CN114628036B

CN114628036B - Brain ischemia risk prediction platform based on neural network

Info

Publication number: CN114628036B
Application number: CN202210531969.6A
Authority: CN
Inventors: 黄清; 贺威; 刘运海; 王淞; 庹佳
Original assignee: Xiangya Hospital of Central South University
Current assignee: Xiangya Hospital of Central South University
Priority date: 2022-05-17
Filing date: 2022-05-17
Publication date: 2022-08-02
Anticipated expiration: 2042-05-17
Also published as: CN114628036A

Abstract

The invention relates to a neural network-based cerebral ischemia risk prediction platform, which comprehensively considers carotid artery plaque data and a cerebral magnetic resonance image, firstly, by analyzing the carotid artery plaque data, obviously non-abnormal plaques (no risk) and obviously abnormal plaques (high risk) are divided, carotid artery plaques which are difficult to judge and are not accompanied with obvious stenosis are left, then by a cerebral magnetic resonance image, the conventional research on the carotid artery plaques per se is broken through, the association between the plaques and a cerebral structure is explored, the carotid artery plaques can be predicted from the early change of the brain based on a neural network module, the height of the cerebral ischemia risk is predicted, the neural network-based cerebral ischemia risk prediction platform can be used for the prediction and prevention of the conditions of acute stroke, chronic cerebral ischemia, vascular dementia and the like, and the neural network-based cerebral ischemia risk prediction platform is used for clinical prediction and has very great positive effect on the prevention of the disease, the diagnosis accuracy can be improved, and the lives of patients can be saved to a greater extent and more timely.

Description

Brain ischemia risk prediction platform based on neural network

Technical Field

The invention relates to an image analysis technology, in particular to a cerebral ischemia risk prediction platform.

Background

Stroke, heart disease, and malignancies collectively make up three major deaths in most countries. Wherein, the stroke is more than malignant tumor and becomes the first consistent death cause in China. In particular, ischemic stroke accounts for 60% -80% of all strokes. Therefore, it is very necessary to develop a method for preventing stroke and stroke-related complications, particularly ischemic stroke and vascular dementia.

Previous studies have focused on changes in brain function in patients with severe stenosis in the carotid artery, and less on patients with carotid artery disease without mild stenosis or plaque. However, more and more evidence shows that: the almond body and the hypothalamus are important brain structures for regulating and controlling the inflammatory response of atherosclerosis, and carotid atherosclerotic lesions with different degrees of human bodies and the subcortical structure-high cortical brain region of the brain are causally and controllably fed back. It is currently believed that the inflammatory immune response within the plaque is critical to the vulnerability of carotid plaque by promoting oxidative stress, lipid metabolism, and the overall process of neonatal trophoblast and even feedback central neurotransmitters in mediating vulnerable plaque progression and the development of thrombotic events. The research proves that; the high cortex ACC (anterior cingulate cortex) and mPF (medial prefrontal cortex) in the brain are mutually associated with the subcortical structure (amygdala body and hypothalamus) for controlling the autonomic nervous system through a large number of nodes, and the central neurotransmitter is regulated to mediate the activity of inflammatory markers in plaques, so that the haemodynamics change near the vulnerable plaques is coordinated, and the rupture of the vulnerable plaques of carotid arteries is promoted. The default network of the brain (DMN) is the brain region where the brain appears to be consistently active in resting state, and the major functional nodes include the medial prefrontal cortex (mPFC), anterior cuneiform, horn gyrus and Anterior Cingulate Cortex (ACC) and Posterior Cingulate Cortex (PCC)), which are involved in the regulation of inflammatory immune response-mediated neuropsychiatric diseases. In particular, ACC and mPFC within the default network are high-level cortex involved in systemic inflammatory activity. Studies have found that ACC and mPFC within the default network regulate brain inflammatory neurotransmitters involved in the process of carotid atherosclerosis. On the contrary, in different degrees of carotid atherosclerosis, even in the early stage of carotid plaque formation, the brain default network (DMN) begins to receive negative feedback regulation, and symptoms such as headache, dizziness, insomnia and the like gradually appear clinically, but cannot pass the existing imaging means for detection and analysis. Therefore, how to explore the early occult change of the brain of patients with different degrees of carotid atherosclerotic plaques, search a specific imaging marker of the early brain function change mediated by the carotid plaques, and construct the association between the two markers is an urgent technical problem to be solved for preventing cerebral ischemic injury, including ischemic stroke, chronic cerebral ischemia and vascular dementia.

Disclosure of Invention

In order to solve the above technical problems, the present invention provides a neural network-based cerebral ischemia risk prediction platform, including: the device comprises an acquisition device, a control device and an output device which are connected in sequence;

the acquisition device is used for acquiring carotid plaque data and a brain magnetic resonance image;

a control device, comprising:

the judging module is connected with the acquisition device and used for judging whether the cerebral ischemia risk is risk-free, high risk or risk to be evaluated according to the carotid plaque data;

the neural network module is connected with the acquisition device and the judgment module and is used for judging whether the cerebral ischemia risk is low risk or second high risk according to the cerebral magnetic resonance image when the cerebral ischemia risk is the risk to be evaluated;

and the output device is connected with the judging module and the neural network module and is used for outputting a judgment result of the cerebral ischemia risk.

Further, the judging module includes:

the first judging unit judges whether the thickening of the local wall of the carotid plaque is less than 1.5 mm;

a first determination unit which determines that the cerebral ischemia risk is no risk if the cerebral ischemia risk is determined to be no risk;

a second judging unit, if not, judging whether the carotid artery stenosis degree is less than 50%;

a second judging unit, if yes, judging the cerebral ischemia risk as the risk to be evaluated;

and a third judging unit, if not, judging the cerebral ischemia risk as high risk.

Further, a neural network module, comprising:

the input unit is connected with the acquisition device and used for inputting the brain magnetic resonance image;

the first feature extraction unit is connected with the input unit and used for extracting brain volume data according to the brain magnetic resonance image;

the second characteristic extraction unit is connected with the input unit and used for extracting functional network topology data according to the brain magnetic resonance image;

the third characteristic extraction unit is connected with the input unit and used for extracting white matter fiber bundle data according to the brain magnetic resonance image;

and the fusion unit is connected with the first feature extraction unit, the second feature extraction unit and the third feature extraction unit and used for judging the cerebral ischemia risk to be low risk or secondary high risk according to the brain volume data, the functional network topology data and the white matter fiber bundle data.

Further, the neural network module further includes:

the system comprises a sample acquisition unit, a data acquisition unit and a data processing unit, wherein the sample acquisition unit is used for acquiring a magnetic resonance sample image and marking a cerebral ischemia risk result on the magnetic resonance sample image so as to construct a training sample;

and the module correcting unit is used for inputting the training samples into the input unit so as to correct the parameters of the neural network module.

Further, the first feature extraction unit includes:

the first segmentation unit is used for segmenting the brain magnetic resonance image to obtain a white matter segmentation image, a gray matter segmentation image and a cerebrospinal fluid segmentation image;

the first calculation unit is used for calculating the brain volume data according to the white matter segmentation image, the gray matter segmentation image and the cerebrospinal fluid segmentation image.

Further, the second feature extraction unit includes:

the second segmentation unit is used for segmenting the brain magnetic resonance image to obtain 90 brain areas;

and the second computing unit is used for judging the functional connection between every two nodes according to each brain area as a node and constructing functional network topology data.

Further, a second calculation unit includes:

the correlation coefficient calculating unit is used for calculating the Pearson correlation coefficients between every two 90 brain areas;

the second small judgment unit is used for judging whether the Pearson correlation coefficient between every two brain areas exceeds a set threshold value or not, judging that the two brain areas are connected if the Pearson correlation coefficient exceeds the set threshold value, and judging that the two brain areas are not connected if the Pearson correlation coefficient does not exceed the set threshold value;

and the network construction small unit is used for constructing functional network topology data according to the judgment result of whether the connection exists between every two brain areas.

Further, a third feature extraction module includes:

the brain area extraction unit is used for calculating the extra scalp stripped before the tensor to obtain a mask image and determine a tensor calculation range;

and the third calculating unit calculates tensor parameters according to the tensor calculation range and is used for analyzing white matter fiber bundle data.

The cerebral ischemia risk prediction platform based on the neural network comprehensively considers carotid artery plaque data and a brain magnetic resonance image, firstly, carotid artery plaque data is analyzed, obvious plaque without abnormality (no risk) and obvious plaque with abnormality (high risk) are divided, carotid artery plaque which is difficult to judge and does not accompany with obvious stenosis (carotid artery plaque with local wall thickening more than or equal to 1.5mm and carotid artery stenosis degree less than 50 percent), namely clinical asymptomatic plaque is remained, conventional research on the carotid artery plaque per se is broken through the brain magnetic resonance image, correlation between the plaque and a brain structure is explored, the carotid artery plaque can be predicted from early (ultra-early) hidden change of the brain based on a neural network module, the cerebral ischemia risk of the carotid artery plaque is predicted, and the carotid artery plaque is a second-high risk plaque with high rupture probability or a low-risk plaque with small rupture probability, or produce chronic cerebral ischemic symptoms, or cause vascular dementia to occur. The method can be used for assessing cerebral ischemia risks, predicting and preventing conditions such as stroke, chronic cerebral ischemia and dementia, has a very great positive effect when being used for clinical prediction and preventing diseases, can improve diagnosis accuracy, and saves the lives of patients to a greater extent and more timely.

Drawings

FIG. 1 is a block diagram of a neural network-based cerebral ischemia risk prediction platform according to an embodiment of the present invention;

FIG. 2 is a block diagram of a judgment module of a neural network-based cerebral ischemia risk prediction platform according to an embodiment of the present invention;

fig. 3 is a structural block diagram of a neural network module of the neural network-based cerebral ischemia risk prediction platform according to an embodiment of the present invention.

Detailed Description

As shown in fig. 1, the present invention provides a brain ischemia risk prediction platform based on a neural network module, including: the device comprises an acquisition device, a control device and an output device which are connected in sequence;

the acquisition device is used for acquiring carotid plaque data and a brain magnetic resonance image; in particular, the acquisition device may be selected from, but not limited to, an ultrasound scanner and a magnetic resonance scanner. More specifically, to obtain carotid plaque data, an ultrasound scanner (such as Philip iu22 intelligent ultrasound diagnostic system, a convex array probe C5-1 combined with a linear array probe L9-3) is optionally but not limited to be used for carotid color ultrasound examination of an examinee, and the carotid artery color ultrasound examination is performed to measure bifurcations of bilateral internal carotid arteries, external carotid arteries, common carotid arteries, and carotid arteries, preferably, the head of the carotid artery color ultrasound probe is tilted backwards to fully expose the neck of the carotid artery, the color ultrasound probe is gently placed on the neck of the patient to clearly see the artery wall and plaque outline, and then detection of each part of each blood vessel is performed. The magnetic resonance image of the brain, optionally but not exclusively by using a magnetic resonance apparatus for head or whole body scanning, is obtained by a doctor familiar with the imaging department. Subsequently, to extract brain volume data, a linehead T1 weighted structure phase, a T2 weighted structure phase, a DWI sequence, a FLAIR sequence, and a three-dimensional craniocerebral volume magnetic resonance imaging (BRAVO) sequence scan are optional but not limited. The BRAVO sequence parameters are: pulse repetition Time (TR): 7.792ms, echo Time (TE): 2.984ms, flip angle: 7 °, number of layers 188, layer thickness 1mm, layer spacing 1mm, matrix: 256' 256, voxel size: 1' 1mm 3. For extracting functional network topology data, optionally but not limited to using rsfMRI, data acquisition uses an EPI sequence, and specific parameters are as follows: TR: 2000ms, TE: 30ms, flip angle: 90 °, number of layers 32, layer thickness 4mm, matrix: 64' 64, voxel size: 3' 3mm3, scan time 360 s. For extracting white matter fiber bundle data, scanning can be performed by selecting but not limited to a line echo imaging sequence, and the DTI parameters are: TR (repetition time) = 12000 ms, TE (echo time) = 72.5 ms, matrix 256 × 256, FOV (field of view) = 230 × 230mm, layer thickness 3mm, diffusion sensitivity gradient direction 32, diffusion sensitivity coefficient (b) = 0, 1000 s/mm 2. Conventional T1WI, T2WI, Flair, Diffusion-weighted imaging (DWI) scans were performed prior to DTI scan to exclude past and acute stroke and other organic lesions in the cranium, as well as to assess intracranial leukogenic conditions. More preferably, the acquiring device, which is optionally but not limited to a device with image-text input function such as a touch screen and image acquisition, directly inputs the carotid plaque data and the brain magnetic resonance image in a manual or automatic manner. More preferably, the acquiring device, which is optionally but not limited to a device with image-text receiving function, directly acquires the carotid plaque data and the brain magnetic resonance image by connecting with a B-ultrasonic device and a magnetic resonance device.

A control device, comprising:

the judging module is connected with the acquisition device and used for judging whether the cerebral ischemia risk is risk-free, high risk or risk to be evaluated according to the carotid plaque data; specifically, according to carotid plaque data, if B-ultrasound shows that local wall thickening is less than 1.5mm, no carotid plaque is substantially defined, and no risk is determined; if B-mode ultrasound shows local wall thickening of more than or equal to 1.5mm and carotid plaque (symptomatic plaque with obvious stenosis) with carotid stenosis degree of more than or equal to 50%, high risk is defined; if the B ultrasonic shows that the local wall thickening is more than or equal to 1.5mm, but the carotid artery plaque (asymptomatic plaque without obvious stenosis) with the carotid artery stenosis degree of less than 50 percent is defined as the risk to be evaluated, clinically, no obvious abnormality is shown at present, but the carotid artery plaque can be ruptured in the future to cause the risks of acute cerebral infarction, chronic cerebral ischemia, vascular dementia and the like or cause no risk, and the plaque is the plaque which is required to be evaluated in the cerebral ischemia risk estimation platform based on the neural network module.

And the neural network module is connected with the acquisition device and the judgment module and is used for judging whether the cerebral ischemia risk is low risk or second high risk according to the cerebral magnetic resonance image when the cerebral ischemia risk is the risk to be evaluated. Specifically, the future condition of the plaque to be evaluated is predicted not from the plaque itself but from the magnetic resonance image of the brain, and the plaque is a low-risk plaque with a small rupture probability or a second-high-risk plaque with a large rupture probability.

And the output device is connected with the judging module and the neural network module and is used for outputting a judgment result of the cerebral ischemia risk. Specifically, the output device may be optionally, but not limited to, indicated by light, color, etc., for example, green for no risk, blue for low risk, yellow for next highest risk, and red for high risk; or directly output in text.

In the embodiment, a cerebral ischemia risk estimation platform based on a neural network is provided, carotid plaque data and a cerebral magnetic resonance image are comprehensively considered, carotid plaque data which is obviously abnormal (no risk) and plaque which is obviously abnormal (high risk) are firstly analyzed, carotid plaque which is difficult to judge and is not accompanied with obvious stenosis (carotid plaque with local wall thickening of more than or equal to 1.5mm and carotid stenosis degree of less than 50 percent), namely clinical asymptomatic plaque, is remained, conventional research on carotid plaque per se is broken through the cerebral magnetic resonance image, association of plaque and a cerebral structure is explored, the carotid plaque can be prejudged from early change of the brain based on a neural network module, the height of the cerebral ischemia risk is estimated, and the carotid plaque is a second-high risk plaque with high rupture probability or a low-risk plaque with small rupture probability, the method can be used for assessing the risk of cerebral ischemia, predicting and preventing the conditions of acute ischemic stroke, chronic cerebral ischemia, vascular dementia and the like, has very great positive effects on clinical prediction and prevention, can improve the diagnosis accuracy, and saves the lives of patients to a greater extent and more timely.

Preferably, as shown in fig. 2, the determining module may optionally but not limited to include:

a first determination unit that determines that the risk of cerebral ischemia is risk-free if the determination unit determines that the risk of cerebral ischemia is risk-free;

and a third judging unit, if not, judging that the cerebral ischemia risk is high risk.

Preferably, as shown in FIG. 3, the neural network module, optionally but not limited to, comprises

and the first feature extraction unit is connected with the input unit and used for extracting brain volume data according to the brain magnetic resonance image. Specifically, the brain volume data may optionally, but not exclusively, include one or more of white matter volume, gray matter volume and cerebrospinal fluid volume, whole brain volume (TCBV, which is the ratio of brain parenchymal volume to total brain volume, wherein brain parenchymal volume is the sum of brain white matter volume and brain gray matter volume after removal of cerebrospinal fluid volume), white matter volume of each brain region, gray matter volume and cerebrospinal fluid volume, and the like.

And the second characteristic extraction unit is connected with the input unit and used for extracting functional network topology data according to the brain magnetic resonance image. Specifically, the functional network topology data is optionally, but not limited to, a connection situation of each brain region.

And the third feature extraction unit is connected with the input unit and used for extracting white matter fiber bundle data according to the brain magnetic resonance image. Specifically, the white matter fiber bundle data, optionally but not limited to DTI parameters, is used to calculate tensor parameters such as FA, MD, AD, and RD for each test for TBSS analysis.

And the fusion unit is connected with the first feature extraction unit, the second feature extraction unit and the third feature extraction unit and used for predicting the cerebral ischemia risk to be low risk or secondary high risk according to the brain volume data, the functional network topology data and the white matter fiber bundle data.

In the embodiment, a specific embodiment of the neural network module is provided, according to theoretical research, feature dimension reduction and feature extraction are carried out on multiple types of data, association between cerebral change and plaque rupture is explored on the basis of three features of brain volume data, functional network topological data and white matter fiber bundle data, the neural network module for predicting the future trend of carotid plaque is constructed, the theoretical research is applied to specific practice, qualitative analysis is carried out on the properties of the carotid plaque, a neural regulation mechanism for the occurrence and development of different types of plaque in carotid is clarified, early warning of the occurrence of atherosclerotic cerebral infarction and the like is achieved, and the neural network module is strong in association and high in accuracy.

More preferably, the neural network module further includes:

the system comprises a sample acquisition unit, a data acquisition unit and a data processing unit, wherein the sample acquisition unit is used for acquiring a magnetic resonance sample image and marking a cerebral ischemia risk result on the magnetic resonance sample image so as to construct a training sample; specifically, the magnetic resonance sample image can be selected but not limited to be acquired as a sample image of a cerebral magnetic resonance image of an examiner who does not have a carotid artery plaque (a carotid artery plaque with a local wall thickening of more than or equal to 1.5mm and a carotid artery stenosis degree of less than 50%), and the cerebral ischemia risk is marked as low risk or secondary high risk according to a later-stage real result, namely whether the plaque is ruptured or not, so as to subsequently return an output result of the same type and represent a risk assessment result corresponding to the cerebral magnetic resonance image. Specifically, the step can be operated during the initial training period of the neural network module, and samples with results and even samples with wrong judgment can be added in the using process of the cerebral ischemia risk estimation platform so as to continuously optimize the neural network module and improve the prediction precision and the judgment accuracy of the neural network module. More specifically, to improve sample reference, the following options are selected, but not limited to: 1) the age is 40-70 years old, and the nature is not limited; (2) MMSE was 27 min, montreal grant ‎ (montreal scientific assessment, MoCA) scored 26 min; (3) there is only one type of neck vascular plaque, hypoechoic, isoechoic or hyperechoic; (4) degree of stenosis of neck vessels < 50%; (5) no obvious abnormality is found by conventional craniocerebral MRI; (6) right handedness; (7) no cerebral infarction or transient ischemic attack in the past; (8) there was no history of serious organic and psychiatric disease and drug dependence.

And the module correcting unit is used for inputting the training samples into the input unit so as to correct the parameters of the neural network module. Specifically, the loss function is continuously reduced in an iterative manner by inputting a training sample (set) to optimize parameters, so that a neural network module with better performance is obtained. More specifically, the optimal model may be optionally, but not limited to, trained by a supervised learning algorithm such as a support vector machine, linear regression, etc.

In the embodiment, the sample acquisition unit and the module correction unit are additionally arranged, so that the neural network module can be continuously corrected and optimized in the training process of the neural network module and the use process of the platform, validity verification can be performed in test data, and the accuracy and timeliness of risk assessment are further improved.

More preferably, the first feature extraction unit, optionally but not limited to, includes a first segmentation unit, configured to segment the brain magnetic resonance image to obtain a white matter segmentation image, a gray matter segmentation image, and a cerebrospinal fluid segmentation image, and specifically, the single image may be obtained by segmentation using a three-dimensional brain image segmentation method; the first calculation unit is used for calculating the brain volume data according to the white matter segmentation image, the gray matter segmentation image and the cerebrospinal fluid segmentation image. Preferably, the first feature extraction unit, optionally but not limited to, further includes a first preprocessing unit, which preprocesses the magnetic resonance image of the brain to obtain a preprocessed magnetic resonance image of the brain, and specifically includes: the first conversion small unit is used for converting the brain magnetic resonance image into an NIFIT format (generally, a DICOM format is adopted, and format conversion is firstly carried out) to obtain a converted first brain magnetic resonance image; the first small clipping unit is configured to clip a target region from the converted first brain magnetic resonance image and perform template matching to obtain a preprocessed brain magnetic resonance image (specifically, the obtained brain magnetic resonance image may be selected but not limited to remove a part below the head to obtain a target region including only the head region, and then the image is adjusted according to a montreal template until the template matching is completed). More specifically, but not limited to, the following may be selected and included: the first detection unit is used for detecting whether the quality of the white matter segmentation image, the gray matter segmentation image and the cerebrospinal fluid segmentation image is qualified or not, and if the quality of the white matter segmentation image, the gray matter segmentation image and the cerebrospinal fluid segmentation image is not qualified, the white matter segmentation image is re-segmented or rejected; and the first smoothing unit is used for further smoothing the qualified white matter segmentation image, the gray matter segmentation image and the cerebrospinal fluid segmentation image, and preferably smoothing by adopting a Gaussian smoothing kernel FWHM 8. More specifically, the first calculating unit, optionally but not limited to, includes: the element calculation small unit is used for calculating a white matter volume, a gray matter volume and a cerebrospinal fluid volume according to the white matter segmentation image, the gray matter segmentation image and the cerebrospinal fluid segmentation image; a whole brain volume calculation small unit for calculating a whole brain volume from the white matter volume, the gray matter volume and the cerebrospinal fluid volume as brain volume data (specifically, the whole brain volume (TCBV) is a ratio of the brain parenchyma volume to the total brain volume. More specifically, the system further comprises a brain region volume calculating small unit, which is used for calculating a white matter volume, a gray matter volume and a cerebrospinal fluid volume of each brain region according to the white matter segmentation image, the gray matter segmentation image and the cerebrospinal fluid segmentation image as brain volume data.

In this embodiment, a specific structure of the first feature extraction unit is given, a specific manner of how to extract brain volume data from a magnetic resonance image of a brain is described, voxels are regarded as basic constituent units of the brain, segmentation of White Matter (WM), Gray Matter (GM) and cerebrospinal fluid (CSF) is performed on the brain structure data, statistical testing, comparison and analysis of differences among groups of constituent voxels of the brain tissue are performed, volume and density differences of the brain tissue are quantitatively detected, and then abnormalities of brain morphology and specific positions of different brain regions are accurately displayed, and white matter volume, gray matter volume, cerebrospinal fluid volume, whole brain volume and volumes of local constituents are extracted.

More preferably, the second feature extraction unit, optionally but not limited to, includes: the second segmentation unit is used for segmenting the brain magnetic resonance image to obtain 90 brain areas; and the second computing unit judges the functional connection between every two nodes according to each brain area as one node, and constructs functional network topology data. Specifically, the second segmentation unit, optionally but not limited to, segments the whole brain into 90 brain regions using an aal (atomic Automatic labeling) brain template. The second computing unit, optionally but not limited to, constructs the functional network according to the average time series correlation between the brain regions. Specifically, the average time series of each brain region of the rs-fMRI image data is obtained by weighting each brain region time series. More specifically, the second calculating unit, optionally but not limited to, includes: a correlation coefficient calculating unit for calculating a Pearson (r) correlation coefficient between every two of the 90 brain regions; the second small judgment unit is used for sequentially judging whether the Pearson correlation coefficient of every two brain areas exceeds a set threshold value, judging that the every two brain areas are connected if the Pearson correlation coefficient exceeds the set threshold value, and judging that the every two brain areas are not connected if the Pearson correlation coefficient does not exceed the set threshold value; and constructing a small unit by the network, and constructing functional network topology data according to the judgment result. Specifically, the functional network topology data is, but not limited to, a 90 × 90 correlation matrix. Preferably, the correlation matrix is, but not limited to, a 90 × 90 binary matrix, and if the pearson correlation coefficient of each two brain regions exceeds a set threshold, i.e., there is a connection therebetween, the value of the corresponding position of the matrix is set to 1, and if the pearson correlation coefficient of each two brain regions does not exceed the set threshold, i.e., there is no connection therebetween, the value of the corresponding position of the matrix is set to 0. More specifically, the functional network topology data may optionally, but not exclusively, include global network parameters, node network parameters, and hub nodes. More specifically, the second extraction module may further include, but is not limited to: the second preprocessing unit is used for preprocessing the brain magnetic resonance image to obtain a preprocessed brain magnetic resonance image; specifically, the second preprocessing unit may be selected from, but not limited to, the following: the second small conversion unit is used for converting the brain magnetic resonance image into an NIFIT format (generally, a DICOM format, the format conversion is firstly carried out); the second screening small unit is used for removing the head portrait data of the first 5 time points and continuously processing the rest 175 time point data so as to reduce the influence of bad early data quality on the result; a second correction cell comprising: time correction is carried out, and the influence of different acquisition times of all parts of the brain on results is eliminated; position correction, namely adjusting each brain magnetic resonance image to be at the same position in the direction of X, Y, Z, and correcting the head position; standardized correction, namely registering the head portrait data of each subject to an EPI template, and reducing the influence on an analysis result due to different head forms; removing linear drift: the rsfMRI image data signal has a time sequence, eliminating the linear drift of the sequence; regression covariates: reducing the effects of WM and CSF signals; filtering and screening: selecting a wave band of 0.01-0.08 Hz; removing: if the head data head of the subject moves and translates for more than 3mm and the rotation angle exceeds 3 degrees, the subject is tried to be rejected.

In this embodiment, a specific structure of the second feature extraction unit is given, and how to extract a specific form of functional network topology data according to a brain magnetic resonance image is described in detail, the brain is divided into 90 brain regions, then a graph theory method is adopted, the brain network is regarded as a stereogram composed of nodes and edges, the brain regions or voxels are nodes, the structures between the nodes are connected as edges, the connection of the brain is quantized, and then the change of the brain structure functional network nodes or edges under various disease states, that is, the functional network topology data, is analyzed. Furthermore, the binary matrix of 01 is adopted to represent functional network topology data, so that the calculation is simple, the operation is easy, and the function is easy to understand.

More preferably, the third feature extraction unit, optionally but not limited to, includes: the brain area extraction unit is used for calculating the extra scalp stripped before the tensor to obtain a mask image (mask) so as to determine the tensor calculation range; and a third calculating unit for calculating tensor parameters (such as FA, MD, AD and RD) of each tested object for analyzing white matter fiber data. More specifically, the third feature extraction unit, which may optionally but not exclusively include a third preprocessing unit, preprocesses the brain magnetic resonance image to obtain a preprocessed brain magnetic resonance image; specifically, the method comprises the following steps of: checking basic data parameters, gradient direction, signal-to-noise ratio and head movement condition; data format conversion: converting the original DICOM format into a 4D NifTi format by adopting Micron software; correcting the head-moving vortex: eliminating head movement in the scanning process and deformation caused by the head movement and vortex; correcting the gradient direction: and adjusting the original gradient direction according to the condition of eddy current correction.

In this embodiment, a specific structure of the third feature extraction unit is given, and how to extract a specific form of white matter fiber bundle data according to the brain magnetic resonance image is explained in detail, so that the structure is simple, and the data is accurate.

The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. A brain ischemia risk prediction platform based on a neural network is characterized by comprising: the device comprises an acquisition device, a control device and an output device which are connected in sequence;

a control device, comprising:

a neural network module, comprising:

the fusion unit is connected with the first feature extraction unit, the second feature extraction unit and the third feature extraction unit and used for judging the cerebral ischemia risk to be low risk or secondary high risk according to the brain volume data, the functional network topology data and the white matter fiber bundle data;

2. The platform of claim 1, wherein the determining module comprises:

3. The platform of claim 1, wherein the neural network module further comprises:

4. The platform for predicting risk of cerebral ischemia according to any one of claims 1-3, wherein the first feature extraction unit comprises:

5. The platform of claim 4, wherein the second feature extraction unit comprises:

6. The platform of claim 5, wherein the second computing unit comprises:

7. The platform of claim 6, wherein the third feature extraction module comprises: