Method for identifying stable subtype of Parkinson's disease based on neuroimaging in combination with machine learning

Abstract

The invention discloses a method for recognizing a stable subtype of Parkinson's disease based on a neural image in combination with machine learning, and the specific recognition method is based on the structural heterogeneity of a PD individual, and a semi-supervised machine learning method of discriminant analysis is adopted to recognize a neural anatomical subtype of a first diagnosis unrecognized medicine taking PD individual. According to the discriminant analysis method, classification and clustering are mainly carried out at the same time, PD individuals and healthy individuals are distinguished through a polyhedron formed by a linear maximum interval classifier, then the PD individuals are clustered according to the distances between patients and different hyperplanes of the polyhedron, and the method is small in influence on individual differences irrelevant to diseases and high in accuracy. The identification of real disease subtypes is facilitated; according to the method, the disease progress is monitored on the basis that the nigra free water value of a PD subtype patient is used as a judgment index for neuroinflammation and dopamine system injury; meanwhile, the clinical decision quality and the individual treatment effect are improved.

Description

Technical Field

[0001] The present invention relates to the field of neuromedicine technology, and in particular to a method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning. Background Art

[0002] Parkinson's disease, abbreviated as PD, is a highly heterogeneous neurodegenerative disease. Patients have significant differences in clinical manifestations, drug efficacy and disease progression. Revealing the biological basis of these differences has become a key goal of PD research. It is generally understood that PD is likely to contain multiple biological subtypes, which are related to the patient's genetic background, environmental exposure and pathological process. Accurately identifying different subtypes of PD is crucial for in-depth pathological research, improving the accuracy of early diagnosis, more accurately predicting disease progression and formulating personalized treatment plans. Despite this, there is currently no unified consensus on the classification of PD subtypes. This situation highlights the urgent need for in-depth research to clarify PD subtypes.

[0003] Currently, clinical and pathological heterogeneity is the main obstacle to Parkinson's disease biomarkers and optimized individualized treatment. With the maturity of artificial intelligence and in vivo imaging technology, as well as easier access to computing resources and large databases, researchers increasingly prefer hypothesis-free, data-driven methods to identify PD subtypes. Researchers have invested tremendous efforts in identifying PD subtypes and have formed two major research directions: hypothesis-based empirical classification and data-based hypothesis-free classification. The former defines PD subtypes based on early hypotheses, usually focusing on age of onset and single motor symptoms; while the latter uses statistical and machine learning methods to identify subtypes. With the rapid development of the field of artificial intelligence, this method has shown great potential in identifying key features of the disease and determining subtypes. Although the two methods have different starting points, after the subtype is determined, they will seek to verify the reliability of the classification results through cerebrospinal fluid analysis, neuroimaging, and genetic variation studies. Some studies also attempt to evaluate the differences in disease progression and their underlying pathological hypotheses among different subtypes. Despite this, there is still no broad consensus on the classification of PD subtypes. This invention mainly focuses on the progress and challenges of PD subtype research, and provides a method for identifying stable subtypes of PD by combining neuroimaging with machine learning, providing a perspective for individualized treatment. Summary of the invention

[0004] In view of the shortcomings of the above-mentioned technologies that it is difficult to classify PD subtypes and form stable PD subtypes, the present invention provides a new method for identifying stable subtypes of PD by combining neuroimaging with machine learning.

[0005] To achieve the above object, the present invention provides a method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning, including the following steps:

[0006] S1, structural MRI and diffusion distribution MRI were performed on the brain of PD individuals, and T1-weighted images and diffusion distribution images were generated by scanning with D gradient echo sequence respectively;

[0007] S2, data acquisition and preprocessing of the scanned T1-weighted image and diffusion distribution image;

[0008] S3, the heterogeneity and intergroup differences of GMV gray matter volume were compared between PD individuals and healthy individuals;

[0009] S4, based on heterogeneity, PD individuals were divided into PD subtypes using discriminant analysis.

[0010] Among them, the T1-weighted image is preprocessed, and the specific steps are:

[0011] S5, T1-weighted images were normalized to the MNI Montreal Neurological Institute space3;

[0012] S6, segmenting the normalized T1-weighted image into gray matter tissue, white matter tissue, and cerebrospinal fluid tissue, and then converting the voxel value of tissue density into volume;

[0013] S7, GMV gray matter volume of the region of interest was calculated based on the predefined AAL automatic anatomical labeling template, and the GMV gray matter volume image was smoothed using an isotropic Gaussian kernel with a full width at half maximum of 8 mm.

[0014] Among them, the diffusion distribution image is preprocessed, and the specific steps are:

[0015] S8, firstly, the diffusion distribution image is subjected to noise reduction, gibbs ring artifacts are eliminated, motion deviation is corrected, and eddy current influence is corrected;

[0016] S9, free water model for estimating diffusion distribution images;

[0017] S10, the free water model is analyzed into free water space and tissue space, where the free water space is used to simulate the isotropic diffusion of water molecules outside the cells, highlighting the situation of water molecules outside the cells; the tissue space is used to reflect the tissue microstructure after removing the free water signal.

[0018] Among them, based on MRI magnetic resonance imaging data, the GMV gray matter volume was obtained using the region of interest defined by the AAL template. The patient's gender, age, years of education and total brain volume were used as covariates, and the region of interest was clustered and subtyped using HYDRA.

[0019] Among them, HYDRA determines the number of hyperplanes according to the number of clusters to form convex polygons to distinguish PD individuals from HCs.

[0020] The distance from the PD individual to each hyperplane is calculated using an extended standard linear maximum margin classifier, and the PD individual is assigned to the nearest hyperplane through the classifier, and the number of K clusters is defined.

[0021] The clustering parameters were set to at least 50 iterations between the hyperplane and the clusters, the number of cluster consistency steps was 20 times, and the regularization parameter was set to 0.25, followed by a 10-fold cross validation.

[0022] Among them, two regions of interest were set on both sides of the substantia nigra respectively, and the section S0 with the largest red nucleus volume was determined in the image. Then, the section was moved ventrally to the section S1 where the red nucleus began to disappear but was vaguely discernible, and then it continued to move ventrally to the section S2 where the red nucleus was completely invisible. On the sections of S1 and S2, the regions of interest of the anterior and posterior substantia nigra were placed in the darkest low-signal areas, and the intra-group consistency coefficient of the marking was calculated to evaluate the marking consistency of the main and auxiliary scorers, and the free level average of the regions of interest of the anterior and posterior substantia nigra on both sides was calculated.

[0023] Among them, the split sample analysis method was used to perform repeatability analysis on PD subtypes. PD individuals and healthy individuals were used as references, and the GMV gray matter volume differences between PD individuals and healthy individuals were calculated and compared through the HYDRA heterogeneity and discriminant analysis method.

[0024] The beneficial effects of the present invention are as follows: compared with the prior art, the present invention provides a method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning. The specific identification method is based on the structural heterogeneity of PD individuals, and adopts a heterogeneous discriminant analysis method, combined with a semi-supervised machine learning method to identify the neuroanatomical subtypes of PD individuals who are not identified as taking medication at the first visit. The discriminant analysis method mainly performs classification and clustering at the same time. The polyhedron formed by the linear maximum margin classifier distinguishes PD individuals from healthy individuals, and then clusters PD individuals according to the distance between the patient and different hyperplanes of the polyhedron. This method is less affected by individual differences unrelated to the disease and helps to identify the true disease subtype; the present invention uses the substantia nigra free water value of PD subtype patients as a judgment indicator of neuroinflammation and dopamine system damage to monitor the progression of the disease; at the same time, in order to improve the quality of clinical decision-making and individual treatment effects, the present invention also uses a linear mixed effects model to predict the development trajectory of motor symptoms in patients with different PD subtypes, providing a variety of treatment possibilities for the subsequent treatment of PD individuals and accurately specifying treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1is a flow chart of the method of the present invention;

[0026] Figure 2 It is the principle diagram of the analysis and discrimination method of the present invention;

[0027] Figure 3 This is a heterogeneity analysis diagram of PD individuals and healthy individuals in the present invention. DETAILED DESCRIPTION

[0028] In order to more clearly illustrate the present invention, the present invention is further described below in conjunction with the accompanying drawings.

[0029] In the following description, the example details are given to provide a deeper understanding of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. It should be understood that the specific embodiments are only used to explain the present invention and are not used to limit the present invention.

[0030] It should be understood that when the terms "include" and / or "comprises" are used in this specification, they indicate the existence of the stated features, integers, steps, operations, elements or components, but do not exclude the existence or addition of one or more other features, integers, steps, operations, elements, components or combinations thereof.

[0031] See also Figure 1 The present invention provides a method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning, comprising the following specific identification methods:

[0032] S1, the brain of PD individuals was scanned with MRI (magnetic resonance imaging) 3D gradient echo sequence to generate T1-weighted images;

[0033] S2, data acquisition and preprocessing of T1-weighted images obtained by MRI (magnetic resonance imaging) scanning;

[0034] S3, the heterogeneity and intergroup differences of GMV gray matter volume were compared between PD individuals and healthy individuals;

[0035] S4, based on heterogeneity, PD individuals were divided into PD subtypes using discriminant analysis.

[0036] First, in the present invention, the PD individuals scanned by MRI should have never taken Parkinson's drugs. Under the influence of drugs, the PD individuals should be classified as early as possible and the course of the disease should be predicted and analyzed. In S1, it is necessary to perform structural and magnetic resonance diffusion tensor imaging on the brain of the PD individual to generate T1-weighted images, and then perform data acquisition and preprocessing on the T1-weighted images. The processed T1-weighted images of the PD individuals are compared with those of healthy individuals to compare the heterogeneity of GMV gray matter volume between the two. Based on the heterogeneity, the subtypes in the PD individuals are identified using the discriminant analysis method.

[0037] Among them, the T1-weighted image is preprocessed, and the specific processing steps are:

[0038] S5, normalize T1-weighted images to the Montreal Neurological Institute space, and resample the voxel size of T1-weighted images to 1.5*1.5*1.5mm 3 ;

[0039] S6, segmenting the normalized T1-weighted images into gray matter, white matter, and cerebrospinal fluid tissues, and then converting the voxel values ​​of tissue density into volumes;

[0040] S7, the GMV gray matter volume of the region of interest was calculated based on the predefined AAL (Automated Anatomical Labeling) template, and the GMV gray matter volume GMV image was smoothed using an isotropic Gaussian kernel with a full width at half maximum of 8 mm.

[0041] Since the size and shape of brain structures vary between individuals, the MRI data must be preprocessed when performing group-level comparisons, that is, the T1-weighted images must be preprocessed, and the individual's brain must be rewritten and cut into pixels with a voxel value of 1.5*1.5*1.5 cubic millimeters. Then, segmentation is performed on the re-cut 3D image to distinguish different brain tissues. Among them, the present invention mainly targets the gray matter part, because the gray matter part mainly represents the part with dense neuronal cell bodies, that is, the darker the color of the gray matter part, the more frequent and intense the brain cell activity. The segmented image is then aligned to the standard space template created by the Montreal Neurological Institute in the prior art, so that different individuals can be compared and comparable; then, the individual's cerebrum and cerebellum are divided into 116 regions through the ALL brain atlas, and the volumes of these regions are extracted for grouping and calculation.

[0042] In another embodiment, all the technical features of the above embodiments are included, and the diffusion distribution image is preprocessed, and the specific steps are:

[0043] S8, firstly, the diffusion distribution image is subjected to noise reduction, gibbs ring artifacts are eliminated, motion deviation is corrected, and eddy current influence is corrected;

[0044] S9, free water model for estimating diffusion distribution images;

[0045] S10, the free water model is analyzed into free water space and tissue space, where the free water space is used to simulate the isotropic diffusion of water molecules outside the cells, highlighting the situation of water molecules outside the cells; the tissue space is used to reflect the tissue microstructure after removing the free water signal.

[0046] Among them, the preprocessing of DTI (diffusion tensor) images was specifically completed using the MRtrix3 toolkit. This preprocessing process involves multiple steps, including noise reduction, elimination of Gibbs ring artifacts, correction of dynamic deviation, and correction of eddy current effects; further, the script provided by the MarkVCID program was used to estimate the free water model, which mainly analyzes two different spaces: one is the free water space, which mainly simulates the isotropic diffusion of water molecules outside the cell to highlight the situation of water molecules outside the cell, and then compare the free water values ​​between groups; the other is the tissue space, which reflects the tissue microstructure after the free water signal is removed. Among them, the volume fraction of the free water space, which is an indicator of the free water content, shows the proportion of free water in each pixel, and its value is between 0 and 1. Through the above preprocessing operations, the noise during magnetic resonance scanning can be removed, the image display can be more accurate, and the image quality can be improved. In addition, the free water value can be extracted based on the image generated by the preprocessing model for inter-group comparison.

[0047] The following is based on a 3.0TVerio Siemens machine. T1-weighted images were scanned using a coronal 3D gradient echo sequence with the following settings: repetition time 2530 milliseconds; echo time 3.34 milliseconds; flip angle 7 degrees; total number of slices 128; slice thickness 1.33 mm, slice gap 0.5 mm; acquisition matrix setting 256 × 192; field of view size 256 mm × 256 mm. DTI data were acquired using a spin echo planar imaging sequence in the axial plane parallel to the anterior-posterior joint line. The parameters were set as follows: repetition time 9000 milliseconds, echo time 104 milliseconds, flip angle 7 degrees, field of view 230 mm × 230 mm, slice thickness 2.5 mm, no gap between slices, matrix size 128 × 128, a total of 49 slices covering the entire brain, b values ​​0 and 1000 seconds per square millimeter (s / mm 2 ), and 64 diffusion gradient directions.

[0048] In another embodiment, all the technical features of the above embodiments are included, based on MRI magnetic resonance imaging data, the GMV gray matter volume is obtained by using the region of interest defined by the AAL template, the patient's gender, age, years of education and total brain volume are used as covariates, and the region of interest is clustered and subtyped through discriminant analysis based on heterogeneity;

[0049] The present invention is based on a semi-supervised machine learning method called heterogeneity through discriminant analysis (HYDRA) to identify neuroanatomical subtypes of drug-naive PD individuals at the first visit. The advantage of HYDRA is that it can perform classification and clustering at the same time: it uses a polyhedron formed by a linear maximum margin classifier to separate patients from HCs, and then clusters patients according to their distance from different hyperplanes of the polyhedron. This method is less susceptible to individual differences unrelated to the disease and helps to identify true disease subtypes. To date, HYDRA has been successfully used in the study of Alzheimer's disease, schizophrenia, autism spectrum disorder and obsessive-compulsive disorder. It is assumed that it can reveal different neuroanatomical subtypes of PD. At the same time, in order to explore the potential pathological mechanism of this structural difference, the free water values ​​of the substantia nigra (SN) of patients with different subtypes were compared. According to previous studies, the free water values ​​of the posterior substantia nigra (PSN) of PD individuals are increased. This phenomenon can be used as an indicator of neuroinflammation and dopamine system damage and used to monitor the progression of the disease. In addition, in order to improve the quality of clinical decision-making and optimize individual treatment, the present invention also uses a linear mixed effects model to predict the development trajectory of motor symptoms in individuals with different subtypes of PD. The linear mixed effects model can assume that different neuroanatomical subtypes show different clinical processes, and then it can observe the disease progression of different types of PD and predict the impact of different independent variables on the dependent variable.

[0050] From the sMRI magnetic resonance imaging data, the GMV gray matter volume (GMV) is obtained according to the region of interest (ROI) defined by the AAL template for further clustering and subtype division, and gender, age, years of education and total intracranial volume (TIV) are used as covariates. HYDRA can identify different subtypes of newly diagnosed PD based on the volume measurements of these ROIs. HYDRA is a hybrid of unsupervised clustering and supervised classification methods. It can simultaneously fit the maximum margin classification boundary and clarify the disease subtype, which is something that unsupervised clustering methods cannot do. This method can limit the influence of confounding variables such as age, gender, race and noise, and only the differences between PD individuals and healthy individuals are clustered together; the number of hyperplanes is determined according to the number of clusters of the K value, forming a convex polygon to distinguish PD individuals from healthy individuals, and the distance from the PD individual to each hyperplane is calculated by an extended standard linear maximum margin classifier. The classifier will assign the PD individual to the nearest hyperplane, thereby defining K clusters; Figure 2 As shown in the figure, the principle of the HYDRA method, the healthy individuals in the control group (represented by green dots) and the PD individuals (represented by red dots) are separated by a convex polygon decision boundary. The solid line represents the classifier, the dotted line represents the boundary, and the red highlighted solid line part represents the convex polygon used to separate the patients and the control group. K is the number of clusters, and the specific discriminant method is to set the cluster range from 2 to 6. The clustering parameters are set as follows: 50 iterations are performed between the estimated hyperplane and the cluster clusters, the number of cluster consistency steps is 20 times, the regularization parameter is set to 0.25, and then a 10-fold cross validation is performed. This method uses the ARI adjusted Rand index to quantify the consistency between clustering results. As an indicator of clustering stability, the ARI value ranges from 0 to 1. Ideally, the highest ARI value is obtained at a specific K value, while the ARI values ​​of other K values ​​are relatively low.

[0051] Among them, the linear mixed effects model can adapt to data imbalance caused by missing data, loss of follow-up, unequal follow-up intervals and different first assessment times, while taking into account the correlation between repeated measurements of each individual, it is often used to analyze the trajectory of longitudinal markers, and the duration of the disease is used as the time scale; in the fixed effect part of the model, the present invention incorporates time, PD subtype and their interaction effects, and uses gender, age of disease onset and LEDD at follow-up as covariates to construct a prediction model for disease progression; in addition, the linear mixed effects model sets a random intercept for ID and a random slope for time for each patient to better adapt to individual differences; the linear mixed effects model generates a predicted value for each observation and uses ggplot to plot the predicted disease progression trajectory, so in the linear mixed effects model, a two-tailed p value less than 0.05 is considered statistically significant.

[0052] In another embodiment, all the technical features of the above embodiments are included, two regions of interest are set on both sides of the substantia nigra, two regions of interest are set on both sides of the substantia nigra, the section S0 with the largest red nucleus volume is determined in the image, and then the section is moved ventrally to the section S1 where the red nucleus begins to disappear but is vaguely discernible, and then the section is continued to move ventrally to the section S2 where the red nucleus is completely invisible, and on the sections S1 and S2, the regions of interest of the anterior substantia nigra and the posterior substantia nigra are placed in the darkest low-signal areas, the intra-group consistency coefficient of the marking is calculated, the marking consistency of the main and auxiliary scorers is evaluated, and the free level average of the regions of interest of the anterior substantia nigra and the posterior substantia nigra on both sides is calculated.

[0053] The specific calculation steps of the free level mean are as follows:

[0054] A. For each region of interest, record its signal strength value.

[0055] B. Calculate the average signal intensity value of the eight voxels within each region of interest.

[0056] C. The average values ​​were calculated for the bilateral anterior and posterior substantia nigra regions of interest.

[0057] First, suppose that S 1i , S 2i , ....S 8i represents the signal intensity values ​​of the eight voxels in the ith region of interest, then the average signal intensity of each region of interest is It can be calculated using the following formula:

[0058]

[0059] Finally, the free level means of the regions of interest for the bilateral substantia nigra anterior and substantia nigra posterior were calculated separately:

[0060]

[0061] Here Represents the average signal intensity of the region of interest in the anterior part of the bilateral substantia nigra, Represents the mean signal intensity of the region of interest in the posterior substantia nigra bilaterally.

[0062]

[0063] In an actual research process, the present invention provides a comparative reference for different PD subtypes. Please refer to Table 1. A total of 283 PD individuals who were not taking medication for the first time were included in the analysis, including 149 subtype 1 and 134 subtype 2. There were no differences in gender, age, years of education, age of onset, disease duration and TIV between the two groups. However, in terms of clinical characteristics, compared with subtype 2, subtype 1 showed more severe motor symptoms (UPDRS-III scores of 22.4±10.3 and 19.5±9.5, P=0.031) and more severe non-motor symptoms (NMSQ scores of 8.2±4.5 and 6.9±4.2, P=0.017). In addition, cognitive impairment was more obvious in subtype 1 (MoCA scores of 21.2±5.3 and 22.3±4.9, P=0.065), although it did not reach a statistically significant level. The free water value of PSN in subtype 1 was significantly higher than that in subtype 2.

[0064]

[0065]

[0066] Table 1

[0067] Wherein, TIV is total intracranial volume; UPDRS-II is the daily life quality part of the unified Parkinson's disease rating scale; UPDRS-III is the motor part of the unified Parkinson's disease rating scale; MoCA is the Montreal Cognitive Assessment; NMSQ is the non-motor symptom questionnaire; HAMA is the Hamilton Anxiety Rating Scale; HAMD is the Hamilton Depression Rating Scale; ASN is the anterior part of the substantia nigra; PSN is the posterior part of the substantia nigra, a is a two-sample t-test, b is a chi-square test, c is a Mann-Whitney U test, and d is a covariance analysis, which was adjusted for age, gender, education level, TIV and disease duration. Based on the information in Table 1, it can be seen that there are no differences between subtype 1 and subtype 2 in terms of gender, age, years of education, age of onset, disease duration and TIV, but in other aspects, subtype 1 has more severe motor symptoms than subtype 2.

[0068] See also Figure 3 , PD subtypes were analyzed for repeatability using the split sample analysis method, and PD individuals and healthy individuals were used as references. The GMV gray matter volume differences between PD individuals and healthy individuals were calculated and compared using the HYDRA heterogeneity discriminant analysis method. Figure 3 It can be seen that PD individuals show higher structural heterogeneity than healthy individuals (PD group, PD individual control group; HC group, healthy individual control group; Variety, heterogeneity).

[0069] The advantages of the present invention are:

[0070] 1. Based on the structural heterogeneity of PD individuals, a semi-supervised machine learning method of discriminant analysis is used to identify the neuroanatomical subtypes of PD individuals who are not identified and taking medication at the first visit, reduce the impact of individual differences unrelated to the disease, and help identify the true disease subtypes.

[0071] 2. This method uses the substantia nigra free water value of PD subtype patients as an indicator of neuroinflammation and dopamine system damage to monitor the progression of the disease.

[0072] 3. The present invention uses a linear mixed effects model to predict the development trajectory of motor symptoms in patients with different PD subtypes, providing a variety of treatment possibilities for the subsequent treatment of PD individuals, accurately specifying treatment plans, and improving the quality of clinical decision-making and individual treatment effects.

[0073] The above disclosures are only several specific embodiments of the present invention, but the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning, characterized in that: The specific identification method includes the following steps: S1, structural MRI and diffusion distribution MRI were performed on the brain of PD individuals, and T1-weighted images and diffusion distribution images were generated by scanning with D gradient echo sequence respectively; S2, data acquisition and preprocessing of the scanned T1-weighted image and diffusion distribution image; S3, the heterogeneity and intergroup differences of GMV gray matter volume were compared between PD individuals and healthy individuals; S4, based on heterogeneity, PD individuals were divided into PD subtypes using discriminant analysis.

2. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 1, characterized in that: Preprocess the T1-weighted image. The specific steps are as follows: S5, T1-weighted images were normalized using the MNI Montreal Neurological Institute space; S6, segmenting the normalized T1-weighted image into gray matter tissue, white matter tissue, and cerebrospinal fluid tissue, and then converting the voxel value of tissue density into volume; S7, GMV gray matter volume of the region of interest was calculated based on the predefined AAL automatic anatomical labeling template, and the GMV gray matter volume image was smoothed using an isotropic Gaussian kernel with a full width at half maximum of 8 mm.

3. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 1, characterized in that: Preprocess the diffusion distribution image. The specific steps are as follows: S8, firstly, the diffusion distribution image is subjected to noise reduction, gibbs ring artifacts are eliminated, motion deviation is corrected, and eddy current influence is corrected; S9, estimating the free water model of the diffusion distribution image; S10, analyzing the free water model into a free water space and a tissue space, wherein the free water space is used to model the free water space. The isotropic diffusion of simulated water molecules outside the cells highlights the situation of water molecules outside the cells; the tissue space is used to reflect the tissue microstructure after the free water signal is removed.

4. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 1, characterized in that: Based on MRI data, the GMV gray matter volume was obtained using the region of interest defined by the AAL template. The patient's gender, age, years of education and total brain volume were used as covariates, and the region of interest was clustered and subtyped using HYDRA.

5. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 4, characterized in that: HYDRA determines the number of hyperplanes according to the number of clusters, forming convex polygons to distinguish PD individuals from HCs.

6. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 5, characterized in that: The distance between the PD individual and each hyperplane is calculated using an extended standard linear maximum margin classifier. The PD individual is assigned to the nearest hyperplane through the classifier, and the number of K clusters is defined.

7. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 6, characterized in that: The number of clusters parameter was set to at least 50 iterations between hyperplanes and clusters, the number of cluster consistency steps was 20, and the regularization parameter was set to 0.25, followed by 10-fold cross validation.

8. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 1, characterized in that: Two regions of interest were set on both sides of the substantia nigra, and the section S0 with the largest red nucleus volume was determined in the image. Then, the section was moved ventrally to the section S1 where the red nucleus began to disappear but was vaguely discernible, and then it continued to move ventrally to the section S2 where the red nucleus was completely invisible. On the sections of S1 and S2, the regions of interest of the anterior and posterior substantia nigra were placed in the darkest low-signal areas, and the intra-group consistency coefficient of the marking was calculated to evaluate the marking consistency of the main and auxiliary scorers, and the free level average of the regions of interest of the anterior and posterior substantia nigra on both sides was calculated.

9. The method for identifying stable subtypes of Parkinson's disease based on neuroimaging combined with machine learning according to claim 1, characterized in that: The split sample analysis method was used to perform repeatability analysis on PD subtypes. PD individuals and healthy individuals were used as references, and the GMV gray matter volume differences between PD individuals and healthy individuals were calculated and compared using the HYDRA heterogeneity discriminant analysis method.

Images