CN114190884B

CN114190884B - Longitudinal analysis method, system and device for brain disease data

Info

Publication number: CN114190884B
Application number: CN202010989643.9A
Authority: CN
Inventors: 雷柏英; 刘冬冬; 汪天富; 杨鹏
Original assignee: Shenzhen University
Current assignee: Shenzhen University
Priority date: 2020-09-18
Filing date: 2020-09-18
Publication date: 2024-05-28
Anticipated expiration: 2040-09-18
Also published as: CN114190884A

Abstract

The invention discloses a longitudinal analysis method, a longitudinal analysis system and a longitudinal analysis device for brain disease data, wherein the longitudinal analysis method comprises the following steps: preprocessing multi-center multi-time rs-fMRI data; adopting a similarity group sparse network to respectively construct brain function connection networks for the preprocessed rs-fMRI data; extracting features from the brain function connection network by using a bidirectional LSTM to obtain representative features; learning features are extracted from the representative features by adopting a self-attention mechanism method, and the learning features are sent into Softmax for classification. The method provided by the invention can be used for conveniently acquiring more physiological changes of the relevant brain region by carrying out deep learning longitudinal analysis on the rs-fMRI data of multiple centers and multiple time points.

Description

Longitudinal analysis method, system and device for brain disease data

Technical Field

The invention relates to the technical field of disease data analysis, in particular to a longitudinal analysis method, a storage medium, a system and a device for brain disease data.

Background

Alzheimer's Disease (AD) is one of the most serious symptoms of neurodegenerative diseases, and about 60-80% of cases of dementia worldwide are caused by AD. Memory and cognitive decline is a major early symptom, and severe people will not be self-care in life and will die within 3-10 years. The world alzheimer's association report in 2018 showed that the total number of AD patients worldwide is currently about 5000 tens of thousands, and this number is expected to increase to about 1.52 billion by 2050. Studies have shown that the global estimated cost of dementia in 2015 will exceed 9755.6 billions of dollars, well above the predicted 8000 billions of dollars reported for 2015 world alzheimer's disease, and this cost is expected to rise to 2.54 trillions dollars by 2030. It follows that AD has become a common worldwide problem. No effective drug has been used to date for the treatment of AD, so early detection of early stages of AD and timely prevention and intervention is necessary. Mild Cognitive Impairment (MCI) is a special condition intermediate between AD and normal people and is of great interest due to its high conversion rate to AD and the optimal diagnosis and intervention period for AD, so clinical research into AD has been focused on MCI.

In recent years, MCI research has focused mainly on brain function connection networks (brain function connected network, BFCN), which are based on studying time series of resting state functional magnetic resonance imaging (rs-fMRI) data, by establishing BFCN between brain regions to achieve the goal of functionally interpreting early brain function changes of MCI. However, most of the existing MCI diagnosis methods are based on data of a single center and a single time point, and neglect the problem of poor model generalization performance caused by obvious differences among different center data scanning parameters, scanning time points and scanning modes, and neglect that the data of multiple time points can provide more learning-aided information and brain diseases as a dynamic change process, so that longitudinal analysis can help clinicians to know about physiological changes of relevant brain regions, and is a more meaningful study.

Disclosure of Invention

The invention aims to solve the technical problems that the prior art is inconvenient to acquire physiological changes of more relevant brain areas because rs-fMRI data based on a single center and a single time point are analyzed.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows:

a method of longitudinal analysis of brain disease data, comprising the steps of:

preprocessing multi-center multi-time rs-fMRI data;

Adopting a similarity group sparse network to respectively construct brain function connection networks for the preprocessed rs-fMRI data;

extracting features from the brain function connection network by using a bidirectional LSTM to obtain representative features;

Learning features are extracted from the representative features by a self-attention mechanism method, and the learning features are sent to Softmax for classification.

The longitudinal analysis method of brain disease data, wherein the preprocessing of the multi-center multi-time-point rs-fMRI data comprises the following steps:

obtaining baseline and annual rs-fMRI data from ADNI-2 and ADNI-3 databases as multicenter multi-time point rs-fMRI data;

Discarding the first 10 rs-fMRI sequences of the multi-center multi-time-point rs-fMRI data to obtain first rs-fMRI data;

Performing scanning difference elimination processing on the first rs-fMRI data by adopting head motion correction, time registration, smoothing and space normalization methods to obtain second rs-fMRI data;

performing noise elimination processing on the second rs-fMRI data by removing baseline drift and filtering high-frequency signal interference to obtain third rs-fMRI data;

And comparing the automatic anatomical labeling template with the third rs-fMRI data, dividing the brain space into 90 regions of interest, and taking an average value as the preprocessed rs-fMRI data.

The method for longitudinally analyzing brain disease data, wherein the step of respectively constructing brain function connection networks for the preprocessed rs-fMRI data by adopting a similarity group sparse network comprises the following steps of:

Constructing a similarity group sparse network, wherein an objective function of the similarity group sparse network is expressed as:

wherein/> AndThe method comprises the steps of respectively obtaining a group sparse regularization term and a similarity constraint term;

Wherein, ，

Wherein/>And/>Representing the group sparse regularization term and similarity constraint term parameters, respectively,/>Representation/>/>Sum of norms,/>The difference between two continuous weighting vectors of the same group is made as small as possible;

Inputting the preprocessed rs-fMRI data into the similarity group sparse network to construct a brain function connection network.

The longitudinal analysis method of the brain disease data further comprises the following steps:

dividing an objective function of the similarity group sparse network into a similarity constraint and a non-similarity constraint, wherein the similarity constraint expression is as follows: ；

The dissimilarity constraint expression is: ；

Using the formula And/>Optimizing the similarity constraint, wherein/>And/>To express/>At/>Gradient and step size at,/>Is determined by line search;

Using the formula And/>Optimizing non-similarity constraints, wherein/>Is a pretrained value,/>And/>Representation/>At/>Gradient and step size at the same time.

The method for longitudinal analysis of brain disease data, wherein the extracting features from the brain function connection network and selecting representative features using a bidirectional LSTM comprises the steps of:

Constructing a bidirectional LSTM comprising a forward LSTM and a backward LSTM, the forward LSTM expressed as: The backward LSTM is expressed as: /(I) Wherein/>And/>Weights of forward LSTM,/>, respectivelyAnd/>Respectively represent the weight of backward LSTM,/>And/>Deviations of forward LSTM and backward LSTM, respectively;

the final output of the addition of forward LSTM and backward LSTM is taken as the final output of the bi-directional LSTM, expressed as: ；

Features are extracted from the brain function connection network and representative features are selected using the bi-directional LSTM expression.

The method for longitudinally analyzing brain disease data, wherein the method for adopting a self-attention mechanism further extracts learning characteristics from the representative characteristics, sends the learning characteristics into Softmax for classification, and the step of longitudinally analyzing brain disease comprises the following steps:

A self-attention mechanism is constructed, and the expression is as follows: Wherein Q is a query vector, and/> ; V is a value vector, and/>; K is a bond vector, and/>；

And taking the representative characteristic as an input characteristic of the self-attention mechanism to obtain an output result, and realizing longitudinal analysis of the brain diseases.

A system for longitudinally analyzing brain disease, comprising:

the preprocessing module is used for preprocessing the rs-fMRI data of the multi-center and multi-time points;

The similarity group sparse network module is used for respectively constructing brain function connection networks for the preprocessed rs-fMRI data;

the bidirectional LSTM module is used for extracting features from the brain function connection network to obtain representative features;

And the classification module is used for extracting learning features from the representative features and sending the learning features to Softmax for classification.

A storage medium, wherein the storage medium stores one or more programs executable by one or more processors to implement steps in a method for longitudinal analysis of brain disease data according to the present invention.

A device for longitudinally analyzing brain diseases, comprising: a processor, a memory, and a communication bus, the memory having stored thereon a computer readable program executable by the processor;

The communication bus realizes connection communication between the processor and the memory;

The processor, when executing the computer readable program, implements the steps in the longitudinal analysis method of brain disease data according to the present invention.

The beneficial effects are that: compared with the prior art, the longitudinal analysis method of brain disease data provided by the invention is used for identifying Early MCI (EMCI) by combining a similarity-group sparse network (SGN) model with a self-attention-based stacked bidirectional LSTM (self attention based stacked long short term memory, SASBiLSTM) feature learning method and analyzing a plurality of central and multiple time points rs-fMRI data. Specifically, the invention firstly uses SGN frame to respectively construct multi-time point brain function connection network for multi-center data; the time-step nature of the bi-directional LSTM, which can learn longitudinally about the similarity and specificity of the multi-time-point features by time-steps, then extracts multiple time-point features from the brain function connection network. Finally, deep features favorable for disease analysis are selected through a self-attention mechanism, and the learned features are sent to a Softmax classifier for classification. The longitudinal analysis method of brain disease data provided by the invention evaluates in the second stage and the third stage (ADNI-2 and ADNI-3) of the public Alzheimer's disease neuroimaging initiative and the database, and the proposed method obtains the classification accuracy of 91.42%, 89.42% and 86.11% on three classification tasks of LMCI vs. NC, EMCI vs. LMCI, LMCI vs. NC.

Drawings

Fig. 1 is a flowchart of a method for longitudinally analyzing brain disease data according to a preferred embodiment of the present invention.

Fig. 2 is a schematic flow chart of a method for longitudinally analyzing brain disease data.

Fig. 3 is a schematic structural diagram of a longitudinal analysis device for brain disease data according to the present invention.

Detailed Description

The invention provides a longitudinal analysis method, a storage medium, a system and a device for brain disease data, which are used for making the purposes, technical schemes and effects of the invention clearer and more definite, and the invention is further described in detail below by referring to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

As used herein, the singular forms "a", "an", "the" and "the" are intended to include the plural forms as well, unless expressly stated otherwise, as understood by those skilled in the art. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Further, "connected" or "coupled" as used herein may include wirelessly connected or wirelessly coupled. The term "and/or" as used herein includes all or any element and all combination of one or more of the associated listed items.

It will be understood by those skilled in the art that all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs unless defined otherwise. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

The invention will be further described by the description of embodiments with reference to the accompanying drawings.

Most of the existing MCI analysis methods are based on data of a single center and a single time point, the method ignores the problem of poor model generalization performance caused by obvious differences among different center data scanning parameters, scanning time points and scanning modes, and ignores the fact that the data of multiple time points can provide more information helpful for learning and brain diseases are dynamic change processes, so that longitudinal analysis can help clinicians to know physiological changes of relevant brain areas, and the method is a more meaningful study. Furthermore, current MCI analysis methods are mostly focused on traditional machine learning methods, while deep learning has been demonstrated in recent years to have good feasibility and high efficiency in medical image processing.

Based on the above problems, the present invention provides a flowchart of a preferred embodiment of a method for longitudinal analysis of brain disease data, as shown in fig. 1, which includes the steps of:

s10, preprocessing multi-center multi-time rs-fMRI data;

S20, adopting a similarity group sparse network to respectively construct brain function connection networks for the preprocessed rs-fMRI data;

s30, extracting features from the brain function connection network by utilizing a bidirectional LSTM and selecting representative features;

And S40, further extracting learning features from the representative features by adopting a self-attention mechanism method, and sending the learning features into a Softmax for classification.

The present embodiment identifies early MCI by analyzing multiple center multiple time points rs-fMRI data by combining a similarity set sparse network model with a self-attention stack based bi-directional LSTM feature learning method. Specifically, as shown in fig. 2, in this embodiment, first, a multi-point-in-time brain function connection network is respectively constructed for multi-center data using an SGN framework; then extracting a plurality of time point features from the brain function connection network by the time step characteristics of the bidirectional LSTM, wherein the bidirectional LSTM can longitudinally learn the similarity and the specificity of the multi-time point features through the time steps; finally, deep features favorable for disease diagnosis are selected through a self-attention mechanism, and the learned features are sent to a Softmax classifier for classification. The longitudinal analysis method of brain disease data proposed in this embodiment evaluates in the second and third stages (ADNI-2 and ADNI-3) of the public alzheimer's neuroimaging initiative and the database, and the proposed method obtains classification accuracy of 91.42%, 89.42% and 86.11% on LMCI vs. NC, EMCI vs. LMCI, LMCI vs. NC classification tasks, respectively.

In some embodiments, the present example obtains data from both time points (baseline and one year later) from the ADNI-2 and ADNI-3 databases, wherein ADNI-2 contains 33 NCs, 39 EMCI and 30 LMCI subjects 'rs-fMRI data, and ADNI-3 contains 25 NCs, 16 EMCI and 10 LMCI subjects' rs-fMRI data. The acquired multi-center multi-time rs-fMRI data are preprocessed, and the method specifically comprises the following steps: the first 10 rs-fMRI sequences of each subject are discarded to achieve the aim of keeping the magnetization unchanged; then eliminating the influence of scanning difference by adopting methods such as head motion correction, time registration, smoothing, space normalization and the like; then, eliminating the influence of noise by removing baseline wander and filtering high frequency signal interference; finally, brain space is divided into 90 regions of interest (ROIs) by using an Automated Anatomical Label (AAL) template to compare to the rs-fMRI data, and they are averaged as pre-processed rs-fMRI data.

In some embodiments, constructing an efficient brain function connection network after preprocessing the rs-fMRI data is an important means to help improve analysis accuracy. In this embodiment, bold uppercase letters, bold lowercase letters, and common italic letters are used to represent matrices, vectors, and scalars, respectively. Assuming K subjects, the present embodiment uses AAL templates to segment the brain into I regions of interest, then the input data of the present embodimentCan be expressed as/>Then the kth subject's/>The M blood oxygen level dependent average time series signals contained in the ROIs can be expressed as. The group sparse brain network (GCS) is to use one of the brain regions to constrain the remaining 89 brain regions, then the data for all ROIs to remove one of them can be expressed as/>. Use/>Sum of norms to accomplish set sparsity constraints, in particular,/>Norms are used for subject number/>Features,/>Norms are used to constrain/>Weight of ROIs. In order to achieve the goal of combining the connection weights of different subjects together, the change in the connection weights of different subjects is utilized, thereby reconstructing the target ROI using the remaining ROIs. Then this embodiment uses/>-Norm constraint target matrix/>Imposing a common connection topology between the principals and on the same elements of the (c). Furthermore, the reconstruction of each ROI is independent of the other ROIs. One advantage of GCS is that it allows all models within the same group to have the same way of connection, helping to obtain learning-friendly features. While the group constraint sparse network can guarantee both group constraint and sparsity, it may also ignore similarity parameters between different objects, resulting in fewer discriminative features. In order to solve the problem, the embodiment explores a similarity group sparse network model, and learns the brain function network of each sample through similarity constraint group sparse learning, and the objective function is expressed as:/>Wherein/>And/>The method comprises the steps of respectively obtaining a group sparse regularization term and a similarity constraint term;

Wherein, ，

Wherein/>And/>Representing the group sparse regularization term and similarity constraint term parameters, respectively,/>Representation/>/>Sum of norms,/>The difference between two continuous weighting vectors of the same group is made as small as possible so as to achieve the purpose of similarity constraint. In the embodiment, regularization items of the model are fused together to finish model optimization. This sparse learning model is referred to as a similarity group sparse network (SGN) in this embodiment. ADNI2 and ADNI3 data set construction of the SGN brain network at multiple time points is performed separately, so that the characteristics of BFCN at different centers can be fully revealed. Inputting the preprocessed rs-fMRI data into the similarity group sparse network to construct a brain function connection network.

In some specific embodiments, the objective function of the similarity group sparse network is divided into a similarity constraint and a non-similarity constraint, and then optimized step by step, wherein the similarity constraint expression is:

；

The dissimilarity constraint expression is: ；

In the optimization of the specific projection gradient descent algorithm, the present embodiment uses two steps to complete the optimization process of the objective function, specifically, in the nth iteration, the first step uses respectively And/>To express/>At/>Gradient and step size at, furthermore/>Is determined by line search, the first and second steps of the specific process can be expressed by the formula/>, respectivelyAnd/>Optimizing the similarity constraint.

After optimization of similarity constraints is completed, the embodiment performs cyclic computation on the non-similarity constraint terms by using near-end operators related to Lasso groups and fusion Lasso constraintsAnd (5) optimizing. To get an approximate solution faster, the present embodiment further accelerates the gradient using an accelerated gradient descent method. Unlike the similarity constraint, the gradient descent calculation search points are based on/>Executing the formula/>And/>Optimizing non-similarity constraints, wherein/>Is a pretrained value,/>And/>Representation/>At/>Gradient and step size at the same time.

In some implementations, after constructing the BFCN, the present example first mixes SGN brain networks from ADNI2 and ADNI3 datasets together randomly and then learns the multi-center data using a feature learning algorithm. The disease analysis result can be improved efficiently due to proper feature learning. Therefore, the embodiment proposes to extract the features from the constructed SGN by using SBiLSTM, and experiments show that the method can effectively improve the disease targeting efficiency, help to obtain better experimental results, and unlike other multi-time-point learning algorithms, the method uses video learning to take each frame as a thought of one time step of LSTM, and takes each time point of multi-time-point data as one time step, so that the system can be helped to learn the similarity between different time points. Before introducing the bidirectional LSTM, we will first introduce the basic LSTM, where the basic LSTM unit has input and output gates to control the input of the feature and the learned feature output, respectively, and may be expressed as:

Wherein/> Respectively represent input and output gates,/>Representing an activation function,/>Representing input and output characteristics/>, respectivelyAnd the previous cell unit/>Is a weight coefficient matrix of (a). /(I)Representing the bias of the inputs and outputs, respectively. The above formula is an input-output function of LSTM, but LSTM can achieve better results because he has a forgetting gate that can help the current neuron forget unimportant features from the previous neuron and retains more important features, so the forgetting gate is critical in LSTM and can be expressed as: identical to the above formula,/> ,/>Weight coefficient matrix used to represent input gate and last cell unit in forgetting gate respectively,/>To forget the bias in the gate. After the forget gate is introduced, the system can introduce the LSTM in its entirety, and the basic LSTM uses a special mechanism to decide which elements to select to remember and forget. Can be expressed by two formulas:/>Wherein/>Is the current state of the neuron. /(I)Is a long-term memory cell consisting of/>,/>、/>And/>Control,/>Is the parameter of the last cell, when the parameter is large/>The previous cell will be remembered, otherwise forgotten. /(I)For inputting data/>Weights of/>For the last output neuron/>Weight of (2), as element multiplier.

After explaining basic LSTM cells, this example will introduce the use of the bidirectional LSTM herein for defining LSTM cells, including forward LSTM and backward LSTM, expressed as: The backward LSTM is expressed as: /(I) Wherein/>And/>Weights of forward LSTM,/>, respectivelyAnd/>Respectively represent the weight of backward LSTM,/>And/>Deviations of forward LSTM and backward LSTM, respectively; the final output of the addition of forward LSTM and backward LSTM is taken as the final output of the bi-directional LSTM, expressed as: ; features are extracted from the brain function connection network and representative features are selected using the bi-directional LSTM expression.

In some embodiments, although LSTM has proven to have good results for brain disease data analysis, there is room for improved detection performance, and attention is directed to a widely used method. Because the self-attention obtains better results in processing the sequence data, the embodiment designs a novel self-attention mechanism for learning tasks, and the self-attention can quickly and effectively capture context information with higher similarity through a query vector, a value vector and a key vector help model, thereby better helping a system to find pathogenic key areas so as to achieve the aim of helping to improve the disease monitoring efficiency.

Specifically, when we get the output characteristics from the last part, we take the output of the last part as the current input and create a query vector) A value vector (/ >)) We then create a key vector (/ >) based on the functional magnetic resonance imaging signals) The present embodiment then multiplies the input query vector Q by the value vector V to obtain a score. Next, the present embodiment divides this by 8/>And then inputting the result into a Softmax function for standardization, and finally obtaining an output result. Self-attention can be expressed as: ; and taking the representative characteristic as an input characteristic of the self-attention mechanism to obtain an output result, and realizing longitudinal analysis of the brain diseases.

The invention provides a brain function network construction method based on similarity group sparse constraint, which can effectively help construct a brain network connection network and fully consider the similarity among individuals, thereby helping to obtain the characteristics favorable for model learning; the invention also designs a bidirectional LSTM framework for effectively utilizing the information of the multi-time-point characteristics to detect the brain diseases, and the method can effectively utilize the bidirectional LSTM time-step learning multi-time-point characteristics, thereby being beneficial to the analysis of disease data; the embodiment also provides a self-care mechanism to find out the most discriminating characteristic in the disease detection so as to improve the analysis performance of the brain disease data.

In some embodiments, based on the above-mentioned method for longitudinally analyzing brain disease data, the present embodiment further provides a system for longitudinally analyzing brain disease, which includes:

And the classification module is used for extracting learning features from the representative features, and sending the learning features into Softmax for classification so as to realize longitudinal analysis of brain diseases.

In some embodiments, the preprocessing module comprises:

The data acquisition unit is used for acquiring the baseline and the rs-fMRI data after one year from the ADNI-2 and ADNI-3 databases and taking the baseline and the rs-fMRI data as multi-center and multi-time-point rs-fMRI data;

the discarding unit is used for discarding the first 10 rs-fMRI sequences of the rs-fMRI data of the multi-center multi-time point to obtain first rs-fMRI data;

The scanning difference eliminating unit is used for carrying out scanning difference eliminating processing on the first rs-fMRI data by adopting head motion correction, time registration, smoothing and space normalization methods to obtain second rs-fMRI data;

the noise elimination unit is used for eliminating baseline drift and filtering high-frequency signal interference to perform noise elimination processing on the second rs-fMRI data to obtain third rs-fMRI data;

And the comparison unit is used for comparing the automatic anatomical marker template with the third rs-fMRI data, dividing the brain space into 90 regions of interest, and taking an average value as the preprocessed rs-fMRI data.

In some embodiments, the affinity group sparse network module comprises:

The similarity group sparse network construction unit is used for constructing a similarity group sparse network, and the objective function of the similarity group sparse network construction unit is expressed as:

Wherein, ，

And the brain function connection network unit is used for inputting the preprocessed rs-fMRI data into the similarity group sparse network to construct a brain function connection network.

In some embodiments, the system for longitudinally analyzing brain disease further comprises:

The optimization module is used for dividing the objective function of the similarity group sparse network into similarity constraint and non-similarity constraint, wherein the similarity constraint expression is as follows: ; the dissimilarity constraint expression is: /(I) ; Using the formula/>And/>Optimizing the similarity constraint, wherein/>And/>To express/>At/>Gradient and step size at,/>Is determined by line search; using the formula/>And/>Optimizing non-similarity constraints, wherein/>Is a pretrained value,/>AndRepresentation/>At/>Gradient and step size at the same time.

In some embodiments, the bi-directional LSTM module includes:

a bidirectional LSTM construction unit configured to construct a bidirectional LSTM, where the bidirectional LSTM includes a forward LSTM and a backward LSTM, and the forward LSTM is expressed as: the backward LSTM is expressed as: wherein/> And/>Weights of forward LSTM,/>, respectivelyAnd/>Respectively represent the weight of backward LSTM,/>And/>Deviations of forward LSTM and backward LSTM, respectively; the final output of the addition of forward LSTM and backward LSTM is taken as the final output of the bi-directional LSTM, expressed as: /(I); Features are extracted from the brain function connection network and representative features are selected using the bi-directional LSTM expression.

In some embodiments, the classification module comprises:

A self-attention mechanism construction unit, configured to construct a self-attention mechanism, where an expression is: Wherein Q is a query vector, and/> ; V is a value vector, and/>; K is a bond vector, and/>；

And the result output unit is used for taking the representative characteristic as the input characteristic of the self-attention mechanism to obtain an output result and realize the longitudinal analysis of the brain diseases.

In some embodiments, a storage medium is also provided, where the storage medium stores one or more programs executable by one or more processors to implement steps in the method for longitudinal analysis of brain disease data according to the present invention.

In some embodiments, an apparatus for longitudinally analyzing brain diseases, as shown in FIG. 3, includes at least one processor (processor) 20; a display screen 21; and a memory (memory) 22, which may also include a communication interface (Communications Interface) 23 and a bus 24. Wherein the processor 20, the display 21, the memory 22 and the communication interface 23 may communicate with each other via a bus 24. The display screen 21 is configured to display a user guidance interface preset in the initial setting mode. The communication interface 23 may transmit information. The processor 20 may invoke logic instructions in the memory 22 to perform the methods of the embodiments described above.

Further, the logic instructions in the memory 22 described above may be implemented in the form of software functional units and stored in a computer readable storage medium when sold or used as a stand alone product.

The memory 22, as a computer readable storage medium, may be configured to store a software program, a computer executable program, such as program instructions or modules corresponding to the methods in the embodiments of the present disclosure. The processor 20 performs functional applications and data processing, i.e. implements the methods of the embodiments described above, by running software programs, instructions or modules stored in the memory 22.

The memory 22 may include a storage program area that may store an operating system, at least one application program required for functions, and a storage data area; the storage data area may store data created according to the use of the terminal device, etc. In addition, the memory 22 may include high-speed random access memory, and may also include nonvolatile memory. For example, a plurality of media capable of storing program codes such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), a magnetic disk, or an optical disk, or a transitory storage medium may be used.

In addition, the specific processes that the storage medium and the plurality of instruction processors in the terminal device load and execute are described in detail in the above method, and are not stated here. Wherein, include: a processor, a memory, and a communication bus, the memory having stored thereon a computer readable program executable by the processor;

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for longitudinal analysis of brain disease data, comprising the steps of:

preprocessing multi-center multi-time rs-fMRI data;

wherein/> And/>Group sparse regularization term and similarity constraint term, respectively,/>Is the mean time series signal of blood oxygen level dependence contained in the ith region of interest of the kth subject,/>To be except/>An average time-series signal dependent on blood oxygen levels for all regions of interest except;

Wherein, ，/>Wherein/>And/>Representing the group sparse regularization term and similarity constraint term parameters, respectively,/>Representation/>/>Sum of norms,/>The difference between two continuous weighting vectors of the same group is made as small as possible;

inputting the preprocessed rs-fMRI data into the similarity group sparse network to construct a brain function connection network;

Constructing a bidirectional LSTM comprising a forward LSTM and a backward LSTM, the forward LSTM expressed as: The backward LSTM is expressed as: /(I) Wherein/>And/>Weights of forward LSTM,/>, respectivelyAnd/>Respectively represent the weight of backward LSTM,/>AndDeviations of forward LSTM and backward LSTM, respectively;

Extracting features from the brain function connection network and selecting representative features using a bi-directional LSTM expression;

extracting learning features from the representative features by adopting a self-attention mechanism method, and sending the learning features into Softmax for classification;

the preprocessing of the multi-center multi-time-point rs-fMRI data comprises the following steps:

2. The method for longitudinal analysis of brain disease data according to claim 1, further comprising the step of:

The dissimilarity constraint expression is: ；

3. The method of claim 1, wherein the step of extracting learning features from the representative features by using a self-attention mechanism, and sending the learning features to Softmax for classification comprises:

4. A system for longitudinally analyzing brain diseases implementing the longitudinal analysis method of brain disease data according to any one of claims 1 to 3, comprising:

the similarity group sparse network module is used for constructing a brain function connection network for the preprocessed rs-fMRI data;

A bi-directional LSTM module for extracting features from the brain function connection network and selecting representative features;

5. A storage medium storing one or more programs executable by one or more processors to perform the steps in the method of longitudinal analysis of brain disease data according to any one of claims 1-3.

6. A device for longitudinally analyzing brain disease, comprising: a processor, a memory, and a communication bus; the memory has stored thereon a computer readable program executable by the processor;

The processor, when executing the computer readable program, implements the steps of the method for longitudinal analysis of brain disease data according to any one of claims 1-3.