A method and system for predicting overall survival time in glioblastoma patients

Abstract

The application provides a glioblastoma patient overall survival time prediction method and system, specifically comprising: acquiring and preprocessing a histopathology image, inputting the histopathology image into a deep cooperative pulse neural membrane system, and acquiring an overall survival time prediction result; wherein the deep cooperative pulse neural membrane system comprises input / output neurons and a plurality of cooperative neuron modules; the cooperative neuron module comprises four kinds of cooperative neurons with different discharge mechanisms, can be self-assembled into different blocks to realize various learning mechanisms, can not only simultaneously meet the OS time prediction and genotype classification, but also further improve the OS time prediction efficiency and OS time prediction accuracy of the GBM patient without ROI.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to a method and system for predicting the overall survival time of glioblastoma patients. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Glioblastoma (GBM) is a highly malignant brain tumor with the worst prognosis. Even with multiple treatment interventions, the median overall survival (OS) of patients is only 12 to 14 months. Overall survival (OS) refers to the time from the onset of the disease to death and is an important indicator for preoperative prognosis assessment of GBM patients. Accurate prediction of OS in GBM patients is crucial for developing personalized treatment plans.

[0004] Currently, some studies use magnetic resonance imaging (MRI) of GBM patients to predict overall survival (OS). However, compared to MRI, whole-slide images (WSI) of histopathology with gigapixel resolution generally provide clearer spatial and morphological information about the tumor, which is more beneficial for cancer prognosis. In recent years, many automated methods for survival prediction using WSI have also been proposed. However, due to the varying survival times of GBM patients and the fact that WSI includes many densely distributed cells with insignificant differences, the accuracy is usually unsatisfactory. Furthermore, most existing works require delineating regions of interest (ROIs), which requires a significant amount of time and expertise, undoubtedly incurring substantial human and time costs.

[0005] The spiking neuron (SN) membrane system is a third-generation neural network model composed of spiking neurons. In the spiking neuron system, pulses are used as communication objects; a nerve fires and emits pulses to communicate with other neurons. Compared with second-generation neural networks, the spiking neuron system can achieve Turing universality with fewer neurons. The distributed structure and parallel computing of the spiking neuron system give it unique advantages. Recently, the spiking neuron system has been applied in fields such as medical image processing, power system fault diagnosis, and engineering optimization.

[0006] In the prior art, using the pulsed nerve membrane system to process whole slide images (WS I) of histopathology to predict the overall survival time of GBM patients is a new attempt of membrane systems in the medical field. A prediction system disclosed in patent application number 202310063513.6 achieves the prediction of overall survival time by introducing hypergraphs into the pulsed nerve membrane system. However, this system uses pulses as objects, converting each pixel of WSI into a 24-bit binary pulse sequence, which is undoubtedly very cumbersome and time-consuming, and its accuracy can be further optimized. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for predicting overall survival (OS) in glioblastoma (GBM) patients, as well as an electronic device and a computer-readable storage medium. By designing a structurally scalable, deep collaborative pulsating nerve (DSSN) membrane system with an attention mechanism, the OS time of GBM patients can be predicted without ROI. This not only further improves the accuracy of OS time prediction and reduces time consumption, but also enables simultaneous prediction of overall survival and genotype classification tasks for GBM patients.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The first aspect of this invention provides a method for predicting overall survival in patients with glioblastoma, comprising:

[0010] Acquire and preprocess histopathological images;

[0011] The pre-processed histopathological images were input into a pre-constructed deep co-pulsed neural membrane system to obtain overall survival prediction results;

[0012] Furthermore, in the first submodule of the deep collaborative pulsed nerve membrane system, the input histopathological images are processed according to phenotype, specifically including: enlarging and slicing each histopathological image and clustering it to obtain several slice clusters;

[0013] Furthermore, several slice clusters are passed to the second submodule of the deep co-pulse neural membrane system through a transfer rule to perform long and short OS time classification to obtain survival-related slice clusters;

[0014] Furthermore, using the transfer rules, survival-related slice clusters that meet the threshold conditions are transferred to the third submodule of the deep co-pulsed neural membrane system to perform OS time prediction and genotype classification.

[0015] Furthermore, the construction process of the deep collaborative pulsed neural membrane system specifically includes:

[0016] The system is formally defined based on four types of cooperating neurons with different firing mechanisms;

[0017] According to the formal definition, impulses are used to activate triggering rules, and the output value is transmitted from the presynaptic neuron to the postsynaptic neuron through the triggering rules to complete the calculation.

[0018] Through calculation, all possible triggering rules in the neuron are triggered. If there are no executable triggering rules, the calculation stops and the calculation result is output.

[0019] A second aspect of the present invention provides a system for predicting overall survival in patients with glioblastoma, comprising:

[0020] The preprocessing module is configured to acquire and preprocess histopathological images;

[0021] The survival prediction module is configured such that the deep co-pulsing neural membrane system includes input / output neurons and multiple co-pulsing neuron modules; the co-pulsing neuron modules include four types of co-pulsing neurons with different firing mechanisms, which can self-assemble into different blocks to realize various learning mechanisms.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] The prediction method provided by this invention benefits from the parallelism of the system. By integrating several dual-task attention networks with different hyperparameters into multiple multi-neuron collaborative sub-modules with different parameters, the accuracy of OS time prediction is further improved and the time consumption is reduced.

[0024] The prediction method provided by this invention achieves simultaneous prediction of OS time and genotype by constructing a multi-task attention network.

[0025] The prediction method provided by this invention selects survival-related clusters by designing four new triggering rules, achieving ROI-free prediction and improving OS time prediction efficiency.

[0026] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0028] Figure 1 A structural diagram of the deep synergistic pulsed nerve membrane system provided in an embodiment of the present invention;

[0029] Figure 2 Two WSIs from two GBM patients and four slides obtained from each WSI are provided for embodiments of the present invention.

[0030] Figure 3 This is the overall working framework of the deep collaborative pulsed nerve membrane system provided in the embodiments of the present invention;

[0031] Figure 4 The diagram shows the functional block connections of the deep synergistic pulsating neural membrane system provided in this embodiment of the invention, which consists of synergistic triggering mechanisms between different neurons. Detailed Implementation

[0032] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all 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 pertains.

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] Example 1:

[0035] During adult brain injury, neurons and glial cells cooperate to reactivate dormant neural progenitor cells and promote regenerative responses, demonstrating that in real neurobiological systems, various types of neurons communicate and cooperate closely to achieve the same goal. Inspired by this biological phenomenon, this invention designs a deep co-pulsing neural membrane system composed of four co-operating neurons with different firing mechanisms to achieve various functions, such as feature extraction, to address the shortcomings of traditional pulsing neural membrane systems in solving practical problems. This invention provides a deep co-pulsing neural membrane system for predicting overall survival (OS) in glioblastoma (GBM) patients. It can more effectively extract survival-related features from pathological images, utilize numerical variables to reduce the consumption of pulses and neurons, and employ four co-operating neurons with triggering rules that can self-assemble into different blocks to achieve various learning mechanisms. By constructing a dual-task attention network to achieve cross-task feature sharing, it can not only simultaneously meet the requirements of predicting OS time and genotype classification, but also improve the efficiency and accuracy of ROI-free OS time prediction for GBM patients.

[0036] like Figure 1 As shown, this invention provides a method for predicting overall survival in patients with glioblastoma:

[0037] like Figure 3As shown, the acquisition and preprocessing of histopathological images specifically included: acquiring two WSIs from two GBM patients in the TCGA-GBM dataset and four slices from each WSI. The OS of patient TCGA-06-0122 was 184 days, and the OS of patient TCGA-02-0006 was 550 days. Each WSI used in this method (usually 100,000 × 100,000) was magnified 40 times and processed into slices of size 1024 × 1024.

[0038] The acquired histopathological images are input into the Deep Co-Pulse Neuron System. The Deep Co-Pulse Neuron System uses input neurons to read the data. Note that various types of data need to be converted into numerical matrices before they can be read. Therefore, when using the Deep Co-Pulse Neuron System to predict the OS time of glioblastoma patients, the pixel matrix containing q WSI images is first input into the input neurons, and the input neurons store these data in variables.

[0039] In this embodiment, the deep co-pulsing neural membrane system, composed of four types of co-pulsing neurons with different firing mechanisms, is defined as follows:

[0040] Definition 1: The formal definition of a deep co-pulsating neural membrane system with degree m≥1 is as follows:

[0041] Π=(O,N,w ij ,β att ,syn,σ in ,σ out (1)

[0042] (1): O = {a} is the alphabet, where a represents a pulse;

[0043] (2): N={σ a ,σ b ,σ l ,σ f} is a set of four types of coordinating neurons. and These are collections of attention neurons, bias neurons, and layer neurons, respectively. Feature neurons are represented as... m1, m2, m3, and m4 represent the neuron σ a ,σ b ,σ l and σ f The number of neurons is given by m1 + m2 + m3 + m4 = m. Cooperative neurons can be uniformly represented as σ. i =(q i ,x i ,r i ), 1≤i≤m where q iIt is the number of pulses, x i It is a variable that stores a numerical matrix, r i It is for x i The triggering rules for performing a certain computation have the following four forms (neuron σ i It is a neuron σ j (presynaptic neurons):

[0044] (2-1): Transmission Rules

[0045]

[0046] If neuron σ i If pulse a is contained in time t, then the propagation rule is executed. At this time, variable x... i The value is x i (t), variable x j The value is x j (t). Neuron σ i Output value PV i (t)=x i (t) is transmitted to neuron σ j variable x j At the same time, the pulse is also transmitted to σ j But neuron σ i The number of pulses in the pulse becomes zero.

[0047] (2-2): Attention Rule

[0048]

[0049] If neuron σ i If time t contains impulse a, then the attention rule is executed. At this time, variable x... i ,x j and attention coefficient β att The values ​​are x i (t),x j (t) and β att (t). Neuron σ i x i The value of (t) and β att Multiply by (t) and combine the output value pv i (t)=x i (t)×β att (t) is transmitted to neuron σ j variable x j At the same time, the pulse is also transmitted to σ. j But neuron σ i The number of pulses in the pulse becomes zero.

[0050] (2-3): Multifunctional Rules

[0051]

[0052] f is the activation mechanism of the multifunctional rule. If neuron σ i If pulse a is contained in time t, then the multi-functional rule is executed. At this time, variable x... i and x j The values ​​are x i (t) and x j (t). Neuron σ i Output value PV i (t)=f(x i (t) is transmitted to neuron σ j variable x j Simultaneously, the pulse is also transmitted to σ. j But neuron σ i The number of pulses in the pulse becomes zero.

[0053] (2-4): Convolution Rules

[0054]

[0055] f is the activation mechanism of the convolution rule. The convolution kernel w... ij The value placed on the synapse (i,j) represents the synaptic weight. If the neuron σ i If time t contains pulse a, then the convolution rule is executed. At this time, variable x... i ,x b ,x j and synaptic weight w ij The values ​​are x i (t),x b (t),x j (t) and w ij (t). w ij (t)x i (t) Perform convolution operations under the influence of three parameters: kernel size, stride, and padding. b (t) comes from biased neuron σ b Neuron σ i Output value PV i (t)=f(w ij (t)x i (t)+x b (t) is transmitted to neuron σ j variable x j Simultaneously, the pulse is also transmitted to σ. j But neuron σ i The number of pulses in the pulse becomes zero.

[0056] (3): representing neuron σi and σ j Synaptic channels between them.

[0057] (4): Input neuron σ in Used to import data from the external environment and output neuron σ. out Used to output the calculation results of the system.

[0058] It is easy to see that, in this embodiment, the deep co-pulsing neural membrane system Π contains four types of co-pulsing neurons, which can self-assemble into different blocks to achieve complex computations. Layer neurons σ l This reflects the structure of the system; pay attention to the neuron σ. a Calculate the attention coefficient β att Biased neuron σ b Regulating the firing of characteristic neurons, characteristic neuron σ f Feature extraction, feature weight allocation, and feature aggregation are performed based on triggering rules. Each neuron σ in the system... i Both contain a certain number of pulses, and a variable x. i And a triggering rule; in addition, Π also contains synaptic weights w ij and attention coefficient β att Two parameters.

[0059] As one implementation method, at time t=0, the system state is called the initial configuration C0=[q i (0),x i (0),w ij (0),β att [0], 1≤i≤m, the transition from one configuration to another is called a transformation, and the computation of the system consists of a series of transformations, i.e.

[0060] Based on the above definition, triggering rules can be triggered by impulse activation. The deep co-pulsing neuromembrane system Π completes the calculation by transmitting the output value from the presynaptic neuron to the postsynaptic neuron through the triggering rules. Assuming neuron σ i It is a neuron σ j Presynaptic neurons and σ i At time t, there is a pulse, σ i Execute trigger rule r i And the output value pv is transmitted through the synaptic channel (i,j). i (t) is transmitted to the postsynaptic neuron σ j The variable. At the same time, σ j Received by σ i The pulse passed to it, variable x j The value at time t is

[0061]

[0062] Where pv(t) is the neuron σ j The sum of the output values ​​of all presynaptic neurons.

[0063] At the next moment, σ i variable x i The value becomes:

[0064]

[0065] Where pv′(t) is the neuron σ i The sum of the output values ​​transmitted by the presynaptic neurons. If σ i If it is triggered at time t, then x i The value of (t) becomes zero; therefore, x i The value of (t+1) is pv′(t); otherwise, x i The value of (t+1) is x i (t)+pv′(t). Under the global clock, system Π performs computation in maximum parallelism, meaning that all triggering rules in the neuron that can be triggered at time t are triggered. If there are no executable triggering rules in the system, the computation stops, and neuron σ... out Output the calculation results.

[0066] Most existing survival prediction methods are based on ROI, which have the disadvantages of being difficult to label, requiring expert experience, and being time-consuming. In this embodiment, the input histopathological images are processed according to phenotype, and each histopathological image is enlarged, sliced, and clustered to obtain several slice clusters.

[0067] In this embodiment of phenotypic clustering: because initially there is only the input neuron σ... in Contains a single pulse, the pulse passes through σ in Transmitted to the postsynaptic neuron without being consumed; neuron σ in Read the pixel matrix of q WSIs and store it in the variable x. in In the middle. When the pixel matrix of all WSIs is read, the neuron σ in Transmission rules in The pixel matrix is ​​transmitted to q feature neurons σ by the impulse activation. 1,1 ,σ 2,1 ,…,σ q,1 variable x 1,1 ,…,x q,1 Since WSIs with gigapixel spatial resolution cannot be directly input into the system for training, the neuron σ... 1,1 ,σ 2,1 ,…,σ q,1 Use multi-functional rules If 1 ≤ i ≤ q, the WSIs are cropped into slices and stored simultaneously in x2. f1 ensures a uniform slice size. Then, neuron σ2 executes the multi-functional rule. Cluster all slices according to phenotype; k∈N * A cluster of slices with different survival discriminations is transmitted to k feature neurons σ. 1,3 ,σ 2,3 ,…,σ k,3 The variables; finally, neuron σ3 transmits the rules. Pass k slice clusters to layer neurons Variables.

[0068] In this embodiment, the present invention transmits several slice clusters to the second submodule of the deep co-pulse neural membrane system through a transmission rule, and performs long and short OS time classification to automatically obtain slice clusters related to survival. Specifically, long OS time and short OS time are classified as follows:

[0069]

[0070] Where x is based on The vector to be computed, x p or x k It is an element in x.

[0071] Among them, layer neurons apply multifunctional rules to optimize the computation process:

[0072]

[0073] Where m represents the number of slices in cluster k, y c (x k ) represents the actual class label of the slice in cluster k.

[0074] like Figure 4 As shown, the assembly of various functional blocks can form deep co-pulsing neural membrane systems with different structures; in deep co-pulsing neural membrane systems, the number and order of neurons are determined according to the specific application; Π ex The model assembles convolutional blocks, attention blocks, and fully connected blocks, and uses input and output neurons to form an attention-based feature extraction model. This model requires four types of coordinating neurons and four rules. Each neuron in the model contains a certain number of spiking kinetics, a variable, and a triggering rule. Furthermore, the deep coordinating spiking neuron system also includes two parameters: synaptic weights and attention coefficients.

[0075] In this embodiment, a DSSN membrane system Π for OS time prediction was constructed. OS The DSSN membrane system consists of input neurons σ inOutput neuron σ out It consists of three multi-neuronal collaborative modules Π1, Π2, and Π3 for OS time prediction and MGMT classification tasks; Note: The three sub-modules of the deep collaborative pulse membrane system mentioned above are the multi-neuronal collaborative modules Π1, Π2, and Π3 in this embodiment.

[0076] Specifically, ∏1 slices and clusters q WSIs in layer neurons σ1 according to phenotype; therefore, k slice clusters with different survival discriminative properties are passed to ∏2 via a transitive rule. ∏2 performs long and short OS time classification in parallel and selects survival-related clusters, where each It includes 25 convolutional blocks and one fully connected block. Then, a propagation rule passes s≤k survival-related slice clusters to ∏3. ∏3 consists of n attention-based dual-task layer neurons. and Composition. Each layer of neurons comprises 58 convolutional blocks, 5 attention blocks, and one fully connected block, used to perform Ov temporal prediction and MGMT classification tasks. Extracted features are transferred between the two tasks via transfer rules to assist each other. Benefiting from the parallelism of the P system, ∏3 uses multifunctional rules to optimize the layer neurons σ. i All predicted values ​​are averaged to perform ensemble learning. Then, the neuron σ... i The OS time prediction results and MGMT classification results are transmitted to the output neuron σ. out Finally, σ out Use transit rules to output the final result to the external environment.

[0077] This embodiment can simultaneously perform OS time prediction and genotype classification: since both OS time prediction and genotype-metamorphic-transfer (MGMT) classification are closely related to survival, an ensemble model Π3 composed of n attention-based multi-neuronal collaborative submodules is designed to jointly learn these two tasks. Specifically, layer neurons simultaneously learn pathological features related to OS time and MGMT through convolutional rules and multifunctional rules. and Features learned from two tasks are interacted through a pass-through rule. To further highlight useful features, features are added to neurons. and The SE attention mechanism is implemented through attention rules; finally, the feature neurons apply multifunctional rules to supervise the entire prediction process using the obtained prediction values.

[0078]

[0079]

[0080] Where m is the neuron σi The number of slices in the middle, y r (x i ) and y c (x i The actual OS time and the weights for different genotypes were respectively assigned to... and To balance two tasks, neurons and Backpropagation updates synaptic weights for optimization.

[0081] Finally, the optimization stops after reaching the predetermined number of iterations, and the optimal value is passed to the output neuron σ. out The variable. Then, σ out Use transit rules The final prediction results are then output to the external environment.

[0082] Experimental setup: The DSSN membrane system proposed in this invention was implemented using the PyTorch framework on an NVIDIA Tesla V100 GPU with 32GB of memory. The Adam algorithm was used to iteratively update the synaptic weights to optimize the DSSN membrane system. Parameter settings are shown in Table 1:

[0083] Table 1: Parameter settings for DSSN membrane system

[0084]

[0085]

[0086] This invention selects 110 WSIs from the GBM cohort of the Cancer Genome Atlas as a dataset. Each WSI is sliced, and 110 WSIs are used for a five-fold cross-validation experiment. The number of WSIs used for training, validation, and testing are 88, 7, and 15, respectively. The DSSN membrane system includes one input neuron, one output neuron, 897 layer neurons, 630 bias neurons, 50 attention neurons, and several feature neurons.

[0087] The dataset includes statistics on the number of patients, WSIs, slides, and the average number of slides extracted per WSI. OS time (in days) is the mean ± standard deviation. "met" and "unmet" represent the methylated and unmethylated states of MGMT, respectively, as shown in Table 2.

[0088] Table 2: Dataset Statistics

[0089]

[0090] This invention uses three indicators, namely root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient (CC), to evaluate the linear correlation between the predicted and true values ​​of OS time. After a five-fold cross-validation experiment, the RMSE, MAE, and CC values ​​of the DSSN membrane system are 212.3±117.6, 191.7±120.2, and 0.541, respectively, which are significantly better than previous methods.

[0091] Example 2

[0092] This invention provides a system for predicting overall survival in patients with glioblastoma, comprising:

[0093] The preprocessing module is configured to acquire and preprocess histopathological images;

[0094] The survival time prediction module is configured such that the deep co-pulsing neural membrane system includes input / output neurons and a multi-neuronal co-operation module; the multi-neuronal co-operation module is self-assembled from four types of co-operating neurons with different firing mechanisms to achieve various learning mechanisms, specifically including: layer neurons, attention neurons, bias neurons, and feature neurons; the deep co-pulsing neural membrane system performs a long-total-survival time classification task in parallel through multiple neuron co-operation sub-modules to select survival-related slice clusters, and then passes a set number of survival-related slice clusters to the next multi-neuronal co-operation module through a transfer rule. It consists of multiple attention-based dual-task layer neurons, used to perform glioblastoma overall survival prediction and genotype classification tasks; the bias neurons are used to regulate the firing of feature neurons; the feature neurons perform feature extraction, feature weight allocation, and feature aggregation according to triggering rules; wherein, each coordinating neuron includes: a certain number of spiking impulses, a variable, and a triggering rule; the triggering rule has four forms, specifically including: transmission rule, attention rule, multifunctional rule, and convolution rule; the feature neurons can also supervise the entire prediction process according to the application of multifunctional rules. In addition, the deep co-pulsing neuromembrane system also includes: synaptic weights and attention coefficients.

[0095] Example 3

[0096] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting the overall survival time of glioblastoma patients using a neuromembrane system as described in Embodiment 1.

[0097] Example 4

[0098] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method for predicting the overall survival time of glioblastoma patients as described in Embodiment 1.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting overall survival in patients with glioblastoma, characterized in that, include: Acquire histopathological images; Histopathological images are input into a pre-constructed deep co-pulsing neural membrane system to obtain overall survival prediction results. The construction process of the deep co-pulsing neural membrane system specifically includes: The deep coordinating pulsating neuron system is formally defined based on four types of coordinating neurons with different firing mechanisms; these four types of neurons with different firing mechanisms are: layer neurons, attention neurons, bias neurons, and feature neurons. According to the formal definition, impulses are used to activate triggering rules, and the output value is transmitted from the presynaptic neuron to the postsynaptic neuron through the triggering rules to complete the calculation. The calculation triggers all the triggering rules in the neuron that can be triggered at a set time. If there are no executable triggering rules, the calculation stops and the calculation result is output. The input histopathological images are processed according to their phenotypes, specifically including: enlarging and slicing each histopathological image and clustering them to obtain several slice clusters; Several slice clusters are passed to the second sub-module of the deep co-pulse neural membrane system through a transfer rule, and long and short OS time classification is performed to obtain slice clusters related to survival. Using the transfer rules, survival-related slice clusters that meet the threshold conditions are transferred to the third submodule of the deep co-pulsing neural membrane system to perform OS time prediction and genotype classification. Transmission rules: If neurons In time Contains pulse If so, the transitive rules are executed; at this time, the variable... The value is ,variable The value is Neuron Output value Transmitted to neurons variables At the same time, the pulse is also transmitted to But neurons The number of pulses in the middle becomes zero; The first, second, and third sub-modules of the deep synergistic pulse neural membrane system are multi-neuron synergistic modules. , ; In layer neurons According to phenotype WSIs slices were clustered to identify different survival discriminants. Each slice cluster is passed to [the relevant entity] via the transit rules. middle; Parallel execution of long and short OS time classification and selection of survival-related clusters, where each It includes 25 convolutional blocks and one fully connected block; the transitive rules will A survival-related slice cluster was passed to middle; Depend on A dual-task layer neuron based on attention and Composition: Each layer of neurons consists of 58 convolutional blocks, 5 attention blocks, and one fully connected block, used to perform OS time prediction and MGMT classification tasks.

2. The method for predicting overall survival in glioblastoma patients as described in claim 1, characterized in that, include: When the deep collaborative pulsed nerve membrane system reads data, various types of data need to be converted into numerical matrices before they can be read. Therefore, histopathological images are converted into pixel matrices and input into the input neurons, which then store these data in variables.

3. The method for predicting overall survival in glioblastoma patients as described in claim 1, characterized in that, When performing OS time prediction and genotype classification, the following are included: Monitor the entire forecasting process: in, To calculate the mean square error between the predicted OS time and the actual OS time, Neuron The number of slices, where i is the index of the slice sample, and t is the time step. Neuron The output value at time t, which is the overall survival time of the i-th sample predicted by the model; Neuron The output value at time t, i.e. the probability that the i-th sample is methylated as predicted by the model; It is the true overall survival time of the i-th sample. The true genotype category label of the i-th sample; and The actual OS time and different genotypes were assigned weights respectively. and To balance two tasks, neurons and Backpropagation updates synaptic weights for optimization; and These are neurons and exist The variable value at any given time; and These are neurons and The output value, and will soon be passed to the variable. and That is, backpropagation updates synaptic weights for optimization.

4. The method for predicting overall survival in glioblastoma patients as described in claim 1, characterized in that, Perform long and short OS time classification to obtain survival-related slice clusters, specifically including classifying long and short OS times: in, It is based on The vector being computed. or yes The elements in Output probabilities for classifying long and short OS times. It is an activation function. At time t, from the neuron arrive Synaptic weights, Neuron The variable at time t, It is a biased neuron The variable at time t.

5. The method for predicting overall survival in glioblastoma patients as described in claim 4, characterized in that, Optimize the computation process by applying multifunctional rules using layer neurons: in Cluster The number of slices in the middle, Cluster The true class tag of the slice, To calculate the cross-entropy loss for classification with long and short OS times, Neuron The output value at time t, Cluster The real class tag of the slice.

6. A system for predicting overall survival in patients with glioblastoma, used to implement the method as described in any one of claims 1-5, characterized in that, The preprocessing module is configured to acquire and preprocess histopathological images; The survival prediction module is configured such that the deep co-pulsing neural membrane system includes input / output neurons and multiple co-pulsing neuron modules; the co-pulsing neuron modules include four types of co-pulsing neurons with different firing mechanisms, which can self-assemble into different blocks to realize various learning mechanisms.

7. The overall survival prediction system for glioblastoma patients as described in claim 6, characterized in that, The coordinating neuron module includes four types of coordinating neurons with different firing mechanisms, which can self-assemble into different blocks to achieve various learning mechanisms. Specifically, these include: layer neurons, attention neurons, bias neurons, and feature neurons. Multiple neuron coordinating sub-modules can perform long and short overall survival classification tasks in parallel to select survival-related slice clusters. It can also be used to perform glioblastoma overall survival prediction and genotype classification tasks. The biased neurons are used to regulate the firing of the characteristic neurons; The feature neurons perform feature extraction, feature weight allocation, and feature aggregation according to the triggering rules.

8. The overall survival prediction system for glioblastoma patients as described in claim 7, characterized in that, Each coordinating neuron includes: a certain number of impulses, a variable, and a triggering rule; The triggering rules have four forms, specifically including: transitive rules, attention rules, multi-functional rules, and convolution rules; In addition, the deep co-pulsing neural membrane system also includes synaptic weights and attention coefficients.

9. A computer device and a readable storage medium, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1-5.

Images