WO2018210233A1

WO2018210233A1 - Intravoxel incoherent motion mri 3-dimensional quantitative detection of tissue abnormality with improved data processing

Info

Publication number: WO2018210233A1
Application number: PCT/CN2018/086871
Authority: WO
Inventors: Yixiang Wang; Weitian Chen; Yao Li; Min Deng; Chi Shun LEUNG
Original assignee: The Chinese University Of Hong Kong
Priority date: 2017-05-15
Filing date: 2018-05-15
Publication date: 2018-11-22
Also published as: CN110785123A; CN110785123B

Abstract

Liver fibrosis can be detected using intravoxel incoherent motion (IVIM) MRI techniques. For example, using a diffusion weighted MRI imaging sequence, scans of a patient's liver can be made. Signal intensity data acquired in MRI scans can be fitted to a bi-exponential model of signal attenuation representing a combination of a "fast" component associated with perfusion and a "slow" component associated with diffusion in the tissue. This allows the extraction of parameters representing the slow and fast diffusion rates, as well as the fractional contributions of the fast and slow components. Analysis of a combination of these parameters in a multi-dimensional space (e.g., in a three-dimensional space) yields a metric that can distinguish healthy liver from fibrotic liver.

Description

INTRAVOXEL INCOHERENT MOTION MRI 3-DIMENSIONAL QUANTITATIVE DETECTION OF TISSUE ABNORMALITY WITH IMPROVED DATA PROCESSING

BACKGROUND

This disclosure relates generally to detection of liver fibrosis and in particular to detection of tissue abnormality, such as liver fibrosis, using intravoxel incoherent motion (IVIM) magnetic resonance imaging (MRI) with a multidimensional, e.g., three-dimensional, analysis tool.

Chronic liver disease is a major public health problem worldwide. The epidemic trend of chronic liver disease is expected to increase owing to an aging population, the growing epidemic of obesity and non-alcoholic steatohepatitis, and chronic viral hepatitis, which can lead to hepatic fibrosis, cirrhosis and hepatocellular carcinoma.

Liver fibrosis, a common feature of almost all chronic liver diseases, involves the accumulation of collagen, proteoglycans, and other macromolecules in the extracellular matrix. Clinically, liver fibrosis usually has an insidious onset and progresses slowly over decades. Originally considered to be irreversible, hepatic fibrosis is now regarded as a dynamic process with the potential for regression, and a number of promising treatments have been developed to expedite the regression of liver fibrosis and promote liver regeneration. Such therapies are more effective in early stages of liver fibrosis; thus, early detection can be beneficial to patient health.

At present, however, the ability to detect early stage liver fibrosis is limited to biopsy, an invasive procedure that has several contraindications and that may cause various complications such as pain, hemorrhage, bile peritonitis, penetration of abdominal viscera, pneumothorax and even death. Noninvasive procedures are therefore desirable.

Magnetic resonance imaging (MRI) offers a noninvasive procedure for imaging liver tissue, and numerous efforts have been made to use MRI to distinguish between healthy and fibrotic liver tissue. One MRI technique that has been studied is the intravoxel incoherent motion (IVIM) technique. IVIM is a form of diffusion-weighted (DW) MRI, where the intensity of the acquired magnetic resonance signal depends on the self-diffusion of the excited spins, i.e., on the microscopic stochastic Brownian molecular motion. The extent and orientation of such molecular motion is influenced by the microscopic structure and organization of biological tissues. One significant contributor is perfusion (motion of blood through the pseudorandom capillary network) ; another contributor is the “true” diffusion of water within the tissue, which depends on the composition of the tissue. IVIM reflects the random microscopic motion that occurs in voxels (volume elements) of MRI images of intracellular and/or extracellular water molecules and involves quantitatively separating tissue diffusivity and tissue microcapillary perfusion.

It is expected that diffusion of water in fibrotic liver tissue (as compared to healthy liver tissue) would be restricted by the presence of collagen fibers in the distorted lobular structure. It has also been observed that liver fibrosis is associated with reduced liver perfusion. Accordingly, there is interest in using IVIM techniques to study diffused liver diseases such as liver fibrosis. However, to date, attempts to detect liver fibrosis using IVIM techniques have not been successful.

SUMMARY

Certain embodiments of the present invention relate to detection of liver fibrosis using IVIM MRI techniques. In particular, using a diffusion weighted MRI imaging sequence, scans of a patient’s liver can be made. Signal intensity data acquired in MRI scans can be fitted to a bi-exponential model of signal attenuation representing a combination of a “fast” component associated with perfusion and a “slow” component associated with diffusion in the tissue. This allows the extraction of parameters representing the slow and fast diffusion rates, as well as the fractional contributions of the fast and slow components. Analysis of a combination of these parameters in a multi-dimensional space (e.g., in a three-dimensional space) yields a metric that can distinguish healthy liver from fibrotic liver, or significantly fibrotic liver from healthy liver.

The following detailed description, together with the accompanying drawings, will provide a better understanding of the nature and advantages of the claimed invention.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows an MRI system that can be used in connection with practicing some embodiments of the present invention.

FIG. 2 is a flow diagram of a process that can be used to detect liver fibrosis according to an embodiment of the present invention.

FIG. 3 shows an example of a ROI for liver tissue that can be selected according to an embodiment of the present invention.

FIGs. 4A-4C show one-dimensional scatter plots of diffusion parameters for liver tissue determined according to an embodiment of the present invention, for a set of test subjects having known stages of liver fibrosis.

FIGs. 5A-5C show three different perspective views of a 3D space into which the diffusion parameters of FIGs. 4A-4C can be mapped according to an embodiment of the present invention, illustrating distributions of points corresponding to tissue with different stages of liver fibrosis.

FIGs. 6A-6C show three different perspective views of the 3D space of FIGs. 5A-5C, with the points corresponding to mildly fibrotic liver tissue removed.

FIG. 7A shows a two-dimensional plot of diffusion parameters for healthy and significantly fibrotic liver tissue.

FIG. 7B shows a two-dimensional plot of diffusion parameters for healthy, mildly fibrotic, and significantly fibrotic liver tissue.

FIGs. 8A-8C show one-dimensional scatter plots of diffusion parameters for liver tissue determined with various threshold b-values according to an embodiment of the present invention, for a set of test subjects having known stages of liver fibrosis.

FIGs. 9A and 9B are bar graphs showing a separation distance parameter computed for each threshold b-value for the data of FIGs. 8A-8C. FIG. 9A shows the separation distance parameter for distinguishing healthy versus mildly and significantly fibrotic livers, and FIG. 9B shows the separation distance parameter for distinguishing healthy versus significantly fibrotic livers.

FIG. 10 shows a graph of results of a pairwise-comparison analysis for the data of FIGs. 8A-8C.

FIG. 11 is a flow diagram of an image cleaning process according to an embodiment of the present invention.

FIG. 12 shows an example of an image series that may be accepted for analysis based on the process of FIG. 11.

DETAILED DESCRIPTION

Overview of IVIM

Intravoxel incoherent motion (IVIM) is a magnetic resonance imaging (MRI) technique whose basic concepts were first developed by Le Bihan et al. (D. Le Bihan et al., “MR imaging of intravoxel incoherent motions: application to diffusion and perfusion in neurologic disorders, ” Radiology 161: 401-7 (1986) ; D. Le Bihan et all, “Separation of diffusion and perfusion in intravoxel incoherent motion MR imaging, ” Radiology 168: 497-505 (1988) ) . IVIM measures signal attenuation responsive to a pair of gradient pulses applied in opposite directions with a temporal separation between them. To the extent that the targeted nuclei (typically hydrogen) do not move between the gradient pulses, the net effect of the gradient pulses on magnetization is expected to be zero. However, if the nuclei move, e.g., due to diffusion, the gradients have a net effect that attenuates the signals. This signal attenuation can be characterized by:

SI (b) =SI ₀ [ (1-PF) exp (-bD _slow) +PF exp (-bD _fast) ] (1)

where b is the IVIM b-value (a standard parameter in the art that characterizes the strength, duration, and temporal spacing between the gradient pulses, also sometimes referred to as “b-factor” ) , SI (b) is the signal intensity measured with gradient pulses having a specified b-value, SI ₀ is the signal intensity measured with no gradient pulses (b = 0 s/mm ²) , PF is the fraction of diffusion linked to microcirculation (perfusion) , D _slow (also sometimes referred to in the art as simply D) is the “true” diffusion coefficient representing pure molecular diffusion, and D _fast (also sometimes referred to in the art as D*) is the pseudo-diffusion coefficient representing the incoherent microcirculation (perfusion-related diffusion) within the voxel.

Measurements of SI (b) can be made using MRI systems, which can be of generally conventional design. FIG. 1 shows an MRI system 100 that can be used in connection with practicing some embodiments of the present invention. MRI system 100 includes a computer 102 communicably coupled to an MRI apparatus 104.

Computer 102 can be of generally conventional design and can include a user interface 106, a processor 108, a memory 110, a gradient controller 112, an RF controller 114, and an RF receiver 116. User interface 106 can include components that allow a user (e.g., an operator of MRI system 100) to input instructions or data and to view information. For example, user interface 106 can include a keyboard, mouse, joystick, display screen, touch-sensitive display screen, and so on. Processor 108 can include one or more general purpose programmable processors capable of executing program code instructions to perform various operations. Memory 110 can include a combination of volatile and nonvolatile storage elements (e.g., DRAM, SRAM, flash memory, magnetic disk, optical disk, etc. ) . Portions of memory 110 can store program code to be executed by processor 108. Examples of the program code can include a control program 118, which can coordinate operations of MRI apparatus 104 as described below in order to acquire data, and an analysis program 120, which can perform analysis algorithms on data acquired from MRI apparatus 104. Gradient controller 112, RF controller 114, and RF receiver 116 can incorporate standard communication interfaces and protocols to communicate with components of MRI apparatus 104 as described below.

MRI apparatus 104 can be of generally conventional design and can incorporate a magnet 130, one or more gradient coils 132, and RF coils 134, 136. Magnet 130 can be a magnet capable of generating a large constant magnetic field B ₀ (e.g., 1.5 T, 3.0 T, or the like) in a longitudinal direction, in a region where a patient can be placed. Gradient coils 132 can be capable of generating gradients along the direction of the constant magnetic field B ₀; operation of gradient coils 132 can be controlled by computer 102 via gradient controller 112. RF coils 134, 136 can include a transmitter (TX) coil 134 and a receiver (RX) coil 136. In some embodiments, a single coil can serve as both transmitter and receiver. In some embodiments, RF transmitter coil 134 can be placed around the portion of the subject’s body that is to be imaged while RF receiver coil 136 is placed elsewhere within MRI apparatus 104. The preferred placement of RF coils 134, 136 may depend on the specific portion of the body that is to be imaged; those skilled in the art with access to the present disclosure will be able to make appropriate selections.

In operation, computer 100 can drive gradient coils 132 using gradient controller 112 to shape the magnetic field around the region being imaged. Computer 100 can drive RF transmitter coil 134 using RF controller 114 to generate RF pulses at a resonant frequency for an isotope of interest, driving nuclear spins into an excited state. RF receiver coil 136 can detect RF waves (or pulses) generated by the spins relaxing from the excited state when RF pulses are not being generated. RF receiver 116 can include amplifiers, digital-to-analog converters, and other circuitry to generate digital data from the RF waves detected by RF receiver coil 136. RF receiver 116 can provide this data to processor 108 for analysis.

MRI system 100 is illustrative, and many variations and modifications are possible. Those skilled in the art will be familiar with a variety of MRI apparatus and with basic principles of MRI data acquisition, including the use of gradient fields and RF pulses, as well as techniques for detecting signals responsive to RF pulses and processing those signals to generate image data.

Detection of Liver Fibrosis using IVIM

In accordance with some embodiments of the present invention, MRI system 100 or other MRI apparatus can be used to generate a pulse sequence suitable for diffusion-weighted (DW) imaging of a specific organ or tissue within a patient, such as the liver. The acquired data can be analyzed using IVIM-based techniques described below to detect an abnormal condition such as liver fibrosis.

FIG. 2 is a flow diagram of a process 200 that can be used to detect liver fibrosis according to an embodiment of the present invention. Process 200 can be implemented using an MRI system such as system 100 of FIG. 1. At block 202, a subject (in this example, a patient) whose tissue is to be imaged is arranged within an MRI apparatus. This can include having the patient assume a supine (or other appropriate) position and aligning the patient within the MRI apparatus. In some embodiments, this may also include positioning of RF and/or gradient coils; the particular positioning will depend on what tissue is being imaged.

At block 204, one or more preparatory pulse sequences can be applied by operating the MRI apparatus. The preparatory pulse sequence (s) can include, e.g., a magnetization reset sequence, a fat suppression sequence such as Spectral Pre-saturation with Inversion Recovery (SPIR) , and/or other appropriate preparatory pulse sequences.

At block 206, a data acquisition sequence can be applied. The data acquisition sequence can be, e.g., a single-shot diffusion weighted (DW) spin-echo type echo-planar imaging sequence as is known in the art. During this sequence, gradient pulses with a number of different b-factors (or b-values) can be applied. For example, in some embodiments, ten b-factors of 10, 20, 40, 60, 80, 100, 15, 200, 400, and 800 s/mm ² are used. The particular sequence, including the b-factors, may be varied as desired, provided that the data produced can be subject to analysis using techniques described below.

At block 208, data gathered during the data acquisition sequence can be analyzed to determine values for a set of diffusion parameters characterizing diffusion in the imaged tissue. In some embodiments, the set of diffusion parameters includes D _slow (the true diffusion coefficient) , D _fast (the perfusion coefficient) , and PF (the perfusion fraction) as defined above with reference to Equation (1) .

Various analysis techniques may be used to determine the parameter values. For example, at block 210, one or more regions of interest (ROIs) within an image can be selected. In some embodiments, the ROI (s) can be selected to minimize artifacts due to cardiac motion or nearness to other tissues. In the case of liver tissue, the ROI (s) may preferably be selected from the right lobe of the liver, since images of the left lobe are more likely to suffer from artifacts due to cardiac motion and are more susceptible to B ₀ inhomogeneity due to proximity to the stomach (which has air inside) . Further, the ROI (s) can be selected to include the tissue of interest, e.g., liver parenchyma, while avoiding areas of vasculature (which may be more susceptible to artifacts) . FIG. 3 shows an example of a ROI for liver tissue that can be selected according to an embodiment of the present invention. Shown is an image of liver tissue in which vasculature appears dark; the ROI is outlined with a light gray line.

At block 212, a mean signal intensity for the ROI can be determined, e.g., by averaging per-pixel signal intensities within the ROI. If multiple ROIs are selected, each ROI can be analyzed separately, and the results across different ROIs can be combined at a later stage. It should be noted that it is also possible to perform parameter extraction (as described below) on a per-pixel basis; however, in cases where the signal-to-noise ratio of the DW images is low, using the mean signal intensity can provide better estimation of relevant parameters.

At block 214, an estimate of the true diffusion parameter D _slow can be obtained. For example, if IVIM signal attenuation is modeled according to Equation (1) , D _slow can be estimated using a least-squares linear fit of the logarithmized image intensity at different b-values to a linear equation. In some embodiments, a subset of the b-values for which image data was acquired are used for the estimation of D _slow; for example, b-values of 200 s/mm ² or greater may be used, to minimize the contribution of the fast (perfusion) component. In some embodiments, the threshold b-value can be selected with the goal of optimizing separation between healthy and fibrotic livers; an example is described below.

At block 216, the perfusion fraction PF can be estimated. For example, the fitted curve from block 214 can be extrapolated to determine an intercept at b = 0, which reflects the expected signal intensity in the absence of perfusion. The ratio between the extrapolated intercept and the measured signal intensity for b = 0 yields an estimate of PF.

At block 218, the fast diffusion parameter D _fast can be estimated. For example, the estimated D _slow (from block 214) and PF (from block 216) can be substituted into Equation (1) , and a nonlinear least-square fit (e.g., using the Levenberg Marquardt algorithm or variations thereof, such as the Trust Region Algorithm) against all b-values can be used to estimate D _fast.

At block 220, the parameter values determined (or estimated) in the analysis at block 208 can be used to determine the condition (or state) of the tissue, such as whether and/or to what degree liver tissue is fibrotic. For example, estimates of D _slow, PF, and D _fast can be mapped to a point in a three-dimensional (3D) space, and the location of that point can serve as an indicator of the tissue condition. A specific example is described below.

In some embodiments, at block 222, each parameter can be normalized to the range [0, 1] using a linear normalization technique where a set of parameter values {x (i) } can be normalized to:

z (i) = (x (i) -x _min) / (x _max-x _min) (2)

where x _min and x _max are the minimum and maximum values for the parameter x. At block 224, the normalized parameters can be mapped to a point in a 3D space. For instance, each normalized parameter can be mapped directly to one of the three coordinate axes of the 3D space.

At block 226, based on the point in the 3D space, the tissue can be classified to indicate its condition. For example, liver tissue can be classified as healthy (e.g., stage F0 as described below) versus fibrotic (e.g., stages F1-F4 as described below) , or as healthy vs. significantly fibrotic (e.g., stages F2-F4 as described below) . For example, as described below, a plane in the 3D space can be defined such that points corresponding to normal liver tissue are on one side of the plane while points corresponding to fibrotic liver tissue ae on the other side of the plane. The plane of separation can be defined based on data from a pool subjects whose liver condition is known from other sources. Once defined, the plane of separation can be used (e.g., in a clinical setting) to assess a subject whose liver condition is not known.

It will be appreciated that process 200 is illustrative and that variations and modifications are possible. The particular pulse sequences, number of b-values, and selection of specific b-values may be modified. In some embodiments, prior to analysis, an image cleaning process can be used to exclude certain images from the analysis based on evidence of motion artifacts that may compromise data quality; examples of image cleaning processes are described below. Different analysis techniques may be used to determine parameter values characterizing diffusion from the acquired MRI data, and different techniques may be used to map the determined parameter values to a point in a space of appropriate dimensionality. Further, mapping to a point in a space is not required; in some embodiments, a characterizing function can be defined such that the determined parameter values are inputs and the output is a value (or set of values) indicating the likelihood that the tissue has a particular degree of fibrosis.

Example 1

A study to determine the feasibility of using a multi-parameter analysis to distinguish between fibrotic and normal liver tissue has been conducted using MRI data from the Shenzhen 2012/2013 dataset (this dataset is described in P. X. Lu et al., “Decreases in molecular diffusion, perfusion fraction and perfusion-related diffusion in fibrotic livers: a prospective clinical intravoxel incoherent motion MR imaging study, ” PLoS One 9 (12) : e113846 (2014) ) . The subjects included sixteen subjects with normal livers (fibrosis stage F0) and thirty-three subjects with viral hepatitis-B and various degrees of liver fibrosis (fibrosis stage F1-F4) . The fibrosis stage was determined for each subject based on a conventional histology-based diagnosis. The labeling of fibrosis stages in this example follows a conventional scheme in which stage F0 indicates no fibrosis; stage F1 indicates mild fibrosis seen only at the portal area; stage F2 indicates fibrosis extending out from the portal areas but with few bridges between portal areas and without destruction of lobular structure; stage F3 indicates severe fibrosis with significant fibrotic bridging between portal areas and between portal areas and center veins; and stage F4 indicates a final stage of cirrhosis with pseudo lobules being formed. Stages F0 and F1 are commonly considered as showing no significant hepatic fibrosis, while stages F2-F4 are considered significant hepatic fibrosis. Hepatic fibrosis may be considered clinically significant (deserving of medical attention) at stages F2 or greater; accordingly it may be useful to distinguish stage F0 livers from stages F2-F4. Detection of stage F1 fibrosis may also be of interest, e.g., for early therapeutic intervention.

MRI imaging was performed on all subjects using a Philips Achieva 1.5-T scanner (available from Philips Healthcare, Best, the Netherlands) . The IVIM DW imaging sequence was based on a single-shot DW spin-echo type echo-planar imaging sequence, with ten b-values of 10, 20, 40, 60, 80, 100, 150, 200, 400, 800 s/mm ². SPIR technique (spectral pre-saturation with inversion recovery) was used for fat suppression. Parameters for MRI imaging included: average TR = 1500 ms; TE = 63 ms; slice thickness = 7 mm; matrix of 124x97; field of view (FOV) of 375 mm x 302 mm; NEX = 2; number of slices = 6.

Analysis to determine diffusion parameters was performed on the acquired data for each subject in the manner described above with reference to FIG. 2, with the algorithms implemented using MATLAB (available from The MathWorks, Inc., of Natick, MA) . Specifically, for each subject, all six image slices were evaluated; slices with notable motion artifacts and those demonstrating notable outlier behavior in the signal-to-b-value relation were discarded. This left between two and five slices for each subject, with an average of three slices per subject. For each slice, a ROI was defined similarly to that shown in FIG. 3. Slow diffusion parameter D _slow was estimated using b-values greater than 200 s/mm ² and a linear least squares fitting algorithm as described above. Perfusion fraction PF was estimated using extrapolation as described above. Fast diffusion parameter D _fast was estimated using Equation (1) and a nonlinear least squares fitting algorithm as described above. The values of the diffusion parameters determined in this manner were found to be comparable to parameter values measured in other studies.

FIGs. 4A-4C show one-dimensional scatter plots of each of the diffusion parameters for the individual subjects (y-axis) versus fibrosis stage (x-axis) . FIG. 4A shows perfusion fraction PF; FIG. 4B shows D _slow, and FIG. 4C shows D _fast. The p-values indicated were determined using ANOVA and Mann-Whitney U test. As can be seen, liver fibrosis cannot be reliably diagnosed using any of these parameters separately.

The diffusion parameters were mapped to a 3D space with axes corresponding to PF, D _fast, and D _slow. FIGs. 5A-5C show three different perspective views of the 3D space. The black dots correspond to healthy (stage F0) tissue, gray dots to mild fibrosis (stage F1) , and white dots to significant fibrosis (stage F2-F4) . As can be seen from the figures, fibrotic liver can be separated from normal liver based on the 3D analysis. For instance, a clear separation between the F0 cases and the significant fibrosis cases (stages F2-F4) can be seen, as indicated by the dashed lines in FIGs. 5A and 5B and by the plane in FIG. 5C. Mild fibrosis cases (F1) are also separated from F0 (healthy) cases. FIGs. 6A-6C show perspective views of the same 3D space as FIGs. 5A-5C but with the F1 cases removed, which makes the separation between healthy and significantly fibrotic cases more clearly visible. For comparison, FIG. 7A shows a two-dimensional (2D) plot of D _slow vs. PF for F0 (black dots) and F2-F4 (white dots) cases; FIG. 7B shows a similar plot including F1 (gray dots) cases. In this 2D plot, a clear separation cannot be made between mild fibrosis and normal liver tissue. Thus, as can be seen from these figures, the 3D analysis provides improved differentiation as compared to the 2D analysis: fibrotic liver can be separated from normal liver based on the 3D analysis.

The Support Vector Machine (SVM) approach was applied to quantitatively separate the various cases. Specifically, SVM can be used to find a plane (parameterized as Ax + By + Cz + D = 0) that is able to separate the data points into two groups by maximizing the sum d ₁ + d ₂, where for group i (i = 1 or 2) , d _i is the distance of the closest point to the plane. In one analysis, SVM was used to find a plane to separate healthy (F0) cases from fibrotic (F1-F4) cases. In another analysis, SVM was used to find a plane to separate healthy (F0) cases from significantly fibrotic (F2-F4) cases.

For the SVM analysis, the parameters PF, D _slow, and D _fast were normalized as described above. Differentiation of F0 cases from F1-F4 cases (shown in FIGs. 5A-5C) was achieved with a plane defined by:

166.58*PF+8.90*D _slow-0.98*D _fast-19.71=0, (3)

with a minimum normalized distance for F0 group of 0.0021 and for F1-F4 group of 0.0026. Differentiation of F0 cases from F2-F4 cases (shown in FIGs. 6A-6C was achieved with a plane defined by:

29.56*PF+4.33*D _slow-0.12*D _fast-6.67=0. (4)

with a minimum normalized distance for F0 group of 0.0149 and for F2-F4 group of 0.0138.

As this example suggests, analysis based on a combination of PF, D _slow, and D _fast has the potential to provide a noninvasive technique for detecting liver fibrosis. Significant fibrosis (stages F2-F4) can be distinguished from healthy liver tissue (stage F0) , and it may also be possible to distinguish mild fibrosis (stage F1) from healthy liver tissue, which may facilitate early therapeutic intervention. For example, in a clinical application, results of an MRI scan of a patient can be used to determine parameter values (PF, D _slow, D _fast) , which can be used to compute metrics such as the left-hand side of Equations (3) and/or (4) ; comparing these metrics to zero (or otherwise assigning the patient’s parameter values to one side or the other of a plane) can provide an indication of the likelihood that the patient has fibrosis and/or the likely stage of fibrosis (mild or significant) .

In addition, the severity of fibrosis can be assessed based on the distance to the plane. On the side of the plane corresponding to fibrotic tissue, points closer to the plane indicate less severe fibrosis and points farther from the plane indicate more severe fibrosis. In general, the larger the distance from the plane, the more severe the disease state.

It should be understood that the specific coefficients shown in Equations (3) and (4) are not necessarily optimal. They are examples derived from an available set of MRI scans of subjects with known stages of fibrosis, and the results may be affected by the small number of participants and/or by the fact that all of the cases of liver fibrosis were in patients with hepatitis-B. Further, the MRI data acquisition used in this analysis was not optimized for IVIM analysis. For example, the set of b-values did not include b = 0. The data may also be less than optimal in other respects, and this may affect the determined parameter values. Studies using the techniques described herein conducted on a larger pool of subjects and with MRI data acquisition optimized for IVIM analysis may yield different parameters. Those skilled in the art will also appreciate that the histological diagnosis that was used to classify the subject tissues is also not clear-cut; for instance, a high-end F1 liver may be similar to a low-end F2 liver, and such classification ambiguity may also affect optimization of a separation plane.

Optimizing IVIM Threshold b-value

In embodiments described above, determination of D _slow from the MRI data (e.g., at block 214 of process 200 of FIG. 2) is based on fitting image intensity data collected using different b-values to Eq. (1) . As noted, a threshold can be applied to select the b-values to be used for this stage of analysis, in order to reduce the contribution of the perfusion component D _fast. In practice, the fitted value of D _slow depends on the threshold, due to varying contribution of D _fast. In the example above, a threshold of b = 200 s/mm ² was used.

Some embodiments of the present invention may exploit the dependence of fitted D _slow on threshold b-value by selecting a threshold b-value for determining D _slow that is expected to maximize a distance in an analysis space between points corresponding to healthy liver tissue and points corresponding to fibrotic liver tissue. A distance metric can be defined in one dimension (any of the PF, D _slow, D _fast axes) or in the normalized 3-D space described above. An optimal threshold b-value, e.g., one that maximizes the distance metric, can be determined empirically using data collected from scans of subjects with normal and fibrotic livers, e.g., during a training process prior to clinical deployment.

For example, an optimal threshold b-value can be determined using a training process based on data collected from subjects whose stage of liver fibrosis is known. The analysis process of FIG. 2 can be repeated with different selections of threshold b-value; for each repetition, a distance metric can be computed. Various distance metrics can be used; examples are described below. A comparison of the values of the distance metric obtained using different threshold b-values can be used to select an optimal threshold b-value. The selected optimal threshold b-value can be used to generate values of diffusion parameters D _slow, PF, and D _fast. These values in turn can be used to determine a separation plane in a normalized 3D space in the manner described above. In clinical application, the same optimal threshold b-value can be applied when performing diagnostic analysis on a patient whose stage of liver fibrosis is to be determined based on the separation plane.

For purposes of illustration, a specific example of optimizing the threshold b-value will now be described.

Example 2

A study to determine an optimal threshold b-value for distinguishing healthy and fibrotic liver tissue has been conducted using MRI data from the Shenzhen 2012/2013 dataset (described above with reference to Example 1) . The right lobe of liver was selected for analysis, and ROIs were defined similarly to that shown in FIG. 3. Specifically, for each subject, an ROI was manually positioned on a b = 10 s/mm ² image to cover a large portion of right liver parenchyma while avoiding large blood vessels; the same ROI mask was propagated to all b-value images for that subject.

For each subject the analysis process of block 208 of FIG. 2 was applied six times, each time with a different threshold b-value for determining D _slow; threshold b values of 40, 60, 80, 100, 150, and 200 s/mm ² were used. In each analysis, images acquired with b-values greater than or equal to the threshold were used to determine D _slow; regardless of threshold, all b-values were used in subsequent stages of the analysis. Specifically, after determining D _slow for a given b-value, extrapolation of the fitted line to obtain an intercept at b = 10 s/mm ² was used to determine PF (as the ratio between the intercept and measured SI (b=10) . The obtained D _slow and PF were substituted into Eq. (1) and a nonlinear least-squares fitting algorithm (the Trust Region Algorithm, a refinement of the Levenberg-Marquardt algorithm, implemented in MATLAB) was used to obtain D _fast.

FIGs. 8A-8C show one-dimensional scatter plots of each of the diffusion parameters of Eq. (1) for the individual subjects (y axis) versus fibrosis stage (x axis) , for each threshold b-value. FIG. 8A shows perfusion fraction (PF) , FIG. 8B shows D _slow, and FIG. 8C shows D _fast. As the threshold b-value increases, PF increases, while D _slow and D _fast decrease.

The rate of increase or decrease is different between F0 and F2-F4 fibrosis stages. By way of illustration, in FIG. 8A, line S1 represents an estimated fit of the dependence of PF on threshold b-value for F0 livers; line S2 represents an estimated fit of the dependence of PF on threshold b-value for F2-F4 livers. As can be seen, the dependence of PF on threshold b-value is stronger for F0 livers than for F2-F4 livers. Using PF alone, the largest threshold b-value tested (b = 200 s/mm ²) provided the best separation between healthy (F0) and significantly fibrotic (F2-F4) livers.

In FIG. 8B, line S3 represents an estimated fit of the dependence of D _slow on threshold b-value for F0 livers; line S4 represents an estimated fit of the dependence of D _slow on threshold b-value for F2-F4 livers. Similarly to PF, the dependence of D _slow on threshold b-value is stronger for F0 livers than for F2-F4 livers. However, using D _slow alone, the smallest threshold b-value tested (b = 40 s/mm ²) provided the best separation between healthy (F0) and significantly fibrotic (F2-F4) livers.

In FIG. 8C, the dependence of D _fast on threshold b-value is similar across different stages of liver fibrosis The best separation is achieved for threshold b-value of 40 s/mm ²; however, as noted above, D _fast alone is generally a weak discriminator between healthy and fibrotic livers.

For 3-D analysis, the obtained parameters D _slow, PF, and D _fast (all of which may depend on threshold b-value) were normalized in the manner described above. Similar to the analysis described above, an SVM approach was applied to determine a plane of separation between either F0 (healthy) livers and F1-F4 (fibrotic) livers or between F0 livers and F2-F4 (significantly fibrotic) livers. The SVM was applied separately to parameters obtained with each threshold b-value.

Based on the SVM results, a distance of each data point from the plane of separation was computed. A mean distance of the healthy-liver data points from the plane was computed, and a mean distance of the fibrotic-liver data points was computed. The sum of the two mean distances was defined as a “separation distance. ” The separation distance was computed separately for each threshold b-value.

FIGs. 9A and 9B are bar graphs showing separation distance computed for each threshold b-value. In FIG. 9A, the separation is between F0 livers and F1-F4 livers; in FIG. 9A, the separation is between F0 livers and F2-F4 livers. As can be seen, for this study, a threshold b-value of 60 s/mm ² provides the largest separation distance, i.e., the best separation between healthy and fibrotic livers, which suggests that a threshold b-value of 60 s/mm ² may be optimal for detection of liver fibrosis. Other studies suggest that a threshold b-value of 200 s/mm ² may provide the best separation between healthy and fibrotic liver, and it should be understood that the present invention is not limited to any particular threshold b-value.

An optimization approach based on comparing data points for individual subjects was also considered. For instance, a distance between the data point for a healthy-liver subject and the data point for a fibrotic-liver subject was computed using the same metric described above. Subjects were arranged into pairs (one healthy, one fibrotic) and numerically ordered, with pair number 1 being the pair with the smallest distance between their data points, pair number 2 being the pair with the smallest distance after excluding both members of pair number 1, and so on. This approach focuses the optimization on the most difficult cases to distinguish.

FIG. 10 shows a graph of results of a pairwise-comparison analysis. The x axis represents the pair number; the y axis represents the mean distance between data points for all pairs up to the indicated pair number. Different choices of threshold b-value are indicated by different line styles. As can be seen, for all pair numbers, threshold b-value of 60 s/mm ² yields the largest mean distance.

Another optimization approach uses different threshold b-values to compute different parameters. For instance, based on the data of FIGs. 8A-8C, it may appear to be desirable to use a threshold b-value of 200 s/mm ² to compute PF and a threshold b-value of 40 s/mm ² to compute D _slow and D _fast. This approach was tested, and mean separation distances were computed from SVM analysis results in the same manner as in the graphs of FIGs. 9A and 9B. The mean separation distance between F0 and F1-F4 data points was determined to be 0.319 relative units (RU) , and the mean separation distance between F0 and F2-F4 data points was determined to be 0.461 RU. Comparing these numbers to FIGs. 9A and 9B, respectively, indicates that using threshold b-value of 60 s/mm ² for determining all parameters provides better separation, at least within this study.

While this example suggests that a specific threshold b-value of 60 s/mm ² is optimal for detecting liver fibrosis using IVIM, those skilled in the art will appreciate that this may not always be the case. For instance, as noted above, the data set used for this analysis is derived from an available set of MRI scans of subjects with known stages of liver fibrosis, and results may be affected by the limitations of the data set. Thus, the same optimization technique repeated with a different (e.g., larger and/or IVIM-optimized) data set may yield a different optimized threshold b-value. Nevertheless, this example illustrates the general principle that optimizing threshold b-value can improve the discriminating power of the process of FIG. 2.

Image Cleaning Process

For clinical application of the techniques described herein, it is desirable that results be repeatable and reproducible. In the context of MRI, a result is considered “repeatable” to the extent that scans of the same tissue within the same session produce the same result and “reproducible” to the extent that repeating the same scan after a time interval produces the same result (assuming that the condition of the tissue being scanned has not changed during that time interval) .

IVIM MRI for abdominal organs, such as the liver, has been observed to suffer from poor repeatability and reproducibility. A number of factors can contribute to scan-to-scan variations. For instance, IVIM imaging typically involves long data acquisition times with images acquired at a series of b-values. Acquisition is usually performed with respiratory gating, a technique in which respiratory movement of the lower chest or abdomen is monitored and data acquisition is synchronized (prospectively or retrospectively) to the movement. However, even with respiratory gating, inter-b-value motion and intra-b-value motion may affect image data. For instance, inter-b-value motion can cause misalignment of anatomical structures on images acquired with different b-values, and intra-b-value motion can cause visual artifacts within a single image. Other factors contributing to poor repeatability and reproducibility can include imperfections in the magnet or pulse sequences, such as B0 inhomogeneity and eddy currents.

To improve repeatability and reproducibility of IVIM-based results, such as the determination of parameters PF, D _slow, and D _fast described above, it may be desirable to identify and exclude suspect or problematic images prior to parameter determination. Accordingly, some embodiments of the present invention incorporate an image cleaning process that may be performed as part of process 200, e.g., prior to block 208. The image cleaning process can result in excluding some of the images obtained during an MRI scan from the subsequent image analysis. In some embodiments, images are excluded based on heuristic criteria suggesting that they contain artifacts or features that may not be repeatable or reproducible.

FIG. 11 is a flow diagram of an image cleaning process 1100 according to an embodiment of the present invention. Process 1100 can be performed as part of process 200 of FIG. 2, e.g., after acquiring data at block 206 and prior to data analysis at block 208. Image cleaning process 1100 proceeds in stages to exclude images that do not satisfy various criteria for reliability. Images that are not excluded by process 1100 are used for further analysis (e.g., at block 208 of process 200) . Image cleaning process 1100 can be performed manually; however, in some embodiments, some or all stages of image cleaning process 1100 can be performed automatically using machine-learned classifiers or other automated techniques to identify objects in images and perform comparisons across images.

At a first stage of cleaning, at block 1102, image slices can be excluded based on covered anatomical structure. For instance, in the case of liver scans, image slices can be excluded if they: (1) cover only the lowest part of segment V-VI (as defined in the commonly used Couinaud classification) , generally below the gall bladder; (2) cover the hepatic dome near the digestive tract; or (3) cover the diaphragmatic surfaces.

At a second stage of cleaning, at blocks 1104 and 1106, image series can be excluded based on absence of evident motion artifacts in the image series. As used herein, an “image series” refers to a set of corresponding image slices obtained using different b-values. In the absence of motion, it is expected that the image slices in an image series should show the same structures in the same locations, and the second stage of cleaning can be based on whether this expectation is met.

For instance, at block 1104, the quality of each image series can be graded, e.g., by a radiologist as “good quality, ” “fair quality, ” or “insufficient quality. ” Specific grading criteria can be used. In some embodiments, the grading entails visually assessing motion-induced imaging data degradation between consecutive images for the same slice at different b-values ( “inter-b-value motion” ) based on the location of specific anatomical structures and also assessing artifacts within a single image slice that may indicate motion during the scan ( “intra-b-value motion” ) . For example, in the case of liver scans, location of the following anatomical structures can be compared between different (e.g., consecutive) image slices in an image series to detect inter-b-value motion: kidneys, gall bladder, spleen, hepatic edges, main hepatic blood vessels (main portal vein, portal veins up to second order, main hepatic veins) . Intra-b-value motion can be detected by identifying and assessing the severity of apparent artifacts in the hepatic parenchyma within a single image slice. If no motion or artifact is noted, the image series can be graded as “good” quality. Image series with only slight displacement or inconspicuous artifacts can be graded as “fair” quality. Image series with significant motion or artifacts can be graded as “insufficient” quality. At block 1106, image series with insufficient quality are excluded from further consideration.

At a third stage of image cleaning, at

blocks

1108 and 1110, image series that generate a poor IVIM diffusion fitted curve can be excluded. At block 1108, a region of interest for the image series can be defined, and a curve defined by Eq. (1) can be fitted to the ROI-mean intensity values for the images in the image series. The fitting procedure described above or other fitting procedures can be used. At block 1110, image series with poor fit can be excluded from further consideration. “Poor fit” can be defined using statistical criteria. For instance, a coefficient of determination (R ²) value can be computed using conventional statistical methods, and image series for which R ² is below a cutoff value (e.g., 0.95) can be excluded. In addition, a plot of signal intensity vs. b-value for each individual image series can be evaluated. Series that demonstrate multiple outliers from the expected MRI signal versus b-value relation can be excluded. Series that resulted in unreasonably high D _fast values (e.g., approaching the upper boundary of 200x10 ^-3 mm ²/s) can also be excluded.

At a final stage of image cleaning, at block 1112, any image series with fewer than three image slices can be excluded. An image series may have fewer than three image slices, e.g., due to exclusion of image slices at block 1102. This stage can be useful, e.g., for analyses where intensity value is being averaged across slices, as in process 200 described above.

To illustrate the application of process 1100, FIG. 12 shows an example of an image series that may be accepted for analysis based on process 1100. No evident motion or artifacts are seen in more than two images. Also shown in FIG. 12 are the ROI (at 1210) used for determining mean signal intensity and generating signal intensity as a function of b-value and a graph (at 1220) showing the fitted curve of signal intensity versus b-value.

Process 1100 is illustrative, and variations and modifications are possible. The specific criteria for excluding images or image slices can be modified, and various stages may be performed in any order, with images not excluded at one stage being considered at the next stage. As noted above, process 1100, or portions thereof, may be automated. For instance, a machine-learning algorithm that has been trained to identify anatomical structures or location represented in a given image slice can be used to perform the first-stage processing of block 1102. Another machine-learning algorithm can be trained to perform image-to-image registration and identify misregistration; this trained algorithm can be used to detect and quantify inter-b-value motion at block 1104. Similarly, a machine-learning algorithm can be trained to identify motion artifacts within an image; this trained algorithm can be used to detect and quantify intra-b-value motion at block 1104. A quality grade for an image series can be computed from the outputs of the machine-learning algorithms, e.g., based on a weighted combination of outputs or other scoring formula. Curve-fitting of intensity data for an image series at block 1108 can also be automated, and specific criteria can be automatically applied to the fitting results identify image series with poor fit to the curve.

It is to be understood that process 1100 need not alter or destroy any image data; image data can be excluded from a particular analysis while being saved for other uses. Further, while process 1100 is described as being used in connection with a process for detecting liver fibrosis, process 1100 can also be used in connection with other IVIM-based image analysis processes.

A study to assess repeatability and reproducibility of determinations of IVIM parameters PF, D _slow, and D _fast for a given subject (e.g., patient) has been conducted and described in O. Chevallier et al., “Removal of evidential motion-contaminated and poorly fitted image data improves IVIM diffusion MRI parameter scan-rescan reproducibility, ” Acta Radiol., doi: 10.1177/0284185118756949 (2018) . Accordingly, it is believed that scan-rescan reproducibility of IVIM parameters can be sufficient for clinical use.

Additional Embodiments

It is believed that embodiments described herein provide a noninvasive technique that can yield a reliable diagnostic indicator for liver fibrosis, including early stage fibrosis that until now has been detectable only via invasive techniques such as liver biopsy.

While the invention has been described with respect to specific embodiments, one skilled in the art will recognize that numerous modifications are possible. The particular data acquisition sequences and analysis processes may be varied. Studies of larger numbers of subjects with known stages of fibrosis (including healthy livers) may allow an optimized set of diagnostic parameters to be determined. Further, it is expected that, in clinical practice, diagnostic techniques described herein would be combined with other information about the patient’s condition to arrive at a diagnosis.

In addition, while the examples described herein make specific reference to liver tissue and detection of liver fibrosis, the techniques described may also be applied to other types of tissue and may be used to aid in detecting other conditions that may affect the diffusion properties of the tissue, including but not limited to fibrosis.

The analysis techniques described herein can be implemented using computer programs that may be executable on a variety of general-purpose or special-purpose computing devices, and those skilled in the art with access to the present disclosure will be capable of writing appropriate program code. The output of the computer programs may include numerical values (e.g., in list or tabular form) , images (e.g., renderings generated using the image data) , graphical output (e.g., graphs such as any or all of FIGs. 4A-7B) , and may be provided on a display, on a paper printout, in an electronic document that can be transmitted via electronic communication channels (e.g., email, secure FTP server, or the like) , or in any other format that can be perceived and interpreted by a clinician. It should be noted that the data analysis can be but need not be performed by the MRI system used to acquire the MRI data. In some embodiments, the MRI system can be used to collect image data that is transferred to a separate computer system for analysis. Computer programs may be stored in any type of computer-readable storage medium (e.g., optical, magnetic, semiconductor-based or other non-transitory storage media) and may also be distributed using transitory computer-readable media (e.g., Internet download) .

Thus, although the invention has been described with respect to specific embodiments, it will be appreciated that the invention is intended to cover all modifications and equivalents within the scope of the following claims.

Claims

A method comprising:

acquiring magnetic resonance imaging (MRI) data for a tissue of a patient, the MRI data corresponding to a diffusion-weighted (DW) MRI scan of the tissue, the MRI data including signal intensity data corresponding to a plurality of different b-values;

analyzing the MRI data to determine a set of diffusion parameter values, the set of diffusion parameter values including a value of a true diffusion parameter (D _slow) , a perfusion fraction (PF) , and a fast diffusion parameter (D _fast) ;

mapping the set of diffusion parameter values to a point in a three-dimensional space; and

determining a condition of the tissue based on the point in the three-dimensional space.
The method of claim 1 wherein the tissue is liver tissue.
The method of claim 2 wherein determining the condition includes classifying the liver tissue as healthy or fibrotic.
The method of claim 3 wherein, for liver tissue classified as fibrotic, determining the condition further includes assessing severity of fibrosis.
The method of claim 1 wherein analyzing the MRI data includes:

selecting a region of interest;

determining a mean signal intensity for the region of interest; and

using the mean signal intensity to determine the set of diffusion parameter values.
The method of claim 1 wherein analyzing the MRI data includes:

determining a value for the true diffusion parameter D _slow using a fitting procedure applied to a portion of the signal intensity data corresponding to b-values greater than a threshold b-value;

extrapolating an exponential curve defined using the determined value of D _slow to a b-value of zero;

determining a value for the perfusion fraction PF from the extrapolated exponential curve; and

using the determined values of D _slow and PF and a fitting procedure applied to the signal intensity data to determine a value for the fast diffusion parameter D _fast.
The method of claim 6 wherein the threshold b-value is determined empirically based on maximizing a distance metric in the three-dimensional space between a first plurality of sample points known to correspond to healthy livers and a second plurality of sample points known to correspond to fibrotic livers.
The method of claim 6 wherein the threshold b-value is approximately 60 s/mm ².
The method of claim 6 wherein the threshold b-value is approximately 200 s/mm ².
The method of claim 1 wherein determining the condition of the tissue is based on a location of the point relative to a plane defined in the three-dimensional space.
The method of claim 10 wherein the plane is defined such that points corresponding to healthy tissue are located on one side of the plane and points corresponding to fibrotic tissue are located on the other side of the plane.
The method of claim 11 wherein, for points corresponding to fibrotic tissue, determining the condition of the tissue further includes determining a severity of fibrosis based on a distance from the point to the plane.
The method of claim 10 wherein the plane is defined such that points corresponding to significantly fibrotic tissue are located on one side of the plane and points corresponding to not significantly fibrotic tissue are located on the other side of the plane.
The method of claim 1 wherein the image data includes DW image data collected for at least ten different b-values.
The method of claim 1 further comprising, prior to analyzing the MRI data to determine the set of diffusion parameter values, performing an image cleaning operation to determine, based on a set of reliability criteria, whether to exclude specific images from the analysis.
The method of claim 15 wherein performing the image cleaning operation includes:

for each of a plurality of image slices of the MRI data, determining whether to exclude the image slice based on presence or absence of a specific anatomical structure;

for each of a plurality of image series of the MRI data, wherein an image series consists of a set of image slices of the same portion of a subject imaged at different b-values:

grading image quality of each image series based on presence or absence of inter-b-value motion and intra-b-value motion within the image series;

determining whether to exclude each of the image series based on the grading; and

for at least some of the plurality of image series of the MRI data:

determining a diffusion fitted curve for each remaining image series;

assessing a fitting quality of the diffusion fitted curve; and

determining whether to exclude each of the image series based on the fit quality.
A computer-readable storage medium storing program instructions that, when executed by a processor of a computer system, cause the computer system to perform the method of any one of claims 1 to 16.
A computer system comprising:

a memory storing program code; and

a processor coupled to the memory, wherein in response to the program code, the processor performs the method of any one of claims 1 to 16.