White matter high-signal focus progress prediction model construction method and application

Abstract

The first aspect of the invention is to disclose a method for constructing a white matter high-signal focus progress prediction model. The second aspect of the invention is to disclose an application method of the white matter high-signal focus progress prediction model constructed by the method. According to the method, clinical characteristics (such as age, BMI and blood sugar) and image indexes (such as FA, MO, ICVF and ISOVF) are fused, so that the limitation of single-modal data prediction is solved, and the accuracy and comprehensiveness of WMH progress prediction are remarkably improved; the prediction algorithm is packaged into executable software (. Exe), the Windows platform off-line operation is supported, the defect that the prior art is lack of mature software tools is overcome, and clinical popularization and application are facilitated; the model can quantify the contribution degree of each feature to a prediction result, provides an intuitive decision basis for doctors, and improves the practical value of the model in clinical practice.

Description

Technical Field

[0001] This invention relates to a method for constructing a white matter hyperintensity lesion (WMH) progression prediction model based on machine learning and integrating multi-dimensional neuroimaging features and clinical indicators, as well as the application of the WMH progression prediction model, which is used to assist in individualized early risk assessment and clinical decision support for cerebral small vessel disease, and belongs to the field of medical artificial intelligence technology. Background Technology

[0002] Small vessel disease of the brain (CSVD) is a clinical, pathological, and imaging syndrome that affects small blood vessels such as arterioles, venules, arterioles, venules, and capillaries due to various causes and mechanisms. It is a major cause of vascular dementia. The earliest and most obvious manifestation of CSVD on magnetic resonance imaging (MRI) is vascular myxomatosis (WMH). The clinical impact of WMH is complex, including cognitive impairment, stroke recurrence, dementia, and an increased risk of death. Early prediction of WMH progression is beneficial for providing a scientific basis for early clinical intervention in vascular cognitive decline.

[0003] Currently, the prediction of WMH progression mainly relies on single-modality data analysis methods, such as using only MRI imaging features or clinical indicators (e.g., age, history of hypertension) to construct predictive models. These methods have significant limitations: 1. The data dimension is too narrow (images or clinical indicators), which leads to insufficient prediction accuracy of the model and fails to fully reflect the complex pathological mechanism of WMH, especially the interaction between metabolic abnormalities and microstructural damage. 2. Traditional statistical methods have limited ability to capture nonlinear relationships and complex interactions, making it difficult to accurately model the dynamic process of WMH progression. Furthermore, they rely on manual feature selection, which can easily miss key predictive factors. 3. Existing machine learning models have shortcomings in multimodal data fusion and lack systematic optimization strategies for integrating imaging features, biochemical indicators, and clinical parameters; 4. Existing prediction models suffer from poor interpretability, making it difficult to quantify the contribution of each feature to the prediction results, which is not conducive to clinical decision support. 5. Existing prediction models suffer from overfitting with small samples. Under small sample conditions, the risk of overfitting is high, the prediction performance is unstable, and the model's generalization ability is affected.

[0004] The aforementioned shortcomings primarily stem from the fact that existing WMH progression prediction schemes do not systematically integrate multi-dimensional data from imaging, biochemistry, and clinical aspects, and lack automated feature selection and interpretable analysis frameworks. Furthermore, the prediction models in existing technologies lack mature software available for download and use, hindering their clinical application. Summary of the Invention

[0005] The technical problem that this invention aims to solve is that existing WMH progression prediction schemes do not systematically integrate multi-dimensional data such as imaging, biochemistry, and clinical data, and lack automated feature selection and interpretable analysis frameworks.

[0006] To address the aforementioned technical problems, the first aspect of this invention discloses a method for constructing a predictive model for the progression of high-signal lesions in white matter, characterized by comprising the following steps: Step 1: Obtain T1WI, FLAIR, SWI, and DTI sequence data from multiple subjects; Step 2: Preprocess the T1WI, FLAIR, SWI, and DTI sequence data, including: Preprocessing of DTI sequence data includes: After distortion correction, brain extraction, and registration, the data is input into two complementary analysis pipelines, which report a series of dMRI-derived parameters in different fiber regions. One analysis pipeline is based on fiber skeleton processing, while the other is based on richer intravoxel fiber structure modeling, followed by probabilistic fiber tracing analysis. dMRI data is input into neurite directional dispersion and density imaging modeling to generate voxel-level microstructure parameters; DTI fitting is performed to obtain fractional anisotropy parameter maps. The fractional anisotropy parameter maps are then input into a fiber bundle-based spatial statistical method. The fractional anisotropy parameter maps are aligned to the white matter fiber skeleton in standard space. The resulting standard space distortion is then applied to all other DTI / NODDI parameter maps. Preprocessing for T1WI, FLAIR, and SWI sequence data includes: Image analysis tools were used to process T1WI, FLAIR and SWI sequence data, and CSVD imaging features, including white matter hyperintensity, lacunar infarcts, perivascular spaces, recent subcortical small infarcts and cerebral microbleeds, were automatically segmented and quantitatively calculated. Step 3: Obtain the imaging phenotypic data and clinical phenotypic data of the subjects. The clinical phenotypic data includes the subjects' baseline data and blood index data. The clinical phenotypic data includes the whole brain average values ​​of FA, MD, OD, MO, ICVF and ISOVF obtained in Step 2. Step 4: After processing all the variable data obtained in Step 3, complete the feature selection. Feature selection includes the following steps: One-way logistic regression was used to initially screen out candidate variables that showed significant differences between the two groups; The candidate variables are input into the model, and stepwise multivariate logistic regression is used to remove them from the full model. The best-fit model is then constructed using the minimum AIC value. Step 5: Divide the feature sample data of the subjects obtained in Step 4 into the WMH growth group and the WMH decrease group, and construct training set and validation set. Among them, the feature sample data of subjects whose white matter high signal foci volume increased during the follow-up period were assigned to the WMH growth group and assigned a positive label, and the feature sample data of subjects whose white matter high signal foci volume decreased during the follow-up period were assigned to the WMH decrease group and assigned a negative label. Step 6: Construct a binary classification machine learning prediction model. The independent variables are the features obtained in Step 4, and the prediction variables are the binary values ​​representing the growth or shrinkage of white matter hypersignal lesions. The binary classification machine learning prediction model is trained using the training set constructed in Step 5, and the trained binary classification machine learning prediction model is validated using the validation set constructed in Step 5 to obtain the white matter hypersignal lesion progression prediction model.

[0007] Preferably, in step 1, subjects with a cerebral small vessel disease burden score greater than or equal to a set threshold are excluded, as are subjects with Alzheimer's disease (AD), brain tumors, and mental illnesses.

[0008] Preferably, in step 2, the dMRI-derived indices include indices based on diffusion tensor modeling and indices based on microstructure model fitting.

[0009] Preferably, in step 2, the voxel-level microstructure parameters include an index of intracellular volume fraction white matter neurite density, isotropic volume fraction, and directional dispersion.

[0010] Preferably, in step 2, the expansion tensor mode parameter map and the average diffusion rate parameter map are also obtained by DTI fitting.

[0011] Preferably, in step 3, the baseline data includes age, gender, education level, smoking status, frequency of alcohol consumption, history of diabetes, history of chronic ischemic heart disease, and body mass index; the blood indicator data includes C-reactive protein, total cholesterol, homocysteine, white blood cell count, red blood cell count, platelet count, vitamin D, fasting blood glucose, eosinophils, basophils, high-density lipoprotein, low-density lipoprotein, hemoglobin count, lymphocyte count, neutrophil count, and glycated hemoglobin.

[0012] Preferably, in step 6, during the model training phase, grid search cross-validation is used to accelerate and optimize the parameter configuration of different models, and then the Logistic Regression model is selected as the final white matter high signal lesion progression prediction model.

[0013] The second aspect of the present invention discloses an application method for a white matter high signal lesion progression prediction model constructed by the above method, characterized in that the white matter high signal lesion progression prediction model constructed by the above method is packaged into executable software.

[0014] Preferably, when the executable software runs on an electronic device, it provides a user interface for inputting Age, BMI, FA, MO, ICVF, ISOVF, Glucose, and Cystatin C. The white matter hypersignal lesion progression prediction model displays the prediction results of the growth or shrinkage of the white matter hypersignal lesion on the same user interface based on the user's input.

[0015] Preferably, the executable software is distributed to users using the following method: When a user installs and uses the executable software for the first time, the background generates a unique machine code based on the hardware of the user's electronic device and presents it to the user. The user sends the machine code to the software owner; After the software owner enters the machine code and the usage period, an authentication code is generated. After the authentication code is returned to the user, the user completes software authentication by entering the authentication code into the backend. The executable software is published to users via a URL link generated by a server deployed in the cloud.

[0016] Compared with existing technical solutions, the present invention has the following beneficial effects: 1. Multimodal data fusion improves prediction accuracy: This invention overcomes the limitations of single-modality data prediction by integrating clinical features (such as age, BMI, blood glucose, etc.) and imaging indicators (such as FA, MO, ICVF, ISOVF), and significantly improves the accuracy and comprehensiveness of WMH progression prediction. 2. Software encapsulation enhances clinical usability: The prediction algorithm is packaged into executable software (.exe), which supports offline operation on the Windows platform, thus solving the problem of the lack of mature software tools in existing technologies and facilitating clinical promotion and application. 3. Explainability supports clinical decision-making: The model can quantify the contribution of each feature to the prediction results, providing doctors with intuitive decision-making basis and enhancing the practical value of the model in clinical practice. Attached Figure Description

[0017] Figure 1 This illustrates the process of filtering data within the UKB system; Figure 2 The model of an embodiment of the present invention is illustrated; Figure 3 This illustrates the performance of the machine learning model constructed in the embodiments of the present invention; Figure 4 This illustrates the user interface. Figure 5 The authorization interface is shown. Detailed Implementation

[0018] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0019] The first aspect of this invention is to address the limitation of single-modal data in existing WMH prediction technologies. It discloses a method for constructing a prediction model for the progression of white matter hyperintensity lesions. This method constructs a high-precision prediction model by integrating clinical features (age, BMI, blood glucose, etc.) with imaging indicators (FA, MO, etc.). Specifically, it includes the following steps.

[0020] Step 1: Sample Data Collection Participants with complete brain imaging data from 2014 and 2019 in the UKB system were selected, mainly including T1WI (T1-weighted Imaging), FLAIR (Fluid Attenuated Inversion Recovery), SWI (Susceptibility Weighted Imaging), and DTI sequences, totaling 4570 cases. This invention acquired the participants' baseline raw brain imaging data and used United Imaging AI software to perform automated cerebral small vessel disease burden scoring. Participants with a score of 2 or higher were included in this study, while participants with Alzheimer's disease (AD), brain tumors, and mental illnesses were excluded. Finally, 1616 participants were included in this embodiment of the invention. Based on the changes in WMH volume at the two time points, the participants were divided into a WMH increase group and a WMH decrease group, as follows: Figure 1 As shown.

[0021] In the aforementioned protocol, UKB was a prospective cohort study that recruited over 500,000 participants from 22 assessment centers in England, Wales, and Scotland, collecting their deep genetic and phenotypic data. The UKB study received ethical approval from the research ethics committee (11 / NW / 0382) and obtained electronic consent forms signed by each participant. An initial subset of participants was recalled to complete brain and cardiac MRI data acquisition. The UKB data is available at https: / / biobank.ctsu.ox.ac.uk / , and this invention has obtained permission to use this data (Research Project No.: 117280).

[0022] Step 2: Sample Data Preprocessing This invention primarily utilizes T1WI, FLAIR, SWI, and DTI sequences. All data have undergone thorough preprocessing and are available for use by researchers with UKB authorization. The data scanner is a standard Siemens Skyra 3T MRI scanner equipped with a standard Siemens 32-channel receiver coil. Specific scanning parameters are as follows: (1) T1WI, 3D-MPRAGE sequence, resolution 1mm*1mm*1mm, matrix: 208*256*256, scan time 5 minutes.

[0023] (2) FLAIR, resolution is 1.05mm*1mm*1mm, matrix: 192*256*256, scanning time is 6 minutes.

[0024] (3) DTI, resolution 2mm*2mm*2mm, matrix: 104*104*72, b value selected as 0 and 1000 s / mm respectively. 2 and 2000 s / mm 2 The scan duration was 7 minutes. For the two diffusion-weighted shells, a total of 50 different diffusion-encoded directions were acquired (and all 100 directions were distinct). The diffusion preparation used a standard (“monopolar”) Stejskal-Tanner pulse sequence. Because this sequence has a shorter echo time (TE = 92 ms) than the two-refocus (“bipolar”) sequence, a higher signal-to-noise ratio (SNR) can be achieved. However, this improvement comes at the cost of stronger eddy current distortion, which is removed during the image processing pipeline.

[0025] (4) SWI sequence: resolution: 0.8mm*0.8mm*0.8mm, matrix: 256*288*48, using 2 echoes, echo times (TE) are 9.42ms and 20ms respectively, scan duration is 2.5 minutes.

[0026] Preprocessing included distortion correction, brain extraction, and registration using the FMRIB software library (FSL, https: / / fsl.fmrib.ox.ac.uk / ). The reference space for data processing was the MNI152 standard template space. Brain extraction was performed on the image data using BET (Brain Extraction Tool), and the images were registered to a 1 mm resolution version of the MNI152 template using the FNIRT tool. Complete DTI image analysis workflow scripts are available at https: / / www.fmrib.ox.ac.uk / ukbiobank / , and currently, these scripts primarily utilize tools from FSL and FreeSurfer. First, for DTI data preprocessing, the Eddy tool (http: / / fsl.fmrib.ox.ac.uk / fsl / fslwiki / EDDY) was used to perform eddy current and head motion correction on the data. Next, the processed data was input into two complementary analysis workflows: one based on fibrous skeleton processing and the other based on richer intravoxel fiber structure modeling, followed by probabilistic fiber tracing analysis.

[0027] Both analysis procedures report a range of dMRI (diffusion MRI) derived parameters for different fiber regions: A) parameters based on diffusion tensor modeling, and B) parameters based on microstructure model fitting. The shell (50 orientations) with b=1000 is input into the DTI fitting tool DTIFIT to generate parameter maps such as fractional anisotropy (FA), diffusion tensor mode (MO), and mean diffusivity (MD). In addition to DTI fitting, the AMICO tool (https: / / github.com / daducci / AMICO) was used to input dMRI data into Neuroite Orientation Dispersion and Density Imaging (NODDI) modeling, generating voxel-level microstructure parameters, including indices of intracellular volume fraction (ICVF), isotropic volume fraction (ISOVF), and orientation dispersion (OD). Subsequently, the FA images from DTI were input into a tract-based spatial statistics (TBSS) method, which aligns the FA images to the white matter fiber skeleton in standard space. The resulting standard space distortion was applied to all other DTI / NODDI parameter maps.

[0028] The original T1, FLAIR, and SWI sequence images were imported into an image analysis tool called uAI ResearchPortal for processing. Using the VB-Net deep learning model, with weighted Dice loss as a constraint, the four most typical CSVD imaging features—WMH, lacunar infarcts (LA), perivascular spaces (PVS), and recent subcortical small infarcts and cerebral microbleeds (CMB)—were automatically segmented and quantitatively calculated. WMH was assessed using the Fazekas scale, evaluating the sum of periventricular and deep WMH, with a score of 0-6. PVS was scored using T1 sequences, and the basal ganglia and centrum semiovale, the regions with the most PVS, were selected for severity grading using a 4-point scale. The CSVD imaging burden scoring method is as follows: 1 point is awarded for each of the following: 1) ≥1 lacunar infarct; 2) Deep WMH score ≥2 on the Fazekas scale and / or periventricular WMH score ≥3; 3) ≥1 deep or infratentorial CMB; 4) Moderate to severe (grade 2-4) PVS in the basal ganglia. A CSVD burden score ≥2 indicates significant CSVD-related manifestations. The above assessment criteria are based on the updated international imaging criteria for cerebral small vessel disease (STROVE-2).

[0029] Step 3: Incorporate imaging and clinical phenotypes This invention enrolled 1616 participants. Baseline data included age, sex, education level, smoking status, alcohol consumption frequency, history of diabetes, history of chronic ischemic heart disease, and body mass index (BMI). Blood parameters included C-reactive protein, total cholesterol, homocysteine, white blood cell count, red blood cell count, platelet count, vitamin D, fasting blood glucose, eosinophils, basophils, high-density lipoprotein, low-density lipoprotein, hemoglobin count, lymphocyte count, neutrophil count, and glycated hemoglobin. Imaging phenotypes included whole-brain averages of FA, MD, OD, MO, ICVF, and ISOVF. A total of 30 variables were ultimately included.

[0030] Step 4: Data Preparation and Feature Filtering Missing values ​​and null values ​​were imputed for 30 variables from 1616 participants. Stepwise regression was used to screen features based on the Akaike information criterion (AIC), a standard for balancing model complexity and goodness of fit, established and developed by Japanese statistician Hiroji Akaike. The Akaike information criterion is based on the concept of information entropy.

[0031]

[0032] in, It is the number of parameters. It is the likelihood function. The formula consists of two parts: the second term represents the goodness of fit of the model, which decreases as the model fits the data better; the first term represents the model complexity, which increases with the number of variables used. Therefore, AIC can easily select models that fit the data well without overfitting.

[0033] First, univariate logistic regression was used to initially screen variables that showed significant differences between the two groups. Then, candidate variables were input into the model, and stepwise multivariate logistic regression was employed to back-select variables from the full model. The best-fit model was then constructed using the minimum AIC value. The entire variable selection and model building process was completed automatically using R programming language, without any human intervention, thus ensuring that the decision to include variables in the model was entirely based on AIC considerations.

[0034] Step 5: Build a machine learning model Participants whose WMH volume increased during the follow-up period were categorized as the WMH growth group (defined as positive), and those whose WMH volume decreased during the follow-up period were categorized as the WMH decrease group (defined as negative). The positive-to-negative ratio of the sample was relatively balanced, with a positive sample:negative sample ratio of 1.26:1. The sample was divided into two datasets in a 7:3 ratio: a training set (Train Cohort) and a validation set (Test Cohort). The positive-to-negative sample ratio within each dataset was maintained consistent with the overall sample size, at 1.26:1.

[0035] We implemented a binary classification machine learning predictive analysis task, with clinical and imaging features selected by AIC as independent variables and WMH changes (increase or decrease) as the predictor variable, based on a Logistic Regression model. During model training, we employed Grid Search Cross-Validation (GCV) to accelerate and optimize the parameter configurations of each model. This series of analyses utilized seven different algorithms to construct the predictive models. Logistic Regression, as a generalized linear regression analysis model, has demonstrated wide application value in data mining, automated disease diagnosis, economic forecasting, and other fields. It accurately estimates the probability of events based on a given dataset of independent variables, such as... Figure 2 As shown.

[0036] In this embodiment of the invention, the constructed prediction model adopts the Logistic Regression algorithm, and its input is the eight selected predictor factors—'Age', 'BMIg', 'CystatinC', 'Glucose', 'MO', 'FA', 'ICVF', and 'ISOVF'. The model with the best performance is trained using Python's scikit-learn library.

[0037] A second aspect of this invention is to provide a method for applying a prediction model constructed using the above method, comprising: The trained machine learning model is saved as a logistic_regression_model.joblib file using Python's joblib library, enabling machine learning and packaging the prediction algorithm into executable software (.exe) that supports offline operation on the Windows platform.

[0038] The resulting executable software provides a simple and practical user interface, such as Figure 4 As shown, there are three groups from top to bottom: the first group is Age and BMI, the second group is FA, MO, ICVF, and ISOVF, and the third group is Glucose and Cystatin C. Users enter data in the eight input boxes and then click the "Predict" button. The result of WMH(-) or WMH(+) will be returned at the bottom of the screen.

[0039] In this embodiment of the invention, the tkinter library in Python is used to implement the user interface. In the background of the user interface, the eight pieces of data input by the user are subjected to rule-based validation, such as checking for non-empty or non-Chinese characters. If the validation passes the rule, the eight pieces of data are combined into an array. The machine learning model saved in the first step is loaded using the joblib library, and then this data is input into the machine learning model. Next, the machine learning model outputs the probability value and result of the validation. Finally, the software interface renders the probability value and result onto the screen, forming a visual presentation that is easy for the user to understand.

[0040] In a preferred embodiment of the present invention, the executable software is distributed to the user using the following method: To accurately count the number of software users, this invention uses a combination of machine code and authentication code to determine the actual number of users. The invention utilizes Python's tkinter library to design two authentication pages for the user and an authorization interface for the software owner. Upon initial installation and use, the software backend generates a unique machine code based on the user's hardware, which is then displayed on the authentication page that pops up on the user's end. Figure 5 As shown. The next step is for the user to send the machine code to the software owner. The software owner then opens the authorization interface, enters the machine code and the usage period, generates an authentication code, and returns the authentication code to the user. The user enters the authentication code into the authentication page to complete software authentication. Only then can the software be opened correctly. After testing, we use the Python pyinstaller library to package and compile all the software code, including the license used for authorization, resulting in two executable (.exe) files. The software's .exe files are then distributed to users for download via a URL generated by an nginx server deployed on Alibaba Cloud.

Claims

1. A method for constructing a predictive model for the progression of high-signal lesions in white matter, characterized in that, Includes the following steps: Step 1: Obtain T1WI, FLAIR, SWI, and DTI sequence data from multiple subjects; Step 2: Preprocess the T1WI, FLAIR, SWI, and DTI sequence data. For DTI sequence data, preprocessing includes distortion correction, brain extraction, and registration. The data is then input into two complementary analysis workflows, which report a series of dMRI-derived parameters in different fiber regions. One workflow is based on fiber skeleton processing, while the other is based on richer intravoxel fiber structure modeling, followed by probabilistic fiber tracing analysis. The dMRI data is then input into neurite orientation dispersion and density imaging modeling to generate voxel-level microstructure parameters. DTI fitting is performed to obtain fractional anisotropy parameter maps. The parametric map is input into a fiber tract-based spatial statistical method, aligning the fractional anisotropy parametric map onto the white matter fiber skeleton in standard space. The resulting standard spatial distortion is applied to all other DTI / NODDI parametric maps. Preprocessing of T1WI, FLAIR, and SWI sequence data includes: processing the T1WI, FLAIR, and SWI sequence data using image analysis tools, automatically segmenting and quantifying CSVD imaging features including white matter hyperintensity, lacunar infarcts, perivascular spaces, recent subcortical small infarcts, and cerebral microbleeds. Step 3: Obtain the subjects' imaging phenotypic data and clinical phenotypic data, including the number of clinical phenotypic data... Based on baseline data and blood parameters of the subjects, clinical phenotypic data include the whole-brain average values ​​of FA, MD, OD, MO, ICVF, and ISOVF obtained in step 2; Step 4: After processing all the variable data obtained in step 3, feature screening is completed. Feature screening includes the following steps: performing univariate logistic regression to initially screen candidate variables with significant differences between the two groups; inputting the candidate variables into the model, using stepwise multivariate logistic regression, back-eliminating from the entire model, and constructing the best-fit model using the minimum AIC value; Step 5: dividing the feature sample data of the subjects obtained in step 4 into a WMH increase group and a WMH decrease group. Step 6: Construct a training set and a validation set. During the follow-up period, feature sample data from subjects whose white matter hyperintensity foci increased in volume are classified as the WMH growth group and assigned a positive label; feature sample data from subjects whose white matter hyperintensity foci decreased in volume during the follow-up period are classified as the WMH decrease group and assigned a negative label. Step 7: Construct a binary classification machine learning prediction model. The independent variables are the features selected in Step 4, and the predictor variables are the binary classification values ​​representing the growth or shrinkage of white matter hyperintensity foci. The binary classification machine learning prediction model is trained using the training set constructed in Step 5, and validated using the validation set constructed in Step 5, thus obtaining a white matter hyperintensity foci progression prediction model.

2. The method for constructing a white matter high-signal lesion progression prediction model as described in claim 1, characterized in that, In step 1, participants with a cerebral small vessel disease burden score greater than or equal to the set threshold were excluded, as were participants with Alzheimer's disease (AD), brain tumors, and mental illnesses.

3. The method for constructing a white matter high-signal lesion progression prediction model as described in claim 1, characterized in that, In step 2, dMRI-derived indices include indices based on diffusion tensor modeling and indices based on microstructure model fitting.

4. The method for constructing a white matter high-signal lesion progression prediction model as described in claim 1, characterized in that, In step 2, the voxel-level microstructure parameters include indices of intracellular volume fraction white matter neurite density, isotropic volume fraction, and directional dispersion.

5. The method for constructing a white matter high-signal lesion progression prediction model as described in claim 1, characterized in that, In step 2, the expansion tensor mode parameter map and the average diffusion rate parameter map are also obtained through DTI fitting.

6. The method for constructing a white matter high-signal lesion progression prediction model as described in claim 1, characterized in that, In step 3, the baseline data includes age, gender, education level, smoking status, frequency of alcohol consumption, history of diabetes, history of chronic ischemic heart disease, and body mass index; the blood indicator data includes C-reactive protein, total cholesterol, homocysteine, white blood cell count, red blood cell count, platelet count, vitamin D, fasting blood glucose, eosinophils, basophils, high-density lipoprotein, low-density lipoprotein, hemoglobin count, lymphocyte count, neutrophil count, and glycated hemoglobin.

7. The method for constructing a white matter high-signal lesion progression prediction model as described in claim 1, characterized in that, In step 6, during the model training phase, grid search cross-validation was used to accelerate and optimize the parameter configuration of different models, and then the Logistic Regression model was selected as the final model for predicting the progression of white matter hyperindication lesions.

8. An application method for a white matter hypersignal lesion progression prediction model constructed by the method described in claim 1, characterized in that, The white matter high-signal lesion progression prediction model constructed by the method described in claim 1 is packaged into executable software.

9. An application method as described in claim 8, characterized in that, When the executable software runs on an electronic device, it provides a user interface for inputting Age, BMI, FA, MO, ICVF, ISOVF, Glucose, and Cystatin C. The white matter high signal lesion progression prediction model displays the prediction results of the growth or shrinkage of the white matter high signal lesion on the same user interface based on the user's input results.

10. The application method as described in claim 8, characterized in that, The executable software is distributed to users using the following method: When a user installs and uses the executable software for the first time, the background generates a unique machine code based on the hardware of the user's electronic device and presents it to the user; the user sends the machine code to the software owner; after the software owner enters the machine code and the usage period, an authentication code is generated. After the authentication code is returned to the user, the user enters the authentication code to complete the software authentication in the background; the executable software is then published to the user via a URL link generated by a server deployed in the cloud.

Images