CN114999627A

CN114999627A - CEST magnetic resonance analysis method and device based on random forest arrangement

Info

Publication number: CN114999627A
Application number: CN202210426923.8A
Authority: CN
Inventors: 宋小磊; 陈一兵; 党旭剑; 朱昌浩; 刘础雨; 贺小伟
Original assignee: NORTHWEST UNIVERSITY; Tsinghua University
Current assignee: NORTHWEST UNIVERSITY; Tsinghua University
Priority date: 2022-04-21
Filing date: 2022-04-21
Publication date: 2022-09-02
Also published as: WO2023202662A1

Abstract

The application discloses a CEST magnetic resonance analysis method and device based on a random forest arrangement, wherein the method comprises the following steps: acquiring a CEST magnetic resonance image to be tested and voxel labels in a lesion area and a non-lesion area; and inputting the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image into a pre-trained permutation random forest model to obtain the importance of each saturation frequency deviation, and determining the contribution of all saturation frequency deviations in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency deviation. Thus, the contribution of the metabolism-related frequency offset signal to the lesion classification and the problem of accurate segmentation of the lesion region can be robustly explained.

Description

CEST magnetic resonance analysis method and device based on random forest arrangement

Technical Field

The application relates to the technical field of chemical exchange saturation transfer magnetic resonance imaging, in particular to a CEST magnetic resonance analysis method and device based on random forest arrangement.

Background

Current clinical routine nuclear magnetic images can more easily delineate diseased regions (e.g., stroke, tumor, etc.), but more detailed metabolic characterization requires advanced metabolic imaging methods such as Chemical Exchange Saturation Transfer (CEST) magnetic resonance imaging. CEST magnetic resonance imaging is an emerging molecular imaging technique that has been used pre-and clinically to assess metabolic impairment in tumors, ischemia and other diseases, to aid in early diagnosis of stroke, to determine if a tumor is benign or malignant, etc. The principle of a typical CEST acquisition sequence is as follows: at the resonance frequency of the specific exchangeable proton of the label, a saturated radio frequency pulse is used to resonate the exchangeable proton to generate a signal saturation phenomenon. This signal saturation is then transferred from the exchangeable protons to the water molecule through various exchange processes, causing a change in the water signal, so that the CEST technique can indirectly detect the signal of a specific labeled proton by detecting the water signal.

For example, Amide Proton Transfer Weighted (APTw) CEST magnetic resonance imaging technique subtracts the signals at normalized-3.5 ppm and 3.5ppm frequency offsets, thereby measuring the changes in Amide protons that are prevalent in the protein, reflecting the pH changes of ischemic stroke and the histological staging of brain tumors. In addition to the amide frequency shift (i.e., 3.5ppm), signals associated with tumor metabolism were also found at the amine frequency shift (about 3ppm), lipid frequency shift (about-3.5 ppm), and creatine frequency shift (about 2 ppm). Adequate analysis of CEST magnetic resonance images acquired at multiple frequency offsets is beneficial to further reflect signal changes of multiple metabolites associated with a lesion.

Various quantification methods are used to reflect the signals of multiple metabolites, including widely used Z-spectrum asymmetry or inverse Z-spectrum analysis, multi-cell Lorentzian fitting, and Lorentzian difference based methods, among others. However, Z-spectral asymmetry analysis cannot separately quantify the lipid frequency shift and its amide frequency shift signal, which is symmetric about the 0 point. The multi-cell lorentzian method requires acquisition of signals at more saturation frequency shifts to accurately fit the Z spectral peak and overestimate the signal of the metabolite. The Lorentz difference is not suitable for the condition of high saturation field intensity (more than or equal to 1 mu T) and the tissue with strong magnetic ring transfer effect. Especially at 3T clinical field strengths, the peaks of the Z spectrum are hardly visible, and its CEST signal analysis is more challenging for the above quantification method. Therefore, a new approach is needed for CEST magnetic resonance analysis that avoids some of the disadvantages of conventional quantification methods and still fully utilizes the information of all the frequency offsets collected to give metabolic-related features at each frequency offset.

Disclosure of Invention

The application provides a CEST magnetic resonance analysis method and device based on a random forest arrangement, which can robustly explain the contribution of a frequency shift signal related to metabolism to the distinguishing of lesion and normal tissues and the problem of accurate segmentation of a lesion area.

An embodiment of a first aspect of the present application provides a CEST magnetic resonance analysis method based on a random forest arrangement, including the following steps: acquiring a CEST magnetic resonance image to be tested and voxel labels in a lesion area and a non-lesion area; inputting the Z spectrum of each voxel in the CEST magnetic resonance image and the voxel label into a pre-trained permutation random forest model to obtain the importance of each saturation frequency deviation, and determining the contribution of all saturation frequency deviations in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency deviation.

Optionally, in an embodiment of the present application, after determining the contribution of all saturation frequency shifts in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency shifts, further comprising: denoising a tested CEST magnetic resonance image of an unknown lesion region; inputting the Z spectrum of the CEST magnetic resonance image before and after denoising and the voxel label into a pre-trained random forest model to classify the lesion at the voxel level to obtain a classification result of each voxel, reducing the classification result of each voxel into a segmentation result at the picture level according to a label generation sequence, filling a cavity in the segmentation result, and completing the lesion segmentation task of the tested CEST magnetic resonance image in an unknown lesion area.

Optionally, in an embodiment of the present application, before acquiring the voxel labels in the lesion region and the non-lesion region, the method further includes: acquiring a multi-contrast magnetic resonance image, and segmenting a lesion region in the multi-contrast magnetic resonance image through a segmentation algorithm to obtain a lesion region and a non-lesion region; and marking the voxels in the lesion area and the non-lesion area respectively to obtain the voxel label.

Optionally, in an embodiment of the present application, before inputting the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image into a pre-trained ranked random forest model, the method further includes: acquiring a training CEST magnetic resonance image; constructing a random forest model, inputting the Z spectrum of the training CEST magnetic resonance image and the voxel label into the constructed random forest model for training, and obtaining the optimal parameters enabling the random forest model to meet preset conditions through a grid search method; reconstructing a random forest model by using the optimal parameters to obtain an optimal random forest model, and determining a classification precision reference value of the optimal random forest model; and combining the optimal random forest model with a ranking importance algorithm to obtain the pre-trained ranking random forest model.

Optionally, in an embodiment of the application, the inputting the Z spectrum of each voxel in the CEST magnetic resonance image and the voxel label into a pre-trained permutation random forest model to obtain an importance of each saturation frequency offset includes: selecting Z spectrums containing m voxels in a parenchymal tissue in the CEST magnetic resonance image, respectively constructing a training data matrix with the matrix size of mxn and a training label matrix with the matrix size of mx1, circulating the CEST magnetic resonance image, judging whether the current voxel belongs to the parenchymal tissue, if the current voxel belongs to the parenchymal tissue, putting the Z spectrums corresponding to the current voxel into the training data matrix according to the sequence, and putting the voxel label in the training label matrix, wherein n is the total number of saturated frequency deviation; randomly disordering the numerical value of the first saturation frequency deviation to obtain a new training set matrix, inputting the new training set matrix into the optimal random forest model to obtain the current classification precision value of the optimal random forest model, and determining the importance of the first saturation frequency deviation according to the current classification precision value and the classification precision reference value; and circulating the training data matrix according to the saturated frequency offset to sequentially obtain the importance of each saturated frequency offset.

The embodiment of the second aspect of the present application provides a CEST magnetic resonance frequency feature extraction device based on a random forest arrangement, including: an acquisition module for acquiring a CEST magnetic resonance image to be tested and voxel labels in a lesion region and a non-lesion region; and the analysis module is used for inputting the Z spectrum of each voxel in the CEST magnetic resonance image and the voxel label into a pre-trained permutation random forest model to obtain the importance of each saturation frequency deviation, and determining the contribution of all saturation frequency deviations in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency deviation.

Optionally, in an embodiment of the present application, the method further includes: a segmentation module, configured to perform denoising on a CEST magnetic resonance image of an unknown lesion region after determining, according to importance of the saturation frequency offsets, contributions of all saturation frequency offsets in the CEST magnetic resonance image to lesion classification; inputting the Z spectrum of the CEST magnetic resonance image before and after denoising and the voxel label into a pre-trained random forest model to classify the lesion at the voxel level to obtain a classification result of each voxel, reducing the classification result of each voxel into a segmentation result at the picture level according to a label generation sequence, filling a cavity in the segmentation result, and completing the lesion segmentation task of the tested CEST magnetic resonance image in an unknown lesion area.

Optionally, in an embodiment of the present application, the method further includes: the preprocessing module is used for acquiring a multi-contrast magnetic resonance image before acquiring voxel labels in a lesion area and a non-lesion area, and segmenting the lesion area in the multi-contrast magnetic resonance image through a segmentation algorithm to obtain the lesion area and the non-lesion area; and marking the voxels in the lesion area and the non-lesion area respectively to obtain the voxel label.

Optionally, in an embodiment of the present application, the method further includes: a training module, configured to obtain a training CEST magnetic resonance image sum before inputting the Z spectrum of each voxel in the CEST magnetic resonance image and the voxel label into a pre-trained permutation random forest model; constructing a random forest model, inputting the Z spectrum of the training CEST magnetic resonance image and the voxel label into the constructed random forest model for training, and obtaining the optimal parameters enabling the random forest model to meet preset conditions through a grid search method; reconstructing a random forest model by using the optimal parameters to obtain an optimal random forest model, and determining a classification precision reference value of the optimal random forest model; and combining the optimal random forest model with a ranking importance algorithm to obtain the pre-trained ranking random forest model.

Optionally, in an embodiment of the present application, the analysis module is further configured to select a Z spectrum of m voxels in a parenchymal tissue in the CEST magnetic resonance image, respectively construct a training data matrix with a matrix size of mxn and a training label matrix with a matrix size of mx 1, circulate the CEST magnetic resonance image, determine whether a current voxel belongs to the parenchymal tissue, if the current voxel belongs to the parenchymal tissue, put the Z spectrum corresponding to the current voxel into the training data matrix according to a sequence, and put the voxel label into the training label matrix, where n is a total saturation frequency offset; randomly disordering the numerical value of the first saturation frequency deviation to obtain a new training set matrix, inputting the new training set matrix into the optimal random forest model to obtain the current classification precision value of the optimal random forest model, and determining the importance of the first saturation frequency deviation according to the current classification precision value and the classification precision reference value; and circulating the training data matrix according to the saturated frequency deviation to sequentially obtain the importance of each saturated frequency deviation.

An embodiment of a third aspect of the present application provides an electronic device, including: a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the method for CEST magnetic resonance analysis based on a rank random forest as described in the above embodiments.

In a fourth aspect of the present application, a computer-readable storage medium stores computer instructions for causing a computer to execute a method for aligned random forest based CEST magnetic resonance analysis according to the foregoing embodiment.

The CEST magnetic resonance analysis method and device based on the arranged random forests have the following beneficial effects:

1) different from the existing quantitative method, the method for analyzing and extracting the contribution of all frequency offsets of CEST to lesion classification is provided by combining a random forest model and a ranking importance algorithm, namely, the frequency importance characteristics of each frequency offset and metabolism can be simply and intuitively acquired.

2) The method and the device have the advantages that the random forest model is suitable for small data volume analysis, and the CEST magnetic resonance image analysis based on a single tested object and less in acquisition frequency deviation can be achieved.

3) Compared with other characteristic importance algorithms (such as the importance of the kini characteristic), the method for calculating the importance of the arranged characteristic is more robust, generates smaller deviation and has more accurate result.

Additional aspects and advantages of the present application will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the present application.

Drawings

The foregoing and/or additional aspects and advantages of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

fig. 1 is a flowchart of a CEST magnetic resonance analysis method based on a random forest arrangement according to an embodiment of the present application;

fig. 2 is a schematic structural framework diagram of a CEST magnetic resonance analysis method based on a arranged random forest according to an embodiment of the present application;

FIG. 3 is a schematic diagram of a ranking feature importance algorithm provided in accordance with an embodiment of the present application;

fig. 4 is a schematic structural framework diagram of another CEST magnetic resonance analysis method based on a arranged random forest according to an embodiment of the present application;

fig. 5 is a graph of CEST frequency significance of simulation data at a field strength of 7 tesla provided in accordance with an embodiment of the present application;

FIG. 6 is a graph of CEST frequency importance of simulation data at a field strength of 3 Tesla, provided in accordance with an embodiment of the present application;

FIG. 7 is a CEST frequency importance graph extracted by the arranged random forest method according to the embodiment of the application under different sample sizes and saturation field strengths;

FIG. 8 is a Z-spectrum and corresponding CEST frequency significance plot for a cerebral ischemic rat provided in accordance with an embodiment of the present application;

fig. 9 is a graph of CEST frequency significance for different lesion markers provided in accordance with an embodiment of the present application;

FIG. 10 is a graph comparing the importance of CEST frequency for patients with glioblastoma and metastases, according to an embodiment of the present invention;

figure 11 is an exemplary diagram of a CEST magnetic resonance analysis apparatus based on a ranked random forest according to an embodiment of the present application;

fig. 12 is a schematic diagram of an electronic device according to an embodiment of the application.

Detailed Description

Reference will now be made in detail to embodiments of the present application, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are exemplary and intended to be used for explaining the present application and should not be construed as limiting the present application.

Fig. 1 is a flowchart of a CEST magnetic resonance analysis method based on a random forest arrangement according to an embodiment of the present application.

As shown in fig. 1, the CEST magnetic resonance analysis method based on the arrangement of random forests comprises the following steps:

in step S101, a CEST magnetic resonance image of a subject and voxel labels within a diseased region and a non-diseased region are acquired.

For CEST magnetic resonance analysis, embodiments of the present application acquire a CEST magnetic resonance image of a subject and divide the CEST magnetic resonance image by an unsaturated reference image S ₀ Obtaining a normalized CEST magnetic resonance image S _sat 。

In the analysis method of the embodiment of the present application, each data analysis requires a labeled region, i.e., a voxel label in a diseased region and a non-diseased region, and the voxel label can be modified or reused. Embodiments of the application may acquire voxel labels using multi-contrast nuclear magnetic images.

As a method for acquiring a voxel label, in the embodiment of the present application, before acquiring a voxel label in a lesion region and a non-lesion region, a multi-contrast magnetic resonance image is acquired, and a segmentation algorithm is used to segment the lesion region in the multi-contrast magnetic resonance image to obtain the lesion region and the non-lesion region; and marking voxels in the lesion area and the non-lesion area respectively to obtain a voxel label.

Specifically, magnetic resonance with high lesion-background contrast, such as CEST quantitative images obtained by algorithms such as T2 weighting, T1 enhancement, lorentz difference or asymmetry analysis, and the like, are obtained. Selecting an image with the highest contrast ratio between the lesion and the background from other mode images and the quantization result image, segmenting the lesion area by using an automatic segmentation algorithm such as threshold segmentation, and enabling the label value of the voxel in the lesion area to be 1 and the label value of the voxel outside the lesion area to be 0 according to the segmented lesion area, thereby obtaining the voxel labels in the lesion area and the non-lesion area.

In step S102, the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image are input into a pre-trained permutation random forest model to obtain the importance of each saturation frequency offset, and the contribution of all saturation frequency offsets in the CEST magnetic resonance image to lesion classification is determined according to the importance of the saturation frequency offsets.

The importance of the saturation frequency offset can reflect the influence of the input frequency characteristics on the model prediction result.

Further, the importance of all frequency shifts is plotted as a histogram, giving a signal contribution plot of all saturation frequency shifts with respect to lesion classification.

Optionally, in an embodiment of the present application, before inputting the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image into the pre-trained ranked random forest model, the method further includes: acquiring a training CEST magnetic resonance image; constructing a random forest model, inputting a Z spectrum and a voxel label of a training CEST magnetic resonance image into the constructed random forest model for training, and obtaining an optimal parameter which enables the random forest model to meet a preset condition through a grid search method; reconstructing the random forest model by using the optimal parameters to obtain an optimal random forest model, and determining a classification precision reference value of the optimal random forest model; and combining the optimal random forest model with the permutation importance algorithm to obtain a pre-trained permutation random forest model.

CEST magnetic resonance imaging is used as a molecular imaging tool with wide application prospect, and can sensitively detect endogenous metabolic change. However, interpretation and analysis of CEST spectrum signals present certain challenges, as CEST spectra do not show sharp peaks like magnetic resonance spectra, especially the Z-spectrum peak is hardly visible at 3 tesla clinical field strengths. The embodiment of the application provides a CEST magnetic resonance analysis method based on a random forest arrangement to extract the importance of CEST frequency of lesions. This method does not require Z-spectrum fitting and can simply and directly account for the contribution of each saturation frequency offset to the lesion classification. As shown in fig. 2, lesion voxels are labeled on the multi-contrast magnetic resonance image, for example, as a graph of the quantification results obtained by a simple Z-spectral asymmetry analysis or a lorentz difference analysis. The voxel labels and corresponding Z spectra are then fed into a ranked random forest model to determine the contribution of each saturation frequency shift to the lesion classification. In addition, the analysis method of the embodiment of the application can be used for analyzing the CEST data of a single tested object.

Specifically, the Z spectrums of all voxels of a training CEST magnetic resonance image and corresponding labels of a multi-contrast magnetic resonance image are used as a training set, a random forest model is trained, and optimal parameters are obtained through a grid search method, so that the model classification accuracy is highest. And inputting the training data set and the trained random forest model into a permutation importance algorithm to obtain the importance of each saturated frequency offset. As shown in fig. 3, the method specifically includes the following steps:

step 1), selecting a Z spectrum containing m voxels in parenchymal tissue as training data. Respectively constructing a training data matrix with the size of mxn and a training label matrix with the size of mx1, circulating the whole image, judging whether the current voxel belongs to a parenchymal tissue, if the current voxel belongs to the parenchymal tissue, putting a Z spectrum of the current voxel into the training data matrix according to the sequence, and putting a generated corresponding label into the training label matrix, wherein n is the total number of the saturated frequency deviation.

Step 2), a random forest model is constructed, the number of trees in the random forest, the maximum depth of the trees, the minimum data volume required by splitting tree nodes and the value of a parameter to be optimized of the minimum data volume parameter required by determining leaf nodes are set, and R predicted by the random forest on a training set is obtained by using grid search ² The highest set of parameters is used as the optimal parameters. R ² For measuring the classification accuracy of the model, the formula is as follows:

wherein, y _j Is a label to be attached to the body,

is the classification result of the model. When R is ² When equal to 1, it means that the classification result is identical to the label, and R ² A decrease in value represents an increase in the error rate of the classification.

And 3) utilizing the optimal parameters to construct the random forest model again to obtain an optimal random forest model, and recording R corresponding to the model ² The value serves as a baseline value R2_ b for subsequent frequency importance calculations.

And 4) randomly disordering the numerical values of the first saturated frequency deviation (namely the first column of the matrix) to obtain a new training set matrix, wherein the size of the matrix is still m multiplied by n, inputting the training set matrix into an optimal random forest model to obtain the accuracy rate R2_1 of model classification, and taking R2_ b-R2_1 as the importance of the first saturated frequency deviation.

And 5) circulating the training set matrix according to the saturated frequency deviation (namely columns), and carrying out the operation of the step 4) to sequentially obtain the importance of each saturated frequency deviation, and explaining the contribution of each saturated frequency deviation to the lesion classification according to the frequency importance.

And obtaining a pre-trained permutation random forest model through the steps, analyzing the importance of each saturation frequency deviation by using the pre-trained permutation random forest model, and determining the contribution of all saturation frequency deviations in the CEST magnetic resonance image to lesion classification.

Optionally, in an embodiment of the present application, inputting the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image into a pre-trained permutation random forest model to obtain an importance of each saturation frequency offset, including: selecting a Z spectrum containing m voxels in a parenchymal tissue in a CEST magnetic resonance image, respectively constructing a training data matrix with the matrix size of mxn and a training label matrix with the matrix size of mx1, circulating the CEST magnetic resonance image, judging whether a current voxel belongs to the parenchymal tissue, if the current voxel belongs to the parenchymal tissue, putting the Z spectrum corresponding to the current voxel into the training data matrix according to the sequence, and putting a voxel label in the training label matrix, wherein n is the total number of saturated frequency shifts; randomly disordering the numerical value of the first saturated frequency deviation to obtain a new training set matrix, inputting the new training set matrix into the optimal random forest model to obtain the current classification precision value of the optimal random forest model, and determining the importance of the first saturated frequency deviation according to the current classification precision value and the classification precision reference value; and circulating the training data matrix according to the saturated frequency offset, and sequentially obtaining the importance of each saturated frequency offset.

Segmentation of the lesion region on the test set may be performed with sufficient magnetic resonance images to alleviate the over-fitting problem, as shown in figure 4.

Optionally, in an embodiment of the present application, after determining the contribution of each saturation frequency offset to the lesion classification according to the importance of the saturation frequency offset, the method further includes: denoising a tested CEST magnetic resonance image of an unknown lesion region; inputting the Z spectrum and the voxel label of the CEST magnetic resonance image before and after denoising treatment into a pre-trained random forest model to classify the lesion at the voxel level to obtain a classification result of each voxel, reducing the classification result of each voxel into a segmentation result at the picture level according to a label generation sequence, filling a hole in the segmentation result, and completing the lesion segmentation task of the CEST magnetic resonance image to be tested in an unknown lesion area.

Specifically, four algorithms of one-dimensional Z spectrum smoothing, image median filtering and two-dimensional and three-dimensional singular value decomposition are used for respectively carrying out noise removal preprocessing on a tested CEST magnetic resonance image of an unknown lesion area to obtain four noise-removed CEST magnetic resonance images.

And 2) constructing a random forest model according to the step 2), and training and optimally selecting parameters of the random forest model by using the Z spectrums and the corresponding labels in the parenchymal tissues of the CEST magnetic resonance image after four types of noise are removed and the original CEST magnetic resonance image to obtain the trained random forest model.

And (3) preprocessing the test set CEST magnetic resonance image of the unknown lesion area for removing noise, inputting the Z spectrum and the corresponding label in the parenchymal tissue of the CEST magnetic resonance image before and after the noise is removed into the trained random forest model to obtain a classification result of 5 voxel levels, and if one or more classification results deem that the voxel is a lesion, determining the final classification result of the voxel as the lesion.

And restoring the classification result of each voxel into a picture-level segmentation result according to the label generation sequence, filling a hole in the final image-level segmentation result by using a flooding filling algorithm, and completing the lesion segmentation task of the test set CEST magnetic resonance image.

By way of introduction, embodiments of the present application use a thresholding-based algorithm to label lesion regions on a multi-contrast magnetic resonance image, i.e., label each lesion voxel as 1 and normal voxel as 0. Then, using the saturated signal sequence for each voxel as a sample, the rank random forest method compares two different sets of voxels (lesion and normal voxels) using a rank importance algorithm to determine the importance of each saturation frequency and rank them. The acquired CEST data can be fully utilized to provide additional metabolic-related features for the diseased region. For example, it can be determined which frequency contributes most to the identification of ischemic lesions, different classes of tumors.

In summary, the ranked random forest method enables a simple and intuitive interpretation and analysis of the importance of all acquired saturation frequencies for lesion classification, which contributes to the clinical generalization of CEST analysis methods.

It should be noted that the method in the embodiment of the present application may be directly applied to post-processing software of a nuclear magnetic resonance instrument, so as to implement an online processing. Namely, when a doctor looks up a scanned image on a computer connected with nuclear magnetic equipment, the doctor can directly mark a lesion area, and a corresponding post-processing program can generate corresponding CEST frequency importance.

The CEST magnetic resonance analysis method based on the arranged random forest according to the present application is described in detail below with reference to the accompanying drawings and specific examples.

Generating a data set:

(1) generating simulated Z spectral data. Five different types of simulated Z spectra were obtained by 4-cell Bloch-McConnel equations, with different exchange rates and concentrations at-3.5 ppm or 3.5ppm, respectively. Wherein, the analog field intensity of the two simulated Z spectrums is 7 Tesla, and the other is 3 Tesla. The four pools comprise water, amide, a nuclear austenite effect and a semi-solid magnetization transfer pool. All ofThe default relaxation times and chemical exchange parameters for the simulations are shown in table 1, and the unique parameters for the different types of simulations are shown in table 2. In addition, B with a size of between + -0.7 ppm/+ -0.05 μ T was randomly added to the simulated Z spectra ₀ /B ₁ Offset, and white gaussian noise with mean 0 and standard deviation of 0.002.

TABLE 1 Default relaxation time and common chemical exchange parameters in all simulations

Note: NOE is an abbreviation for the nuclear austenite effect. f. of _s Is the solute concentration, k _SW Is the water-solute exchange rate, Δ is the frequency shift of the solute protons relative to the water protons, T ₁ Is the longitudinal relaxation time, T ₂ Is the transverse relaxation time.

TABLE 2 chemical exchange parameters unique to different types of simulated Z spectra

Note-indicates that the default value in Table 1, k, is used _AW Is the exchange rate between amide and water, k _NW Is the exchange rate between the nuclear austenite effect and water, f _A Denotes the amide concentration, f _N Indicating the concentration of the nuclear austenite effect.

(2) Brain MR data was acquired for an ischemic stroke rat model. Adult Sprague-Dawley rats (male, body weight 240 to 270 g) induced transient cerebral ischemia by 2 hours of mid-cerebral artery occlusion, followed by magnetic resonance imaging on a 7T magnetic resonance scanner (Bruker Biospec, germany) after reperfusion for 2 hours (5 mice) and 24 hours (4 mice). Single layer CEST imaging was performed using a fast acquisition and relaxation enhancement sequence with the following saturation parameters: the saturation time was 2.5s, the saturation field strength was 1. mu.T, and 50 frequency shifts of-10 ppm to 10ppm were collected. Other imaging parameters were as follows: the fast acquisition and relaxation enhancement factor is 32, the repetition Time (TR) is 5s, the echo time (echo t)ime, TE) 4ms, layer thickness 1mm, matrix size 96 × 64, field of view 34 × 28mm ² . Correction of B using water saturation offset reference method ₀ Field inhomogeneity. The diffusion-weighted imaging parameters are as follows: b value of 2000s/mm ² TR 2.5s and TE 55 ms.

(3) Magnetic resonance image data were collected for enrolled patients with brain tumors, namely two histologically confirmed patients with glioblastoma (1 female 59 and 1 male 63) and two patients with brain metastases (rectal cancer and lung cancer metastases, 1 female 66 and 1 male 68). The experiment was approved by the institutional review board and written informed consent was obtained for each patient. Before operation, a 3T nuclear magnetic scanner (Ingenia, Philips Healthcare) performs magnetic resonance imaging, which has a 32-channel phased array coil, a three-dimensional amide proton transfer CEST sequence as a scanning sequence, and a fast spin echo sequence as a readout sequence.

Processing the simulation data:

for the simulated data at 7 tesla, the simulated Z spectrum with the higher exchange rate at 3.5ppm (or-3.5 ppm) is taken as a positive sample, labeled 1; the Z spectrum of the lower exchange rate is taken as a negative sample, labeled 0. For the simulated data at 3 Tesla, the simulated Z spectrum with higher exchange rate and concentration at 3.5ppm and lower concentration at-3.5 ppm is taken as a positive sample, labeled 1.

And inputting each type of Z spectrum and the corresponding label into a random forest model, and obtaining optimal model parameters and a trained optimal random forest model by using a grid search method.

And sending the optimal model and the training data set into a ranking importance algorithm to obtain the importance of all frequency offsets of the CEST data.

The rat brain ischemia data was processed as follows:

and automatically marking the lesion area by using the CEST magnetic resonance image and the CEST magnetic resonance quantification result image. The mean lorentz delta dispersion result plot and the mean normalized CEST magnetic resonance plot from 3.5ppm to 3.75ppm were used to label lesions in the ischemic rat brain. The method comprises the following specific steps: and respectively generating two candidate lesion areas for the same mouse by using a threshold segmentation algorithm on the two average images. The optimal threshold is the threshold that maximizes the difference between the lesion and the background voxel average, and can be automatically obtained by searching between all voxel values, with a search step size of 0.01. The two candidate lesion regions are intersected to generate a more accurate lesion region. A flood fill algorithm is used to fill the holes in the lesion area.

And training a random forest model to obtain the importance of CEST frequency of the cerebral data of the ischemic rat. To assess the impact of different lesion labeling methods on the PRF method, CEST frequency significance obtained 2 hours after rat brain ischemia was compared when the lesion labels were different, including the above-described automatically generated labels using lorentz difference and saturation images, and the lesions were artificially labeled according to T2 weighted images and diffusion weighted images.

If the training set data volume is sufficient to mitigate overfitting, the method of the embodiment of the application can also perform lesion segmentation on a CEST magnetic resonance image of an unknown lesion region. Taking rat brain data as an example, firstly, respectively performing noise-removing pretreatment on a CEST magnetic resonance image sequence by using four algorithms of one-dimensional Z spectral smoothing, image median filtering and two-dimensional and three-dimensional singular value decomposition to obtain four noise-removed CEST magnetic resonance images as amplification data; constructing a random forest model, and pre-training the model by using the amplification data and the original data to relieve overfitting; and performing the same denoising treatment on the test data, predicting the test data by using a pre-trained random forest, and voting by using five prediction results to obtain a final classification result of the voxel level. And reducing the voxel classification result into a picture-level segmentation result in sequence, filling the holes in the picture by using a flooding filling algorithm, and completing the lesion segmentation of the CEST magnetic resonance image of the test set.

Tumor patient data were processed:

and (3) marking a lesion area by taking each anatomical layer of the CEST magnetic resonance image as an analysis object, and training a random forest model to obtain the importance of the CEST frequency. Lesion areas are labeled as follows: the APT weighted picture (APTw) is calculated as follows:

a lesion label is then generated using a semi-automatic segmentation algorithm. Firstly, manually drawing a candidate region in a lesion hemibrain region of an APTw image to remove high signals related to cerebrospinal fluid and artifacts; secondly, a threshold segmentation algorithm is adopted to segment the candidate region to generate a tumor label. Wherein the threshold (glioblastoma: 0.02, metastatic brain tumor: 0.01) is the threshold that maximizes the contrast-to-noise ratio (CNR) difference between lesion and background APTw values, which can be automatically obtained by searching between all values, with a search step size of 0.01. The CNR is calculated as:

when an analysis object is selected, if the number of voxels in the lesion region marked in a certain layer of data is less than 200, the voxels are excluded from subsequent data analysis.

The processed data were subjected to a simulation test, and the results are as follows.

Accuracy of calculating CEST frequency importance of the arranged random forest:

the simulation results are shown in FIG. 5, and the simulation results (B) are shown at a field strength of 7 Tesla ₁ 1 μ T): in FIG. 5, (A), (B), (C) are two groups having different exchange rates at 3.5ppm in the Z spectrum, and (D), (E) and (F) are two groups having different exchange rates at-3.5 ppm. As shown in FIGS. 5 (A) and (B), the two Z spectra showed the greatest difference at 3.5ppm (B) ₀ ＝7T，B ₁ 1 μ T), the permutation random forest correctly revealed that the CEST frequency at 3.5ppm was the most important (as shown in fig. 5 (C). Also, fig. 5 (D), (E), (F) show scatter plots, mean Z spectra and CEST frequency importance for two additional sets of Z spectra, with the greatest difference between the two sets of Z spectra being at-3.5 ppm. In the simulation results at 3 Tesla field intensity (B) ₀ ＝3T，B ₁ 2 μ T), with B ₀ Reduction of (A) and (B) ₁ Has been difficult to distinguish from the scatter plot which of them is increasingThe frequency has the greatest contribution to the classification of the two sets of Z spectra. Also, as shown in fig. 6 (a) and (B), the CEST peak at 3.5ppm becomes "invisible", but with the aligned random forest method, the frequency shift with the greatest difference in the two sets of Z-spectra (as in fig. 6 (C)), i.e., the frequency shift with the highest importance (3.5ppm), can still be determined. In addition to the two sets of Z spectra having different exchange rates at 3.5ppm, (D), (E), (F) of FIG. 6 show the scattergram, the average Z spectrum, and the CEST frequency importance of the two sets of Z spectra having different concentrations at 3.5ppm, respectively, and (G), (H), (I) of FIG. 6 show the scattergram, the average Z spectrum, and the CEST frequency importance of the two sets of Z spectra having different concentrations at-3.5 ppm.

Simulation experiments have shown that despite a high B at a field strength of 3 Tesla ₁ (2 μ T), CEST peaks become invisible, but the arrangement of random forests still allows CEST frequency importance to be correctly analyzed, determining important frequency offsets that can distinguish the two sets of Z spectra.

As shown in FIG. 7, to test the universality of the arranged random forest under different conditions, the arranged random forest was verified at different sample sizes and different B ₁ The following properties. The other simulation parameters were the same as (c) in Table 2, B ₁ Values were randomly selected from the range of 0.5. mu.T to 2. mu.T. As shown in fig. 7 (a), although the sample size varies over a wide range (198 to 10010), the importance of the CEST frequency derived from the aligned random forest method always correctly identifies the important frequency offset that distinguishes the two sets of Z spectra, i.e., 3.5 ppm. Also, when the sample size is as low as about 200, the data size is sufficient to analyze the CEST signal using a spread random forest. FIG. 7 (B) shows that there are 12 differences B ₁ Of the experiments of (3.5ppm) CEST frequency is of highest importance, which indicates that the method of the present application example is in different B ₁ (0.5 to 2 μ T). With B ₁ The highest CEST frequency importance drops from about 1.8 to about 1.2.

To further evaluate the rank random forest approach, in vivo magnetic resonance image data of rats after transient cerebral ischemia induced by MCAO was used, as shown in figure 8. For data obtained 2 hours after transient cerebral ischemia, fig. 8 (a) shows a Z-spectrum scatter plot of voxels labeled lesion and voxels within the mirrored contralateral region of interest. The importance of the CEST frequency for the aligned random forest extraction clearly indicates a peak at 3.6ppm, the center frequency of the amide proton ((B) of fig. 8). The above conclusion is confirmed to some extent by the mean Z spectra within the 2 × 2 lesion and contralateral area (fig. 8 (C) and (D)). Similarly, fig. 8 (E), (F) and (H) show scatter plots, CEST frequency importance and mean Z spectra of the regions of interest of the rats 24 hours after transient cerebral ischemia. Fig. 8 (F) shows that CEST frequency importance increases at multiple frequency shifts, such as amide, water and lipid proton frequency shifts, after 24 hours of ischemia.

Figure 9 shows the CEST frequency importance obtained for two different markers compared on CEST magnetic resonance data 2 hours after rat brain ischemia, including markers automatically generated using lorentz difference and saturation images, and artificial markers weighted according to T2 and diffusion weighted images, to assess the effect of different lesion labeling methods on the rank random forest method. Fig. 9 (a1) - (a4), (B1) - (B4) show the above two different markers on a representative rat image. Since the principles of different magnetic resonance imaging methods are different, the labeled lesion regions are also different, and fig. 9 (a5) and (B5) show that CEST frequency importance distributions corresponding to the two labels are also slightly different. But the most important frequencies extracted on different markers are all the same, demonstrating that the rank random forest method can account for the importance of each saturated frequency offset signal after roughly delineating the lesion area.

The statistical results of the importance of CEST frequency for brain tumor patients are shown in fig. 10. And (3) obtaining the CEST frequency importance of 9 glioblastoma multiforme objects and 10 metastatic brain tumor objects by taking an anatomical layer image as an analysis object, and performing statistical analysis on all frequency importance through a two-tailed unpaired student t test. The results of the statistical analysis showed that glioblastoma and metastatic brain tumors were in the amide (3.5ppm) (p)<0.01) and amine frequency shift (2.7ppm) showed significant difference (p)<0.05) without statistically significant difference (p) in the importance of the frequency shift associated with the nuclear austenite effect (-3.5ppm) and the semi-solid magnetic ring transfer (4.3ppm)>0.05). It can be seen that in the clinical 3 Tesla, high B of 2. mu.T ₁ In the following, the method of the application can effectively increase the information about the contribution interpretation of the lesion signal.

According to the CEST magnetic resonance analysis method based on the arranged random forest, disclosed by the embodiment of the application, voxels of a lesion area are marked by using other modality magnetic resonance images, a classification task of lesions is learned by using a random forest model through a Z spectrum and a label of each voxel, a classification mechanism of the model is further explained by using an arrangement importance algorithm, and frequency importance is generated to determine the contribution of a CEST frequency offset signal to lesion classification. If the training data volume is sufficient to alleviate the overfitting problem, the random forest model can accurately segment the lesion area of the test set. The signal interpretation method can robustly interpret the contribution of the CEST frequency offset signal under various acquisition conditions to the lesion classification, and further simply and intuitively provide more disease-related metabolic information for the CEST clinical imaging sequence.

Next, a CEST magnetic resonance analysis apparatus based on a ranked random forest according to an embodiment of the present application is described with reference to the drawings.

Fig. 11 is an exemplary diagram of a CEST magnetic resonance analysis apparatus based on a ranked random forest according to an embodiment of the present application.

As shown in fig. 11, the CEST magnetic resonance analysis apparatus 10 based on the arranged random forest includes: an acquisition module 100 and an analysis module 200.

Wherein the acquiring module 100 is configured to acquire a CEST magnetic resonance image of a subject and voxel labels in a diseased region and a non-diseased region. And the analysis module 200 is configured to input the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image into a pre-trained permutation random forest model, obtain the importance of each saturation frequency offset in the Z spectrum of each voxel, and determine the contribution of all saturation frequency offsets in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency offsets.

Optionally, in an embodiment of the present application, the arrangement random forest based CEST magnetic resonance analysis apparatus 10 further includes: the segmentation module is used for de-noising the tested CEST magnetic resonance image of the unknown lesion area after determining the contribution of all saturation frequency shifts in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency shifts; inputting the Z spectrum and the voxel label of the CEST magnetic resonance image before and after denoising treatment into a pre-trained random forest model to classify the lesion at the voxel level to obtain a classification result of each voxel, reducing the classification result of each voxel into a segmentation result at the picture level according to the label generation sequence, filling the holes in the segmentation result, and completing the lesion segmentation task of the tested CEST magnetic resonance image in an unknown lesion area.

Optionally, in an embodiment of the present application, the arrangement random forest based CEST magnetic resonance analysis apparatus 10 further includes: the preprocessing module is used for acquiring a multi-contrast magnetic resonance image before acquiring voxel labels in a lesion area and a non-lesion area, and segmenting the lesion area in the multi-contrast magnetic resonance image through a segmentation algorithm to obtain the lesion area and the non-lesion area; and marking the voxels in the lesion area and the non-lesion area respectively to obtain a voxel label.

Optionally, in an embodiment of the present application, the arrangement random forest based CEST magnetic resonance analysis apparatus 10 further includes: the training module is used for acquiring a training CEST magnetic resonance image before inputting the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image into a pre-trained permutation random forest model; constructing a random forest model, inputting a Z spectrum and a voxel label of a training CEST magnetic resonance image into the constructed random forest model for training, and obtaining an optimal parameter which enables the random forest model to meet a preset condition through a grid search method; reconstructing the random forest model by using the optimal parameters to obtain an optimal random forest model, and determining a classification precision reference value of the optimal random forest model; and combining the optimal random forest model with the permutation importance algorithm to obtain a pre-trained permutation random forest model.

Optionally, in an embodiment of the present application, the analysis module 200 is further configured to select a Z spectrum of m voxels in a parenchymal tissue in a CEST magnetic resonance image, respectively construct a training data matrix with a matrix size of mxn and a training label matrix with a matrix size of mx 1, circulate the CEST magnetic resonance image, determine whether a current voxel belongs to the parenchymal tissue, if the current voxel belongs to the parenchymal tissue, put the Z spectrum corresponding to the current voxel into the training data matrix according to a sequence, put a voxel label in the training label matrix, where n is a total saturation frequency offset; randomly disordering the numerical value of the first saturated frequency deviation to obtain a new training set matrix, inputting the new training set matrix into the optimal random forest model to obtain the current classification precision value of the optimal random forest model, and determining the importance of the first saturated frequency deviation according to the current classification precision value and the classification precision reference value; and circulating the training data matrix according to the saturated frequency offset, and sequentially obtaining the importance of each saturated frequency offset.

It should be noted that the foregoing explanation on the embodiment of the CEST magnetic resonance analysis method based on the arranged random forest is also applicable to the CEST magnetic resonance analysis device based on the arranged random forest of this embodiment, and details are not repeated here.

According to the CEST magnetic resonance analysis device based on the arranged random forest, disclosed by the embodiment of the application, voxels of a lesion area are marked by using other modality magnetic resonance images, a classification task of lesions is learned by using a random forest model through a Z spectrum and a label of each voxel, a classification mechanism of the model is further explained by using an arrangement importance algorithm, and frequency importance is generated to determine the contribution of a CEST frequency offset signal to lesion classification. If the training data volume is sufficient to alleviate the overfitting problem, the random forest model can accurately segment the lesion area of the test set. The signal interpretation method can robustly interpret the contribution of the CEST frequency offset signal under various acquisition conditions to the lesion classification, and further simply and intuitively provide more disease-related metabolic information for the CEST clinical imaging sequence.

Fig. 12 is a schematic structural diagram of an electronic device according to an embodiment of the present application. The electronic device may include:

a memory 1201, a processor 1202, and a computer program stored on the memory 1201 and executable on the processor 1202.

The processor 1202 when executing the program implements the arranged random forest based CEST magnetic resonance analysis method provided in the above embodiments.

Further, the electronic device further includes:

a communication interface 1203 for communication between the memory 1201 and the processor 1202.

A memory 1201 for storing computer programs executable on the processor 1202.

The memory 1201 may comprise high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as at least one disk memory.

If the memory 1201, the processor 1202 and the communication interface 1203 are implemented independently, the communication interface 1203, the memory 1201 and the processor 1202 may be connected to each other through a bus and perform communication with each other. The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended ISA (EISA) bus, or the like. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is shown in FIG. 12, but this is not intended to represent only one bus or type of bus.

Alternatively, in a specific implementation, if the memory 1201, the processor 1202 and the communication interface 1203 are integrated on one chip, the memory 1201, the processor 1202 and the communication interface 1203 may complete communication with each other through an internal interface.

Processor 1202 may be a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or one or more Integrated circuits configured to implement embodiments of the present Application.

The present embodiment also provides a computer-readable storage medium having stored thereon a computer program which, when being executed by a processor, implements the above arranged random forest based CEST magnetic resonance analysis method.

In the description herein, reference to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or N embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the feature. In the description of the present application, "N" means at least two, e.g., two, three, etc., unless specifically limited otherwise.

Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more N executable instructions for implementing steps of a custom logic function or process, and alternate implementations are included within the scope of the preferred embodiment of the present application in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of implementing the embodiments of the present application.

The logic and/or steps represented in the flowcharts or otherwise described herein, e.g., an ordered listing of executable instructions that can be considered to implement logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or N wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.

It should be understood that portions of the present application may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods may be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.

It will be understood by those skilled in the art that all or part of the steps carried out in the method of implementing the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, and the program, when executed, includes one or a combination of the steps of the method embodiments.

In addition, functional units in the embodiments of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a stand-alone product, may also be stored in a computer readable storage medium.

The storage medium mentioned above may be a read-only memory, a magnetic or optical disk, etc. Although embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application, and that variations, modifications, substitutions and alterations may be made to the above embodiments by those of ordinary skill in the art within the scope of the present application.

Claims

1. A CEST magnetic resonance analysis method based on a random forest arrangement is characterized by comprising the following steps:

acquiring a CEST magnetic resonance image to be tested and voxel labels in a lesion area and a non-lesion area;

inputting the Z spectrum of each voxel in the CEST magnetic resonance image and the voxel label into a pre-trained permutation random forest model to obtain the importance of each saturation frequency deviation, and determining the contribution of all saturation frequency deviations in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency deviation.

2. The method of claim 1, further comprising, after determining the contribution of all saturation frequency shifts in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency shift:

denoising a tested CEST magnetic resonance image of an unknown lesion region;

inputting the Z spectrum of the CEST magnetic resonance image before and after denoising and the voxel label into a pre-trained random forest model to classify the lesion at the voxel level to obtain a classification result of each voxel, reducing the classification result of each voxel into a segmentation result at the picture level according to a label generation sequence, filling a cavity in the segmentation result, and completing the lesion segmentation task of the tested CEST magnetic resonance image in an unknown lesion area.

3. The method of claim 1, further comprising, prior to obtaining voxel labels within the diseased region and the non-diseased region:

acquiring a multi-contrast magnetic resonance image, and segmenting a lesion region in the multi-contrast magnetic resonance image through a segmentation algorithm to obtain a lesion region and a non-lesion region;

and marking the voxels in the lesion area and the non-lesion area respectively to obtain the voxel label.

4. The method of claim 1, further comprising, prior to inputting the Z-spectrum and the voxel label for each voxel in the CEST magnetic resonance image into a pre-trained ranked random forest model:

acquiring a training CEST magnetic resonance image;

constructing a random forest model, inputting the Z spectrum of the training CEST magnetic resonance image and the voxel label into the constructed random forest model for training, and obtaining the optimal parameters enabling the random forest model to meet preset conditions through a grid search method;

reconstructing a random forest model by using the optimal parameters to obtain an optimal random forest model, and determining a classification precision reference value of the optimal random forest model;

and combining the optimal random forest model with a ranking importance algorithm to obtain the pre-trained ranking random forest model.

5. The method of claim 4, wherein the inputting the Z spectrum and the voxel label of each voxel in the CEST magnetic resonance image into a pre-trained ranked random forest model, resulting in the importance of each saturation frequency shift, comprises:

selecting Z spectrums of m voxels in a parenchymal tissue contained in the CEST magnetic resonance image, respectively constructing a training data matrix with the matrix size of mxn and a training label matrix with the matrix size of mx1, circulating the CEST magnetic resonance image, judging whether the current voxel belongs to the parenchymal tissue, if the current voxel belongs to the parenchymal tissue, putting the Z spectrums corresponding to the current voxel into the training data matrix according to the sequence, and putting the voxel labels into the training label matrix, wherein n is the total saturated frequency deviation;

randomly disordering the numerical value of the first saturated frequency deviation to obtain a new training set matrix, inputting the new training set matrix into the optimal random forest model to obtain the current classification precision value of the optimal random forest model, and determining the importance of the first saturated frequency deviation according to the current classification precision value and the classification precision reference value;

and circulating the training data matrix according to the saturated frequency offset to sequentially obtain the importance of each saturated frequency offset.

6. A CEST magnetic resonance analysis device based on a random forest arrangement is characterized by comprising:

an acquisition module for acquiring a CEST magnetic resonance image to be tested and voxel labels in a lesion region and a non-lesion region;

and the analysis module is used for inputting the Z spectrum of each voxel in the CEST magnetic resonance image and the voxel label into a pre-trained permutation random forest model to obtain the importance of each saturation frequency deviation, and determining the contribution of all saturation frequency deviations in the CEST magnetic resonance image to lesion classification according to the importance of the saturation frequency deviation.

7. The apparatus of claim 6, further comprising:

a segmentation module, configured to perform denoising processing on a CEST magnetic resonance image of an unknown lesion region after determining, according to importance of the saturation frequency offsets, contribution of all saturation frequency offsets in the CEST magnetic resonance image to lesion classification; inputting the Z spectrum of the CEST magnetic resonance image before and after denoising and the voxel label into a pre-trained random forest model to classify the lesion at the voxel level to obtain a classification result of each voxel, reducing the classification result of each voxel into a segmentation result at the picture level according to a label generation sequence, filling a hole in the segmentation result, and completing the lesion segmentation task of the tested CEST magnetic resonance image in an unknown lesion area.

8. The apparatus of claim 6, further comprising:

the preprocessing module is used for acquiring a multi-contrast magnetic resonance image before acquiring voxel labels in a lesion area and a non-lesion area, and segmenting the lesion area in the multi-contrast magnetic resonance image through a segmentation algorithm to obtain the lesion area and the non-lesion area; and marking the voxels in the lesion area and the non-lesion area respectively to obtain the voxel label.

9. The apparatus of claim 6, further comprising:

a training module, configured to obtain a training CEST magnetic resonance image before inputting the Z spectrum of each voxel in the CEST magnetic resonance image and the voxel label into a pre-trained permutation random forest model; constructing a random forest model, inputting the Z spectrum of the training CEST magnetic resonance image and the voxel label into the constructed random forest model for training, and obtaining the optimal parameters enabling the random forest model to meet preset conditions through a grid search method; reconstructing a random forest model by using the optimal parameters to obtain an optimal random forest model, and determining a classification precision reference value of the optimal random forest model; and combining the optimal random forest model with a ranking importance algorithm to obtain the pre-trained ranking random forest model.

10. The apparatus of claim 9, wherein the analysis module is further configured to,

11. An electronic device, comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement a method of arranged random forest based CEST magnetic resonance analysis as claimed in any one of claims 1 to 5.

12. A computer-readable storage medium, having a computer program stored thereon, the program being executable by a processor for implementing a method for rank random forest based CEST magnetic resonance analysis as claimed in any one of claims 1 to 5.