WO2023123380A1 - 基于弱监督对比学习的动态成瘾神经环路生成方法及系统 - Google Patents

基于弱监督对比学习的动态成瘾神经环路生成方法及系统 Download PDF

Info

Publication number
WO2023123380A1
WO2023123380A1 PCT/CN2021/143741 CN2021143741W WO2023123380A1 WO 2023123380 A1 WO2023123380 A1 WO 2023123380A1 CN 2021143741 W CN2021143741 W CN 2021143741W WO 2023123380 A1 WO2023123380 A1 WO 2023123380A1
Authority
WO
WIPO (PCT)
Prior art keywords
brain
dynamic
connection
addiction
brain connection
Prior art date
Application number
PCT/CN2021/143741
Other languages
English (en)
French (fr)
Inventor
王书强
宫长威
Original Assignee
深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Priority to PCT/CN2021/143741 priority Critical patent/WO2023123380A1/zh
Priority to US18/111,876 priority patent/US20230215006A1/en
Publication of WO2023123380A1 publication Critical patent/WO2023123380A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G06T7/0014—Biomedical image inspection using an image reference approach
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R33/00—Arrangements or instruments for measuring magnetic variables
    • G01R33/20—Arrangements or instruments for measuring magnetic variables involving magnetic resonance
    • G01R33/44—Arrangements or instruments for measuring magnetic variables involving magnetic resonance using nuclear magnetic resonance [NMR]
    • G01R33/48—NMR imaging systems
    • G01R33/54—Signal processing systems, e.g. using pulse sequences ; Generation or control of pulse sequences; Operator console
    • G01R33/56—Image enhancement or correction, e.g. subtraction or averaging techniques, e.g. improvement of signal-to-noise ratio and resolution
    • G01R33/5608—Data processing and visualization specially adapted for MR, e.g. for feature analysis and pattern recognition on the basis of measured MR data, segmentation of measured MR data, edge contour detection on the basis of measured MR data, for enhancing measured MR data in terms of signal-to-noise ratio by means of noise filtering or apodization, for enhancing measured MR data in terms of resolution by means for deblurring, windowing, zero filling, or generation of gray-scaled images, colour-coded images or images displaying vectors instead of pixels
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0475—Generative networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/62—Extraction of image or video features relating to a temporal dimension, e.g. time-based feature extraction; Pattern tracking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R33/00—Arrangements or instruments for measuring magnetic variables
    • G01R33/20—Arrangements or instruments for measuring magnetic variables involving magnetic resonance
    • G01R33/44—Arrangements or instruments for measuring magnetic variables involving magnetic resonance using nuclear magnetic resonance [NMR]
    • G01R33/48—NMR imaging systems
    • G01R33/4806—Functional imaging of brain activation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10088—Magnetic resonance imaging [MRI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20076—Probabilistic image processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30016—Brain
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images

Definitions

  • the present invention relates to the field of artificial intelligence, in particular to a method and system for generating a dynamic brain addiction neural circuit based on weakly supervised contrastive learning.
  • Addiction is a disease characterized by seeking compulsive drugs. Take smoking addiction as an example. At present, there are more than 300 million tobacco users in my country, and as many as 1 million people die from tobacco-related diseases every year. From the perspective of inducing smoking behavior, nicotine addiction is the main inducement of smoking and the main obstacle for smokers to quit smoking. Addiction is also regarded as a chronic relapsing functional brain disease.
  • the detection of the neural circuit of nicotine addiction is to analyze and calculate the abnormal neural circuit from fMRI (functional magnetic resonance imaging).
  • fMRI functional magnetic resonance imaging
  • traditional statistics-based methods require complex preprocessing operations on image data, which has a large amount of redundancy; moreover, fMRI image data has the defect of small samples, which makes training difficult and low in training accuracy.
  • a method for generating a dynamic addiction neural circuit based on weakly supervised contrastive learning provided by the present invention adopts the following technical scheme:
  • a method for generating dynamic addiction neural circuits based on weakly supervised contrastive learning including:
  • each set of fMRI includes preset Multiple fMRI images acquired at a fixed frequency over time;
  • the step of extracting the spatio-temporal features of the brain connection in the dynamic brain connection map it also includes:
  • the standard brain connection map is used as a real sample, and the dynamic brain connection map is used as a fake sample to generate adversarial learning, and the dynamic brain connection map is updated.
  • a dynamic brain connection map that not only conforms to the spatial position information of the standard brain template, but also has the original fMRI voxel spatial information is obtained, which is conducive to the accuracy of training; and, as a standard space for prior knowledge The amount of brain template data is small, and the learning difficulty is low.
  • the step of extracting the spatio-temporal features of brain connections in the dynamic brain connection map specifically includes:
  • the temporal features of the brain connections in the dynamic brain connection map are extracted by using the gated recurrent unit based on the time-series brain connection attention mechanism.
  • the spatial and temporal features of the dynamic brain connection map are extracted respectively, so as to obtain the spatiotemporal characteristics of the brain connection, which fully characterizes the properties of the dynamic addiction brain connection.
  • inputting the spatiotemporal features into the abnormal connection detection network, and calculating the abnormal brain connection probability based on comparative learning specifically includes;
  • F score MLP(H t );
  • h i is the hidden state feature vector of the i-th node
  • h j is the hidden state feature vector of the j-th node
  • ⁇ ( ) is the sigmoid function
  • a and b are the optimization parameters of the output layer
  • ⁇ and ⁇ are A pair of hyperparameters
  • ⁇ represents the weight of the edge e ij ;
  • L TOTAL L S + ⁇ a ⁇ A, a' ⁇ A L ⁇ ;
  • HS is the spatio-temporal feature of the saline injection group
  • H a is the spatio-temporal feature of the nicotine injection group
  • A ⁇ 1,2,...,A′ ⁇
  • A′ is the nicotine
  • ⁇ is a hyperparameter.
  • the comparison learning between the normal saline injection group and all the nicotine injection groups is carried out, and the pairwise comparison learning is carried out on the multiple groups injected with nicotine, which can fully analyze the differences in different control groups.
  • Individual differences in dynamic brain connectivity maps facilitate the discovery of abnormal brain connectivity in addiction.
  • the step of generating a dynamic addiction neural circuit based on neuroscience prior knowledge and the brain connection with the highest abnormal probability at each moment specifically includes:
  • the brain connection with the highest abnormal probability is corrected and integrated to obtain the addiction neural circuit at each moment, and a dynamic addiction neural circuit is generated according to the time series.
  • the brain connection with the highest abnormal probability at each moment is corrected and normalized, thereby ensuring the accuracy of the generated addiction neural circuit.
  • the present invention provides a dynamic addiction neural circuit generation system based on weakly supervised contrastive learning, which adopts the following technical solutions:
  • a dynamic addiction neural circuit generation system based on weakly supervised contrastive learning including:
  • the dynamic brain connection map generation module is used to reduce the voxels of multiple sets of fMRI images to the attributes of brain region nodes based on the convolutional neural network, and generate multiple groups of dynamic brain connection maps including time series according to the attributes of the brain region nodes ; Wherein, each group of fMRI includes multiple fMRI images collected at a fixed frequency within a preset time period;
  • a spatiotemporal feature extraction module for extracting spatiotemporal features of brain connections in each of the dynamic brain connectivity graphs
  • the most abnormal brain connection acquisition module is used to input the spatiotemporal features into the abnormal connection detection network, calculate the abnormal probability of brain connection based on comparative learning, and obtain the brain connection with the highest abnormal probability at each moment;
  • the dynamic addiction neural circuit generation module is used to generate a dynamic addiction neural circuit based on neuroscience prior knowledge and the brain connection with the highest abnormal probability at each moment.
  • the present invention provides an electronic device, which adopts the following technical solution:
  • the electronic device includes a memory and a processor, and the memory stores a computer program that can be loaded by the processor and execute the method.
  • the present invention provides a computer-readable storage medium, which adopts the following technical solution:
  • a computer-readable storage medium stores a computer program that can be loaded by a processor and execute the method.
  • the present invention includes at least one of the following beneficial technical effects:
  • Figure 1 is a flowchart of a method for generating dynamic addiction neural circuits based on weakly supervised contrastive learning.
  • Fig. 2 is a flowchart of a method for extracting spatiotemporal features of brain connections.
  • Figure 3 is a structural block diagram of a dynamic addiction neural circuit generation system based on weakly supervised contrastive learning.
  • Figure 4 is a structural block diagram of the spatio-temporal feature extraction module.
  • Fig. 5 is a schematic diagram of an electronic device.
  • the embodiment of the present invention discloses a method for generating a dynamic addiction neural circuit based on weakly supervised contrastive learning.
  • the method for generating dynamic addiction neural circuits based on weakly supervised contrastive learning includes:
  • the voxels of multiple sets of fMRI images are reduced to the attributes of brain region nodes, and multiple sets of dynamic brain connection maps including time series are generated according to the attributes of the brain region nodes.
  • fMRI images of rat animal models in nicotine addiction experiments to generate dynamic addiction neural circuits based on weakly supervised contrastive learning.
  • Multiple groups of fMRI images include the fMRI images of the rats in the saline injection group and the nicotine injection group.
  • the number of rats in the saline group is 1, and the number of rats in the nicotine injection group can be determined according to the actual situation.
  • Set preferably 1-4.
  • Each group of fMRI images includes multiple fMRI images collected at a fixed frequency within a preset time period, where the preset time period is the length of time the rats are injected with saline or nicotine, which can be set to two weeks, three weeks, four weeks or other The duration is not limited here.
  • the rats in the saline injection group and the nicotine injection group were used for the experiment.
  • the rats in the nicotine group can be divided into multiple groups according to different concentrations, or only one group can be set, and each group of rats is continuously injected with the preset
  • the time period is long to simulate nicotine ingestion and addiction.
  • the fMRI images of rats in each group were collected at a fixed frequency, and the fMRI images of rats in each group reflected the changes of brain nerves in the time dimension after injection of normal saline or corresponding concentrations of nicotine.
  • the dynamic brain connection map G is: Specifically, in the dynamic brain connection graph G, fMRI time segments are used as time series, each time series is regarded as a time point, brain regions are nodes, and brain connections are edges. Brain connections refer to functional connections between brain regions, and T is The total number of time points, G t is a snapshot of the dynamic brain connection graph at time t, V t and E t respectively represent the node set and edge set in it, and the edge e ij ⁇ E t means that at time t, the i-th
  • the brain function connection between node j and node j, and its weight is ⁇ , Represents the attribute set of the node, N represents the total number of nodes, is the attribute vector of the i-th node at time t.
  • fMRI images are directly input into the weakly supervised adversarial learning model to obtain a dynamic brain connection map, eliminating redundant and complex preprocessing calculations.
  • the standard brain connection map is the brain function connection map preprocessed by the standard spatial brain template containing the prior information of the brain atlas.
  • the standard brain connection map is used as the real sample G True , and the dynamic brain connection map obtained in the previous step is used as the fake sample G Fake to generate adversarial learning and update the dynamic brain connection map.
  • D is a discriminator, representing a standard brain connection map
  • G is a generator, representing an abnormal brain connection map.
  • G specifically refers to a dynamic brain connection map.
  • step S12 is optional. If step 12 is performed, through confrontational learning, a dynamic brain connection map that not only conforms to the spatial position information of the standard brain template, but also has the original fMRI voxel spatial information is obtained, which improves the accuracy of training; The preprocessing of the standard spatial brain template is used to obtain the labeled brain functional connection map.
  • the standard spatial brain template has a small amount of data and low learning difficulty.
  • Each group of dynamic brain connection maps is input into the neural network for learning, and the spatial and temporal features of brain connections in the dynamic brain connection maps are extracted.
  • step S13 includes the following sub-steps:
  • the graph convolutional neural network based on the topological space brain connection focus mechanism is composed of multi-head anomalous connection focus block (Multi-head Anomalous Connection Focus Block, MACFB), which has a powerful graph learning ability and can learn the connection between nodes in the network. Geometry features.
  • Multi-head Anomalous Connection Focus Block MACFB
  • the total number of abnormal connection attention blocks is C, and the value of C is preferably 2, 3 or 4.
  • the multi-head abnormal connection attention block of layer l, t ⁇ 1,2,...,T ⁇ , T is the total number of time points
  • i represents the node
  • N is the total number of nodes.
  • i a node
  • c an abnormal connection attention block
  • is a learnable weight matrix for computing query-node attribute vectors is the offset of the calculation query-node attribute vector
  • j represents the node, is a learnable weight matrix for computing key-node attribute vectors, is the offset to calculate the key-node attribute vector;
  • e ij is an edge in the dynamic brain connection graph, which represents the brain connection between the i-th node and the j-th node at time t, and its weight is ⁇ .
  • W c,e is the learnable weight matrix for computing additional brain connectivity feature vectors
  • b c,e is the offset for computing additional brain connectivity feature vectors.
  • ⁇ q,k> is an exponential dot product function
  • N(i) is a value other than i.
  • the message transmission in the graph is carried out to output the hidden state feature vector of the next layer of nodes;
  • the short-term and long-term hidden state features are learned, so that the hidden state features of each time-series brain functional connection map include the hidden state information of other time-series context brain connections, so as to learn the dynamic Temporal features of brain connectograms.
  • Ht represents the output of the graph convolutional neural network, that is, the spatial features of the abnormal brain connection patterns in the extracted dynamic brain connection graph.
  • the window size is w+1, and i represents the node
  • r is the weight vector and Q is the weight matrix
  • H t+1 GRU(Current t+1 ,Short t+1 );
  • Short t+1 represents the short-term hidden state feature vectors of all nodes
  • the set of i ⁇ 1,2,...,N ⁇ , N is the total number of nodes.
  • U P and W P are both learnable weight matrices for calculating P t+1
  • b P is an offset for calculating P t+1 .
  • R t+1 ⁇ (U R Current t+1 +W R Short t+1 +b R );
  • U R and W R are both learnable weight matrices for calculating R t+1
  • b R is an offset for calculating R t+1 .
  • U H , W H are calculated The learnable weight matrix of .
  • the input is spatial features. Therefore, when extracting temporal features using the gated recurrent unit based on the time-series brain connection attention mechanism, the hidden state feature vector set H t output at time t represents the extracted The spatio-temporal signature of abnormal brain connectivity patterns in the dynamic brain connectogram of .
  • the output H t of the hidden state feature vector set is calculated through the fully connected layer, and the abnormal probability of brain connection at time t is obtained as:
  • the abnormal probability indicates the abnormal probability of the edge e ij ⁇ E t at time t, that is, the abnormal probability of each edge in the dynamic brain connection graph G S and G a .
  • h i is the hidden state feature vector of the i-th node
  • h j is the hidden state feature vector of the j-th node
  • ⁇ ( ) is the sigmoid function
  • a and b are the optimization parameters of the output layer
  • ⁇ and ⁇ are A pair of hyperparameters
  • ⁇ represents the weight of the edge e ij ; input the spatio-temporal features of each group of dynamic brain connection maps into the above function f( ), and obtain f(H S ) and f(H a ).
  • the contrastive loss (Contrastive Loss) function of abnormal brain connection learning is defined as:
  • L TOTAL L S + ⁇ a ⁇ A, a' ⁇ A L ⁇ ;
  • HS is the spatio-temporal feature of the saline injection group
  • H a is the spatio-temporal feature of the nicotine injection group
  • A ⁇ 1,2,...,A′ ⁇
  • A′ is the nicotine
  • ⁇ is a hyperparameter.
  • the saline injection group and all nicotine injection groups are compared and studied, and the multiple groups injected with nicotine are pairwise compared and learned, which can fully analyze the dynamic brain connection maps in different control groups. Individual differences facilitate the discovery of abnormal brain connections in addiction.
  • the brain connection with the highest abnormal probability is obtained in multiple groups of dynamic brain connection graphs at each time point.
  • the abnormal brain connection probability is input to a preset classifier for iterative training until the classification accuracy converges to the preset accuracy.
  • the abnormal probability of each brain connection in each group of dynamic brain connection graphs is input to a preset classifier for iterative training.
  • the preset classifier can be selected according to actual needs, and conventional classifiers can realize the above functions, such as SVM, KNN, naive Bayesian, decision tree, logistic regression, neural network algorithm, etc., which are not specifically limited here.
  • the preset accuracy is set according to actual requirements, and is not limited here.
  • the brain functional connection with the highest abnormal probability is corrected and integrated to obtain the addiction neural circuit at each moment, and then according to the entire time series, a dynamic addiction neural circuit is generated.
  • a weakly supervised contrastive learning model is formed, and the training difficulty is low.
  • fMRI images are directly input into the weakly supervised contrastive learning model, eliminating redundant and complex preprocessing calculations.
  • the differences in brain connections between different groups of samples can be obtained, combined with a small amount of neuroscience prior knowledge, to reveal the neural circuit mechanism of addiction, which is easy to train and has high training accuracy .
  • the embodiment of the invention also discloses a dynamic addiction neural circuit generation system based on weakly supervised contrastive learning.
  • the dynamic addiction neural circuit generation system based on weakly supervised contrastive learning includes:
  • the dynamic brain connection map generation module 10 is used to reduce the voxels of multiple groups of fMRI images to the attributes of brain region nodes based on the convolutional neural network, and generate multiple groups of dynamic brain connections including time series according to the attributes of the brain region nodes Figure; wherein, each group of fMRI includes multiple fMRI images collected at a fixed frequency within a preset time period;
  • the dynamic brain connection update module 20 is used to use the standard brain connection map as a real sample and the dynamic brain connection map as a fake sample to generate confrontation learning and update the dynamic brain connection map;
  • Spatial-temporal feature extraction module 30 used to extract the spatio-temporal features of brain connections in each of the dynamic brain-connectivity graphs;
  • the most abnormal brain connection acquisition module 40 is used to input the spatio-temporal features into the abnormal connection detection network, calculate the abnormal probability of brain connections based on comparative learning, and obtain the brain connection with the highest abnormal probability at each moment;
  • the dynamic addiction neural circuit generation module 50 is used to generate a dynamic addiction neural circuit according to neuroscience prior knowledge and the brain connection with the highest abnormal probability at each moment.
  • the dynamic brain connection updating module 20 is optional.
  • spatio-temporal feature extraction module 30 specifically comprises:
  • the spatial feature extraction module 3010 is used to extract the spatial features of the brain connection in the dynamic brain connection graph by using the graph convolutional neural network based on the topological space brain connection attention degree mechanism.
  • the temporal feature extraction module 3020 is used for extracting the temporal features of the dynamic brain connection map by using the gated recurrent unit based on the time-series brain connection attention degree mechanism.
  • an embodiment of the present disclosure also provides an electronic device 500 .
  • an electronic device 500 includes a processor 501 , a memory 502 and a bus.
  • the memory 502 is used to store computer programs, including internal memory 5021 and external memory 5022;
  • the memory 5021 exchanges data with the external memory 5022 .
  • the memory 502 is specifically used to store a computer program for executing the technical solution of the present application, and the execution is controlled by the processor 501 . That is, when the electronic device 500 is running, the processor 501 communicates with the memory 502 through the bus, so that the processor 501 executes the computer program stored in the memory 502, and then executes the method described in any of the foregoing embodiments.
  • memory 502 can be, but not limited to, random access memory (Random Access Memory, RAM), read-only memory (Read Only Memory, ROM), programmable read-only memory (Programmable Read-Only Memory, PROM), can Erase read-only memory (Erasable Programmable Read-Only Memory, EPROM), etc.
  • RAM Random Access Memory
  • ROM Read Only Memory
  • PROM Programmable Read-Only Memory
  • EPROM Erasable Programmable Read-Only Memory
  • the processor 501 may be an integrated circuit chip and has signal processing capability.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC) , field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
  • DSP digital signal processor
  • ASIC application-specific integrated circuit
  • FPGA field programmable gate array
  • a general-purpose processor may be a microprocessor, or the processor may be any conventional processor, and the like.
  • the structure illustrated in the embodiment of the present application does not constitute a specific limitation on the electronic device 500 .
  • the electronic device 500 may include more or fewer components than shown in the figure, or combine certain components, or separate certain components, or arrange different components.
  • the illustrated components can be realized in hardware, software or a combination of software and hardware.
  • the present embodiment also provides a computer readable storage medium, such as floppy disk, CD, hard disk, flash memory, U disk, SD (Secure Digital Memory Card, safe digital card) card, MMC (Multimedia Card, multimedia card) card etc.
  • a computer readable storage medium such as floppy disk, CD, hard disk, flash memory, U disk, SD (Secure Digital Memory Card, safe digital card) card, MMC (Multimedia Card, multimedia card) card etc.
  • Computer programs for implementing the above steps are stored in the readable storage medium, and the computer programs can be executed by one or more processors to implement the methods in the above embodiments.
  • the disclosed devices and methods may be implemented in other ways.
  • the device embodiments described above are only illustrative.
  • the division of the units is only a logical function division.
  • multiple units or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.
  • each functional module in each embodiment of the present application may be integrated to form an independent part, each module may exist independently, or two or more modules may be integrated to form an independent part.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Medical Informatics (AREA)
  • Quality & Reliability (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Signal Processing (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Condensed Matter Physics & Semiconductors (AREA)
  • Probability & Statistics with Applications (AREA)
  • Databases & Information Systems (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

一种基于弱监督对比学习的动态成瘾神经环路生成方法及系统，属于人工智能技术领域；其中，生成方法包括：基于卷积神经网络，将多组fMRI的体素降维到脑区节点属性，并根据所述脑区节点属性生成包含时间序列的多组动态脑连接图（S11）；提取每个所述动态脑连接图中脑连接的时空特征（S13）；将时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，并获取各时刻异常概率最大的脑连接（S14）；根据神经科学先验知识和异常概率最大的脑连接，生成动态成瘾神经环路（S15）。直接将fMRI影像输入模型中，免去了冗余且复杂的预处理计算。通过对比学习，获得不同样本脑连接的差异性，并结合少量神经科学先验知识，揭示成瘾神经环路机制，易于训练，且精度高。

Description

基于弱监督对比学习的动态成瘾神经环路生成方法及系统 技术领域
本发明涉及人工智能的领域，尤其是涉及一种基于弱监督对比学习的动态脑成瘾神经环路生成方法及系统。
背景技术
成瘾是一种以寻求强迫性药物为特征的疾病，以吸烟成瘾为例，目前我国烟草使用者已经超过三亿人口，每年死于使用烟草而导致相关疾病的人数高达100万之多。从诱发吸烟行为的角度来看，尼古丁成瘾是吸烟的主要诱因，是吸烟者戒烟的主要障碍，成瘾也被看作是一种慢性复发的功能性脑疾病。
目前，对于尼古丁成瘾神经环路的检测是从fMRI(功能性磁共振成像)中分析计算出功能异常的神经环路。然而，传统基于统计学的方法需要对影像数据进行复杂的预处理操作，冗余量大；并且，fMRI影像数据具有小样本的缺陷，导致训练难度大，训练精度低。
发明内容
本发明提供的一种基于弱监督对比学习的动态成瘾神经环路生成方法，采用如下的技术方案：
一种基于弱监督对比学习的动态成瘾神经环路生成方法，包括：
基于卷积神经网络，将多组fMRI影像的体素降维到脑区节点属性，并根据所述脑区节点属性生成包含时间序列的多组动态脑连接图；其中，每组fMRI包括预设时间段内以固定频率采集的多张fMRI影像；
提取每个所述动态脑连接图中脑连接的时空特征；
将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，并获取各时刻异常概率最大的脑连接；
根据神经科学先验知识和各时刻所述异常概率最大的脑连接，生成动态成瘾神经环路。
通过采用上述技术方案，根据每组样本的组别标签以及少量的神经科学先验知识，结合对比学习，构成了弱监督对比学习模型，通过直接将fMRI影像输入到弱监督对比学习模型中，免去了冗余且复杂的预处理计算。并且，在小样本的前提下，通过对比学习，获得不同组样本间的脑连接差异性，并结合少量的神经科学先验知识，揭示成瘾神经环路机制，训练难度低，且训练精度高。
可选的，在所述提取所述动态脑连接图中脑连接的时空特征的步骤之前，还包括：
将标准脑连接图作为真样本，将所述动态脑连接图作为假样本，生成对抗学习，对所述动态脑连接图进行更新。
通过采用上述技术方案，通过对抗学习，得到既符合标准脑模板空间位置信息，又具有原始fMRI体素空间信息的动态脑连接图，有助于训练的精度；并且，作为先验知识的标准空间脑模板数据量小，学习难度低。
可选的，所述提取所述动态脑连接图中脑连接的时空特征的步骤，具体包括：
利用基于拓扑空间脑连接关注度机制的图卷积神经网络提取所述动态脑连接图中脑连接的空间特征；
利用基于时序脑连接关注度机制的门控循环单元提取所述动态脑连接 图中脑连接的时间特征。
通过采用上述技术方案，分别提取动态脑连接图的空间特征和时间特征，从而获得脑连接的时空特征，充分表征了动态成瘾脑连接的属性。
可选的，所述将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接异常概率，具体包括；
脑连接异常概率为：F score＝MLP(H t)；
Figure PCTCN2021143741-appb-000001
其中，h i是第i个节点的隐状态特征向量，h j是第j个节点的隐状态特征向量，σ(·)是sigmoid函数，a和b是输出层的优化参数，β和μ是一对超参数，ω表示边e ij的权重；
对比损失函数：L S＝min(max{0,f(H S)-∑ a∈Af(H a)+γ})；
L Δ＝min(max{0,f(H a)-f(H a′)+γ})，a∈A，a′∈A，a≠a′；
L TOTAL＝L S+∑ a∈A,a′∈AL Δ；
其中，max{0,·}是合页损失，H S是盐水注射组的时空特征，H a是尼古丁注射组的时空特征，A＝{1,2,…,A′}，A′是尼古丁注射组的组别总个数，γ是超参数。
通过采用上述技术方案，基于对比学习的策略，将注射生理盐水组和所有注射尼古丁组进行对比学习，并对注射尼古丁的多个组别进行两两对比学习，能够充分分析出不同对照组中的动态脑连接图的个体差异，便于发现成瘾异常脑连接。
可选的，所述根据神经科学先验知识和各时刻所述异常概率最大的脑连接，生成动态成瘾神经环路的步骤，具体包括：
根据神经科学先验知识，对异常概率最大的脑连接进行修正和整合，获得各个时刻的成瘾神经环路，并根据所述时间序列，生成动态成瘾神经环路。
通过采用上述技术方案，根据神经科学先验知识，对各时刻异常概率最大的脑连接进行修正和正常，从而保证了生成的成瘾神经环路的精确度。
第二方面，本发明提供一种基于弱监督对比学习的动态成瘾神经环路生成系统，采用如下的技术方案：
一种基于弱监督对比学习的动态成瘾神经环路生成系统，包括：
动态脑连接图生成模块，用于基于卷积神经网络，将多组fMRI影像的体素降维到脑区节点属性，并根据所述脑区节点属性生成包含时间序列的多组动态脑连接图；其中，每组fMRI包括预设时间段内以固定频率采集的多张fMRI影像；
时空特征提取模块，用于提取每个所述动态脑连接图中脑连接的时空特征；
最具异常性脑连接获取模块，用于将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，并获取各时刻异常概率最大的脑连接；
动态成瘾神经环路生成模块，用于根据神经科学先验知识和各时刻所述异常概率最大的脑连接，生成动态成瘾神经环路。
第三方面，本发明提供一种电子设备，采用如下的技术方案：
电子设备，包括存储器和处理器，存储器上存储有能够被处理器加载并执行所述的方法的计算机程序。
第四方面，本发明提供一种计算机可读存储介质，采用如下的技术方案：
计算机可读存储介质，存储有能够被处理器加载并执行所述的方法的计算机程序。
综上所述，本发明包括以下至少一种有益技术效果：
1.根据每组样本的组别标签以及少量的神经科学先验知识，结合对比学习，构成了弱监督对比学习模型，通过直接将fMRI影像输入到弱监督对比学习模型中，免去了冗余且复杂的预处理计算。
2.在小样本的前提下，通过对比学习，获得不同组样本间的脑连接差异性，并结合少量的神经科学先验知识，揭示成瘾神经环路机制，易于训练，且训练精度高。
3.采用对比学习的策略，将注射生理盐水组和所有注射尼古丁组进行对比学习，并对注射尼古丁的多个组别进行两两对比学习，能够充分分析出不同对照组中的动态脑连接图的个体差异，便于发现成瘾异常脑连接。
附图说明
图1是基于弱监督对比学习的动态成瘾神经环路生成方法流程图。
图2是脑连接时空特征提取方法流程图。
图3是基于弱监督对比学习的动态成瘾神经环路生成系统结构框图。
图4是时空特征提取模块结构框图。
图5是电子设备示意图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图 1-4及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本发明，并不用于限定本发明。
本发明实施例公开一种基于弱监督对比学习的动态成瘾神经环路生成方法。参照图1，基于弱监督对比学习的动态成瘾神经环路生成方法包括：
S11、基于卷积神经网络，将多组fMRI影像的体素降维到脑区节点属性，并根据所述脑区节点属性生成包含时间序列的多组动态脑连接图。
利用尼古丁成瘾实验中大鼠动物模型的fMRI影像来生成基于弱监督对比学习的动态成瘾神经环路。多组fMRI影像包括注射生理盐水组大鼠和注射尼古丁组大鼠的fMRI影像，其中，生理盐水组大鼠的组别个数是1，注射尼古丁组大鼠的组别数量可根据具体实际情况进行设定，优选为1-4。每组fMRI影像包括预设时间段内以固定频率采集的多张fMRI影像，其中，预设时间段是大鼠注射生理盐水或尼古丁的时长，可设定为两周、三周、四周或其它时长，在此不作限定。
具体的，采用注射生理盐水组大鼠和注射尼古丁组大鼠进行实验，尼古丁组大鼠可根据不同的浓度分为多组，也可只设定一组，每组大鼠均连续注射预设时间段时长，以模拟尼古丁摄入并成瘾的情况。在预设时间段内，以固定频率采集每组大鼠的fMRI影像，每组大鼠的fMRI影像反映了注射生理盐水或对应浓度的尼古丁后，大脑神经在时间维度上的变化情况。
利用卷积神经网络对每组fMRI的体素进行降维，能够在保持时间分辨率不变的情况下，将fMRI的时间片段以具有一定意义的时间分别率来分割，每段作为一个时间点生成动态脑连接图，充分提取了fMRI的时序动态信息。
动态脑连接图G为：
Figure PCTCN2021143741-appb-000002
具体的，在动态脑连 接图G中，以fMRI时间片段为时序，每个时序作为一个时间点，脑区为节点，脑连接为边，脑连接是指脑区之间的功能连接，T是时间点总个数，G t是t时刻动态脑连接图中的一幅快照，V t和E t分别表示其中的节点集和边集，边e ij∈E t，表示在t时刻，第i个与第j个节点之间的脑功能连接，且其权重是ω，
Figure PCTCN2021143741-appb-000003
表示节点的属性集合，N表示节点总个数，
Figure PCTCN2021143741-appb-000004
是第i个节点在t时刻的属性向量。将注射生理盐水组I S和注射尼古丁组I a的fMRI影像分别输入上述动态脑连接图G，得到动态脑连接图G S、G a，其中，a∈A，A＝{1,2,…,A′}，A′是尼古丁注射组的组别总个数。
在该步骤中，直接将fMRI影像输入到弱监督对抗学习模型中，获得动态脑连接图，免去了冗余且复杂的预处理计算。
S12、将标准脑连接图作为真样本，将动态脑连接图作为假样本，生成对抗学习，对动态脑连接图进行更新。
标准脑连接图是含有脑图谱先验信息的标准空间脑模板预处理出的脑功能连接图。将标准脑连接图作为真样本G True，将上一步骤得到动态脑连接图作为假样本G Fake，生成对抗学习，对动态脑连接图进行更新。
生成对抗学习的公式是：
Figure PCTCN2021143741-appb-000005
其中，D是判别器，表示标准脑连接图，G是生成器，表示异常脑连接图，在本申请中，G具体是指动态脑连接图。对三组类别的动态脑连接图分别生成对抗学习，对动态脑连接图进行更新，也就是将G S和G a分别输入上述 公式进行更新。更新后的动态脑连接图是近似于脑图谱空间信息的动态脑连接图，表示不同类别下尼古丁成瘾的异常脑连接模式。
需要说明的是，步骤S12是可选的。若执行步骤12，通过对抗学习，得到既符合标准脑模板空间位置信息，又具有原始fMRI体素空间信息的动态脑连接图，提高了训练的精度；并且，通过对含有脑图谱先验信息的标准空间脑模板的预处理来获取标注脑功能连接图，标准空间脑模板数据量小，学习难度低。
S13、提取每个所述动态脑连接图中脑连接的时空特征。
将每组动态脑连接图输入神经网络进行学习，提取动态脑连接图中脑连接的空间特征和时间特征。
具体的，参照图2，步骤S13包括以下子步骤：
S1311、利用基于拓扑空间脑连接关注度机制的图卷积神经网络提取动态脑连接图中脑连接的空间特征。
基于拓扑空间脑连接关注度机制的图卷积神经网络由多头异常连接关注度块(Multi-head Anomalous Connection Focus Block，MACFB)组成，具有强大的图学习能力，可学习到连接网络中节点间的几何空间特征。
具体的，异常连接关注度块总个数为C，C的取值优选为2、3或4。
隐状态特征向量集合：
Figure PCTCN2021143741-appb-000006
Figure PCTCN2021143741-appb-000007
表示在t时刻，第l层多头异常连接关注度块，t∈{1,2,…,T}，T是时间点总个数，i表示节点，i∈{1,2,…,N}，N是节点总个数。
Figure PCTCN2021143741-appb-000008
表示第i个节点在t时刻第l层的多头异常连接关注度块。
查询-节点属性向量：
Figure PCTCN2021143741-appb-000009
其中，i表示节点，c表示异常连接关注度块，
Figure PCTCN2021143741-appb-000010
是计算查询-节点属性向量的可学习权重矩阵，
Figure PCTCN2021143741-appb-000011
是计算查询-节点属性向量的偏移量；
键-节点属性向量：
Figure PCTCN2021143741-appb-000012
其中，j表示节点,
Figure PCTCN2021143741-appb-000013
是计算键-节点属性向量的可学习权重矩阵，
Figure PCTCN2021143741-appb-000014
是计算键-节点属性向量的偏移量；
额外的脑连接特征向量：e c,ij＝W c,ee ij+b c,e；
其中，e ij是动态脑连接图中的边，表示在t时刻，第i个与第j个节点之间的脑连接，且其权重为ω。W c,e是计算额外的脑连接特征向量的可学习权重矩阵，b c,e是计算额外的脑连接特征向量的偏移量。
脑连接关注度系数：
Figure PCTCN2021143741-appb-000015
其中，<q,k>是指数点积函数，N(i)是取除i以外的值。
在计算到脑连接关注度系数后，再进行图中的消息传递，以输出下一层节点隐状态特征向量；
Figure PCTCN2021143741-appb-000016
Figure PCTCN2021143741-appb-000017
Figure PCTCN2021143741-appb-000018
是计算
Figure PCTCN2021143741-appb-000019
的可学习权重矩阵，
Figure PCTCN2021143741-appb-000020
是计算
Figure PCTCN2021143741-appb-000021
的偏移量。
将下一层节点隐状态特征向量
Figure PCTCN2021143741-appb-000022
的集合
Figure PCTCN2021143741-appb-000023
作为下一层多头异常连接关注度块的输入，或者作为图卷积神经网络的最终输出结果。若作为图卷积神经网络的最终输出结果，表示提取到的动态脑连接图中的异常脑连接模式的空间特征。
S1312、利用基于时序脑连接关注度机制的门控循环单元提取动态脑连接图的时间特征。
结合脑连接关注度机制的门控循环单元，学习短期和长期的隐状态特征，使每个时序的脑功能连接图的隐状态特征包涵了其它时序的上下文脑连接隐状态信息，从而学习到动态脑连接图的时间特征。
在门控循环单元中，需由t时刻的隐状态特征向量去计算下一时刻t+1的隐状态特征向量。
长期隐状态特征向量集合：Current t+1＝MACFB(H t)；
在初始状态，H t表示图卷积神经网络的输出结果，即提取到的动态脑连接图中的异常脑连接模式的空间特征。
根据脑连接关注度窗口计算短期隐状态特征向量；在脑功能连接关注度窗口内，节点隐状态特征向量的输出：
Figure PCTCN2021143741-appb-000024
其中，窗口大小为w+1，i表示节点；
时序脑连接关注度系数：
Figure PCTCN2021143741-appb-000025
其中，r是权重向量，Q是权重矩阵；
Figure PCTCN2021143741-appb-000026
节点短期隐状态特征向量：
Figure PCTCN2021143741-appb-000027
下一时刻脑连接隐状态特征向量集合：H t+1＝GRU(Current t+1,Short t+1)；
Short t+1表示所有节点的短期隐状态特征向量
Figure PCTCN2021143741-appb-000028
的集合，i∈{1,2,…,N}，N是节点总个数。
进一步地，H t+1＝GRU(Current t+1,Short t+1)的具体计算过程为：
脑连接状态更迭门：P t+1＝σ(U PCurrent t+1+W PShort t+1+b P)；
其中，U P、W P均是计算P t+1的可学习权重矩阵，b P是计算P t+1的偏移量。
脑连接状态重组门：R t+1＝σ(U RCurrent t+1+W RShort t+1+b R)；
其中，U R、W R均是计算R t+1的可学习权重矩阵，b R是计算R t+1的偏移量。
Figure PCTCN2021143741-appb-000029
其中，U H、W H是计算
Figure PCTCN2021143741-appb-000030
的可学习权重矩阵。
Figure PCTCN2021143741-appb-000031
由于在计算时间特征时，输入的是空间特征，因此，在利用基于时序脑连接关注度机制的门控循环单元提取时间特征时，在t时刻输出的隐状态特征向量集合H t，表示提取到的动态脑连接图中的异常脑连接模式的时空特征。
S14、将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，并获取各时刻异常概率最大的脑连接。
在异常连接检测网络(Anomaly Connectivities Detection Network)中，在t时刻，将隐状态特征向量集合的输出H t通过全连接层计算，得到t时刻脑连接的异常概率为：
F score＝MLP(H t)；
异常概率表示在t时刻，边e ij∈E t的异常概率，也就是动态脑连接图G S和G a中每个边的异常概率。
Figure PCTCN2021143741-appb-000032
其中，h i是第i个节点的隐状态特征向量，h j是第j个节点的隐状态特征向量，σ(·)是sigmoid函数，a和b是输出层的优化参数，β和μ是一对超参数，ω表示边e ij的权重；将每组动态脑连接图的时空特征输入上述函数f(·)，得到f(H S)和f(H a)。
采用对比学习的策略，定义异常脑连接学习的对比损失(Contrastive Loss)函数为：
对比损失函数：L S＝min(max{0,f(H S)-∑ a∈Af(H a)+γ})；
L Δ＝min(max{0,f(H a)-f(H a′)+γ})，a∈A，a′∈A，a≠a′；
L TOTAL＝L S+∑ a∈A,a′∈AL Δ；
其中，max{0,·}是合页损失，H S是盐水注射组的时空特征，H a是尼古丁注射组的时空特征，A＝{1,2,…,A′}，A′是尼古丁注射组的组别总个数，γ是超参数。
采用上述对比学习的策略，将注射生理盐水组和所有注射尼古丁组进行对比学习，并对注射尼古丁的多个组别进行两两对比学习，能够充分分析出不同对照组中的动态脑连接图的个体差异，便于发现成瘾异常脑连接。
通过对比学习计算出每一动态脑连接图中的脑连接的异常概率后，针对各个时刻，在多组动态脑连接图中获取异常概率最大的脑连接。
作为一种实施方式，在获取各时刻异常概率最大的脑连接之前，将脑连接异常概率输入至预设分类器进行迭代训练，直至分类精度收敛至预设精度。
具体的，在时刻t，将每组动态脑连接图中的每个脑连接异常概率输入到预设分类器进行迭代训练。预设分类器可以根据实际需求进行选择，常 规分类器均可实现上述功能，如SVM、KNN、朴素贝叶斯、决策树、逻辑回归、神经网络算法等，在此不作具体限定。预设精度是根据实际需求设定的，在此不作限定。
S15、根据神经科学先验知识和各时刻异常概率最大的脑连接，生成动态成瘾神经环路。
结合神经科学先验知识，针对每个时刻t，对异常概率最大的脑功能连接进行修正和整合，获得各个时刻的成瘾神经环路，进而根据整个时间序列，生成动态成瘾神经环路。对各时刻异常概率最大的脑连接进行修正，是指根据神经科学先验知识，删除错误的脑连接和/或对偏差较大的脑连接进行调节；对最具异常性的脑功能连接进行整合，是指针对每个时刻，将修正后的脑功能整合成一个完整的成瘾神经环路。
在本实施例中，根据每组样本的组别标签以及少量的神经科学先验知识，结合对比学习，构成了弱监督对比学习模型，训练难度低。并且，直接将fMRI影像输入到弱监督对比学习模型中，免去了冗余且复杂的预处理计算。进一步地，在小样本的前提下，通过对比学习，获得不同组样本间的脑连接差异性，并结合少量的神经科学先验知识，揭示成瘾神经环路机制，易于训练，且训练精度高。
本发明实施例还公开一种基于弱监督对比学习的动态成瘾神经环路生成系统。参照图3，基于弱监督对比学习的动态成瘾神经环路生成系统包括：
动态脑连接图生成模块10，用于基于卷积神经网络，将多组fMRI影像的体素降维到脑区节点属性，并根据所述脑区节点属性生成包含时间序 列的多组动态脑连接图；其中，每组fMRI包括预设时间段内以固定频率采集的多张fMRI影像；
动态脑连接更新模块20，用于将标准脑连接图作为真样本，将动态脑连接图作为假样本，生成对抗学习，对动态脑连接图进行更新；
时空特征提取模块30，用于提取每个所述动态脑连接图中脑连接的时空特征；
最具异常性脑连接获取模块40，用于将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，并获取各时刻异常概率最大的脑连接；
动态成瘾神经环路生成模块50，用于根据神经科学先验知识和各时刻所述异常概率最大的脑连接，生成动态成瘾神经环路。
其中，动态脑连接更新模块20是可选的。
参照图3，时空特征提取模块30具体包括：
空间特征提取模块3010，用于利用基于拓扑空间脑连接关注度机制的图卷积神经网络提取动态脑连接图中脑连接的空间特征。
时间特征提取模块3020，用于利用基于时序脑连接关注度机制的门控循环单元提取动态脑连接图的时间特征。
所述的系统实施例可以用于执行上述方法实施例，其原理和技术效果类似，此处不再赘述。
基于同一技术构思，本公开实施例还提供了一种电子设备500。参照图5所示，电子设备500包括处理器501、存储器502和总线。其中，存储器502用于存储计算机程序，包括内部存储器5021和外部存储器5022；内部存 储器5021用于暂时存放处理器501中的运算数据，以及与硬盘等外部存储器5022交换的数据，处理器501通过内部存储器5021与外部存储器5022进行数据交换。
本申请实施例中，存储器502具体用于存储执行本申请技术方案的计算机程序，并由处理器501来控制执行。也即，当电子设备500运行时，处理器501与存储器502之间通过总线通信，使得处理器501执行存储器502中存储的计算机程序，进而执行前述任一实施例中所述的方法。
其中，存储器502可以是，但不限于，随机存取存储器(Random Access Memory，RAM)，只读存储器(Read Only Memory，ROM)，可编程只读存储器(Programmable Read－Only Memory，PROM)，可擦除只读存储器(Erasable Programmable Read－Only Memory，EPROM)等。
处理器501可能是一种集成电路芯片，具有信号的处理能力。上述的处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Network Processor，NP)等；还可以是数字信号处理器(DSP)、专用集成电路(ASIC)、现场可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本发明实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
可以理解的是，本申请实施例示意的结构并不构成对电子设备500的具体限定。在本申请另一些实施例中，电子设备500可以包括比图示更多或更少的部件，或者组合某些部件，或者拆分某些部件，或者不同的部件布置。图示的部件可以以硬件，软件或软件和硬件的组合实现。
本实施例还提供了一种计算机可读存储介质，如软盘、光盘、硬盘、闪存、U盘、SD(Secure Digital Memory Card，安全数码卡)卡、MMC(Multimedia Card，多媒体卡)卡等，在该可读存储介质中存储有实现上述各个步骤的计算机程序，该计算机程序可被一个或者多个处理器执行，以实现上述实施例中的方法。
在本申请所提供的实施例中，应该理解到，所揭露装置和方法，可以通过其它的方式实现。以上所描述的装置实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，又例如，多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些通信接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
另外，在本申请各个实施例中的各功能模块可以集成在一起形成一个独立的部分，也可以是各个模块单独存在，也可以两个或两个以上模块集成形成一个独立的部分。
以上均为本发明的较佳实施例，并非依此限制本发明的保护范围，本说明书(包括摘要和附图)中公开的任一特征，除非特别叙述，均可被其他等效或者具有类似目的的替代特征加以替换。即，除非特别叙述，每个特征只是一系列等效或类似特征中的一个例子而已。

Claims (10)

  1. 一种基于弱监督对比学习的动态成瘾神经环路生成方法，其特征在于，包括：
    基于卷积神经网络，将多组fMRI影像的体素降维到脑区节点属性，并根据所述脑区节点属性生成包含时间序列的多组动态脑连接图；其中，每组fMRI包括预设时间段内以固定频率采集的多张fMRI影像；
    提取每个所述动态脑连接图中脑连接的时空特征；
    将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，并获取各时刻异常概率最大的脑连接；
    根据神经科学先验知识和各时刻所述异常概率最大的脑连接，生成动态成瘾神经环路。
  2. 根据权利要求1所述的方法，其特征在于，在所述提取所述动态脑连接图中脑连接的时空特征的步骤之前，还包括：
    将标准脑连接图作为真样本，将所述动态脑连接图作为假样本，生成对抗学习，对所述动态脑连接图进行更新。
  3. 根据权利要求1或2所述的方法，其特征在于，所述提取所述动态脑连接图中脑连接的时空特征的步骤，具体包括：
    利用基于拓扑空间脑连接关注度机制的图卷积神经网络提取所述动态脑连接图中脑连接的空间特征；
    利用基于时序脑连接关注度机制的门控循环单元提取所述动态脑连接图中脑连接的时间特征。
  4. 根据权利要求3所述的方法，其特征在于，所述将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，具体包括；
    脑连接的异常概率为：F score＝MLP(H t)；
    Figure PCTCN2021143741-appb-100001
    其中，h i是第i个节点的隐状态特征向量，h j是第j个节点的隐状态特征向量，σ(·)是sigmoid函数，a和b是输出层的优化参数，β和μ是一对超参数，ω表示边e ij的权重；
    对比损失函数：L S＝min(max{0,f(H S)-∑ a∈Af(H a)+γ})；
    L Δ＝min(max{0,f(H a)-f(H a′)+γ})，a∈A，a′∈A，a≠a′；
    L TOTAL＝L S+∑ a∈A,a′∈AL Δ；
    其中，max{0,·}是合页损失，H S是盐水注射组的时空特征，H a是尼古丁注射组的时空特征，A＝{1,2,…,A′}，A′是尼古丁注射组的组别总个数，γ是超参数。
  5. 根据权利要求4所述的方法，其特征在于，所述获取各时刻异常概率最大的脑连接之前，还包括：
    将所述脑连接的异常概率输入至预设分类器进行迭代训练，直至分类精度收敛至预设精度。
  6. 根据权利要求4所述的方法，其特征在于，所述根据神经科学先验知识和各时刻所述异常概率最大的脑连接，生成动态成瘾神经环路的步骤，具体包括：
    结合神经科学先验知识，对异常概率最大的脑连接进行修正和整合，获得各个时刻的成瘾神经环路，并根据所述时间序列，生成动态成瘾神经环路。
  7. 一种基于弱监督对比学习的动态成瘾神经环路生成系统，其特征在 于，包括：
    动态脑连接图生成模块，用于基于卷积神经网络，将多组fMRI影像的体素降维到脑区节点属性，并根据所述脑区节点属性生成包含时间序列的多组动态脑连接图；其中，每组fMRI包括预设时间段内以固定频率采集的多张fMRI影像；
    时空特征提取模块，用于提取每个所述动态脑连接图中脑连接的时空特征；
    最具异常性脑连接获取模块，用于将所述时空特征输入异常连接检测网络，基于对比学习计算脑连接的异常概率，并获取各时刻异常概率最大的脑连接；
    动态成瘾神经环路生成模块，用于根据神经科学先验知识和各时刻所述异常概率最大的脑连接，生成动态成瘾神经环路。
  8. 根据权利要求7所述的系统，其特征在于，还包括动态脑连接更新模块；
    所述动态脑连接更新模块，用于将标准脑连接图作为真样本，将所述动态脑连接图作为假样本，生成对抗学习，对所述动态脑连接图进行更新。
  9. 一种电子设备，其特征在于：包括存储器和处理器，存储器上存储有能够被处理器加载并执行如权利要求1-6任一项所述的方法的计算机程序。
  10. 一种计算机可读存储介质，其特征在于：存储有能够被处理器加载并执行如权利要求1-6任一项所述的方法的计算机程序。
PCT/CN2021/143741 2021-12-31 2021-12-31 基于弱监督对比学习的动态成瘾神经环路生成方法及系统 WO2023123380A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2021/143741 WO2023123380A1 (zh) 2021-12-31 2021-12-31 基于弱监督对比学习的动态成瘾神经环路生成方法及系统
US18/111,876 US20230215006A1 (en) 2021-12-31 2023-02-20 Method and system for generating a dynamic addictive neural circuits based on weakly supervised contrastive learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/143741 WO2023123380A1 (zh) 2021-12-31 2021-12-31 基于弱监督对比学习的动态成瘾神经环路生成方法及系统

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/111,876 Continuation US20230215006A1 (en) 2021-12-31 2023-02-20 Method and system for generating a dynamic addictive neural circuits based on weakly supervised contrastive learning

Publications (1)

Publication Number Publication Date
WO2023123380A1 true WO2023123380A1 (zh) 2023-07-06

Family

ID=86991981

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/143741 WO2023123380A1 (zh) 2021-12-31 2021-12-31 基于弱监督对比学习的动态成瘾神经环路生成方法及系统

Country Status (2)

Country Link
US (1) US20230215006A1 (zh)
WO (1) WO2023123380A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3948682A1 (en) * 2019-05-23 2022-02-09 Google LLC Connection weight learning for guided architecture evolution
CN117593594B (zh) * 2024-01-18 2024-04-23 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) 基于一致性对齐的脑部mri图像分类方法、设备和介质
CN117653147B (zh) * 2024-01-31 2024-04-26 长春理工大学 一种基于脑电信号特征的分类方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110889496A (zh) * 2019-12-11 2020-03-17 北京工业大学 一种基于对抗生成网络的人脑效应连接识别方法
CN111067522A (zh) * 2019-12-16 2020-04-28 中国科学院深圳先进技术研究院 大脑成瘾结构图谱评估方法及装置
CN112233086A (zh) * 2020-10-14 2021-01-15 南京工业大学 基于脑区功能连接的fMRI数据分类识别方法及装置
WO2021067464A1 (en) * 2019-10-01 2021-04-08 The Board Of Trustees Of The Leland Stanford Junior University Joint dynamic causal modeling and biophysics modeling to enable multi-scale brain network function modeling
US20210201119A1 (en) * 2019-12-31 2021-07-01 X Development Llc Artificial neural network architectures based on synaptic connectivity graphs

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021067464A1 (en) * 2019-10-01 2021-04-08 The Board Of Trustees Of The Leland Stanford Junior University Joint dynamic causal modeling and biophysics modeling to enable multi-scale brain network function modeling
CN110889496A (zh) * 2019-12-11 2020-03-17 北京工业大学 一种基于对抗生成网络的人脑效应连接识别方法
CN111067522A (zh) * 2019-12-16 2020-04-28 中国科学院深圳先进技术研究院 大脑成瘾结构图谱评估方法及装置
US20210201119A1 (en) * 2019-12-31 2021-07-01 X Development Llc Artificial neural network architectures based on synaptic connectivity graphs
CN112233086A (zh) * 2020-10-14 2021-01-15 南京工业大学 基于脑区功能连接的fMRI数据分类识别方法及装置

Also Published As

Publication number Publication date
US20230215006A1 (en) 2023-07-06

Similar Documents

Publication Publication Date Title
WO2023123380A1 (zh) 基于弱监督对比学习的动态成瘾神经环路生成方法及系统
WO2020238293A1 (zh) 图像分类方法、神经网络的训练方法及装置
US11842487B2 (en) Detection model training method and apparatus, computer device and storage medium
US20230037908A1 (en) Machine learning model training method and device, and expression image classification method and device
CN111401281B (zh) 基于深度聚类和样例学习的无监督行人重识别方法及系统
WO2021008328A1 (zh) 图像处理方法、装置、终端及存储介质
WO2019228317A1 (zh) 人脸识别方法、装置及计算机可读介质
CN110363210B (zh) 一种图像语义分割模型的训练方法和服务器
WO2019218410A1 (zh) 图像分类方法、计算机设备和存储介质
Wang et al. An effective image representation method using kernel classification
Ouyang et al. Forgetmenot: Memory-aware forensic facial sketch matching
Li et al. On improving the accuracy with auto-encoder on conjunctivitis
WO2021155792A1 (zh) 一种处理装置、方法及存储介质
WO2021147325A1 (zh) 一种物体检测方法、装置以及存储介质
Xing et al. Towards robust and accurate multi-view and partially-occluded face alignment
Wang et al. Topogan: A topology-aware generative adversarial network
CN106296734B (zh) 基于极限学习机和boosting多核学习的目标跟踪方法
WO2017174982A1 (en) Method of matching a sketch image to a face image
Arce et al. Dendrite ellipsoidal neurons based on k-means optimization
CN111666976B (zh) 基于属性信息的特征融合方法、装置和存储介质
CN111274919A (zh) 基于卷积神经网络的五官检测方法、系统、服务器及介质
Hu et al. An integrated classification model for incremental learning
CN113158971A (zh) 一种事件检测模型训练方法及事件分类方法、系统
Pillai et al. Applying deep learning kernel function for species identification system
Zou et al. Deep learning and its application in diabetic retinopathy screening

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21969730

Country of ref document: EP

Kind code of ref document: A1