Method for constructing a multi-parametric post-processing model for magnetic resonance diffusion weighted imaging

Abstract

The present invention provides a method for constructing a multi-parameter post-processing model for magnetic resonance diffusion weighted imaging, which is aimed at DWI signals with different diffusion sensitivity coefficients. b The changes under the focus on reflecting the characteristics of different aspects of biological tissue structure, by using a unified mathematical description, the current post-processing model only targets a certain b The model is based on the DWI signal of the high value range, which cannot fully reflect the limitations of the structural characteristics of biological tissues. b Under the value of , the diffusion movement of water molecules restricted by cell walls, extracellular matrix, etc. in biological tissues is no longer Gaussian. b Under different values, the microcirculation of blood flow in the capillaries causes the proton phase incoherence within the voxel, which affects the change of DWI signal value. The constructed IVIM-VDC model uniformly describes the DWI signal value under different b The value of each b The signal segmentation of the DWI image with the corresponding value is fitted using the post-processing model IVIM-VDC, and the corresponding parameters of the IVIM-VDC model are output, which can more comprehensively reflect the structural characteristics of biological tissues.

Description

Technical Field

[0001] The present invention belongs to the field of image processing and medical technology, and relates to a method for constructing a multi-parameter post-processing model (IntraVoxel Incoherent Motion and Varying Diffusion Curvature, IVIM_VDC) for magnetic resonance diffusion-weighted imaging (DWI). Background Art

[0002] Magnetic resonance diffusion-weighted imaging (DWI) uses motion-sensitive gradients to detect the diffusion-restricted motion of water molecules, thereby indirectly reflecting changes in tissue microstructure, including changes in cell density, cell size, and tissue composition. It is currently the only imaging technology that can reflect the motion state of water molecules in living tissues. The traditional DWI analysis model assumes that human tissue is a single uniform structure and the microscopic motion of water molecules follows a Gaussian distribution. In this case, if a diffusion-weighted gradient exists, the signal attenuation caused by diffusion motion can be described by a single exponential attenuation model (6). By fitting the signals collected by the DWI method at different diffusion sensitivity coefficients, the apparent diffusion coefficient (ADC) can be obtained through formula (7). The ADC can reflect the macroscopic diffusion motion ability of water molecules and can thus be used as a quantitative indicator to distinguish benign and malignant tumors.

[0003] S b =S0*exp(-b*ADC) (6)

[0004] Among them, S b is the DWI signal value when the diffusion sensitivity coefficient is b, and S0 is the DWI signal value when the diffusion sensitivity coefficient is 0 (no diffusion sensitivity gradient). By acquiring two DWI images corresponding to different diffusion sensitivity coefficient values, the DWI image signal values ​​acquired under two different b values ​​b1 and b2 are calculated using formula (7). and Perform fitting and calculate ADC value.

[0005]

[0006] However, in biological tissues, water molecules interact with surrounding structures (cell walls, extracellular matrix, etc.), which act as barriers, resulting in restricted diffusion movement, and the diffusion movement of water molecules no longer satisfies the Gaussian distribution. At the same time, the microcirculation of blood flow in the capillaries will also cause proton phase incoherence within the voxel. These factors cause the DWI signal to no longer decay in a single exponential manner with the diffusion sensitivity coefficient b. Studies have found that when the diffusion sensitivity coefficient b is <200s / mm2 When b>1000s / mm 2 When the non-Gaussian distribution of water molecule diffusion is more obvious in the signal. The ADC obtained by fitting the single exponential decay method (7) reflects the average diffusion coefficient within the voxel and cannot provide more information reflecting the changes in composition and structure. In addition, the sensitivity of ADC values ​​to b-values ​​limits the consistency of ADC values ​​in clinical applications. Therefore, the classic single exponential DWI analysis model has certain limitations.

[0007] To characterize the diffusion of water molecules within different b-value ranges, studies have proposed models such as the intravoxel incoherent motion (IVIM), the diffusion kurtosis imaging (DKI), the fractional order calculus (FROC), the stretched exponential model (SEM), and the varying diffusion curvature (VDC). These targeted diffusion models are based on physical laws, tissue structure, and biophysical constraints.

[0008] The VDC model is based on the assumption that due to the complexity of tissue structure and composition within the subvoxel, the diffusion weighting coefficient D(b) changes with the diffusion sensitivity coefficient b.

[0009] D(b)=D0*exp(-b*D1) (8)

[0010] The diffusion weighted coefficient D0 in equation (8) corresponds to hindered diffusion, while the diffusion weighted coefficient D1 is more sensitive to restricted diffusion and can reflect the complexity of the tissue.

[0011] DWI signal S b decays with the value of b in the manner given by equation (9)

[0012]

[0013] It can reflect the changes in D(b) caused by the tissue structure and composition within the subvoxel.

[0014] The IVIM model is a bi-exponential model that uses the pseudo-diffusion coefficient D to represent the blood flow effect in capillaries (volume fraction f). * The water diffusion effect is characterized by the diffusion coefficient D, and the DWI signal S b The decay of the model given by equation (10) with the value of b

[0015] S b =S0*(f*exp(-b*D * )+(1-f)*exp(-b*D)) (10)

[0016] The changes in DWI signals under different diffusion sensitivity coefficients (b) will focus on reflecting the characteristics of different aspects of biological tissue structure. However, the current DWI post-processing model can only process DWI signals within a certain b value range and cannot fully reflect the tissue structure characteristics, which has certain limitations. Summary of the Invention

[0017] The present invention aims to provide a method for constructing a multi-parameter post-processing model (IVIM_VDC) for diffusion-weighted magnetic resonance imaging (DWI). This method combines the interaction between water molecules in biological tissues and surrounding structures (cell walls, extracellular matrix, etc.) at high b-values. These structures act as barriers, resulting in restricted diffusion motion. The diffusion motion of water molecules is no longer Gaussian, resulting in the DWI signal no longer decaying with a single exponential. At low b-values, the microcirculation of blood flow in capillaries causes proton phase incoherence within the voxel. The diffusion of water molecules in living tissues and the microcirculatory perfusion in the capillary network all affect the changes in DWI signal values. The changes in DWI signals at different b-values ​​are uniformly described using the IVIM_VDC model. The parameters generated by the model reflect the diversity of biological tissue structures as comprehensively as possible. Therefore, the constructed model has a better goodness of fit for DWI signals and can be used for auxiliary diagnosis of benign and malignant tumors.

[0018] The specific technical solution for achieving the purpose of the present invention is:

[0019] A method for constructing a multi-parameter post-processing model (IVIM_VDC) for diffusion-weighted magnetic resonance imaging (DWI) comprises the following steps:

[0020] Step 1: Acquire a set of DWI data using multiple b-values, where the b-value is the diffusion sensitivity coefficient;

[0021] Step 2: Remove Rician noise from the collected DWI data;

[0022] Step 3: Construct DWI multi-parameter post-processing model IVIM_VDC:

[0023]

[0024] Where b is the diffusion sensitivity coefficient, S b is the DWI signal value when the diffusion sensitivity coefficient is b value, S0 is the DWI signal value when the diffusion sensitivity coefficient b value is 0, f is the volume fraction of the perfusion effect, D *is the pseudo-diffusion weighting coefficient. The diffusion weighting coefficient D0 corresponds to hindered diffusion, and the diffusion weighting coefficient D1 is more sensitive to restricted diffusion, reflecting the complexity of the tissue.

[0025] Step 4: Fit each b value to its corresponding DWI signal using the IVIM-VDC model to calculate the parameters S0, f, and D of the IVIM-VDC model. * , D0 and D1; where:

[0026] The constructed DWI multi-parameter post-processing model IVIM_VDC is a model that uses a mathematical expression to comprehensively describe the changes in DWI signals at different b-values. DWI data should be collected within a range of different b-values. The multiple b-values ​​mentioned in step 1 are at least 4 with a maximum of 200 s / mm 2 b value, with at least 2 minimum 200s / mm 2 Maximum 1000s / mm 2 The b value and there are at least 2 minimum 1000s / mm 2 The b value of .

[0027] In step 2, the DWI data is processed to remove Rician noise. The specific processing process uses the following denoising equation:

[0028]

[0029] Among them S raw is the DWI signal acquired by magnetic resonance scanning, S noise is the background noise value, S corr is the DWI signal after Rician noise removal; a noise region of interest of 20 pixels * 20 pixels is selected from the background area of ​​the DWI image corresponding to each b value, and the average signal S of the noise region of interest of the DWI image corresponding to each b value is calculated. noise , use the above denoising equation to perform pixel-by-pixel denoising calculation on the DWI image corresponding to each b value, and obtain the DWI signal data S after Rician noise removal corr .

[0030] The step 4 specifically includes: in the method of performing nonlinear fitting point by point (each b value and its corresponding DWI signal), in order to reduce the long reconstruction time caused by the multi-parameter IVIM_VDC model and the problem that the parameters may have abnormal values, a segmented fitting method is used for IVIM-VDC fitting. First, the b value with the lowest value of 200s / mm is selected. 2 And the highest is 1000s / mm 2 DWI data, according to the formula:

[0031] S b=S0*(1-f)*exp(-b*D0) (3)

[0032] Use nonlinear least squares method to fit and estimate f, D0, and S0 values;

[0033] Secondly, select the b value with the lowest value of 1000s / mm 2 DWI data, according to the formula:

[0034]

[0035] Substitute f and D0 estimated by equation (3) into equation (4), use S0 estimated by equation (3) as the initial value to fit using the nonlinear least squares method to estimate D1 and S0;

[0036] Finally, keeping the estimated f, D0, and D1 unchanged, and the estimated S0 from equation (4) unchanged, the IVIM-VDC model equation (1) is fitted to estimate S0 and D * value;

[0037] After the three fitting processes, f, D0, D1, S0 and D * value.

[0038] The IVIM_VDC model of the present invention was compared with the single exponential decay DWI model, IVIM model, VDC model, and DKI model in terms of goodness of fit. The Akaike information criterion (AIC) was calculated for each model, and the goodness of fit of the model was ensured by comparing the AIC values. The AIC is expressed as:

[0039] AIC=N*ln(SSE)-N·ln(N)+2k (5)

[0040] Among them, N is the number of fitting points, corresponding to the total number of b values, SSE is the residual sum of squares of model fitting, and k is the number of parameters to be fitted in the model. By comparing the size of the AIC value of the single exponential DWI model, IVIM, VDC, DKI and IVIM-VDC model, the optimal model is the model with the smallest AIC value.

[0041] The proposed method for constructing a multi-parameter post-processing model for diffusion-weighted magnetic resonance imaging (DWI) combines the advantages of the IVIM and VDC models through a unified mathematical description. This method not only better fits DWI signals but also simultaneously obtains information on tissue blood perfusion, as well as subvoxel tissue structure and composition. This information can be used to distinguish different tissue types and assess the impact of disease on tissue microstructure. Furthermore, this method reduces the cost of using different models, significantly improving the efficiency of clinical auxiliary diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of the present invention;

[0043] Figure 2 The figure shows the fitting results of the IVIM_VDC model, VDC model, IVIM model, DKI model and single exponential DWI model of the present invention for a breast tumor DWI signal;

[0044] Figure 3 for Figure 2 At low b values ​​(0-400s / mm 2 ) The fitting results of the DWI signal in the range are shown in FIG;

[0045] Figure 4 for Figure 2 At high b value (800~3000s / mm 2 )The fitting results of the DWI signal in the range are shown in the figure. DETAILED DESCRIPTION

[0046] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0047] See Figure 1 , the present invention comprises the following specific steps:

[0048] (1) Diffusion-weighted imaging (DWI) data were acquired using multiple gradient directions and multiple b-values;

[0049] (2) Remove Rician noise from DWI image data;

[0050]

[0051] Among them S raw is the original DWI signal, S noise is the background noise intensity, S corr is the DWI signal after Rician noise removal;

[0052] (3) Constructing the DWI multi-parameter post-processing model IVIM_VDC:

[0053]

[0054] Where b is the diffusion sensitivity coefficient, S b is the DWI signal value when the diffusion sensitivity coefficient is b value, S0 is the DWI signal value when the diffusion sensitivity coefficient b value is 0, f is the volume fraction of the perfusion effect, D * The pseudo-diffusion weighting coefficient D0 corresponds to hindered diffusion, similar to the ADC of the single exponential decay model, while the diffusion weighting coefficient D1 is more sensitive to restricted diffusion and reflects the complexity of the tissue. For the IVIM-VDC fitting process, a piecewise fitting method is used.

[0055] First, select a b value greater than 200s / mm 2 And less than 1000s / mm 2 DWI image, according to the formula:

[0056] S b =S0*(1-f)*exp(-b*D0) (3)

[0057] Use nonlinear least squares method to fit and estimate f, D0, and S0 values;

[0058] Secondly, select greater than 1000s / mm 2 DWI image, according to the formula:

[0059]

[0060] Substitute f and D0 estimated by equation (3) into equation (4), use S0 estimated by equation (3) as the initial value to fit using the nonlinear least squares method to estimate D1 and S0.

[0061] Finally, keeping the estimated f, D0 and D1 unchanged, and the estimated S0 of equation (4) unchanged, the IVIM-VDC model equation (2) is fitted to estimate S0 and D * After the above three fitting processes, we can obtain f, D0, D1, S0 and D * value.

[0062] (4) Evaluate the goodness of fit of the single exponential decay model, DKI model, IVIM model, VDC model, and IVIM-VDC model, calculate the AIC value for each model, and by comparison, the model with the lowest AIC value is the best fitting model.

[0063] AIC=N*ln(SSE)-N·ln(N)+2k (5)

[0064] Among them, N is the number of fitting points, corresponding to the total number of b values, SSE is the residual sum of squares of model fitting, and k is the number of fitting parameters in the model. By comparing the size of the AIC value of the single exponential model, DKI, IVIM, VDC and IVIM-VDC models, the optimal model is the model with the smallest AIC value.

[0065] Example

[0066] 1.1 Multi-b-value diffusion data acquisition

[0067] In this example, the data used were derived from 64 breast DWI data acquired using a 3T MRI system (Siemens 3T MAGNETOM Prisma). The acquisition parameters were as follows: 3 orthogonal diffusion gradients, and DWI images acquired at 16 b-values: b = 0, 10, 20, 30, 50, 70, 100, 150, 200, 400, 800, 1200, 1500, 2000, 2500, and 3000 s / mm. 2 , Repetition time (TR) / Echo time (TE) = 4000ms / 67ms, FOV = 190*350mm 2 , acquisition matrix size: 104×192, slice thickness: 4 mm (the parameters can be adjusted according to the specific scanning area).

[0068] 1.2 Data Denoising

[0069] Since noise in MRI can cause deviations in the acquired DWI signals, we first calibrate the signal intensity in the diffusion data to eliminate noise interference. The specific process is as follows:

[0070] A noise region of interest (ROI) of 20*20 pixels is selected from the background area of ​​the DWI image corresponding to each b value, and the average signal S of the noise region of interest of the DWI image corresponding to each b value is calculated. noise , use the following formula to get the corrected signal strength S corr ;

[0071]

[0072] The DWI image corresponding to each b value is denoised pixel by pixel to obtain the DWI signal data S after Rician noise removal. corr .

[0073] 1.3 IVIM-VDC model analysis

[0074] The DWI signal data after removal S corr Use the multi-parameter post-processing model IVIM_VDC:

[0075]

[0076] Where b is the diffusion sensitivity coefficient, S b is the signal value of the DWI image corrected by formula (2) when the diffusion sensitivity coefficient is b value, S0 is the signal value of the DWI image corrected by formula (2) when the diffusion sensitivity coefficient b value is 0, f is the volume fraction of the perfusion effect, D * is the pseudo-diffusion weighting coefficient, D0 corresponds to the diffusion weighting coefficient of hindered diffusion, and the diffusion weighting coefficient D1 is more sensitive to restricted diffusion and reflects the complexity of the tissue. For the IVIM-VDC fitting, a piecewise fitting method is used.

[0077] First, select the b value 200s / mm 2 ,400s / mm 2 , 800s / mm 2 DWI image, according to the formula:

[0078] S b =S0*(1-f)*exp(-b*D0) (3)

[0079] Use nonlinear least squares method to fit and estimate f, D0, and S0 values;

[0080] Then, select 1200s / mm 2 The above DWI image, according to the formula:

[0081]

[0082] Substitute D0 into the equation, use the nonlinear least squares method to fit the estimated S0 and f from equation (3) as the initial values, and estimate D1, S0, and f. Finally, keep the estimated D0 and D1 unchanged, fit the IVIM-VDC model equation (1), and estimate S0 and D * , and f value. After the above three fitting processes, f, D0, D1, S0 and D * value.

[0083] 1.4 Model testing and verification

[0084] To evaluate the fit between the IVIM-VDC analysis model and the DWI data, the IVIM-VDC model was compared with the IVIM, VDC, DKI models, and the single exponential model, as shown below. The AIC value was calculated for each model, and the model with the lowest AIC value was identified as the best fitting model.

[0085] AIC=N*ln(SSE)-N·ln(N)+2k (5)

[0086] Where N is the number of fitting points, corresponding to the total number of b values, SSE is the residual sum of squares of the model fitting, and k is the number of fitting parameters in the model.

[0087] In this embodiment, the AIC values ​​of IVIM_VDC, VDC, IVIM, DKI and single-exponential DWI models are shown in Table 1. The AIC value of IVIM_VDC is the lowest and is significantly different from the AIC values ​​of other models (P<0.01). For this embodiment, the IVIM_VDC analysis model is the best multi-parameter analysis model among the five diffusion-weighted imaging post-processing models. Figure 2 The figure shows the fitting results of the IVIM_VDC model of the present invention and the VDC model, IVIM model, DKI model and single exponential DWI model for a breast tumor DWI signal. It can be seen from the figure that the signal value fitted by the IVIM_VDC model has the highest consistency with the original DWI value. The fitting signal values ​​of the other models deviate from the DWI signal in different b-value ranges. In order to make a clearer comparison, Figure 3 、 Figure 4 From low b value (0~400s / mm 2 ) and high b value (800~3000s / mm 2 ) The signal fitting diagrams of IVIM_VDC, VDC, IVIM, DKI and single exponential DWI models are compared. Figure 3 for Figure 2 At low b values ​​(0-400s / mm 2 ) range DWI signal fitting results are displayed in the figure; the figure shows that at low b-values, the signal fitting effects of IVIM_VDC and IVIM models are better than those of other models, among which the signal fitting effect of the single exponential DWI model is the worst. The results shown in the figure are consistent with the phenomenon that blood perfusion will affect DWI values ​​at low b-values. Figure 4 for Figure 2 At high b value (800~3000s / mm 2 ) range DWI signal fitting results display diagram; the figure shows that when the b value is high, the difference between the IVIM_VDC model fitting signal and the DWI value is the smallest, and the fitting effect is better than other models. 2 The difference between the DKI and VDC model fitting signals and DWI values ​​decreased, which is consistent with the fact that these two models can explain the non-Gaussian phenomenon of diffusion motion at high b values. Therefore, the IVIM_VDC model, which comprehensively considers blood perfusion and non-Gaussian diffusion motion, has the highest degree of consistency with DWI values ​​and can best reflect changes in DWI values.

[0088] The above example is only one application of the invention and does not limit its scope of application. Analysis of the examples demonstrates that, compared with traditional single-exponential decay DWI, VDC, and IVIM, the DKI and IVIM_VDC analysis models provide a better fit to DWI signals and are more conducive to distinguishing benign from malignant tumors. The present invention is not limited to a specific body part and can be widely applied to tumor imaging in other locations.

[0089] Table 1

[0090] AIC P value IVIM_VDC -130.82±13.63 VDC -111.56±9.45 <0.0001 DKI -127.37±12.93 0.0012 IVIM -117.54±9.37 <0.0001 DWI -80.10±8.47 <0.0001

Claims

1. A method for constructing a multi-parameter post-processing model for magnetic resonance diffusion-weighted imaging, characterized in that: The method specifically comprises the following steps: Step 1: Acquire a set of DWI data using multiple b-values, where the b-value is the diffusion sensitivity coefficient; Step 2: Remove Rician noise from the collected DWI data; Step 3: Construct DWI multi-parameter post-processing model IVIM_VDC: Where b is the diffusion sensitivity coefficient, S b is the DWI signal value when the diffusion sensitivity coefficient is b value, S0 is the DWI signal value when the diffusion sensitivity coefficient b value is 0, f is the volume fraction of the perfusion effect, D * is the pseudo-diffusion weighting coefficient. The diffusion weighting coefficient D0 corresponds to hindered diffusion, and the diffusion weighting coefficient D1 is more sensitive to restricted diffusion, reflecting the complexity of the tissue. Step 4: Fit each b value to its corresponding DWI signal using the IVIM-VDC model to calculate the parameters S0, f, and D of the IVIM-VDC model. * , D0 and D1; where: The constructed DWI multi-parameter post-processing model IVIM_VDC is a model that uses a mathematical expression to comprehensively describe the changes in DWI signals at different b-values. DWI data should be collected within a range of different b-values. The multiple b-values ​​mentioned in step 1 are at least 4 with a maximum of 200 s / mm 2 b value, with at least 2 minimum 200s / mm 2 Maximum 1000s / mm 2 The b value and there are at least 2 minimum 1000s / mm 2 The b value of .

2. The construction method according to claim 1, characterized in that In step 2, the DWI data is processed to remove Rician noise. The specific processing process uses the following denoising equation: Among them S raw is the DWI signal acquired by magnetic resonance scanning, S noise is the background noise value, S corr is the DWI signal after Rician noise removal; a noise region of interest of 20 pixels * 20 pixels is selected from the background area of ​​the DWI image corresponding to each b value, and the average signal S of the noise region of interest of the DWI image corresponding to each b value is calculated. noise , use the above denoising equation to perform pixel-by-pixel denoising calculation on the DWI image corresponding to each b value, and obtain the DWI signal data S after Rician noise removal corr .

3. The construction method according to claim 1, characterized in that The step 4 specifically includes: Use the segmented fitting method. First, select the b value with the lowest value of 200s / mm 2 And the highest is 1000s / mm 2 DWI data, according to the formula: S b =S0*(1-f)*exp(-b*D0) (3) Use nonlinear least squares method to fit and estimate f, D0, and S0 values; Secondly, select the b value with the lowest value of 1000s / mm 2 DWI data, according to the formula: Substitute f and D0 estimated by equation (3) into equation (4), use S0 estimated by equation (3) as the initial value to fit using the nonlinear least squares method to estimate D1 and S0; Finally, keeping the estimated f, D0, and D1 unchanged, and the estimated S0 from equation (4) unchanged, the IVIM-VDC model equation (1) is fitted to estimate S0 and D * value; After the three fitting processes, f, D0, D1, S0 and D * value.

Images