Intelligent disease diagnosis device based on brain structure connection identifier and application thereof

Abstract

The invention provides an intelligent disease diagnosis device based on a brain structure connection identifier and application thereof. The device comprises (1) a data acquisition module; (2) the data processing module is used for reconstructing a spin distribution function of the data and enabling the spin distribution function to be distributed in a standard space; (3) a data projection module for projecting the spin distribution functions of the patient and the healthy individual into a standard space to obtain a Z value in the maximum direction in voxels; and (4) a Connectome Identifier extraction module is used for carrying out dimensionality reduction on the Z value to form a one-dimensional characteristic value. The device can be used for diagnosis of various brain diseases and judgment of focus positions and severity.

Description

Technical Field

[0001] The present application belongs to the field of medical imaging data processing. Specifically, the present application provides a disease intelligent diagnosis device based on brain structure connection identifiers and its application. Background Art

[0002] The clinical application of diffusion tensor imaging (DTI) technology has been confirmed by many studies. For example, through cadaver anatomical experiments, the validity of the tracking results of medium and large white matter fiber bundles obtained using DTI technology can be confirmed. In subsequent studies, DTI was used for preoperative evaluation of neurosurgery, and the distribution and morphology of the main white matter structures can be depicted before surgery. Common white matter fiber bundles reconstructed by DTI include: pyramidal tract, visual conduction pathway, and language conduction pathway. DTI can further achieve the purpose of preserving function for patients by presenting these structures, and has an important impact on the formulation of surgical strategies for complex lesions. However, DTI technology has its limitations. It cannot collect and analyze crossed fibers, cannot accurately map the starting and ending points of white matter fibers, and false positive and false negative fibers often appear during the tracking process, which is determined by the technical principle of DTI. Technically speaking, DTI can only collect the main restricted diffusion direction within a voxel, that is, the diffusion direction of water molecules after being wrapped by the capsule, and it is a relative value of the restricted diffusion direction, not an absolute measurement value. Specifically, the tissue characteristics are measured by eigenvalues ​​(λ1, λ2 and λ3). Among them, the axial diffusion capacity, represented by λ1, reflects the ability of water molecules to diffuse along the long axis of the fiber; the radial eigenvalue is divided into two perpendicular eigenvalues ​​λ2 and λ3. Fractional anisotropy (FA) is a diffusion scalar parameter commonly used in DTI technology. It reflects the degree of partial anisotropy, which is the proportion of the ability of water molecules to diffuse along a certain direction in the total diffusion capacity. The formula is:

[0003]

[0004] The value is between 0 and 1. Generally speaking, fibers with better myelination have higher FA values. When the degree of myelination decreases or neurons are lost, the FA value decreases. In tissues with dense fibers, the diffusion rate perpendicular to the fiber direction is inhibited, and the FA value increases; when the fiber structure is damaged by degenerative diseases or other diseases, the diffusion rate perpendicular to the fiber direction increases, and the FA value decreases. Therefore, evaluating the FA value of the brain can evaluate the structural changes of the tissue. Therefore, for complex fiber structures such as crossing and angulation within voxels, DTI does not collect this information from the beginning of data acquisition, so it will be mixed with a large number of variations in neural tissue structure. When reflecting histological characteristics, a large sample size of data is required to reveal the network function state of the brain. The deterministic tracking formed on this basis cannot reproduce the anatomical characteristics of white matter fibers found in fiber bundle anatomy and histological techniques. By the same principle, for white matter fibers embedded in edematous tissue, since DTI only considers the average direction of the voxel, it cannot be fully tracked and presented. Therefore, it is urgent to start from the acquisition, that is, to fully consider the acquisition technology of the non-restricted diffusion direction within the voxel.

[0005] In recent years, new fiber bundle mapping technologies such as high-angular-resolution diffusion imaging (HARDI) and DSI technology have emerged and can be used to solve these problems. Diffusion spectrum imaging (DSI) is a diffusion-based white matter imaging technology. Compared with traditional diffusion tensor imaging (DTI), DSI has more imaging directions and higher b-values. It is based on the probability density function (PDF) framework and can describe the three-dimensional distribution of microscopic displacements of spin subs in voxels in units of voxels, which is more accurate for the analysis of brain white matter fiber bundles. Brain white matter research is an important part of neuroscience research. DTI technology makes it possible to map white matter in vivo and can be widely used in the diagnosis and treatment of brain diseases. DTI technology uses the generalized q-sampling imaging (GQI) algorithm, the core model of which is the GQI model, and the corresponding diffusion index parameter is quantitative anisotropy (QA). QA is calculated from the peak value of the spin distribution function (SDF). The algorithm analyzes all fiber directions within a voxel, thereby solving the problem of tracking crossed fibers. Specifically, it can be analyzed from two aspects. First, in terms of the evaluation content, the tensor model of DTI evaluates diffusivity. If it is based on diffusivity, due to the characteristics of Brownian motion, restricted and unrestricted diffusion cannot be completely separated. The analysis of crossed fibers, turning fibers, and fiber starting and ending points can only be derived by post-processing reconstruction models (such as probabilistic tracking, etc.). However, for DTI with poor signal-to-noise ratio, it is often impossible to effectively restore it by relying on the model. The GQI algorithm used by DSI can distinguish restricted and unrestricted motion from signal acquisition. Subsequent multi-directional identification only requires the simplest linear judgment. Therefore, in principle, DSI will have higher accuracy than DTI; secondly, in terms of voxel signals, FA evaluates the average diffusion signal of all fibers in the voxel, while QA evaluates the diffusion signal of each single fiber subpopulation in the voxel, that is, for each fiber subpopulation in a specific direction in the voxel, QA will independently measure the diffusion direction and size of the subpopulation. For fiber crossings or corners, FA will be reduced, while QA may be almost unaffected.

[0006] For brain network diseases such as epilepsy, Alzheimer's disease, and Parkinson's disease, the location diagnosis of artificial intelligence still needs to be studied. For example, other non-invasive methods of epilepsy location, such as scalp EEG, cannot record EEG signals in deep brain areas and can only detect 6cm2 The coordinated discharge of 18 F-FDG PET is also subject to interference from low metabolism in distant parts and physiological low metabolism; invasive deep intracranial EEG has a limited number of sampling points, so the accuracy of lesion detection can only be guaranteed when there is a priori inference in non-invasive assessment. Therefore, there is currently no single modality that can achieve non-invasive localization of epilepsy. Summary of the invention

[0007] The present invention uses quantitative DSI acquisition to construct a template for the connection of healthy human brain structures, and obtains the intensity distribution of water molecule diffusion in each direction in each voxel (the unit of the smallest signal in 3D space, similar to the pixel after a camera is photographed, specifically the value corresponding to a coordinate in a 3D matrix). Then the patient's DSI data is collected, and its 3D matrix is ​​projected onto the standard template to obtain the distribution matrix of the patient's water molecule diffusion data in the standard space, completing the standardization of the space.

[0008] The data in the same voxel of the patient in the standard space is decomposed into vectors according to the direction of the corresponding voxel in the standard template. The multiple of the standard deviation of the patient's vector value and the vector value of the population standard template in the most important direction (i.e., Z transformation) is taken as the eigenvalue in that direction. The distribution of the whole brain eigenvalue is the ConnectomeIdentifier.

[0009] In previous studies, the inventors found that the Connectome Identifier of the same healthy subject at different time periods has individual specificity, that is, the Connectome Identifier based on the brain connection group can accurately lock the subject without unblinding. The inventors found that the Connectome Identifier is used in patients with temporal lobe epilepsy, combined with machine learning, to accurately identify whether the epilepsy is left-sided, right-sided or bilateral. In Alzheimer's patients, ConnectomeIdentifier can identify the severity of Alzheimer's lesions at different stages.

[0010] In one aspect, the present application provides a disease intelligent diagnosis device based on brain structure connection identifiers, the device comprising the following modules:

[0011] (1) Data acquisition module: used to obtain DSI data;

[0012] (2) Data processing module: Perform Q-space differential isomorphism reconstruction on the DSI data obtained in step (1), so that the spin distribution function is reconstructed in the individual space and the distribution of the diffusion direction in each voxel in the standard space is projected;

[0013] (3) Data projection module: The spin distribution functions of one or more healthy individuals in the standard space are projected together and averaged to obtain the mean and standard deviation of the diffusion vector length in the direction with the largest diffusion in the voxel; the spin distribution function of the patient individual is also projected into the standard space, the largest direction in the voxel of a single subject is taken, and the mean length at the group level is subtracted. mean , divided by the standard deviation length std , get the Z value of the maximum direction in the voxel: Z = (l-length mean ) / length std ;

[0014] (4) Connectome Identifier extraction module: The Z values ​​arranged in a three-dimensional matrix are encoded in the order of left-back-front to form a one-dimensional feature value, which is the Connectome Identifier.

[0015] Furthermore, the data acquisition module includes an MRI instrument.

[0016] Furthermore, the DSI data is DSI data of a patient.

[0017] Furthermore, the DSI data are DSI data of the patient and DSI data of one or more healthy individuals.

[0018] Furthermore, the DSI data is brain DSI data.

[0019] Furthermore, the functions performed by the data processing module include: processing the DSI data obtained in step (1) using DSI-Studio software, entering the DICOM format data in the original image, and packaging it together with the b-value table into an SRC format file; reading the SRC file to observe the range of brain parenchyma in the influence space, selecting a threshold that can cover the brain parenchyma but not include the skull, and performing Q-space differential isomorphism reconstruction.

[0020] Furthermore, the functions performed by the data processing module include: before the Q-space differential isomorphism reconstruction begins, TOPUP is used to correct the distortion caused by the magnetic field, and Eddy current is used to correct the head movement artifact.

[0021] Furthermore, in the data processing module, the diffusion sampling length ratio is 1.1 during Q-space differential isomorphism reconstruction, the fitting method is CDM, the spatial template is HCP1021, the directional distribution function interpolation parameter is 20 fold, the number of fiber processing within the voxel is 10, and the number of threads is 4.

[0022] Furthermore, the device also includes (5) a Connectome Identifier reading module for reading the Connectome Identifier.

[0023] In another aspect, the present application provides use of the above device in diagnosing brain diseases.

[0024] On the other hand, the present application provides the use of the above-mentioned device in determining the location of brain disease lesions.

[0025] On the other hand, the present application provides the use of the above device in distinguishing the severity of brain diseases.

[0026] Furthermore, the brain disease is epilepsy, Parkinson's disease, or Alzheimer's disease.

[0027] On the other hand, the present application provides a storage medium in which a program for executing the functions of the above modules is recorded.

[0028] The DSI data of healthy individuals in the present application can be obtained by detection by an MRI device in a diagnostic apparatus, or can be obtained from a database or DSI data of healthy individuals stored in a device or a storage medium.

[0029] Beneficial effects:

[0030] Previous artificial intelligence recognition was mainly based on diffusion scalars, which considered the average value of all directions of water molecule diffusion in the voxel, without considering the water molecule diffusion vector direction and structural connection factors. The present invention tracks the most important vector direction in the voxel, and then analyzes the ratio difference between patients and healthy people in this direction, taking into account both the diffusion direction and the diffusion strength, which is more in line with the law of disease development. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Show the basic principle of this application.

[0032] Figure 2 An example of a connection identifier for a patient;

[0033] Figure 3 An example of the effectiveness of artificial intelligence recognition for connection identifiers. DETAILED DESCRIPTION

[0034] Example 1 Data Acquisition

[0035] SIGNA Premier 3.0T MRI (GE, USA) with 64-channel head coil was used. The subjects were asked to lie on their backs, be quiet and keep their heads still. The hands were placed on both sides of the body during the scan. The scanning range was the entire head. The scanning sequence and parameters were as follows: sagittal T1MPRAGE: TE 2.69ms, TR 2477ms, flip angle 8°, 1.0mm isovoxel; DSI sequence: TE 84.2ms, TR 5589ms, FOV 224mm×224mm, voxel 2×2×2mm 3 , the maximum b value is 7000s / mm 2 , the total number of diffusion directions is 258; T2W FLAIR sequence: TE 105ms, TR 6300ms, 1.0mm equal voxels.

[0036] Example 2 Data Processing

[0037] The data of healthy subjects and test subjects were obtained using the method described in Example 1.

[0038] The data obtained in Example 1 were processed using DSI-Studio software, and the DICOM format data in the original image was entered and packaged into an SRC format file together with the b-value table. The SRC file was read to observe the range of brain parenchyma in the impact space, and the threshold that can cover the brain parenchyma but not include the skull was selected for reconstruction.

[0039] Before reconstruction, TOPUP was used to correct the distortion caused by the magnetic field, and Eddy current was used to correct the head motion artifact.

[0040] Q-space diffeomorphic reconstruction (QSDR) is performed to reconstruct the spin distribution function (SDF) in the individual space and further project the distribution of the diffusion direction in each voxel in the standard space (such as Figure 1 As shown), that is, the brain's morphological energy is distributed into the space of the standard brain. In this process, the diffusion sampling length ratio is 1.1, the fitting method is CDM, the spatial template is HCP1021, and the SDF is distributed in the standard space. The directional distribution function (ODF) interpolation parameters are 20-fold (20-Fold), the number of fiber processing within the voxel is 10, and the number of threads is 4; check the half-Q space diffusion imaging, select ODF for the output value, and get the SDF in the standard space.

[0041] Example 3 Data Projection and Connectome Identifier Extraction

[0042] The SDFs in the standard space of several healthy individuals were projected together and averaged to obtain the mean and standard deviation of the diffusion vector length in the direction of maximum diffusion within the voxel.

[0043] The SDF of the individual patient is also projected into the standard space, and the maximum direction within the voxel of a single subject is taken, minus the mean of the group level (length mean ), divided by the standard deviation (length std ), remember the Z value of the maximum direction in the voxel: Z = (l-length mean ) / length std .

[0044] The Z values ​​arranged in a 3D matrix are encoded in the order of left-back-front to form a 1-dimensional feature value (such as Figure 2 As shown), it is the Connectome Identifier.

[0045] Example 4 Application example of the connection identifier of the present application

[0046] A total of 74 subjects' DSI data were included, and Connectome Identifier was extracted respectively. A support vector machine model was used, with 55 cases in the training set and 19 cases in the validation set, to detect the Connectome Identifier indicator's ability to identify left and right temporal lobe epilepsy and healthy volunteers. The actual categories included 25 healthy people, 29 people with left temporal lobe epilepsy, and 20 people with right temporal lobe epilepsy. The test was performed according to the method of the above embodiment, and the results showed that in the validation set, healthy people and patients with right temporal lobe epilepsy were completely correctly identified, 3 patients with left temporal lobe epilepsy were incorrectly identified, and 7 patients were correctly identified (such as Figure 3 The consistency rate with the surgical resection result was 84.2%, reaching the level of manual interpretation, and the interpretation time was only 21 seconds.

Claims

1. A disease intelligent diagnosis device based on brain structure connection identifiers, the device comprising the following modules: (1) Data acquisition module: used to obtain DSI data; (2) Data processing module: Perform Q-space differential isomorphism reconstruction on the DSI data obtained in step (1), so that the spin distribution function is reconstructed in the individual space and the distribution of the diffusion direction in each voxel in the standard space is projected; (3) Data projection module: The spin distribution functions of one or more healthy individuals in the standard space are projected together and averaged to obtain the mean and standard deviation of the diffusion vector length in the direction with the largest diffusion in the voxel; the spin distribution function of the patient individual is also projected into the standard space, the largest direction in the voxel of a single subject is taken, and the mean length at the group level is subtracted. mean , divided by the standard deviation length std , get the Z value of the maximum direction in the voxel: Z = (l-length mean ) / length std ; (4) Connectome Identifier extraction module: The Z values ​​arranged in a three-dimensional matrix are encoded in the order of left-back-front to form a one-dimensional feature value, which is the Connectome Identifier.

2. The device according to claim 1, wherein the data acquisition module comprises an MRI instrument.

3. The device according to claim 1 or 2, wherein the functions performed by the data processing module include: The DSI data obtained in step (1) is processed using DSI-Studio software, and the DICOM format data in the original image is entered and packaged into an SRC format file together with the b-value table; the SRC file is read to the range of the brain parenchyma in the influence space, and a threshold that can cover the brain parenchyma but not include the skull is selected to perform Q-space differential isomorphism reconstruction.

4. The device according to any one of claims 1 to 3, wherein the functions performed by the data processing module include: Before the Q-space differential isomorphism reconstruction begins, TOPUP is used to correct the distortion caused by the magnetic field, and Eddy current is used to correct the head motion artifact.

5. According to the device according to any one of claims 1-4, during the operation of the data processing module, the diffusion sampling length ratio is 1.1, the assembly method is CDM, the spatial template is HCP1021, the directional distribution function interpolation parameter is 20 fold, the number of fiber processing within the voxel is 10, and the number of threads is 4.

6. The device according to any one of claims 1-5, further comprising (5) a Connectome Identifier interpretation module for interpreting a Connectome Identifier.

7. The device according to any one of claims 1 to 6, wherein the DSI data is DSI data of a patient.

8. The device according to any one of claims 1 to 6, wherein the DSI data is DSI data of a patient and DSI data of one or more healthy individuals.

9. The device according to claim 7 or 8, wherein the DSI data is brain DSI data.

10. The device according to any one of claims 1 to 9, for use in diagnosing brain diseases.

11. The device according to any one of claims 1 to 9, wherein the device is used for determining the location of brain disease lesions.

12. The device according to any one of claims 1 to 9, for use in differentiating the severity of brain diseases.

13. The device according to any one of claims 10-12, wherein the brain disease is epilepsy, Parkinson's disease, or Alzheimer's disease.

14. A storage medium having recorded therein a program for executing the following steps: (1) Obtain DSI data; (2) Data processing: DSI data are reconstructed in Q space using differential isomorphism, which allows the spin distribution function to be reconstructed in individual space and the distribution of the diffusion direction within each voxel in the standard space to be obtained by projection; (3) Data projection: Project the spin distribution functions of one or more healthy individuals in the standard space together and average them to obtain the mean and standard deviation of the diffusion vector length in the direction with the largest diffusion in the voxel; project the spin distribution function of the patient individual into the standard space, take the largest direction in the voxel of a single subject, and subtract the mean length at the group level. mean , divided by the standard deviation length std , get the Z value of the maximum direction in the voxel: Z = (l-length mean ) / length std ; (4) Connectome Identifier Extraction: The Z values ​​arranged in a 3D matrix are encoded in the order of left-back-front to form a one-dimensional feature value, which is the Connectome Identifier.

15. According to the storage medium of claim 14, the data processing step comprises processing the acquired data using DSI-Studio software, entering the DICOM format data in the original image, and packaging it together with the b-value table into an SRC format file; reading the SRC file to observe the range of brain parenchyma in the influence space, selecting a threshold that can cover the brain parenchyma but not include the skull, and performing Q-space differential isomorphism reconstruction.

16. The storage medium according to claim 14 or 15, wherein the data processing comprises: Before the Q-space differential isomorphism reconstruction begins, TOPUP is used to correct the distortion caused by the magnetic field, and Eddy current is used to correct the head motion artifact.

17. The storage medium according to any one of claims 14-16, wherein in the data processing process, the diffusion sampling length ratio is 1.1, the fitting method is CDM, the spatial template is HCP1021, the directional distribution function interpolation parameter is 20 fold, the number of fiber processing within the voxel is 10, and the number of threads is 4.

Images