment RSI model. The corresponding D_i of each model compartment is listed in parentheses next to the compartment label.

Figure 2: Model-fitting residual at the voxel- and ROI-level. The top row shows voxel-wise maps of fitting residual in a coronal plane of the same patient using different models. A T2-weighted image of the same plane is included for reference. The bottom figure graphs the fitting residual within all lesion and tissue-specific ROIs. A better fit to the data was observed with the 4-compartment RSI model than with the conventional monoexponential model or the lower-order RSI models.

Figure 1 A 69-year-old male with PCa.ROI was manually placed on the local lesion of MK map (A). The Ka, Kr, FAk, MD, Da, Dr and FA maps of DKI were shown(B-H).

Table 2 Results of correlation analysis of DKI parameters and PSA

* P value is statistically significant.

Figure 1: Segmentation methods as used in 52 recent papers on sarcoma shows a large variety in important aspects of the segmentation strategy.

Choice of the segmentation strategy can have large impact on mean and minimum ADC (apparent diffusion coefficient) measured over the segmented area. Segmentation performed by two experienced radiologists (rater 1 and 2, outline) and a trained researcher (circular and 3D) on an ADC map of pediatric rhabdomyosarcoma. Segmented areas are shown in figure 3.

Figure 4. Cumulative incidence analysis for pretreatment apparent diffusion coefficient, ADC (mm²/s), true diffusion coefficient, D (mm²/s), perfusion fraction, f, volume transfer constant, K^trans (min^-1), the volume fraction of extravascular extracellular, v_e, and mean lifetime of intracellular water protons, τ_i (sec). Gray’s test revealed a significant difference for ADC, D, and f (P <0.05), and a borderline significance for τi (sec) (P=0.098).

Figure 3. Left: Representative histogram plot of true diffusion coefficient, D, and kurtosis coefficient, K), in patients with and without LRF of nasopharyngeal cancer (Figure 2). D values spread out towards higher in a patient without LRF than compared to with LRF. In contrast, K values spread to a higher value in a patient with LRF than without LRF. Right: K^trans values spread out towards higher in a patient without LRF as compared to with LRF. In contrast, τ_i values spread to higher in a patient with LRF as compared without LRF.

Figure 2. Colormaps of cell radius index (R), intracellular volume fraction (f_IC), vascular volume fraction (f_VASC), and extracellular extravascular volume fraction (f_EES) as estimated by VERDICT in the central slice of an example tumour. Colormaps are shown for two repeated measurements (M1 and M2) and the difference between the repeated measurements (Δ)

Figure2 (a) SUV map, (b) ADC map, (c, k) DKI parameter maps, (d, e) diffusion parameters and (g-j) perfusion parameters from IVIM analysis of a 32-year-old male was diagnosed for NHL; the tumour was present in sternum encircled in red.

Figure1 PET images (red) overlapped onto ADC images of a patient with NHL, where the tumour is encircled (green). (a) to (g) represents baseline scans at axial view. (h) Representative 3D sagittal view with lines indicating slices (a-g) selected for visualization.

IVIM DWI parametric maps of ADC, D, D* and f values and T2WI images for SW480 (A-E)and SW480/5-FU(F-J).

The Comparison of The ADC(a), D(b), D*(c) and f(d) values between SW480/5-FU and SW480 xenografts.

Fig1. Exemplary Sagittal IVIM parametric maps of lumbar spine in NDMM group and healthy group.

Fig3. Analysis of ROC curve between ADC, D and D*. The area under the receiver operating characteristic (ROC) curve of ADC, D and D* were 0.935± 0.037, 0.944± 0.034, and 0.935± 0.036, respectively. Among these, the IVIM-dfast showed the highest sensitivity was 90.0%, and its Specificity was 94.7%, with the cutoff value of 141.8×10-3㎜2/s.

Figure 1: Representative pre-SRS and post-SRS (within 72 hours of treatment) MR images from an 80 year-old female with right occipital lobe brain metastasis, which demonstrated progressive disease (PD). Top panel: D(×10^-3 mm²/s), D*(×10^-3 mm²/s), and f maps are zoomed in at the locations of ROIs (pre-SRS b=0 s/mm² image). The lesion volume was unchanged at post-SRS. Bottom panel: K^trans (min^-1), v_e, and v_p maps are zoomed in at the locations of ROIs (pre-SRS T₁-postcontrast image).

Figure 4: Lesions that progressed had a higher f mean and median than those that did not. Lesions that responded to SRS had a lower pre-SRS K^trans mean, K^trans median, and v_e median than those that did not. ADC mean and median trended higher in lesions post-SRS than pre-SRS. v_p was lower in lesions post-SRS than pre-SRS. ADC, apparent diffusion coefficient. K^trans, index of tumor vascular permeability. v_e, volume fraction of extracellular extravascular space. v_p, vascular volume.

Figure 1: High grade glioma. a) DSC CBF map; b) IVIM fD* map; c) SWI map; d) Correlation between fD* and CBF at the tumor ROI in yellow at the tumor border; e) Correlation between fD* and CBF at the tumor ROI in green at the tumor interior.

Figure 2: Low grade glioma. a) DSC CBF map; b) DSC CBV map; c) IVIM fD* map; d) FLAIR; e) Post contrast T1 image; f) SWI; g) Correlation between fD* and CBF at the tumor ROI at the tumor border; h) Correlation between fD* and CBF at the tumor ROI at the tumor interior.

Figure 1: Gd-enhanced T1w images and the maps of ADC(OGSE-17Hz, OGSE-33Hz, PGSE), cellularity, intracelluar fraction and cell diameter as obtained from the IMPULSED model. The ADC and microstructural maps were overlaid on the b0 image for reprehensive low-grade glioma, high-grade glioma, and medulloblastoma cases.

Figure 4: (A) Correlation between ADC (PGSE) and cellularity. Regression analysis revealed correlation was strongest for low-grade glioma (r = -0.8623, p = 0.0003), followed by high-grade glioma (r = -0.7246, p = 0.0117); whereas not significant correlation were found for medulloblastoma (p=0.4095). (B) Similar correlations patterns were found between ADC (PGSE) and Intracellular fraction.

(A to I) T1w, T2w, FLAIR, DWI, ADC, SWI, post-contrast T1w images, rCBV map, and perfusion curve of a bithalamic H3K27M mutant DMG.

(A to I) T1w, T2w, FLAIR, DWI, ADC, SWI, post-contrast T1w images, rCBV map, and perfusion curve of a left thalamic H3K27M wild-type DMG.

Fig. 1. Bar graphs of AD_maximum, MD_maximum, RD_maximum, QIV_maximum and QIV_range values averaged across grade Ⅱ (n=33), grade Ⅲ (n=11) and grade Ⅳ (n=54) gliomas. All parameters are significant with P＜0.05.

Fig. 2. ROC analysis for AD_maximum, MD_maximum, RD_maximum, QIV_maximum and QIV_range in three comparisons (GradeⅡ VS GradeⅢ, GradeⅡVS Grade Ⅳ and GradeⅢ VS Grade Ⅳ)

The distribution of mean and SD of useful parameters

The list for the parameters of diffusion models and histogram analysis

Three representative subjects demonstrating the CST changes in HGG patients. The subject A was a 39-year-old healthy male. All the CST features have good bilateral symmetry. The patient B was a 72-year-old male with glioblastoma, and his motor function was normal. The ICVF of the affected CST decreased obviously, but changes of the affected CST were not obvious in the other diffusion parameters. The patient C was a 62-year-old female with glioblastoma, and she had motor weakness (grade 2). Changes of the affected CST were obvious in FA, MD and ICVF, but not in ISOVF and ODI.

Changes of relative CST features in HGG patients. Box and whisker plot (A-C) showed that compared with the relative CST features in HGG_NMF, the relative MD was significantly higher, where the relative FA and ICVF were significantly lower (*) in HGG_MW; compared with the relative CST features in HC, the relative ICVF was significantly lower (*) in HGG_NMF; in contrast, no significant changes were found in the relative CST morphological characteristics, ISOVF or ODI.

HGG_MW = HGG patients with motor weakness, HGG_NMF = HGG patients with normal motor function, HC = healthy subject.

Figure 1. Analysis of microcystic/angiomatous meningioma. The upper row shows measured (A) and predicted mean diffusivity based on cellularity (B): a second-degree polynomial yielded R² of 0.19 (C). The error map (D) shows large negative (green area) and positive (red) errors where the predicted diffusivity is over/under-estimated. Overestimation was related to high prevalence of microcysts (a; bottom row) because these provide restrictions. Underestimation was found in regions that contained macrocysts and vessels (b and c) because these provide free water compartments.

Figure 3. Quantitative analysis. Part A shows predicted diffusivity omitting and part B including microcysts. The error maps reveal large systematic error areas (yellow arrows) that disappears when cysts are included in the fit already as an additional linear term. In addition, R² increases from 0.19 to 0.41. The blue part marks newly emerged error area that was not previously there. Part C highlights the changes in the error maps. The changes are strongest in the middle part where the cyst density is also the highest. The changes overall strongly correlate with cyst density (R² = 0.81).

Figure 2. A 62-year-old-male with a PCNSL. A: The post-contrast T1-weighted image shows a ring-like enhancing mass lesion in the right frontal lobe (arrow). The enhancing lesion shows high signal intensity on the DWI (B) and a low ADC (0.70×10⁻³ mm²/sec, C). This lesion shows a small κ (1.76, D), a large θ (4.85×10⁻⁶ mm²/sec, E), a large f1 (0.626, F), a small f2 (0.270, G), and a small f3 (0.104, H).

Figure 1. Comparisons of the GD model-derived parameters between the PCNSLs and GBs in the gadolinium-enhancing lesion. A: The κ was significantly smaller in the PCNSL group than in the GB group. B: The θ was not significantly different between the groups. C–E: The f1 was significantly larger and the f2 and f3 were significantly smaller in the PCNSL group than in the GB group.