Training method of brain disease auxiliary diagnosis model, auxiliary diagnosis method, equipment and medium

Abstract

The invention discloses a brain disease auxiliary diagnosis model training method, an auxiliary diagnosis method, equipment and a medium, and belongs to the technical field of human brain disease auxiliary diagnosis. The training method comprises the following steps: acquiring multi-modal image data I of healthy testees with the same gender, age group and educated degree; preprocessing the first multi-modal image data, and constructing a human brain structure connection matrix SC used for reflecting the brain structure connection condition and a human brain function connection matrix FC used for reflecting the brain function activity condition; constructing a path reachability network feature WA; dividing the first multi-modal image data into a training sample and a test sample, and establishing a prediction model of a path reachability characteristic WA of the healthy testee and a human brain function connection matrix FC of the healthy testee. According to the path reachability feature provided by the invention, the influence of the shortest path and the surrounding paths on the formation of the FC is comprehensively considered, and the problem that multiple paths with different distances and different synchronization lengths are difficult to fuse in the existing network communication feature is solved.

Description

Technical Field

[0001] The present application relates to the technical field of human brain disease assisted diagnosis, and particularly relates to a training method of a brain disease assisted diagnosis model, an assisted diagnosis method, a device, and a medium. Background Art

[0002] The brain is the most important and complex organ of the human body and the material basis of human intelligence. It contains approximately 86 billion neurons, and each neuron is interconnected with other neurons through thousands of synapses to form an extremely complex neural network. Currently, the research on the human brain covers scales from the microscopic biomolecular and nerve cell levels to the macroscopic level of the brain, including the research on abnormalities of proteins and genes, the research on brain neurodynamics, etc. Nevertheless, for some brain diseases, such as neurodegenerative diseases (Parkinson's disease, Alzheimer's disease, etc.), mental diseases (schizophrenia, depression, etc.), their early causes and influencing factors are not very clear, the pathogenesis is still unknown, and specific and sensitive early warning and accurate and effective diagnosis methods have not been established. For example, the clinical diagnosis and treatment of Parkinson's disease mainly rely on motor symptoms. However, many of its non-motor symptoms occur earlier than motor symptoms, such as hyposmia, sleep disorders, depressive symptoms, etc. The clinical diagnosis and treatment relying on motor symptoms lag far behind pathological changes. The diagnosis of schizophrenia is mainly based on the evaluation of the patient's disease course review, mental state, and symptoms by clinicians to draw conclusions, lacking effective, reliable, and unified objective biomarker-assisted diagnosis. This makes it difficult to substantially improve the disease outcome in the clinical diagnosis and treatment of some brain diseases, and there is an urgent need to find effective biomarker quantitative evaluation for early assisted diagnosis of brain diseases.

[0003] In recent years, more and more neuroscientists have characterized the structural network map of the brain at the macroscopic or microscopic level through the human connectome, and used the brain's structural network and the functional network formed by neural activities to understand how the brain works. Exploring the functions and cognition of the brain through network analysis methods has always been one of the commonly used methods in the field of neuroscience. Its advantages are as follows. Firstly, the brain network is an abstract expression of the brain, which can reduce the complexity of neuron-based network analysis. Secondly, although there are generally differences in the volume and surface shape of individual brains, by using the same reference template, that is, the same method of defining network nodes, it is helpful to compare different individuals or different types of brain networks, such as the brain networks of the elderly and the young, the brain networks of healthy people and patients, etc. In the prior art, network topological properties such as node degree, shortest path length, local efficiency, small world, modularity, etc. are used to measure the ability of healthy people or patients' brains or parts of brain regions to resist external interference, the ability of information transmission, and the ability of functional separation and functional integration, etc., providing effective features for the characterization of brain functions and brain diseases. However, the above-mentioned network topological properties are not effective in describing brain network communication and the construction of brain functions, which is not conducive to establishing an auxiliary diagnosis mechanism for healthy people and brain disease patients. It is urgent to improve the network features in the existing technology, solve the problem of inaccurate brain function construction, and establish effective, reliable, and unified objective biomarkers for the auxiliary diagnosis and treatment of brain diseases. Summary of the Invention

[0004] This application aims to solve at least one of the technical problems in the related art to some extent. For this purpose, this application provides a training method, an auxiliary diagnosis method, a device, and a medium for a brain disease auxiliary diagnosis model. The path reachability feature provided by this application comprehensively considers the influence of the shortest path and its surrounding paths on the formation of FC, making up for the problem of difficult integration of multi-paths with different distances and different step lengths in the existing network communication features.

[0005] To achieve the above object, in the first aspect, this application provides a training method for a brain disease auxiliary diagnosis model, including the following steps:

[0006] Obtain the first multi-modal imaging data of healthy subjects with the same gender, age group, and education level. The first multi-modal imaging data includes T1 data, dMRI data, and fMRI data;

[0007] Preprocess the first multi-modal imaging data to construct a human brain structural connection matrix SC for reflecting the brain's structural connection situation and a human brain functional connection matrix FC for reflecting the brain's functional activity situation;

[0008] Construct the path reachability network feature WA, and calculate its corresponding path reachability network feature WA based on the human brain structural connection matrix SC, where the path reachability network feature WA is used to characterize the probability that the source node accesses the destination node through the shortest path and its adjacent preset number of steps of paths;

[0009] Divide the multi-modal image data into training samples and test samples, and establish a prediction model between the path reachability feature WA of healthy subjects and their human brain functional connection matrix FC.

[0010] Preferably, the steps of preprocessing the multi-modal image data include: based on the brain partition template at the standard space macroscopic scale, perform partition processing on the multi-modal image data to obtain multiple brain regions.

[0011] Preferably, the steps of constructing the human brain structural connection matrix SC for reflecting the brain structural connection situation and the human brain functional connection matrix FC for reflecting the brain functional activity situation include:

[0012] Based on the preprocessed multi-modal image data, obtain its fiber tracking data and BOLD data reflecting the brain functional activity. The fiber tracking data includes the number of fiber connections between two brain regions, and / or fiber connection density, and / or fiber connection probability;

[0013] Construct the human brain structural connection matrix SC based on the fiber tracking data;

[0014] Construct the human brain functional connection matrix FC based on the BOLD data. The human brain functional connection matrix FC is the Pearson correlation coefficient of the BOLD time series between two brain regions.

[0015] Preferably, the steps of calculating the corresponding path reachability network feature WA based on the human brain structural connection matrix SC include:

[0016] Normalize the network weights of the human brain structural connection matrix SC. The normalization method is determined according to the weights of the human brain structural connection matrix SC. If the human brain structural connection matrix SC is the number of fiber connections, then use log normalization. If the human brain structural connection matrix SC is fiber connection density or fiber connection probability, then use min-max normalization;

[0017] Construct a distance connection matrix to quantify the path distance between each brain region;

[0018] Construct the path reachability network feature WA to characterize the probability that the source node accesses the destination node through the shortest path and its adjacent two-step paths.

[0019] Preferably, the path reachability network feature WA introduces a balance matrix T to represent the weight coefficient for obtaining information from the shortest path, and uses 1 - T to represent the weight coefficient for obtaining information from the surrounding paths. The path reachability network feature WA introduces a scale parameter α to limit the length of the two-step peripheral path.

[0020] Preferably, dividing the multi-modal image data into training samples and test samples, the steps of establishing a prediction model for the path reachability feature WA of healthy subjects and their human brain functional connectivity matrix FC include:

[0021] Based on the training samples, calculate the path reachability feature WA of healthy subjects and establish a linear model between it and the human brain functional connectivity matrix FC;

[0022] Use the mean square error minimum constraint condition to obtain the optimal scale parameter α and the optimal linear model parameter W m ;

[0023] To prevent overfitting, add an L2-regularization term to the linear model to constrain the weights;

[0024] Based on the test samples, use the optimal scale parameter α and the optimal linear model parameter W m Calculate the predicted functional connectivity matrix, and calculate the Pearson correlation coefficient between the predicted functional connectivity matrix and the true human brain functional connectivity matrix FC to describe the effect of human brain function construction. The larger the Pearson correlation coefficient, the better the effect of human brain function construction.

[0025] In a second aspect, the present application provides a method for assisting in the diagnosis of brain diseases. Based on the above prediction model, a brain disease classification model is established to achieve the auxiliary diagnosis of brain diseases, including:

[0026] Obtain multi-modal image data two of brain disease patients and healthy subjects with the same gender, age group, and education level. The multi-modal image data two includes T1 data, dMRI data, and fMRI data;

[0027] Preprocess the multi-modal image data two to construct a human brain structural connectivity matrix SC for reflecting the brain's structural connection situation and a human brain functional connectivity matrix FC for reflecting the brain's functional activity situation;

[0028] Calculate the path reachability feature WA of each healthy subject and brain disease patient, and input them into the prediction model in sequence to obtain the corresponding predicted simulated functional connectivity matrix FCs;

[0029] Calculate the mean square error between the predicted simulation functional connectivity matrix FCs and the true functional connectivity matrix FCe, and the Pearson correlation coefficient between the predicted simulation functional connectivity matrix FCs and the true functional connectivity matrix FCe. Use these two as classification features for healthy subjects and brain disease patients, and establish a classification model for healthy subjects and brain disease patients using a machine learning algorithm to achieve the auxiliary diagnosis of brain diseases.

[0030] Preferably, the steps of calculating the path reachability feature WA for each healthy subject and brain disease patient include: based on the optimal scale parameter α and the optimal linear model parameter W m Calculate the path reachability feature WA for each healthy subject and brain disease patient.

[0031] In a third aspect, the present application provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the program to implement the steps of any one of the above training methods or brain disease auxiliary diagnosis methods.

[0032] In a fourth aspect, the present application provides a computer-readable storage medium, including a computer program, which when running on an electronic device, causes the electronic device to execute the steps of any one of the above training methods or brain disease auxiliary diagnosis methods.

[0033] Based on the above technical solutions, the training method, auxiliary diagnosis method, device, and medium of the brain disease auxiliary diagnosis model of the present application, compared with the prior art, at least have one of the following beneficial effects:

[0034] 1. The path reachability feature provided by the training method of the brain disease auxiliary diagnosis model of the present application comprehensively considers the influence of the shortest path and its surrounding paths on the formation of FC, and introduces a balance matrix to describe the weight coefficient of obtaining information from the shortest path and the weight coefficient of obtaining information from the surrounding paths, making up for the problem that it is difficult to integrate multi-paths with different distances and different step lengths existing in the existing network communication features.

[0035] 2. The path reachability feature provided by the training method of the brain disease auxiliary diagnosis model of the present application selects two-step paths that have a greater impact on the formation of the human brain functional connectivity matrix FC as the surrounding paths, and introduces a scale parameter to limit the length of the two-step peripheral paths to ensure the efficiency of information propagation through the peripheral paths, so as to describe the diffusion and integration of information along the paths by finding a balance between efficiency and network stability, and improve the accuracy of brain function construction.

[0036] 3. The training method of the brain disease auxiliary diagnosis model of the present application can extract path reachability based on any type of structural network, with strong applicability. And the prediction model of the present application is a simple linear model, which can avoid the problems that are not conducive to feature extraction and model training caused by the large dimensions of human brain structure and functional connection data in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0038] Figure 1 is a flowchart of a training method for a brain disease auxiliary diagnosis model provided in Embodiment 1 of the present application;

[0039] Figure 2 is a flowchart of a brain disease auxiliary diagnosis method provided in Embodiment 2 of the present application;

[0040] Figure 3 is a structural block diagram of a brain disease auxiliary diagnosis device based on path reachability provided in Embodiment 5 of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] To make the objectives, technical solutions, and advantages of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the drawings.

[0042] The terms "first", "second", "third", "fourth", "fifth", "sixth", "seventh", and "eighth", etc. (if any) in the specification, claims, and drawings of the present application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here.

[0043] In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those clearly listed steps or units, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0044] The basic idea of this application is to comprehensively consider the influence of the shortest path and its surrounding paths on the formation of the human brain functional connectivity matrix FC by constructing path reachability features, so as to make up for the problem that it is difficult to integrate multi-paths with different distances and different step lengths in the existing network communication features, and improve the accuracy of brain function construction. Furthermore, it realizes the early diagnosis and adjuvant treatment of brain diseases with unclear exact etiology and pathogenesis, and diseases with similar clinical manifestations, such as schizophrenia, Parkinson's disease, Alzheimer's disease and its subtypes.

[0045] Example 1

[0046] As Figure 1 shown, in order to construct an accurate prediction model of the human brain functional connectivity matrix, the inventor of the present invention has conducted in-depth research on brain structural connectivity and proposed a training method for a brain disease auxiliary diagnosis model, including the following steps:

[0047] S1. Obtain the first multi-modal image data of healthy subjects with the same gender, age group and education level, where the first multi-modal image data includes T1 data, dMRI data and fMRI data;

[0048] S2. Preprocess the first multi-modal image data to construct a human brain structural connectivity matrix SC for reflecting the brain structural connectivity and a human brain functional connectivity matrix FC for reflecting the brain functional activity;

[0049] S3. Construct a path reachability network feature WA, and calculate its corresponding path reachability network feature WA based on the human brain structural connectivity matrix SC, where the path reachability network feature WA is used to describe the probability that a source node accesses a destination node through the shortest path and its adjacent preset number of step paths;

[0050] S4. Divide the first multi-modal image data into training samples and test samples, and establish a prediction model between the path reachability feature WA of healthy subjects and their human brain functional connectivity matrix FC.

[0051] Specifically, in step S1, diffusion magnetic resonance imaging (dMRI) is a special magnetic resonance imaging technique that can measure the diffusion motion of water molecules in biological tissues. In the brain, the diffusion direction of water molecules is closely related to the orientation of white matter fibers. Therefore, the trajectory of white matter fibers can be inferred from dMRI data. The acquisition of dMRI data requires setting appropriate parameters, such as the diffusion gradient direction (usually 64 or more directions), the diffusion sensitivity coefficient (b value, usually between 1000 - 3000 s / mm 2 ²) etc., to ensure that the diffusion information of water molecules can be accurately captured.

[0052] Functional magnetic resonance imaging (fMRI) indirectly reflects neural activity by detecting changes in blood oxygenation levels in the brain. When neuronal activity in a certain area of the brain increases, local blood flow increases, leading to a change in the ratio of oxyhemoglobin to deoxyhemoglobin, which is then detected by fMRI. The acquisition of fMRI data requires setting appropriate parameters, such as repetition time (TR, usually between 2 - 3 seconds), echo time (TE), number of scan slices, etc., to ensure that the functional activity signals of the brain can be captured.

[0053] T1 data refers to magnetic resonance imaging (MRI) data obtained through T1-weighted imaging (T1WI). It is a commonly used sequence in MRI scans, mainly used to obtain anatomical structure information of the brain. T1-weighted imaging mainly reflects the differences in the longitudinal relaxation time (T1 value) of tissues. In T1 images, different tissue types such as brain tissue (e.g., gray matter, white matter) and cerebrospinal fluid will show different signal intensities, where white matter usually has a stronger signal than gray matter, and cerebrospinal fluid has a weaker signal. T1 data is mainly used for:

[0054] 1. Visualization of brain anatomical structure: Provide detailed images of high-resolution brain structures such as gray matter, white matter, and ventricles.

[0055] 2. Brain tissue segmentation: Divide the brain into different tissue types (such as gray matter, white matter, cerebrospinal fluid) and regions of interest (ROIs) through image segmentation algorithms, providing a basis for subsequent functional and structural analysis.

[0056] 3. Spatial normalization: Register individual brain images to a standard space (such as the MNI template) for group analysis and comparison.

[0057] In step S2, the steps of preprocessing the multimodal image data one include: Based on a brain partition template at the macroscopic scale of the standard space, performing partition processing on the multimodal image data to obtain multiple brain regions, and the number of partitions is Nr. Usually, standard brain partition templates (such as the AAL template, Desikan-Killiany template, etc.) are used.

[0058] Preferably, the steps of constructing a human brain structural connection matrix SC for reflecting the brain structural connection situation and a human brain functional connection matrix FC for reflecting the brain functional activity situation include:

[0059] Based on the preprocessed multimodal image data one, obtaining its fiber tracking data and BOLD data reflecting the brain functional activity, and the fiber tracking data includes the number of fiber connections, and / or fiber connection density, and / or fiber connection probability between two brain regions;

[0060] Construct the human brain structural connection matrix SC based on fiber tracking data;

[0061] Construct the human brain functional connection matrix FC based on BOLD data, where the human brain functional connection matrix FC is the Pearson correlation coefficient of BOLD time series between pairwise brain regions.

[0062] Preferably, the steps of calculating the corresponding path reachability network feature WA based on the human brain structural connection matrix SC include:

[0063] Normalize the network weights of the human brain structural connection matrix SC. The normalization method is determined according to the weights of the human brain structural connection matrix SC. If the human brain structural connection matrix SC is the number of fiber connections, logarithmic normalization is used. If the human brain structural connection matrix SC is the fiber connection density or fiber connection probability, min-max normalization is used;

[0064] Construct a distance connection matrix to quantify the path distances between various brain regions;

[0065] Construct the path reachability network feature WA to characterize the probability that the source node accesses the destination node through the shortest path and its adjacent two-step paths.

[0066] Preferably, the path reachability network feature WA introduces a balance matrix T to represent the weight coefficient of obtaining information from the shortest path, and 1 - T is used to represent the weight coefficient of obtaining information from the surrounding paths. The path reachability network feature WA introduces a scale parameter α to limit the length of the two-step peripheral paths.

[0067] Preferably, the steps of dividing the multimodal image data into training samples and test samples and establishing a prediction model of the path reachability feature WA of healthy subjects and their human brain functional connection matrix FC include:

[0068] Based on the training samples, calculate the path reachability feature WA of healthy subjects and establish a linear model between it and the human brain functional connection matrix FC;

[0069] Obtain the optimal scale parameter α and the optimal linear model parameter W using the minimum mean square error constraint condition m ;

[0070] To prevent overfitting, add an L2-regularization term to the linear model to constrain the weights;

[0071] Based on the test samples, use the optimal scale parameter α and the optimal linear model parameter W mCalculate the predicted functional connectivity matrix, and calculate the Pearson correlation coefficient between the predicted functional connectivity matrix and the true human brain functional connectivity matrix FC to describe the effect of human brain function construction. The larger the Pearson correlation coefficient, the better the effect of human brain function construction.

[0072] Thus, the path reachability feature provided by the training method of the brain disease auxiliary diagnosis model in this embodiment comprehensively considers the influence of the shortest path and its surrounding paths on the formation of FC, and introduces a balance matrix to describe the weight coefficient of obtaining information from the shortest path and the weight coefficient of obtaining information from the surrounding paths, making up for the problem that it is difficult to integrate multi-paths with different distances and different step lengths existing in the existing network communication features. In addition, the path reachability feature provided by the training method of the brain disease auxiliary diagnosis model in this embodiment selects two-step paths that have a greater impact on the formation of the human brain functional connectivity matrix FC as the surrounding paths, and introduces a scale parameter to limit the length of the two-step peripheral paths to ensure the efficiency of information propagation through the peripheral paths, so as to describe the diffusion and integration of information propagation along the path by finding a balance between efficiency and network stability, improve the accuracy of brain function construction, and provide an accurate brain functional connectivity prediction model for subsequent disease auxiliary diagnosis.

[0073] Embodiment 2

[0074] As Figure 2 shown, the embodiment of the present application provides a brain disease auxiliary diagnosis method, which establishes a brain disease classification model based on the above prediction model to realize the auxiliary diagnosis of brain diseases, including:

[0075] S1. Obtain multi-modal image data II of brain disease patients and healthy subjects with the same gender, age group, and education level. The multi-modal image data II includes T1 data, dMRI data, and fMRI data;

[0076] S2. Preprocess the multi-modal image data II to construct a human brain structural connectivity matrix SC for reflecting the brain structural connection situation and a human brain functional connectivity matrix FC for reflecting the brain functional activity situation;

[0077] S3. Calculate the path reachability feature WA of each healthy subject and brain disease patient, and input it into the prediction model in sequence to obtain the corresponding predicted simulated functional connectivity matrix FCs;

[0078] S4. Calculate the mean square error between the predicted simulation functional connectivity matrix FCs and the true functional connectivity matrix FCe, and the Pearson correlation coefficient between the predicted simulation functional connectivity matrix FCs and the true functional connectivity matrix FCe. Use these two as the classification features of healthy subjects and brain disease patients, and adopt a machine learning algorithm to establish a classification model of healthy subjects and brain disease patients to achieve the auxiliary diagnosis of brain diseases.

[0079] The preprocessing of the multimodal image data two includes: obtaining a brain partition template at the standard space macroscopic scale, and the number of partitions is Nr. Based on neuroimaging processing software, preprocess the acquired T1 data, dMRI data, and fMRI data to obtain fiber tracking data and BOLD data reflecting brain functional activities. Construct the human brain structural connectivity matrix SC = [W ij (i, j ∈ [1, Nr]) and the human brain functional connectivity matrix FC = [fc ij (i, j ∈ [1, Nr]), where the human brain structural connectivity matrix SC can be the number of fiber connections, fiber connection density, or fiber connection probability between two brain regions i and j, and the FC is the Pearson correlation coefficient of the BOLD time series between two brain regions i and j.

[0080] Normalize the structural network weights. The normalization method is determined according to the SC weights. If SC is the number of fiber connections, use log normalization. If SC is the fiber connection density or connection probability, use min-max normalization. Obtain the distance connectivity matrix D = [d ij , where the connection distance d ij between nodes (brain regions) i and j can be calculated by 1 / W ij or -log(W ij ), W ij represents the normalized connection weight. The larger W ij , the longer the connection distance between nodes. If there is no connection edge between nodes, the connection distance is ∞.

[0081] Adopt the shortest path algorithm to obtain the path π i→j = {w ik , w km , …, w zj} with the shortest distance from the source node to the destination node. Then the shortest path length is SPL i→j = d ik + d km + … d zj , and the nodes passed by the shortest path are Ω i→j={i, k, m, …z, j}, and the shortest path algorithm can be Dijkstra algorithm, Floyd-Warshall algorithm, Johnson algorithm, Bellman-Ford algorithm, etc.

[0082] Preferably, the step of calculating the path reachability feature WA of each healthy subject and brain disease patient includes: calculating the path reachability feature WA of each healthy subject and brain disease patient based on the optimal scale parameter α and the optimal linear model parameter Wm.

[0083] Use a machine learning algorithm to establish a classification model for healthy subjects and brain disease patients to realize the auxiliary diagnosis of brain diseases. The machine learning method can be deep learning, clustering, neural network, decision tree and other methods.

[0084] The construction of the path reachability feature WA is specifically defined as follows:

[0085]

[0086]

[0087] Among them, represents the probability that information propagates along the shortest path; w i is the node weight of node i, representing the sum of the weights of all SCs connected to the i brain region; if there are multiple intermediate nodes between the source node s and the destination node t, the shortest path can be expressed as: π s→t ={w si , w ij , …, w rt}, and the corresponding nodes passed through are Ω s→t ={s, i, j, …r, t}, the number of steps of the shortest path is K s→t =|π s→t |, |·| is used to calculate the number of connection edges between nodes, and the number of nodes is |Ω s→t | = K s→t +1; represents the node set that does not include the destination node, is the first connection edge of the shortest path between and, and represents a two-step path with a shorter distance, that is, the path {w sk , w kt} composed of the source node s, the destination node t and the intermediate nodes therebetween; func_d(s, k, t) is used to limit the distance of the two-step path to be less than α times the length of the shortest path: The selected scale parameter α ∈ [1, 10]; the balance matrix It is used to balance the influence of the shortest path and its surrounding path pairs on forming FC. The value ranges from 0 to 1, representing the weight coefficient for obtaining information from the shortest path; It represents the sum of the two-step path distances that meet the func_d condition; the Θ(·) function is a binarization function. If w sk > 0, then Θ(w sk ) = 1, otherwise Θ(w sk ) = 0.

[0088] Optionally, the human brain functional connectivity prediction model based on path reachability features is specifically:

[0089] vec(FCs) = XW m

[0090] X = [vec(ones(n,n))|vec(WA)]

[0091]

[0092] where W m represents the linear weight of the path reachability network feature WA, and FCs is the predicted FC. To obtain the optimal linear weight, it is necessary to minimize the error between the predicted FCs and the measured FCe. ||·|| 2 is the vector 2-norm, and vec(·) represents the vectorization of the matrix. According to the least squares method, To prevent overfitting, an L2-regularization term is added to the formula to constrain the weight.

[0093] Optionally, calculate the Pearson correlation coefficient r between the predicted functional connectivity matrix and the measured functional connectivity matrix, specifically:

[0094]

[0095] where X and Y respectively represent the predicted functional connectivity matrix and the measured functional connectivity matrix, n represents the number of elements in the functional connectivity matrix, and respectively represent the average values of the elements in the predicted functional connectivity matrix and the measured functional connectivity matrix. The larger the correlation coefficient r, the greater the correlation between the predicted functional connectivity matrix and the measured functional connectivity matrix, indicating a better effect of human brain function construction.

[0096] The brain disease assisted diagnosis method of this embodiment comprehensively considers the influence of the shortest path and its surrounding paths on the formation of the human brain functional connection matrix FC by constructing path reachability features, so as to make up for the problem that it is difficult to fuse multi-paths with different distances and different step lengths existing in the existing network communication features, improve the accuracy of brain function construction, and use the mean square error {E ind} n1 and {E ind} n2 and Pearson correlation coefficient {r ind} n1 and {r ind} n2 as classification features, where n1 and n2 are the numbers of brain disease patients and healthy subjects respectively, {E ind} n1 represents the mean square error between the predicted simulation functional connection matrix FCs and the true functional connection matrix FCe of brain disease patients, and {E ind} n2 represents the mean square error between the predicted simulation functional connection matrix FCs and the true functional connection matrix FCe of healthy subjects; {r ind} n1 represents the Pearson correlation coefficient between the predicted simulation functional connection matrix FCs and the true functional connection matrix FCe of brain disease patients, and {r ind} n2 represents the Pearson correlation coefficient between the predicted simulation functional connection matrix FCs and the true functional connection matrix FCe of healthy subjects. A classification model of healthy subjects and brain disease patients is established by using machine learning algorithms to realize the assisted diagnosis of brain diseases. The specific machine learning methods can be deep learning, clustering, neural network, decision tree and other methods. Furthermore, it realizes the early assisted diagnosis and treatment of brain diseases with unclear exact etiology and pathogenesis, and diseases with similar clinical manifestations, such as schizophrenia, Parkinson's disease, Alzheimer's disease and its subtypes.

[0097] This embodiment of this application extracts the mean square error and Pearson correlation coefficient as classification features of healthy subjects and brain disease patients for assisting in the diagnosis of brain diseases. It should be noted that this embodiment can also extract other features of path reachability, such as the path reachability between specific nodes, the sum of the path reachabilities between specific nodes and other nodes, etc. for the assisted diagnosis of brain diseases. The selection of specific features depends on the type of brain disease and its prior clinical knowledge. Those skilled in the art can extract one or more features based on path reachability to construct a brain disease diagnosis model.

[0098] Embodiment III

[0099] An embodiment of the present application provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the program to implement the steps of the training method or the brain disease assisted diagnosis method described in any one of the above, and implement the following functions: The path reachability feature provided by the training method of the brain disease assisted diagnosis model comprehensively considers the influence of the shortest path and its surrounding paths on the formation of FC, and introduces a balance matrix to characterize the weight coefficients for obtaining information from the shortest path and from the surrounding paths, making up for the problem of multi-path with different distances and different step lengths that existing network communication features have difficulty in fusing. In addition, the path reachability feature provided by the training method of the brain disease assisted diagnosis model in this embodiment selects two-step paths that have a greater impact on the formation of the human brain functional connection matrix FC as the surrounding paths, and introduces a scale parameter to limit the length of the two-step peripheral paths to ensure the efficiency of information propagation through the peripheral paths, thereby describing the diffusion and integration of information along the path by finding a balance between efficiency and network stability, improving the accuracy of brain function construction, and providing an accurate brain functional connection prediction model for subsequent disease assisted diagnosis.

[0100] Embodiment 4

[0101] An embodiment of the present application provides a computer-readable storage medium, including a computer program. When the computer program runs on an electronic device, the electronic device is caused to execute the steps of the training method or the brain disease assisted diagnosis method described in any one of the above, and implement the following functions: The path reachability feature provided by the training method of the brain disease assisted diagnosis model comprehensively considers the influence of the shortest path and its surrounding paths on the formation of FC, and introduces a balance matrix to characterize the weight coefficients for obtaining information from the shortest path and from the surrounding paths, making up for the problem of multi-path with different distances and different step lengths that existing network communication features have difficulty in fusing. In addition, the path reachability feature provided by the training method of the brain disease assisted diagnosis model in this embodiment selects two-step paths that have a greater impact on the formation of the human brain functional connection matrix FC as the surrounding paths, and introduces a scale parameter to limit the length of the two-step peripheral paths to ensure the efficiency of information propagation through the peripheral paths, thereby describing the diffusion and integration of information along the path by finding a balance between efficiency and network stability, improving the accuracy of brain function construction, and providing an accurate brain functional connection prediction model for subsequent disease assisted diagnosis.

[0102] Embodiment 5

[0103] As Figure 3 shown, an embodiment of the present application provides a brain disease assisted diagnosis device based on path reachability. The device includes:

[0104] The first acquisition module is used to acquire multi-modal imaging data of brain disease patients {PAT} (where {...} represents a set) and healthy subjects {CON} with no significant differences in gender, age, and educational level from a pre-set human brain neuroimaging database. The multi-modal imaging data includes T1 data, dMRI data, and fMRI data. Here, n1 and n2 are the numbers of brain disease patients and healthy subjects respectively. A brain partition template at the standard space macro scale is acquired, and the number of partitions is Nr. n1 ({...} represents a set) and healthy subjects {CON} n2 The multi-modal imaging data of brain disease patients {PAT} (where {...} represents a set) and healthy subjects {CON} includes T1 data, dMRI data, and fMRI data. Here, n1 and n2 are the numbers of brain disease patients and healthy subjects respectively. A brain partition template at the standard space macro scale is acquired, and the number of partitions is Nr.

[0105] The first processing module is used to preprocess the acquired T1 data, dMRI data, and fMRI data based on neuroimaging processing software to obtain fiber tracking data and BOLD data reflecting brain functional activities. A human brain structural connection matrix SC = [W ij (i, j ∈ [1, Nr]) and a human brain functional connection matrix FC = [fc ij (i, j ∈ [1, Nr]) are constructed. Among them, the human brain structural connection matrix SC can be the number of fiber connections, fiber connection density, or fiber connection probability between two brain regions i and j, and the FC is the Pearson correlation coefficient of the BOLD time series between two brain regions i and j.

[0106] The second processing module is used to normalize the structural network weights. The normalization method is determined according to the SC weights. If SC represents the number of fiber connections, log normalization is used. If SC is fiber connection density or connection probability, min-max normalization is used. A distance connection matrix D = [d ij is obtained, where the connection distance d ij between nodes (brain regions) i and j can be calculated by 1 / W ij or -log(W ij ). W ij represents the normalized connection weight. The larger W ij , the longer the connection distance between nodes. If there is no connection edge between nodes, the connection distance is ∞.

[0107] The third processing module is used to obtain the shortest path π i→j = {w ik , w km , …, w zj} from the source node to the destination node. Then the shortest path length is SPL i→j = d ik + d km + … d zj , and the nodes passed by the shortest path are Ω i→j={i, k, m, . . . z, j}, the shortest path algorithm may be Dijkstra algorithm, Floyd-Warshall algorithm, Johnson algorithm, Bellman-Ford algorithm, etc.

[0108] The fourth processing module constructs the path reachability network feature WA, which characterizes the probability that the source node accesses the destination node through the shortest path and its adjacent two-step path. This feature assumes that paths with shorter distances and fewer steps play a greater role in information transmission, and only considers the influence of the shortest path and its two-step peripheral paths with shorter distances, in order to describe the diffusion and integration of information along the path by finding a balance between efficiency and network stability. In order to balance the influence of the shortest path and its two-step peripheral paths on the formation of FC, WA introduces a balance matrix T to represent the weight coefficient of obtaining information from the shortest path, and uses 1-T to represent the weight coefficient of obtaining information from surrounding paths. WA introduces a scale parameter α to limit the length of the two-step peripheral path, that is, the path length is less than in It represents the shortest path length from the source node s to the destination node.

[0109] The second acquisition module is used to obtain a portion of data from the pre-set healthy subjects (FC train , SC train ) as training samples, and obtain a portion of data (FC test , SC test ) were used as test samples to establish a prediction model of path accessibility characteristics WA and functional connectivity FC for healthy subjects.

[0110] The fifth processing module is used to process the training sample SC obtained by the second obtaining module. train , the path accessibility feature WA of the healthy subjects is calculated using the method described in the fourth processing module, and the functional connection FC is established train The linear model is constructed by using the minimum mean square error constraint to obtain the optimal scale parameter α and the optimal linear model parameter W. m And the corresponding predicted human brain functional connectivity matrix. To prevent overfitting, L2-regularization terms are added to the linear model to constrain the weights. In the test sample, the optimal scale parameter α and the optimal linear model parameter W obtained from the training sample are used. m Calculate the predicted functional connectivity matrix FCp and calculate the predicted simulation functional connectivity matrix FCp and the human brain functional connectivity matrix FC test The Pearson correlation coefficient r between the two is used to describe the effect of human brain function construction. The larger the correlation coefficient r, the better the effect of human brain function construction.

[0111] A loop module, which is used to repeat the second acquisition module and the fifth processing module N times, and take the average value α of the optimal scale parameters mean as the final scale parameter, and take the average value W of the optimal linear model parameters mean as the final model parameter. The number of repetitions N is determined by the number of samples of the structural and functional connectivity data in the human brain neuroimaging database. The larger the number of samples, the larger the number of repetitions N.

[0112] A sixth processing module, which is used to calculate the path reachability feature WA of each healthy subject and brain disease patient by using the scale parameter α mean and the linear model parameter W mean , and predict the mean square error E between the simulated functional connectivity matrix FCs and the real functional connectivity matrix FCe ind ; ind

[0113]

[0114] The Pearson correlation coefficient r between the simulated functional connectivity matrix FCs and the real functional connectivity matrix FCe ind .

[0115] A seventh processing module, which is used to use the mean square errors of functional connectivity {E ind} n1 and {E ind} n2 and the Pearson correlation coefficients {r ind} n1 and {r ind} n2 obtained by the sixth processing module as classification features, and use a machine learning algorithm to establish a classification model for healthy subjects and brain disease patients to achieve the auxiliary diagnosis of brain diseases. The machine learning method can be methods such as deep learning, clustering, neural network, decision tree, etc.

[0116] It should also be noted that: when the above-mentioned brain disease auxiliary diagnosis device based on path reachability performs brain disease diagnosis, only the above-mentioned division of each functional module is used for illustration. In practical applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the above-mentioned brain disease auxiliary diagnosis device based on path reachability and the embodiment of a brain disease auxiliary diagnosis method based on path reachability belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0117] The serial numbers of the above embodiments of the present application are only for description and do not represent the advantages and disadvantages of the embodiments.​

[0118] Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware, or can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium. The above-mentioned storage medium can be a read-only memory, a magnetic disk, an optical disc, or the like.

[0119] The above embodiments are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A training method for an auxiliary diagnosis model of brain diseases, characterized in that, It includes the following steps: Obtain multimodal imaging data one of healthy subjects with the same gender, age group, and education level. The multimodal imaging data one includes T1 data, dMRI data, and fMRI data; Preprocess the multimodal imaging data one to construct a human brain structural connectivity matrix SC for reflecting the brain's structural connection situation and a human brain functional connectivity matrix FC for reflecting the brain's functional activity situation; Construct a path reachability network feature WA, and calculate its corresponding path reachability network feature WA based on the human brain structural connectivity matrix SC, where the path reachability network feature WA is used to characterize the probability that the source node accesses the destination node through the shortest path and its adjacent preset number of step paths; Divide the multimodal imaging data one into training samples and test samples, and establish a prediction model between the path reachability feature WA of healthy subjects and their human brain functional connectivity matrix FC.

2. The training method according to claim 1, wherein The steps for preprocessing the multimodal imaging data one include: based on a brain partition template at the standard space macroscale, perform partition processing on the multimodal imaging data to obtain multiple brain regions.

3. The training method according to claim 2, characterized in that The steps for constructing a human brain structural connectivity matrix SC for reflecting the brain's structural connection situation and a human brain functional connectivity matrix FC for reflecting the brain's functional activity situation include: Based on the preprocessed multimodal imaging data one, obtain its fiber tracking data and BOLD data reflecting the brain's functional activity. The fiber tracking data includes the number of fiber connections between two brain regions, and / or fiber connection density, and / or fiber connection probability; Construct a human brain structural connectivity matrix SC based on the fiber tracking data; Construct a human brain functional connectivity matrix FC based on the BOLD data. The human brain functional connectivity matrix FC is the Pearson correlation coefficient of the BOLD time series between two brain regions.

4. The training method according to claim 3, wherein The steps for calculating its corresponding path reachability network feature WA based on the human brain structural connectivity matrix SC include: Normalize the network weights of the human brain structural connectivity matrix SC. The normalization method is determined according to the weights of the human brain structural connectivity matrix SC. If the human brain structural connectivity matrix SC is the number of fiber connections, then use log normalization. If the human brain structural connectivity matrix SC is fiber connection density or fiber connection probability, then use min-max normalization; Construct a distance connectivity matrix to quantify the path distance between each brain region; Construct a path reachability network feature WA to characterize the probability that the source node accesses the destination node through the shortest path and its adjacent two-step paths.

5. The training method according to claim 4, wherein The path reachability network feature WA introduces a balance matrix T to represent the weight coefficient for obtaining information from the shortest path, and uses 1 - T to represent the weight coefficient for obtaining information from the surrounding paths. The path reachability network feature WA introduces a scale parameter α to limit the length of the two-step peripheral paths.

6. The training method according to claim 5, characterized in that The steps for dividing the multimodal imaging data one into training samples and test samples and establishing a prediction model between the path reachability feature WA of healthy subjects and their human brain functional connectivity matrix FC include: Based on the training samples, calculate the path reachability feature WA of healthy subjects, and establish a linear model between it and the human brain functional connectivity matrix FC. Obtain the optimal scale parameter α and the optimal linear model parameter W using the least mean square error constraint condition m ; To prevent overfitting, an L2-regularization term is added to the linear model to constrain the weights; Based on the test samples, using the optimal scale parameter α and the optimal linear model parameter W m Calculate the predicted functional connectivity matrix, and calculate the Pearson correlation coefficient between the predicted functional connectivity matrix and the true human brain functional connectivity matrix FC to describe the effect of human brain function construction. The larger the Pearson correlation coefficient, the better the effect of human brain function construction.

7. A method for auxiliary diagnosis of brain diseases, characterized in that, Based on the prediction model described in any one of claims 1-6, a brain disease classification model is established to achieve the auxiliary diagnosis of brain diseases, including: Obtaining multi-modal imaging data II of brain disease patients and healthy subjects with the same gender, age group, and education level, where the multi-modal imaging data II includes T1 data, dMRI data, and fMRI data; Preprocessing the multi-modal imaging data II to construct a human brain structural connection matrix SC for reflecting the brain's structural connection situation and a human brain functional connection matrix FC for reflecting the brain's functional activity situation; Calculating the path reachability feature WA of each healthy subject and brain disease patient, and sequentially inputting it into the prediction model to obtain the corresponding predicted simulated functional connection matrix FCs; Calculating the mean square error between the predicted simulated functional connection matrix FCs and the true functional connection matrix FCe, and the Pearson correlation coefficient between the predicted simulated functional connection matrix FCs and the true functional connection matrix FCe. Using these two as the classification features of healthy subjects and brain disease patients, a classification model of healthy subjects and brain disease patients is established using a machine learning algorithm to achieve the auxiliary diagnosis of brain diseases.

8. The brain disease assisted diagnosis method according to claim 7, wherein, The steps for calculating the path accessibility feature WA for each healthy subject and brain disease patient include: based on the optimal scale parameter α and the optimal linear model parameter W m Calculate the path accessibility feature WA for each healthy subject and brain disease patient.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the training method described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, Including a computer program, when the computer program runs on an electronic device, it causes the electronic device to execute the steps of the training method described in any one of claims 1 to 6.

Images