Fast high snr cest quantification parameter map calculation method based on k-means clustering

Abstract

The application discloses a kind of fast high signal-to-noise ratio CEST quantification parameter map calculation methods based on K-means clustering, method includes: according to the pixel number and standard deviation of CEST-MRI image Setting clustering termination condition;K-means clustering is carried out to the target to be clustered, and the average Z spectrum of each class is fitted respectively;For the class that does not meet the condition, the class is taken as the target to be clustered, and the fitting parameter of the average Z spectrum of the class is taken as the initial value of fitting after next clustering, turn to clustering step;For the class that meets, the fitting value of the class is taken as the quantification signal output of all pixels in the class, and when all classes meet the clustering termination condition, the fitting value of all classes is output based on class fitting Quantification map according to the fitting value of all classes.The method can improve the signal-to-noise ratio and fitting accuracy of CEST-MRI quantification map and significantly reduce the calculation time, which is beneficial to the promotion and application in preclinical research and clinical practice.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of CEST MRI, and particularly relates to a fast high signal-to-noise ratio CEST quantification parameter map calculation method based on K-means clustering. BACKGROUND

[0002] Chemical Exchange Saturation Transfer (CEST) magnetic resonance imaging (MRI) is a new emerging molecular imaging technology, which is usually used for metabolic detection of diseases such as cancer, tumor, and PH detection of diseases such as stroke.

[0003] Because the concentrations of chemical substances such as glucose, protein, creatine, and salicylic acid analogues in the human body are low, and the frequencies of protons in these chemical substances are different from the frequency of hydrogen protons in water, a certain saturation radio frequency pulse (RF) can be applied to these specific chemical substances to make the protons in the chemical substances reach a saturated state, thereby affecting the signal strength of water through chemical exchange, and the exchange rate and concentration information of macromolecular substances in the body can be indirectly reflected by detecting the change of water signal strength. Therefore, CEST technology can be used as a signal amplifier for low-concentration biomolecules. Generally, exchange information at a specific frequency offset can be studied by plotting a Z-spectrum (water signal as a function of saturation frequency offset), and the required CEST signal can be extracted by observing the signal peak at a specific frequency offset of the plotted Z-spectrum.

[0004] However, the Z-spectrum not only contains the required CEST effect, but also contains other effects such as direct water saturation effect (DS), magnetization transfer effect (MT), etc. Due to the existence of these interference effects, it will have a certain impact on the analysis and extraction of the required CEST effect. Therefore, in order to quantify the CEST signal in the Z-spectrum, it is necessary to remove the influence of the interference effect on the Z-spectrum as much as possible. How to extract the CEST effect and weaken the influence of the interference effect as much as possible to realize the accurate quantification of specific molecules and further more accurately diagnose diseases has always been a hot spot in the research of CEST MRI.

[0005] To this end, a Lorentzian Difference (LD) analysis method is proposed in the related art, and a contrast map is generated using a voxel fitting method, but the method has high calculation cost, and the initial value of Lorentz fitting is difficult to determine. A method based on image downsampling and accelerated least squares is proposed in the related art to improve the Lorentz fitting of Z spectrum, wherein the fitting initial value of the image with larger resolution is given by the fitting result of the down-sampled image with smaller resolution, but the method is irrelevant to the content of the image, ignores the consistency between tissues, and thus has low quantification accuracy. SUMMARY

[0006] The present application aims to at least partially solve one of the technical problems in the related art. To this end, one object of the present application is to provide a fast high signal-to-noise ratio CEST quantification parameter map calculation method based on K-means clustering.

[0007] In a first aspect, the present application provides a fast high signal-to-noise ratio CEST quantification parameter map calculation method based on K-means clustering, the method comprising: obtaining a CEST-MRI image, and setting a clustering termination condition according to the number of pixels and the standard deviation of the CEST-MRI image, and determining a target K value; performing K-means clustering on a target to be clustered based on the target K value, and performing parameter fitting based on a preset function on the average Z spectrum of each class, wherein the initial target to be clustered is the CEST-MRI image; respectively judging whether each class meets the clustering termination condition; for the class that does not meet the clustering termination condition, taking the class as the target to be clustered, and taking the fitting parameter of the average Z spectrum of the class as the initial value of the fitting after the next clustering, and turning to the step of performing K-means clustering on the target to be clustered; for the class that meets the clustering termination condition, taking the fitting value of the class as the quantification signal output of all pixels in the class, and when all classes meet the clustering termination condition, outputting a class-based quantification map according to the fitting values of all classes.

[0008] According to one embodiment of the present application, before the clustering termination condition is set according to the number of pixels and the standard deviation of the CEST-MRI image and the CEST-MRI image is clustered, the method further comprises: pre-processing the CEST-MRI image.

[0009] According to one embodiment of the present application, the pre-processing of the CEST-MRI image comprises: standardizing the CEST-MRI image, and correcting the non-uniformity of the B0 field of the standardized CEST-MRI image using a WASSR method.

[0010] According to one of the embodiments of the present application, the K-means clustering of the to-be-clustered targets comprises: performing K-means clustering based on the Z-spectrum signal values of each pixel point in the to-be-clustered targets.

[0011] According to one of the embodiments of the present application, the clustering termination condition is that the number of pixels in a class is less than a pixel number threshold, or the standard deviation of pixels in a class is less than a standard deviation threshold; wherein the pixel number threshold is the pixel number of the CEST-MRI image divided by the square of a first preset parameter, and the standard deviation threshold is the standard deviation of pixels in the CEST-MRI image divided by a second preset parameter.

[0012] According to one of the embodiments of the present application, for a class satisfying the clustering termination condition, the fitting value of the class is taken as the quantized signal of all pixels in the class, and when all classes satisfy the clustering termination condition, a class-fitting-based quantization map is output according to the fitting values of all classes.

[0013] According to one of the embodiments of the present application, the obtaining of the quantized signal from the class comprises: performing pixel-by-pixel parameter fitting based on the preset function with the quantized values in each class as the initial values of fitting, to obtain a fine quantization map combining class fitting and pixel-by-pixel fitting.

[0014] According to one of the embodiments of the present application, the preset function comprises: single-pool Lorentz function, multi-pool Lorentz function, super-Lorentz function, Gaussian function, linear combination of at least two of the above functions, and Z-spectrum function calculated from Bloch equation containing exchange term; wherein when the multi-pool Lorentz function is used for Lorentz fitting, the fitting boundary is limited in the clustering transmission process, and the limiting method is that the fitting peak value is limited to the fitting initial value plus or minus a first preset multiple of the maximum standard deviation in the class, the fitting peak width is limited to the fitting initial value plus or minus a second preset multiple of the maximum standard deviation in the class, and the fitting frequency offset is limited to the fitting initial value plus or minus a third preset multiple of the maximum standard deviation in the class, the first preset multiple and the second preset multiple are both greater than 1, and the first preset multiple is less than the second preset multiple, and the third preset multiple is greater than 0 and less than 1.

[0015] According to one of the embodiments of the present application, the determination of the target K value comprises: performing a pre-experiment on a preset image of the same type as the CEST-MRI image to obtain a first relationship between the K value and the final class number, and a second relationship between the final class number and the quantization effect; determining a K value meeting the quantization effect as a candidate K value according to the first relationship and the second relationship, and taking the smallest candidate K value as the target K value.

[0016] According to one embodiment of the present application, the quantification effect includes one or more of signal intensity of amide protons, amine protons, guanidinium protons, hydroxyl protons, and nuclear Overhauser effect, and goodness of fit R 2 , contrast noise ratio and coefficient of variation of the quantification map.

[0017] The K-means clustering-based fast high signal-to-noise ratio CEST quantification parameter map calculation method of the embodiment of the present application adopts hierarchical clustering to provide a classified update of the fitting parameters, specifically, the voxels in the CEST MRI image are clustered into multiple classes according to their Z-spectrum characteristics, and then the loosely set initial values are used to fit the average spectrum in the class to obtain a class fitting spectrum and a set of fitting parameters; each class is further clustered into subclasses until the number of pixels or the intensity error meets the standard, and the fitting parameters of the fine clustering are used to add voxel fitting to further improve the results, thereby improving the signal-to-noise ratio and fitting accuracy of the CEST-MRI quantification map and significantly reducing the calculation time, which is conducive to the promotion and application in preclinical research and clinical practice. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the K-means clustering-based fast high signal-to-noise ratio CEST quantification parameter map calculation method of one embodiment of the present application;

[0019] Figure 2 is a flowchart of the K-means clustering-based fast high signal-to-noise ratio CEST quantification parameter map calculation method of one specific embodiment of the present application;

[0020] Figure 3 is a schematic diagram of cluster parameter optimization of one example of the present application;

[0021] Figure 4 is a schematic diagram of robustness evaluation of noise of one example of the present application;

[0022] Figure 5 is a schematic diagram of analysis of the Lorenz difference of ischemic stroke rats of one example of the present application;

[0023] Figure 6 is a schematic diagram of the relationship between the number of classes and the quantification effect of ischemic stroke rats of one example of the present application;

[0024] Figure 7 is a schematic diagram of the data outliers of ischemic stroke rats on KALE of one example of the present application;

[0025] Figure 8 is a comparison diagram of the parameter map of the KALE model of five-pool Lorenz fitting and the traditional parameter map of one example of the present application;

[0026] Figure 9is a traditional five-pool Lorentzian fit of an ischemic rat example of the present application compared with KALE statistics plot;

[0027] Figure 10 is a traditional five-pool Lorentzian fit R 2 Figure and fitted spectrum plot;

[0028] Figure 11 is a comparison plot of traditional method, IDEAL and KALE method of an example of the present application;

[0029] Figure 12 is a structural block diagram of an electronic device of an embodiment of the present application. DETAILED DESCRIPTION

[0030] Embodiments of the present application are described in detail below with reference to examples illustrated in the accompanying drawings, in which the same or like reference numerals represent the same or like elements throughout. The embodiments described below are examples and are intended to explain the present application, and should not be understood as limiting the present application.

[0031] Reference is made below to the accompanying Figure 1 - the accompanying Figure 12 A fast high signal-to-noise ratio CEST quantification map calculation method based on K-means clustering is described for an embodiment of the present application.

[0032] Figure 1 is a flowchart of a fast high signal-to-noise ratio CEST quantification map calculation method based on K-means clustering of an embodiment of the present application.

[0033] As shown in the figure, the fast high signal-to-noise ratio CEST quantification map calculation method based on K-means clustering includes: Figure 1

[0034] S1, obtain a CEST-MRI image, and set a clustering termination condition according to the number of pixels and the standard deviation of the CEST-MRI image, and determine a target K value.

[0035] The CEST-MRI image is an image to be quantified, which can be an animal brain image, such as a stroke rat image 64x64x51, a healthy human brain image 256x256x31, etc.

[0036] In some embodiments, the clustering termination condition is that the number of pixels in the class is less than a pixel number threshold, or the standard deviation of the pixels in the class is less than a standard deviation threshold. The pixel number threshold is the number of pixels of the CEST-MRI image divided by the square of a first preset parameter, and the standard deviation threshold is the standard deviation of the pixels in the CEST-MRI image divided by a second preset parameter. ​

[0037] Specifically, regarding the number of pixels of the CEST-MRI image, which is set in the CEST-MRI image acquisition process, can be regarded as known, for example, a rat image is 64x64, and a human brain image is 256x256. Regarding the standard deviation of the pixels in the CEST-MRI image, for example, the standard deviation of all pixel points in 50 images can be calculated using the std function of MATLAB.

[0038] Specifically, in order to reduce the difference in the tissue, the clustering termination condition can be set for the number of pixels and the standard deviation, as shown in the following formulas (1) and (2):

[0039]

[0040]

[0041] wherein σ cluster represents the standard deviation of the pixels in the class, σ brain represents the standard deviation of all pixels in the image; N cluster represents the number of pixels in the class, N image represents the number of pixels of a single frequency offset Z spectrum image. R N and R s are respectively a first preset parameter and a second preset parameter, which are parameters for controlling the clustering termination, and the values thereof can be obtained according to experimental verification. For example, R s = 120, R N = 50 on a human brain image, and R s = 72, R N = 13 on a rat image.

[0042] In some embodiments, determining the target K value includes: performing a pre-experiment on a preset image of the same type as the CEST-MRI image to obtain a first relationship between the K value and the final class number and a second relationship between the final class number and a quantification effect; determining a K value meeting the quantification effect as a candidate K value according to the first relationship and the second relationship, and taking the smallest candidate K value as the target K value.

[0043] The quantification effect can include but is not limited to: signal intensity of one or more of amide protons, amine protons, guanidino protons, hydroxyl protons, and nuclear Overhauser effect, and goodness of fit R 2 , contrast noise ratio of the quantification image, and coefficient of variation.

[0044] For example, when the quantification effect includes the signal intensity, in the pre-experiment, the following relationship can be obtained: Figure 6The second relationship shown in (a) is used to obtain all final class numbers corresponding to signal strengths satisfying preset conditions (e.g., similar signal strengths as those of traditional methods). According to the final class numbers and the first relationship, the corresponding K values are obtained as candidate K values, and the smallest candidate K value is selected as the target K value.

[0045] It should be noted that the preset image and the pre-experiment can be preselected and performed, and the corresponding target K value can be obtained for each type of image, i.e., the corresponding relationship between the image type and the target K value is obtained, and the corresponding relationship is stored. In actual application, the corresponding target K value can be obtained by calling the pre-stored corresponding relationship according to the type of the CEST-MRI image to be calculated.

[0046] S2, K-means clustering is performed on the target to be clustered based on the target K value, and parameter fitting based on a preset function is performed on the average Z spectrum of each class, wherein the initial target to be clustered is a CEST-MRI image.

[0047] In some embodiments, the K-means clustering is performed on the target to be clustered, including: performing K-means clustering based on the Z spectrum signal values of each pixel point in the target to be clustered.

[0048] Specifically, the CEST-MRI image is three-dimensional, and taking a rat image 64x64x50 as an example, the third dimension is the frequency offset. For example, the amide with a distance of 3.5 ppm from the water frequency corresponds to the 39th image in the third dimension, and the amide proton signal is contained in this image. When K-means clustering is performed on the target to be clustered, K-means clustering can be performed based on the frequency offset value.

[0049] After clustering, a plurality of different classes are obtained, each class contains a different number of pixels on the image, and the average value of the pixels contained in each class is calculated. Taking the rat image as an example, there are 50 images in total, and an average value is calculated for each image in a class. Therefore, the average Z spectrum is a 1x50 vector.

[0050] In the embodiments of the present application, the preset function can include but is not limited to: single-pool Lorentz function, multi-pool Lorentz function, super Lorentz function, Gaussian function, linear combination of at least two types of functions (i.e., at least two of single-pool Lorentz function, multi-pool Lorentz function, super Lorentz function, and Gaussian function), and Z spectrum function calculated by Bloch equation containing exchange term. It should be noted that when any preset function is used for initial fitting, the initial fitting value can be a preset value.

[0051] Taking the Lorentz fitting as an example, the Lorentz fitting can be performed by using a single-pool Lorentz fitting function or a multi-pool Lorentz function, and the fitting parameters and fitting values of each class are obtained.

[0052] In the multi-pool Lorentz fitting, the fitting boundary is limited in the clustering transmission process, and the limiting method is that the fitting peak value is limited to the fitting initial value plus or minus the first preset multiple of the maximum standard deviation in the class, the fitting peak width is limited to the fitting initial value plus or minus the second preset multiple of the maximum standard deviation in the class, and the fitting frequency deviation is limited to the fitting initial value plus or minus the third preset multiple of the maximum standard deviation in the class. The first preset multiple and the second preset multiple are both greater than 1, and the first preset multiple is less than the second preset multiple. The third preset multiple is greater than 0 and less than 1. For example, the first preset multiple is 10 times, the second preset multiple is 100 times, and the third preset multiple is 0.5 times.

[0053] Specifically, the single-pool Lorentz function is shown in the following formula (3), and the five-pool Lorentz function is shown in the following formula (4):

[0054]

[0055]

[0056] where Δω represents the frequency deviation from the water saturation frequency, a, b, and c represent the fitting frequency deviation, the fitting peak value, and the fitting peak width, respectively, and d represents the fitting reference peak value. z and M0 are the Z spectrum and the Z spectrum without the saturation pulse, a i , b i , and c i represent the fitting frequency deviation, the fitting peak value, and the fitting peak width, respectively, i represents the i-th proton pool, and i from 1 to 5 represents the amide proton pool (amide), the nuclear Overhauser effect (NOE) pool, the water pool (water), the semi-solid macromolecule (MT) pool, and the amine proton (amine) pool.

[0057] S3, respectively, determine whether each class satisfies the clustering termination condition.

[0058] S4, for the class that does not satisfy the clustering termination condition, the class is taken as a target to be clustered, and the fitting parameters of the average Z spectrum of the class are taken as the initial value of the next clustering and fitting, and the step of K-means clustering for the target to be clustered is turned to.

[0059] It should be noted that, as Figure 2 shown, for the multi-pool Lorentz fitting function, the fitting parameters of the class can also be used as the boundary limit of the fitting initial value in the next clustering and Lorentz fitting.

[0060] S5, for the class satisfying the clustering termination condition, outputting the fitting value of the class as the quantization signal of all pixels in the class, and when all classes satisfy the clustering termination condition, outputting a quantization map based on class fitting according to the fitting values of all classes.

[0061] Specifically, the flow of the fast high signal-to-noise ratio CEST quantification parameter map calculation method based on K-means clustering is as shown in the figure Figure 2 Taking the Lorentz fitting as an example, the entire CEST image sequence is input into the K-means++ clustering algorithm, and K=2 is taken as an example in the figure. The dark gray line represents the class satisfying the clustering termination condition, and the iteration will continue. The light gray line represents the class satisfying the clustering termination condition, and the output is performed. Output 1 represents the class-level output, that is, group-k, and output 2 is the voxel-level output, that is, voxel-k. The single-pool strategy refers to the KALE model of the Lorentz difference (LD), and the parameter transfer process is illustrated. The multi-pool strategy refers to the KALE model of the multi-pool Lorentz fitting, and the updated parameters and boundaries in the multi-pool Lorentz fitting process are illustrated, and the final parameters are taken as the output.

[0062] The fast high signal-to-noise ratio CEST quantification parameter map calculation method based on K-means clustering adopts hierarchical clustering to provide a classified update of the fitting parameters, specifically: the voxels in the CEST MRI image are clustered into multiple classes according to their Z-spectrum characteristics, and then the average spectrum in the class is fitted with loosely set initial values to obtain a class fitting spectrum and a set of fitting parameters; each class is further clustered into subclasses until the number of pixels or the intensity error meets the standard, and the fitting parameters of the fine clustering are used to add voxel fitting to further improve the results. This method can improve the signal-to-noise ratio and fitting accuracy of the CEST-MRI quantification map and significantly reduce the calculation time, which is beneficial to the promotion and application in preclinical research and clinical practice.

[0063] In some embodiments, before clustering the CEST-MRI image, the method further comprises: preprocessing the CEST-MRI image according to the number of pixels and the standard deviation of the CEST-MRI image.

[0064] Specifically, the preprocessing of the CEST-MRI image can include: standardizing the CEST-MRI image, and correcting the B0 field inhomogeneity of the standardized CEST-MRI image by using the WASSR method.

[0065] Specifically, the B0 field inhomogeneity will cause a frequency shift, for example, the image at the amide proton (3.5ppm), some of the pixels in the image are shifted to 3.6ppm due to the B0 field inhomogeneity, and the B0 image collected shows that the shift of these pixels is-0.1ppm. When preprocessed, the frequency of these pixels is added to the value of the corresponding pixel in the B0 image, and the correction is completed.

[0066] In some examples, the normalized CEST-MRI image is obtained by dividing pixel values of each position in the corrected CEST-MRI image by pixel values of corresponding positions in a preset reference image.

[0067] Specifically, the normalization processing, which can also be referred to as standardization processing, can be implemented by using the formula M = S sat / S0. Wherein, S0 is a preset reference image, which can be determined from the acquired CEST-MRI image. For example, S0 can be an image far from the water frequency, which does not occur chemical exchange saturation transfer, i.e., does not contain CEST information; S sat is the corrected CEST-MRI image, and M is the normalized CEST-MRI image. The CEST-MRI image is three-dimensional, and for example, the rat brain image is 64x64x51, wherein the first 50 images are from -10ppm to 10ppm of the water frequency, denoted as M z , and the last image is an image far from the water frequency (which is considered to be far from the water and does not occur chemical exchange saturation transfer), which does not contain CEST information, denoted as M0. The normalization processing is performed on the first 50 frequency offset images, specifically: the pixel value of each position in each image M z is divided by the pixel value of the corresponding position of the image M0.

[0068] In some embodiments, after obtaining the class-based fitting quantization map, the method further includes: for the class-based fitting quantization map, performing a pixel-by-pixel parameter fitting based on a preset function on each pixel point with the quantization value in each class as the initial value of fitting, to obtain a fine quantization map combining class fitting and pixel-by-pixel fitting.

[0069] It should be noted that after obtaining the class-based fitting quantization map, class average fitting signal extraction can be performed, and the extracted signal (i.e., the quantization signal) is denoted as LD group-k . Taking Lorentz fitting as an example, LD group-k is equal to the Lorentz fitting of the intra-class average Z spectrum minus the intra-class average Z spectrum, i.e., LD group-k = LF group-k - Z avg_k , where Z avg_k represents the intra-class average Z spectrum, and LF group-k represents the Lorentz fitting of the intra-class average Z spectrum, from which the amide signal and the nuclear Overhauser effect (NOE) signal can be obtained.

[0070] After obtaining the fine quantization map combining class fitting and pixel-by-pixel fitting, pixel-by-pixel fitting signal extraction can be performed, and the extracted signal (i.e., the quantization signal) is denoted as LD voxel-k. Take the Lorentz fitting as an example, the Lorentz fitting parameters of the class average Z spectrum are taken as initial values for Lorentz fitting of each pixel in the class, and the fitted value is subtracted from each pixel point Z spectrum to obtain LD voxel-k , i.e. the quantized amide proton (amide) signal and nuclear Overhauser effect (NOE) signal.

[0071] The beneficial effects of the fast high signal-to-noise ratio CEST quantification parameter map calculation method based on K-means clustering of the application are illustrated by experiments as follows:

[0072] (I) Experimental data

[0073] 1) Stroke rat data. The animals used in the experiment are adult male Sprague-Dawley rats weighing between 240-250 grams, and the stroke model used is transient focal cerebral ischemia. The rats were reperfused within 2 hours after MCAO (middle cerebral artery occlusion), and 2 hours of middle cerebral artery occlusion occurred after the nylon suture was pulled out, and the rat body temperature was maintained at 37-37.5°C during the operation.

[0074] 2 hours after MCAO surgery, brain MR imaging was performed on a Biospec 7T horizontal scanner (Bruker, Germany) using a transmit-receive volume coil (diameter = 40 mm). The rats were first anesthetized with 4% isoflurane, and then maintained at 1-2% during data acquisition. First, multi-slice T2w images were acquired, and then 1 mm thick coronal slices were selected in the center of the striatum to collect WASSR and CEST images. The imaging sequence included continuous wave presaturation pulses (T sat = 2500 ms) and fast-acquisition relaxation enhancement (RARE) readouts (RARE factor = 32, TR / TE = 5000 ms / 4 ms). When B1sat = 0.7 μT, 51 Z-spectrum images were collected from -10 ppm to 10 ppm.

[0075] 2) Normal human brain data. A number of healthy volunteers were recruited for brain data acquisition to verify KALE on clinical data. Magnetic resonance data were collected on a 3.0 Tesla Philips scanner (Ingenia, Philips Healthcare) with a 32-channel phased array.

[0076] CEST images were acquired using a 3D APT sequence, and 7 mm thick axial slices were selected from T2w images for turbo spin echo readout. Other imaging parameters for CEST images are: T sat= 2 s, B1sat= 0.7 pT, 31 frequency offsets (i.e. ±0.5, 0.8, ±1, 1.2, 1.7, 1.9, ±2.1, 2.3, 2.5, ±3, +3.4, ±3.5, +3.6, +3.7, ±4, +4.5, ±5, ±6, ±8, ±10 ppm), TR = 5 s, TE = 7.8 ms, FOV = 220 x 186 mm2, voxel size 2 x 2 x 7 mm3, matrix size 256 x 256, 5 slices. M0 image at -1560 ppm was acquired for data normalization. WASSR images (38) were acquired for B0 correction, with the same saturation time and power as the animal scans. Multi-slice T2w images were acquired with a slice thickness of 1.5 mm, TR = 5000 ms, TE = 336 ms. T1w images were acquired with a slice thickness of 1.25 mm, TR / TE of 6.5 / 3.2 ms.

[0077] (ii) Data analysis

[0078] All data were processed using custom-written MATLAB code on an Intel(R) Core(TM) i7-9700K processor, 3.60 GHz CPU, and 32 G RAM computer. First, the WASSR images were used to correct the saturation data for B0 inhomogeneity, then the MATLAB function "spline" was used for interpolation with a step size of 0.1 ppm. Then the M0 images without saturation pulses were used to normalize the B0-corrected image series. The calculation time from start to finish (excluding data preprocessing) was obtained using the MATLAB functions "tic" and "toc".

[0079] 1) Determination of parameters R s and r n

[0080] To determine the appropriate R s and R n , the values of r s and R N were verified to be greater than 130 and 60, respectively, as the pixel number threshold and standard deviation threshold were too small, resulting in too many iterations of clustering. Therefore, R s was calculated from 0 to 130 with an interval of 10 and R N ​The two-by-two combination of intervals from 0 to 60 by 5, the indicators are the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR) between the Lorentz fitting and the original input saturated data, since SSIM and PSNR can only measure two images at a single frequency offset, and cannot represent the fitting effect of the entire Z spectrum, so the SSIM and PSNR of the frequency offset far from water and the frequency offset near the water signal are calculated, and the average value is taken as the indicator, so as to more comprehensively measure the fitting effect of the entire Z spectrum, and the specific frequency offset is -10ppm to -6.25ppm, -2ppm to 2ppm and 6.25ppm to 10ppm. The calculation formulas of SSIM and PSNR are as follows:

[0081]

[0082] Wherein, x represents the fitting value, y represents the standardized original Z spectrum data; μ x and μ y represent the average value of x and y; σ x and σ y represent the standard deviation of x and y; σ xy represents the covariance of x and y; c1=(k1L) 2 ,c2=(k2L) 2 are two weak stability coefficients, usually k1=0.01, k2=0.03, and L is the dynamic range of pixels.

[0083]

[0084] Wherein, peakvalue is set to 1, and MSE represents the mean square error of x and y.

[0085] 2) Details of Lorentz fitting

[0086] For single-pool fitting, the initial parameters of KALE (i.e. the method of the application) and traditional Lorentz fitting are set according to the existing related literature. Z spectrum is used for fitting in the range of [-6.25, -10], [-2, 2] and [6.25, 10] ppm. In addition, different levels of Gaussian noise (σ is 0-0.03) are added to the human brain image to evaluate the robustness of the KALE model of Lorentz difference (LD) to noise, and in addition, the outlying points are tested on the rat data, i.e. using the stroke rat brain image with muscle and blood vessels to test the performance of KALE on outlying pixels.

[0087] For five-pool fitting, the entire Z spectrum is input for fitting. The initial values and boundaries of the fitting parameters of KALE and the traditional method are shown in Table 1.

[0088] Table 1

[0089]

[0090] Note: astart represents the starting point of the offset for the five-pool Lorentz fitting, bstart represents the starting point of the amplitude, cstart represents the starting point of the line width, low and high represent the minimum and maximum boundaries, and the units of peak offset and line width are ppm.

[0091] For the traditional Lorentz fitting method and the KALE method, optimization was performed using the Matlab function lsqcurvefit (a nonlinear least squares solver) and the Lengenberg-Marquardt algorithm. The termination tolerance of the function value was set to 1e-10, and the maximum number of function evaluations was set to 10000. Numerical solutions were obtained using the finite difference method, without using the Jacobian matrix.

[0092] 3) Evaluation of Fitting Performance

[0093] The quantification performance of traditional methods versus KALE is measured using the coefficient of variation (COV) within the region of interest (ROI):

[0094]

[0095] Where σ(brain) and μ(brain) represent the standard deviation and mean within the brain.

[0096] Using the coefficient of determination (R) 2 Calculate the goodness of fit:

[0097]

[0098] Among them, y i It is the Z-spectrum to be fitted. It is the average value of the Z spectrum, y i ′ is the fitted spectrum, and N is the number of pixels in the brain.

[0099] The significance of each statistical item was analyzed using the Prism embedded paired t-test. A p-value ≤ 0.05 was considered to indicate a difference between groups.

[0100] (III) Experimental Results

[0101] 1) Determine the K value and the weighting parameter R for K-means clustering. N ,R s

[0102] Figure 3 This is a schematic diagram illustrating clustering parameter optimization as an example of the present invention. Figure 3 (a) shows the relationship between the K value and the computation time, and (b) shows the relationship between R... N ,R s The relationship between R and computation time is shown in (c).N The relationship between Rs and the Structural Similarity Index (SSIM) measure, (d) shows R N ,R s Relationship with peak signal-to-noise ratio (PSNR).

[0103] To determine the appropriate K-value and clustering termination criteria for clinically sized data (256×256), KALE was tested on healthy human brain data. The choice of K-value was first evaluated based on the algorithm's computation time, with empirical R... N =50,R s =120. For example... Figure 3 As shown in (a), when K is greater than 3, the computation time increases sharply, so K can be set to 3. When K = 3, considering both computation time and fitting performance, including SSIM and PSNR, R0 is determined. N ,R s Because of R N Greater than 60, R s If the value is greater than 130, clustering cannot converge, therefore, for R... N and R s Perform cluster tests for 0-60 and 130. For example... Figure 3 As shown in (b)-(d), when R N Greater than 40 and R s When R is greater than 50, the computation time group increases significantly, while when R... N Less than 40 and R s When the value is less than 50, SSIM and PSNR increase slowly. To obtain the best fitting performance in a sufficiently short computation time, R is chosen. s =120 and R N =50, where SSIM=0.989, PSNR=35.65, and the calculation time is 16.84 seconds / piece.

[0104] The parameters of the rat image were estimated using a linear relationship between the size of the human brain image (256×256) and the size of the rat image (64×64) and their standard deviations. Therefore, a coarse estimate of the rat image [R] was obtained. N ,R s [13,76].

[0105] 2) Noise Resistance Analysis

[0106] Figure 4 This is a schematic diagram illustrating an example of the robustness evaluation of noise according to the present invention. Figure 4Figure (a) shows the B0 plot, and (b) shows the T1w plot. The white lines drawn on the shell and core are the regions of interest (ROI) analyzed in Figures (f) and (g); (c) shows the T2w plot, and (d) and (e) show the amide (3.5 ppm) and nuclear Orthauine effect NOE (-3.5 ppm) plots (LD) of the conventional method and the method of this invention. voxel-k This represents pixel-level LD results. group-k (Representing class-level LD results). (f) and (g) show the coefficients of variation (COV) for amide (3.5 ppm) and NOE (-3.5 ppm) in the ROI of the three methods, where black represents the conventional method and gray represents LD. group-k ,

[0107] Light gray is LD voxel-k (h) shows the statistical results of the calculation time for five subjects (the volunteers mentioned above) (5 pieces per subject). *** indicates p < 0.001.

[0108] By introducing varying degrees of Gaussian noise, clinically sized human brain data was further used to evaluate the quantification effect and noise robustness of KALE with LD. Figure 4 For noise-free Z-spectrum, KALE's LD contrast plot is similar to conventional results. When the CEST-MRI image sequence is subjected to Gaussian noise of σ = 0.01, LD... group-k The method outperforms other methods, achieving contrast similar to clear images. To evaluate the performance of the method affected by noise, the COV within the shell and core was calculated. Figure 4 (b) shows the regions of interest (ROIs) drawn in the middle. Figure 4 From (f) and (g), it can be seen that LD group-k The method's COV is unaffected by noise, while LD voxel-k The COV of the conventional method increases with increasing noise level. The fitting results of 5 subjects (5 pieces per subject) were included in the statistical analysis of computation time. group-k and LD voxel-k The average computation time per subject (5 tablets) was 2.18 minutes and 4.82 minutes, respectively, significantly shorter than the 9.48 minutes of the traditional method. Figure 4 (h), p<0.001).

[0109] 3) Evaluation of brain function images in stroke rats

[0110] Figure 5 This is a schematic diagram of Lorenz difference analysis in ischemic stroke rats, an example of the present invention. Figure 5(a) shows T2w images of the region of interest (ROI1, black circle) on the lesion side and the region of interest (ROI2, white circle) on the contralateral side. (b) shows the B0 map obtained by WASSR. (c) The top shows the results of conventional LD ​​analysis and voxel-level KALE (LD) analysis, respectively. voxel-k ) and class-level KALE (LD) group-k (d) Comparison of amide protons at 3.5 ppm; the bottom shows the NOE comparison at -3.5 ppm, with the same layout as the top row. *** indicates a p < 0.001 in the t-test, indicating a highly significant difference between the two methods. (e)-(g) show the traditional method, voxel-level KALE (LD) method, and other methods, respectively. voxel-k ) and class-level KALE (LD) group-k ) spectral lines.

[0111] First, the KALE model of Lorenz difference was evaluated in the brains of stroke-affected rats, such as... Figure 5 As shown. For the Z-spectral image sequence, KALE clustered 4096 pixels in the CEST image into 107 classes, reducing the computation time for pixel-by-pixel fitting from 7 seconds to 2 seconds. Due to the limitation of the number of pixels, the saving in computation time is not significant. However, statistically, the computation time for 6 rats was significantly reduced in both group-k and voxel-k KALE (see...). Figure 5 (d)). Since only the initial parameters were optimized, the image ( Figure 5 (c) and spectrum ( Figure 5 In (e) and (f)), the contrast and spectra of KALE in group-k are very similar to those of conventional methods. However, the KALE results in group-k show lower contrast intensity in the amide and NOE plots.

[0112] Figure 6 This is a schematic diagram illustrating the relationship between the number of categories and the quantification effect in ischemic stroke rat data, as an example of the present invention. Figure 6 (a) shows the relationship between the final class number of ROI and the average signal intensity on the comparison plots of amide and the nuclear Austmann effect (NOE). (b) shows the traditional method and the LD method with K=35 for amide and the nuclear Austmann effect (NOE), respectively. group-k The comparison chart shows that the white curved area represents the ROI.

[0113] For group-k KALE, the intra-class average reduces image resolution; increasing the final number of classes can improve image resolution. Experiments were conducted on the final number of clusters and the image contrast of group-k KALE data for rat data, such as... Figure 6The final number of clusters increases by adjusting the K value, and when K = 35, the final number of clusters is 330, and the average signal intensity of amide and NOE within the ROI is similar, as shown in FIG. 3B. Figure 6 The final number of clusters increases by adjusting the K value, and when K = 35, the final number of clusters is 330, and the average signal intensity of amide and NOE within the ROI is similar, as shown in FIG. 3B.

[0114] Figure 7 FIG. 4 is a schematic diagram of the outlying points of ischemic stroke rat data in an example of the present application, which is shown on KALE, Figure 7 The first row from left to right in FIG. 5 is a saturation image at 3.5 ppm, traditional LD, KALE of voxel-k, KALE of group-k, and a plotted APT contrast map. The second row is a saturation image at -3.5 ppm, traditional LD, KALE of voxel-k, KALE of group-k, and a plotted NOE contrast map.

[0115] A brain image with muscle and blood vessels was used to test the performance of KALE on outlying voxels. As shown in FIG. 6, Figure 7 As shown in FIG. 7, both the amide proton (APT) contrast map of traditional Lorentz difference (LD) and the amide proton (APT) map of KALE show high signal in the internal carotid artery and vertebral artery (marked by white circles in the figure). The signal intensity of the unmarked internal carotid artery is lower, which may be related to the partial proton effect. For KALE of group-k and voxel-k, the clustering is accurate for outlying voxels.

[0116] 4) KALE of five-pool Lorentz fitting

[0117] Figure 8 FIG. 8 is a contrast map of the parameter map of the KALE model of five-pool Lorentz fitting and the traditional parameter map in an example of the present application. Each column from left to right is the amide pool (3.5 ppm), the NOE pool (-3.5 ppm), the amine pool (2 ppm), the water pool, and the MT pool. Figure 8 (a) shows the amplitude of traditional five-pool Lorentz fitting (first row) and KALE (second row). (b) shows the line width map of traditional five-pool Lorentz fitting (first row) and KALE (second row). (c) shows the offset map of traditional five-pool Lorentz fitting (first row) and KALE (second row).

[0118] Figure 9 FIG. 9 is a statistical contrast map of traditional five-pool Lorentz fitting and KALE in an example of the present application. The black bar represents the traditional five-pool Lorentz fitting, and the gray bar represents the KALE of five-pool Lorentz fitting. Figure 9(a) shows the contrast-to-noise ratio (CNR) of each pool (two ROIs per rat) assessed by the two methods on amplitude maps of 6 rats (n=6). (b) shows the coefficient of variation (COV) of each pool assessed by the two methods on amplitude maps of 6 rats (n=6). (c) shows the mean goodness of fit (R²) of the two methods on 6 rats (n=6). 2 The average R of the KALE method 2 The average R-value of the traditional method is 0.9935. 2 The mean value is 0.9934. (d) shows the computation time for the two methods on 6 rats (n=6). The average computation time for the conventional method was 8.27 minutes, and the average computation time for the KALE method was 0.41 minutes. * indicates p<0.05, ** indicates p<0.01, and *** indicates p<0.001.

[0119] Figure 10 This is an example of a five-pool fitting R-squared method using two different approaches from this invention. 2 Figure and fitted spectrum. Figure 10 (a) shows the R values ​​of KALE and the traditional method. 2 Figure. ROI1, plotted on KALE, marks the R on the KALE plot. 2 Regions with higher values. ROI2 is plotted in two methods. 2 (b) ROI1: Original data points (circled), KALE fitted spectrum (light gray line), traditional method fitted spectrum (dark gray line), and the spectral difference between the two methods plotted on the right vertical axis (dashed line). (c) Spectral lines in the range of 1–5 ppm. (d) ROI2: Original data points (circled), KALE fitted spectrum (light gray line), traditional method fitted spectrum (dark gray line), and the spectral difference between the two methods plotted on the right vertical axis (dashed line). (e) Spectral lines in the range of 1–5 ppm.

[0120] KALE was further integrated into a five-pool Lorentz fit and tested on stroke rats with the same K value (K=3) and clustering termination criteria as the KALE animal experiments with Lorentz difference (LD). Figure 8 , Figure 9 As shown, KALE improved the image quality of amplitude maps, and the contrast-to-noise ratio (CNR) between lesions and contralateral tissues was significantly improved in all six rats. Furthermore, the fitted offset and linewidth maps in KALE were more uniform, with less noise, particularly in the MT pool. Figure 9 As shown, COV was tested using a normal region ROI and a whole brain ROI. 2 The mean values ​​were used to evaluate the goodness of fit in 6 rats. Compared with the traditional method, KALE had a smaller COV and a higher R value. 2 R values ​​for both methods 2 Comparison chart as followsFigure 10 As shown, two ROIs were drawn in the regions with higher values in the KALE contrast map. Since the difference in the two methods in fitting the spectra showed greater variance in the range of 1 to 5 ppm, the zoomed-in spectra of 1 to 5 ppm were drawn as indicated by the arrows, where KALE fitted the original data better than the traditional method in both ROIs by about 2 ppm. In addition to improving the fitting performance, as shown in Figure 10 Figure 9 As shown, the average computation time for the 6 rats was 0.41 minutes (24.6 seconds), which was substantially shorter than the 8.27 minutes for the traditional method.

[0121] Compared to the traditional Lorentzian fitting using fixed initial values and boundaries for all pixels, KALE had different initial values and boundaries for classes, and these values were constantly optimized and updated as the classes became smaller. Figure 10 As shown in the zoomed-in spectra in (b), the traditional Lorentzian fitting in pixel units can have inaccurate fitting in some voxels. In contrast, KALE achieved stronger adaptability and more accurate fitting through multi-level voxel clustering.

[0122] 5) KALE vs. IDEAL

[0123] Figure 11 The comparison of the traditional method, IDEAL, and KALE method in one example of the present application, the amplitude maps of the three methods and the corresponding R 2 maps are shown in FIG. 6, where each row forms the amide (3.5 ppm), nuclear Overhauser effect (NOE, -3.5 ppm), amine (2 ppm), water, and MT amplitude maps, and the R 2 maps from left to right. Figure 11 In (a), the results of IDEAL are shown. In (b), the results of the traditional Lorentzian fitting are shown. In (c), the results of KALE are shown. In (d), four sets of ROIs drawn on the lesion hemisphere (solid line) and the normal hemisphere (dashed line) are shown. In (e), the CNRs of the three methods on the amide, NOE, and amine contrast maps are shown. The blank bar represents the traditional method, the gray bar represents IDEAL, and the black bar represents KALE.

[0124] The amplitude maps produced by the IDEAL method are smoother than those of the traditional method. In the NOE contrast map, the CNR of IDEAL is comparable to that of KALE, the CNR of amide is lower by ~1.47, and the CNR of amine is higher by ~0.40 Figure 10 (e)). The principle of improving the signal-to-noise ratio by updating the region of the fitting parameters is the same as that of IDEAL. However, KALE is more accurate, as it clusters based on the structural correlation of voxels, while the smoothness of IDEAL is square in some regions. The R 2 ​0.9605, 0.9924 and 0.9928, respectively. The computation time of IDEAL and KALE is 65.14 s and 20.98 s, respectively.

[0125] Based on the above experiments, the CEST quantification method based on K-means clustering and accelerated Lorentz fitting proposed in the application adopts data-driven pixel classification and hierarchical parameter updating, and provides a fast and accurate method for CEST quantification map. At each level, the voxels in the class are clustered according to the similarity of the Z spectrum of the voxels. Then, only one fitting process is performed in a subclass, and the parameters of the batch optimization are used. Compared with the traditional voxel fitting, the ischemic rat brain and healthy human brain show comparable or even better CEST contrast maps, and the computation time is significantly reduced. Intelligent hierarchical data-driven pixel clustering and parameter updating may contribute to more preclinical and clinical CEST applications.

[0126] Since the traditional voxel-based fitting method separately fits the spectrum of each voxel, the computation time increases with the size of the image. In addition, the fitting process uses the same fixed parameters for all pixels without considering the correlation and diversity between voxels. Therefore, for a part of voxels, multiple iterations are needed to reach the optimization goal. Since KALE updates the parameters multiple times using different initial values, the number of iterations experienced by the fitting process is less. For rat data, the traditional LD average iteration is 14-41 times, and the KALE average iteration is 5-10 times. In addition to the computation time of the fitting process, KALE also needs clustering time. But compared with the long fitting process, the average clustering time is almost negligible, and for the final clustering of the human brain slice, only about 17 seconds is needed. Therefore, KALE greatly reduces the computation time by reducing the number of fittings and the number of iterations of each fitting.

[0127] For each clustering level, KALE fits the average spectrum of all voxels in the cluster, thereby improving the anti-noise performance compared with the traditional voxel LD. Compared with the previous IDEAL which uses fixed ratio square downsampling for fitting, KALE uses the same voxel grouping idea to maintain the anti-noise performance. But compared with IDEAL, KALE is more intelligent, which clusters according to the similarity of the Z spectrum of the voxels. Therefore, KALE avoids the reduction of image contrast from the average value of the voxels, thereby obtaining more accurate quantification maps with less structural details. It should be noted that the final cluster number will affect the overall quantification accuracy and image details of KALE, and the final class number is mainly determined by the K value and the clustering termination condition. Since adjusting the termination condition may cause clustering failure, the application studies different K values to determine the optimal final cluster number. The rat data set shows that K = 35 can provide LD group-k Results.

[0128] When using the multi-pool Lorentzian fitting, KALE only uses 1 / 20 of the computation time (0.41 min vs. 8.27 min) compared to the traditional voxel-wise fitting, while achieving higher CNR and better fitting results. This further demonstrates the advantage of KALE in complex multi-parameter fitting, where the five-pool model has 15 parameters, and the initial values of the fitting parameters and their boundaries are updated at each level of K-means clustering. The boundary constraint mainly considers the standard deviation of the voxels within each cluster in all the shifts in the original Z-spectrum. Therefore, the anomalies that may occur in voxel fitting can be constrained by the information of adjacent voxels.

[0129] Currently, several machine learning-based CEST quantification methods have been proposed. Although the online step of the machine learning-based method is fast, the contrast images produced are highly dependent on the training set. Model-based fitting is the most widely used method in current in vivo experiments. The basic idea of KALE is to update the fitting parameters in a clustered down-sampled manner based on the similarity within the same tissue or structure, which can be implemented on other quantification methods such as the polynomial (PLOF), omega analysis (LAREX omega-plot) and Bloch-McConnell equation fitting. Although KALE is only evaluated on brain images in this work, as Figure 7 shown, it demonstrates its robustness to outliers (such as blood vessels), which indicates the potential application prospect of KALE on other structures.

[0130] Based on the K-means clustering-based fast high signal-to-noise ratio CEST quantification parameter map calculation method of the above embodiment, the present application further proposes a computer readable storage medium.

[0131] In this embodiment, the computer readable storage medium stores a computer program, which, when executed by a processor, implements the K-means clustering-based fast high signal-to-noise ratio CEST quantification parameter map calculation method of the above embodiment.

[0132] Figure 12 is a structural block diagram of an electronic device according to an embodiment of the present application.

[0133] As Figure 12 shown, the electronic device 500 includes a processor 501 and a memory 503. The processor 501 and the memory 503 are connected, such as through a bus 502. Optionally, the electronic device 500 can also include a transceiver 504. It should be noted that in actual applications, the transceiver 504 is not limited to one, and the structure of the electronic device 500 does not constitute a limitation on the embodiments of the present application.

[0134] The processor 501 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic device, a transistor logic device, a hardware component, or any combination thereof. It can implement or execute various exemplary logical blocks, modules and circuits described in conjunction with the present disclosure. The processor 501 can also be a combination of computing functions, such as one or more microprocessor combinations, combinations of DSP and microprocessor, etc.

[0135] The bus 502 can include a path for transmitting information between the above-mentioned components. The bus 502 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 502 can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 8 In the figure, only one thick line is used to represent, but it does not mean that there is only one bus or one type of bus.

[0136] The memory 503 is used to store a computer program corresponding to the K-means clustering-based fast high signal-to-noise ratio CEST quantification parameter map calculation method of the above-mentioned embodiments of the present application, which is controlled and executed by the processor 501. The processor 501 is used to execute the computer program stored in the memory 503 to realize the content shown in the foregoing method embodiments.

[0137] Among them, the electronic device 500 includes but is not limited to: mobile phone, notebook computer, digital broadcast receiver, PDA (Personal Digital Assistant), PAD (Tablet Personal Computer), PMP (Portable Multimedia Player) and other mobile terminals such as mobile terminals such as digital TV, desktop computer and the like. Figure 8 The electronic device 500 shown is only an example, and should not bring any limitation to the function and use range of the embodiments of the present application.

[0138] It is to be appreciated that the above description and the examples that follow are intended to be illustrative only and that changes can be made to the description and examples without departing from the scope of the application. Note also that the use of particular brand names in the description is solely for illustration and should not be construed as an endorsement of such brands.

[0139] It should be understood that aspects of the application can be implemented in hardware, software, firmware or a combination thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, known in the art, can be used: a hybrid of the technologies mentioned above, discrete logic circuitry having logic gates for implementing logic functions upon data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so forth.

[0140] In the description of the present application, reference has been made to the use of terms such as "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means that a particular feature, structure, material or characteristic is included in at least one embodiment or example of the present application. The illustrative examples given are not necessarily to be construed as preferred or advantageous or with the exclusion of other equally valid examples that can be particularly adapted to a given application. Furthermore, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and operation described. Accordingly, all suitable modifications and equivalents should be included within the scope of the present application.

[0141] In the description of the application, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application.

[0142] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise explicitly specified and limited.

[0143] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0144] In the present application, unless otherwise explicitly specified and limited, the first feature "on" or "under" the second feature can be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, the first feature "above", "above" and "above" the second feature can be that the first feature is directly above or obliquely above the second feature, or only indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "below" and "below" the second feature can be that the first feature is directly below or obliquely below the second feature, or only indicates that the horizontal height of the first feature is less than that of the second feature.

[0145] Although the embodiments of the application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as a limitation on the application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the application.

Claims

1. A fast, high signal-to-noise ratio CEST quantization parameter plot calculation method based on K-means clustering, characterized in that, The method includes: Acquire CEST-MRI images, and set clustering termination conditions based on the number of pixels and standard deviation of the CEST-MRI images, and determine the target K value; K-means clustering is performed on the target K value, and the average Z spectrum of each class is fitted with parameters based on a preset function. The initial target to be clustered is the CEST-MRI image. Determine whether each class satisfies the clustering termination condition; For classes that do not meet the clustering termination condition, the class is taken as the target to be clustered, and the fitting parameters of the average Z spectrum of the class are used as the initial values ​​for the fitting after the next clustering, and then the process proceeds to the step of performing K-means clustering on the target to be clustered. For a class that satisfies the clustering termination condition, the fitted value of the class is output as the quantization signal of all pixels in the class. When all classes satisfy the clustering termination condition, a quantization map based on the fitted values ​​of all classes is output.

2. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 1, characterized in that, Before setting the clustering termination condition based on the number of pixels and standard deviation of the CEST-MRI images and clustering the CEST-MRI images, the method further includes: The CEST-MRI images were preprocessed.

3. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 2, characterized in that, The preprocessing of the CEST-MRI images includes: The CEST-MRI images were standardized, and the B0 field inhomogeneity was corrected using the WASSR method.

4. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 1, characterized in that, The process of performing K-means clustering on the target clustering object includes: K-means clustering is performed based on the Z-spectrum signal values ​​of each pixel in the target to be clustered.

5. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 1, characterized in that, The clustering termination condition is: the number of pixels within a cluster is less than the number of pixels threshold, or the standard deviation of pixels within a cluster is less than the standard deviation threshold. Wherein, the pixel count threshold is the number of pixels in the CEST-MRI image divided by the square of a first preset parameter, and the standard deviation threshold is the standard deviation of pixels in the CEST-MRI image divided by a second preset parameter.

6. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 1, characterized in that, The method also includes, For the quantization map based on class fitting, the quantization values ​​within each class are used as the initial values ​​for fitting. Then, pixel-by-pixel parameter fitting is performed on each pixel based on the preset function to obtain a fine quantization map that combines class fitting and pixel-by-pixel fitting.

7. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 1, characterized in that, The preset functions include: Single-pool Lorentz function, multi-pool Lorentz function, super-Lorentz function, Gaussian function, linear combination of at least two of the above functions, and Z-spectral function calculated by Bloch equations containing commutative terms; In the Lorentz fitting process using a multi-pool Lorentz function, the fitting boundary is restricted during the clustering transfer process. The restriction method is as follows: the fitting peak value is restricted to a first preset multiple of the initial fitting value plus or minus the maximum standard deviation within the class; the fitting peak width is a second preset multiple of the initial fitting value plus or minus the maximum standard deviation within the class; and the fitting frequency offset is a third preset multiple of the initial fitting value plus or minus the maximum standard deviation within the class. The first preset multiple and the second preset multiple are both greater than 1, and the first preset multiple is less than the second preset multiple. The third preset multiple is greater than 0 and less than 1.

8. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 1, characterized in that, Determining the target K value includes: Preliminary experiments were conducted on preset images of the same type as the CEST-MRI images to obtain the first relationship between the K value and the final number of classes, and the second relationship between the final number of classes and the quantization effect. Based on the first relationship and the second relationship, a K value that satisfies the quantization effect is determined as a candidate K value, and the smallest candidate K value is taken as the target K value.

9. The method for calculating the fast, high signal-to-noise ratio CEST quantization parameter map based on K-means clustering according to claim 8, characterized in that, The quantization effect includes the signal intensity of one or more of the following: amide protons, amine protons, guanidine protons, hydroxy protons, and the nuclear Orstauld effect, as well as the goodness of fit R. 2 The contrast-to-noise ratio and coefficient of variation of the quantization plot.

Images