CN115040147A

Movatterモバイル変換

Info

Publication number: CN115040147A
Application number: CN202210616879.7A
Authority: CN
Inventors: 彭莉玲; 高欣; 李伟凯
Original assignee: Shanghai Universal Medical Imaging Diagnosis Center Co ltd
Current assignee: Shanghai Universal Medical Imaging Diagnosis Center Co ltd
Priority date: 2022-06-01
Filing date: 2022-06-01
Publication date: 2022-09-13

Abstract

The invention discloses a Parkinson's disease prediction method based on an 18F-FDG PET metabolic network, and belongs to the technical field of preventive medicine. The method comprises the steps of collecting and processing brain CT data of an object, constructing a metabolic network to obtain links and graph attribute, fusing different connection information based on the multi-core SVM, training and classifying, establishing a PD diagnosis prediction model, and predicting by using the trained model. Most consensus junctions in the frontal, parietal and occipital regions of PD patients were found to decrease, while frontal, temporal and subcortical regions increased based on metabolic networks. These abnormal functional network measurements show ideal classification performance in identifying PD individuals from HC individuals, with accuracy rates as high as 91.84%.

Description

Parkinson's disease prediction method based on 18F-FDG PET metabolic network

Technical Field

The invention belongs to the technical field of preventive medicine, and particularly relates to a Parkinson's disease prediction method based on an 18F-FDG PET metabolic network.

Background

Currently, about 620 million people are affected worldwide, and Parkinson's Disease (PD) is the second most common neurodegenerative disorder, the prevalence of which may increase in the next decades (Vos et al, 2017). Unfortunately, the diagnosis of PD and disease severity assessment are primarily assessed through clinical examination and follow-up. An effective method for predicting the follow-up diagnosis is urgently needed to improve the diagnostic performance of PD. Positron emission tomography using 18F-fluorodeoxyglucose (18F-FDG PET) is a functional neuroimaging technique that can measure metabolic abnormalities of PD at the system level. The existing PET-based method directly utilizes PET data for diagnosis, usually fails to consider the metabolic interaction among regions, may lose the relevant information about metabolic topology or individual difference in the network, and greatly reduces the diagnosis performance.

Through retrieval, the Chinese invention patent is as follows: the application discloses a Parkinson prediction method based on nuclear magnetic resonance (with the application number of CN201810983951.3 and the application date of 2018.08.27), which comprises the following steps: acquiring a craniocerebral nuclear magnetic resonance image to be diagnosed; processing the craniocerebral nuclear magnetic resonance image to be diagnosed to obtain a white matter image to be diagnosed; and predicting the Parkinson according to the white matter image to be diagnosed based on a pre-trained Parkinson prediction model. Correspondingly, the invention also discloses a Parkinson prediction system based on nuclear magnetic resonance, a computer readable storage medium and terminal equipment. By adopting the technical scheme of the invention, the prediction of the Parkinson can be realized, and the prediction accuracy is improved. However, this application has a disadvantage that the judgment is performed only by morphological analysis based on a white matter image of the brain, which is not comprehensive enough to cause a problem of low recognition rate.

Disclosure of Invention

1. Technical problem to be solved by the invention

The invention aims to solve the problems that the existing PET-based method directly utilizes PET data to diagnose, usually fails to consider the metabolic interaction among regions, possibly loses relevant information about metabolic topology or individual difference in a network, and greatly reduces the diagnosis performance.

2. Technical scheme

In order to achieve the purpose, the technical scheme provided by the invention is as follows:

the Parkinson's disease prediction method based on the 18F-FDG PET metabolic network comprises the steps of collecting and processing brain CT data of an object, constructing the metabolic network to obtain links and graph attributes, fusing, training and classifying different connection information based on the multi-core SVM, establishing a PD diagnosis prediction model, and predicting by using the trained model.

Preferably, the method specifically comprises the following steps:

s100, data acquisition;

s200, preprocessing data;

s300, constructing a metabolic network;

s400, extracting features, and acquiring link and graph theory attributes;

s500, establishing a PD diagnosis prediction model;

s600, predicting a result.

Preferably, the data acquisition of step S100 is to build a 3-dimensional (3D) model after scanning the object by using a PET/CT scanner, and reconstruct the image by using an ordered subset expectation maximization algorithm with 6 iterations and 6 subset methods.

Preferably, the data preprocessing of step S200 is specifically data preprocessing by using SPM-based toolkit to normalize the single 18F-FDG PET image volume space to a standard stereotactic Montreal Neurology Institute (MNI) space with linear and non-linear 3D transforms, and to apply an Automatic Anatomical Labeling (AAL) atlas to segment the cerebral cortex into 90 regions.

Preferably, the metabolic network construction of step S300 is specifically to use JS divergence to capture statistical relationships of brain glucose metabolic similarities in any 2 regions, and further to depict individual metabolic connections, use 90 ROIs from AAL atlas to represent brain nodes to depict individual metabolic networks;

extracting voxel intensities in each ROI and estimating a Probability Density Function (PDF) of the corresponding ROI by using a kernel density estimation; metabolic networking was obtained according to Jensen-Shannon (JS) divergence as follows:

JSs(P||Q)＝exp(-D_JS (P||Q))；

wherein,

p and Q represent probability density functions of different ROIs, M is 0.5 (P + Q), D_kL (. |. cndot.) represents KL divergence.

Preferably, the step S400 is specifically to extract global and local attributes based on the constructed metabolic network, where the global attributes include a clustering coefficient (C _ p), a feature path length (L _ p), a normalized clustering coefficient (γ), a normalized feature path length (λ), a small world (σ), a global efficiency (E _ global), and a modularization score (Q); the local attributes include degree centrality, node efficiency, intermediary centrality, shortest path length, and node clustering coefficients.

Preferably, the step S500 is specifically to fuse, train and classify different connection information based on the multi-core SVM, sparsify the network under different sparse thresholds (0.02-0.5, step length is 0.01) to obtain a sum of 49 values of the attribute of the corresponding node under the sparse threshold, and then train the classifier by using the sum of 49 values of each node (area under the curve, AUC) as the input of the attribute to establish the PD diagnosis prediction model.

3. Advantageous effects

Compared with the prior art, the technical scheme provided by the invention has the following beneficial effects:

the Parkinson's disease prediction method based on the 18F-FDG PET metabolic network comprises the steps of collecting and processing brain CT data of an object, constructing the metabolic network to obtain links and graph attributes, fusing, training and classifying different connection information based on the multi-core SVM, establishing a PD diagnosis prediction model, and predicting by using the trained model. Most consensus junctions in the frontal, parietal and occipital regions of PD patients were found to decrease, while frontal, temporal and subcortical regions increased based on metabolic networks. These abnormal functional network measurements show ideal classification performance in identifying PD individuals from HC individuals, with accuracy as high as 91.84%.

Drawings

FIG. 1 is a flow chart of a method of the present invention;

FIG. 2 is a diagram of the consensus connection of example 2.

Detailed Description

In order to make the technical solutions of the present application better understood by those skilled in the art, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only some embodiments of the present application, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict. The present application will be described in detail below with reference to the embodiments with reference to the attached drawings.

Example 1

Referring to fig. 1, according to the parkinson's disease prediction method based on the 18F-FDG PET metabolic network, brain CT data of an object is collected and processed, after metabolic network acquisition links and graph attribute are constructed, different connection information is fused, trained and classified based on the multi-core SVM, a PD diagnosis prediction model is established, and the trained model is used for prediction. Most consensus junctions in the frontal, parietal and occipital regions of PD patients were found to decrease, while frontal, temporal and subcortical regions increased based on metabolic networks. These abnormal functional network measurements show ideal classification performance in identifying PD individuals from HC individuals, with accuracy as high as 91.84%.

The method specifically comprises the following steps:

s100, data acquisition;

s200, preprocessing data;

s300, constructing a metabolic network;

s400, extracting features, and acquiring link and graph theory attributes;

s500, establishing a PD diagnosis prediction model;

and S600, predicting a result.

The data acquisition in step S100 is specifically to use a PET/CT scanner to scan an object, establish a 3-dimensional (3D) model, and reconstruct an image by using an ordered subset expectation-maximization algorithm with 6 iterations and 6 subset methods.

The data preprocessing of step S200 is specifically to utilize SPM-based toolkit to perform data preprocessing, normalize the single 18F-FDG PET image volume space to a standard stereotactic Montreal Neurology Institute (MNI) space with linear and nonlinear 3D transforms, and apply an Automatic Anatomical Labeling (AAL) atlas to segment the cerebral cortex into 90 regions.

The metabolic network construction of the step S300 is specifically to use JS divergence to capture statistical relationships of brain glucose metabolic similarity in any 2 regions, and further to depict individual metabolic connections to use 90 ROIs from the AAL map set to represent brain nodes so as to depict individual metabolic networks;

JSs(P||Q)＝exp(-D_JS (P||Q))；

wherein,

p and Q represent probability density functions of different ROIs, M0.5 × (P + Q), D_kL (. |. cndot.) represents KL divergence.

The step S400 is specifically to extract global and local attributes based on the constructed metabolic network, where the global attributes include a clustering coefficient (C _ p), a feature path length (L _ p), a normalized clustering coefficient (γ), a normalized feature path length (λ), a small world (σ), a global efficiency (E _ global), and a modularization score (Q); the local attributes include degree centrality, node efficiency, intermediary centrality, shortest path length, and node clustering coefficients.

The step S500 is specifically to fuse, train and classify different connection information based on the multi-core SVM, sparsify the network under different sparse thresholds (0.02-0.5, step length is 0.01), obtain the sum of 49 values of the corresponding node attribute under the sparse threshold, train the classifier using the sum of 49 values of each node (area under the curve, AUC) as the input of the attribute, and establish the PD diagnosis prediction model.

Example 2

A total of 49 patients (33 male and 16 female participants, 53.94 ± 11.16 years) diagnosed as idiopathic PD according to the international association of parkinsonism and dyskinesias (MDS) diagnostic criteria and continuously enrolled in the study during months 1 and 12 in 2018 to 2019 using an 18F-FDG PET scan. Patients with a history of head injury, stroke, intracranial surgery, psychiatric disease, and substance use disorder were excluded. Detailed clinical participant information is presented in table 1. In addition, 49 HC's of matched age, education, and gender distribution were enrolled randomly to obtain normative data. None of the HCs has a history of cognitive disorders, psychiatric disorders, central nervous system disorders or head injuries. All participants provided written informed consent.

TABLE 1 demographic and clinical characteristics of PD patients and HC

A single 18F-FDG PET image volume space is normalized to a standard MNI space with linear and non-linear 3D transforms. To facilitate comparisons among all participants, a de-baseline operation was performed. After that time, the user can use the device,an Automatic Anatomical Labeling (AAL) map was applied to segment the cerebral cortex into 90 regions (45 regions per hemisphere without cerebellum). After the corresponding image is acquired, the JSSE (JSs (P | | Q) ═ exp (-D) is used_JS (P | | Q))) to construct a metabolic network. The constructed metabolic network further extracts global and local graph theory attributes, wherein the global graph theory attributes are as follows:

TABLE 2 Global graph attribute of the constructed metabolic network

A_r ,assortativity；C_p ,clustering coefficient；E_global ,global efficiency；E_local ,local efficiency；Hr,hierarchy；L_p ,characteristic path length；NC,normal control；PD,Parkinson's disease；Q,modularity score；S_r ,synchronization；γ,normalized clustering coefficient；λ,normalized characteristic path length；σ,small-world.*p-value<0.05.

The consensus junction was obtained as in FIG. 2:

finally, the classification results by using various modes and multiple cores are as follows: it can be seen that the PD classification diagnosis can be effectively performed.

TABLE 3 accuracy of classification

Note:C+G+N methods are significantly superior to Connection,Global,and Nodal under 95％confidence interval with p-value equals to 0.0482,4×10^-6 and0.0115 respectively。

The above embodiments only express a certain implementation manner of the present invention, and the description is specific and detailed, but not to be understood as limiting the scope of the present invention; it should be noted that, for those skilled in the art, without departing from the concept of the present invention, several variations and modifications can be made, which are within the protection scope of the present invention; therefore, the protection scope of the present patent should be subject to the appended claims.

Claims

1. A Parkinson's disease prediction method based on an 18F-FDG PET metabolic network is characterized by comprising the following steps: collecting and processing brain CT data of an object, constructing a metabolic network to obtain links and graph attribute, fusing different connection information based on the multi-core SVM, training and classifying, establishing a PD diagnosis prediction model, and predicting by using the trained model.

2. The method for predicting the Parkinson's disease based on the 18F-FDG PET metabolic network according to claim 1, which is characterized by comprising the following steps:

s100, data acquisition;

s200, preprocessing data;

s300, constructing a metabolic network;

s400, extracting features, and acquiring link and graph theory attributes;

s500, establishing a PD diagnosis prediction model;

s600, predicting a result.

3. The Parkinson's disease prediction method based on the 18F-FDG PET metabolic network according to claim 2, which is characterized in that: the data acquisition of step S100 is specifically to use a PET/CT scanner to scan an object, then build a 3-dimensional (3D) model, and use an ordered subset expectation-maximization algorithm with 6 iterations and 6 subset methods to reconstruct an image.

4. The Parkinson's disease prediction method based on the 18F-FDG PET metabolic network according to claim 2, which is characterized in that: the data preprocessing of step S200 is specifically to utilize SPM-based toolkit to perform data preprocessing, normalize the single 18F-FDG PET image volume space to a standard stereotactic Montreal Neurology Institute (MNI) space with linear and nonlinear 3D transforms, and apply an Automatic Anatomical Labeling (AAL) atlas to segment the cerebral cortex into 90 regions.

5. The Parkinson's disease prediction method based on the 18F-FDG PET metabolic network according to claim 2, which is characterized in that: the metabolic network construction of the step S300 is specifically to use JS divergence to capture statistical relationships of brain glucose metabolic similarity in any 2 regions, and further to depict individual metabolic connections to use 90 ROIs from the AAL map set to represent brain nodes so as to depict individual metabolic networks;

JSs(P||Q)＝exp(-D_JS (P||Q))；

wherein,

6. The Parkinson's disease prediction method based on the 18F-FDG PET metabolic network according to claim 2, which is characterized in that: the step S400 is specifically to extract global and local attributes based on the constructed metabolic network, where the global attributes include a clustering coefficient (C _ p), a feature path length (L _ p), a normalized clustering coefficient (γ), a normalized feature path length (λ), a small world (σ), a global efficiency (E _ global), and a modularization score (Q); the local attributes include degree centrality, node efficiency, intermediary centrality, shortest path length, and node clustering coefficients.

7. The Parkinson's disease prediction method based on the 18F-FDG PET metabolic network according to claim 2, characterized in that: the step S500 is specifically to fuse, train and classify different connection information based on the multi-core SVM, sparsify the network under different sparse thresholds (0.02-0.5, step length is 0.01), obtain the sum of 49 values of the corresponding node attribute under the sparse threshold, train the classifier using the sum of 49 values of each node (area under the curve, AUC) as the input of the attribute, and establish the PD diagnosis prediction model.