Method and system for predicting lifetime of glioblastoma patient

Abstract

The invention provides a glioblastoma patient lifetime prediction method and system, and relates to the technical field of intelligent medical treatment, and the method comprises the steps: obtaining and coding a digital full-slice image of a patient; screening high-information-content image blocks based on information entropy; executing a dendritic calculation rule and a structural plasticity rule through internal learnable dendritic neurons by using a dendritic morphological neural calculation model, and extracting global and local features from the image blocks; fusing the features through feature fusion neurons to obtain depth features; and finally, predicting the total lifetime through the output neurons. By simulating hierarchical calculation and plasticity of neuron dendrites, a double-level feature learning model capable of integrating global context and local details of a full-slice image at the same time is constructed, and the technical problems that an existing prediction method is insufficient in feature integration and limited in model calculation structure are effectively solved. And more comprehensive and accurate prediction of the lifetime of the glioblastoma patient is realized.

Description

Technical Field

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

[0002] Glioblastoma (GBM) is the most aggressive primary brain tumor, characterized by high incidence, high mortality, and diverse clinical manifestations. The five-year survival rate after diagnosis is only about 5%, resulting in an extremely poor prognosis. Its severity and complexity make it a focal point in the field of neuro-oncology. Developing accurate overall survival prediction models is crucial for individualized treatment plans and improving disease prognosis. Currently, deep learning-based computational pathology is working to predict cancer patient survival from digital whole-slice images (WSI). Typical automated survival prediction methods usually involve three key steps: image patch sampling, block-level feature extraction, and slice-level representation learning.

[0003] However, existing model architectures suffer from the following problems: On the one hand, mainstream feature extraction models, such as multi-instance learning frameworks and multi-scale fusion networks, are limited by local perspectives and struggle to effectively integrate global contextual information and local nucleus details from whole-slice images; on the other hand, while novel bio-inspired computational models, such as the spiking neural (SN) membrane system, offer advantages in parallel computing, they model neurons as simple computational units, neglecting the powerful, hierarchical computational potential inherent in the dendritic structure of biological neurons. This results in existing models being insufficiently comprehensive and accurate in capturing complex pathological morphological features for precise, individualized survival prediction. Summary of the Invention

[0004] To address the technical problems of existing prediction models in the background art, such as their inability to integrate global and local image information and their failure to fully utilize the internal computational structure of neurons, this invention provides a method and system for predicting the survival of glioblastoma patients. Its core lies in introducing a dendritic morphology neural computation model. By simulating the hierarchical computation and plasticity of dendrites in biological neurons, complex feature processing is achieved within a single neuron, and a two-level learning framework integrating global and local features is constructed, thereby achieving a more comprehensive and accurate prediction of the survival of GBM patients.

[0005] To achieve the above objectives, a first aspect of the present invention provides a method for predicting the survival of patients with glioblastoma, comprising: Digital whole-section images of glioblastoma patients were acquired and encoded into numerical matrices. Based on information entropy, select the image patches with the highest information content from the numerical matrix; The image patch is input into a dendritic morphology neural computation model for processing. The model contains at least one learnable dendritic neuron. Each neuron has a multi-layered dendritic tree structure. Each dendritic branch serves as an independent computational unit. By executing preset communication rules, global and local features are extracted from the image patch. The global and local features are fused using the feature fusion neurons in the model to generate fused deep features. The fused deep features are input into the output neuron to calculate the predicted overall survival value of the patient.

[0006] Furthermore, the preset communication rules include dendrite calculation rules and structural plasticity rules: The dendrite calculation rule is as follows: ,in, For dendritic branching variables, It is a learnable linear or nonlinear function; The structural plasticity rule is as follows: ,in, The weights are values ​​in the range [0,1] and are used to dynamically adjust dendritic connections.

[0007] Furthermore, the step of selecting the multiple image patches with the highest information content from the numerical matrix based on information entropy includes: The information entropy value of the image region is calculated according to the dendrite calculation rule, and a preset number of regions with the highest information entropy value are selected as the image blocks.

[0008] Furthermore, the feature fusion neuron generates the fused deep features according to the following rules. Firstly, according to Generate attention feature map ; and then according to Feature weighting and concatenation are performed; among them, and These are the global features and local features, respectively. Represents matrix multiplication. This represents element-wise multiplication. This indicates a splicing operation.

[0009] Furthermore, the dendritic morphology neural computation model also includes a self-supervised learning architecture, which comprises an encoder and a decoder. The encoder performs multi-level downsampling, feature extraction, and attention enhancement on the image patch, while the decoder reconstructs the image based on the depth features output by the encoder and minimizes the mean squared error loss between the reconstructed image and the original image. To optimize model parameters, model parameters according to To be updated, among which, This is the learning rate.

[0010] Furthermore, the calculation rule for the output neuron is as follows: Preliminary predicted values ​​were obtained. ,in, The fused depth features are referred to as such. and These are learnable parameters; Based on the preliminary predicted values Compared to actual lifespan According to the loss function Calculate the predicted loss, where the predicted total survival value is... ,in, For the number of patients.

[0011] A second aspect of the present invention provides a survival prediction system for glioblastoma patients, comprising: The image acquisition and encoding unit is used to acquire digital whole-slice images of glioblastoma patients and encode them into a numerical matrix; An image patch filtering unit is used to filter out the multiple image patches with the highest information content from the numerical matrix based on information entropy; The feature extraction unit is used to input the image patch into the dendritic morphology neural computing model for processing. The model contains at least one learnable dendritic neuron. Each neuron has a multi-layered dendritic tree structure. Each dendritic branch is an independent computing unit. By executing preset communication rules, it extracts global and local features from the image patch. The feature fusion unit is used to fuse the global features and local features through the feature fusion neurons in the model to generate fused deep features; The survival prediction unit is used to input the fused deep features into the output neuron and calculate the predicted total survival value of the patient.

[0012] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the glioblastoma patient survival prediction method as described in the first aspect of the present invention.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the glioblastoma patient survival prediction method as described in the first aspect of the present invention.

[0014] A fifth aspect of the present invention provides a computer program product comprising software code, wherein a program in the software code performs steps in the method for predicting the survival of glioblastoma patients as described in the first aspect of the present invention.

[0015] Compared with existing technologies, the survival prediction method and system for glioblastoma patients provided by this invention have the following beneficial effects: (1) Based on the dendritic morphology neural computation model and feature fusion neurons, this invention processes image blocks of different resolutions in parallel through dendritic branches, and uses rules such as matrix multiplication and element-wise multiplication to perform feature interaction and splicing, mapping the global organizational context and local cell nucleus details to different feature streams respectively, and performing explicit attention weighting and fusion calculation within the dendritic structure, thereby generating deep features that simultaneously contain macroscopic and microscopic information, significantly improving the comprehensiveness of survival prediction and the accuracy of context awareness.

[0016] (2) Based on the dendritic computation rules and structural plasticity rules, this invention models the dendrites of neurons as independently executable computations (through functions). ) and dynamic reconstruction connections (through weights) The functional unit (regulation) endows a single neuron with powerful internal computation and structural adaptation capabilities. Specifically, computation on dendritic branches is performed hierarchically and orderly, and the connection topology can be optimized and adjusted according to the input signal. This allows a single neuron to simulate the hierarchical integration and learning process of biological dendrites, achieving feature transformation and adaptive expression capabilities comparable to multilayer networks while maintaining higher parameter efficiency.

[0017] (3) This invention is based on information entropy filtering and a self-supervised learning architecture. It automatically focuses on information-rich key regions in the image through information entropy and uses an encoder-decoder for image reconstruction pre-training. Specifically, information entropy filtering filters out low-information background noise at the source, while the self-supervised reconstruction task drives the encoder to learn the general pathological morphological features that are crucial for reconstruction, reducing the dependence on a large amount of finely labeled survival data. This enables the model to learn from massive amounts of unlabeled WSI, enhancing its stability and generalization ability in the face of data noise and sample differences. Attached Figure Description

[0018] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0019] Figure 1 This is a flowchart of a method for predicting the survival of glioblastoma patients provided in Embodiment 1 of the present invention; Figure 2This is a schematic diagram of matrix calculation in the SN membrane system and LDN membrane system provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of neuronal dendrite layering calculation in the LDN membrane system provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of four learnable neurons based on dendritic computation provided in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the LDN membrane system encoding WSI provided in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the neuronal structure of the LDN system for self-supervised survival prediction of GBM patients provided in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram showing the global-local feature extraction results of the self-supervised learning module provided in Embodiment 1 of the present invention with and without the attention mechanism enabled; Figure 8 This is an architecture diagram of the glioblastoma patient survival prediction system provided in Embodiment 2 of the present invention. Detailed Implementation

[0020] 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 in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0023] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0024] Example 1 like Figure 1 This embodiment provides a method for predicting the survival of patients with glioblastoma, including: Digital whole-section images of glioblastoma patients were acquired and encoded into numerical matrices. Based on information entropy, select the image patches with the highest information content from the numerical matrix; The image patch is input into a dendritic morphology neural computation model for processing. The model contains at least one learnable dendritic neuron. Each neuron has a multi-layered dendritic tree structure. Each dendritic branch serves as an independent computational unit. By executing preset communication rules, global and local features are extracted from the image patch. The global and local features are fused using the feature fusion neurons in the model to generate fused deep features. The fused deep features are input into the output neuron to calculate the predicted overall survival value of the patient.

[0025] The dendritic morphology neural computation model provided in this embodiment is a learnable dendritic neural (LDN) membrane system to simulate the hierarchical computation and plasticity of dendrites in biological neurons. Inspired by the dendritic integration and nonlinear computation principles of biological neurons, the LDN membrane system visualizes the dendritic structure of neurons as a tree-like topology with multi-level branches. Each dendritic branch acts as an independent computational unit, performing complex calculations using numerical variables and mathematical functions, thereby endowing a single neuron with computational capabilities comparable to multi-layer networks. The system also incorporates specialized plasticity rules that adaptively adjust dendritic morphology and function.

[0026] Specifically, the LDN membrane system is defined by the following quintuple: ; (1) A set of neurons, where, ≥1 indicates a neuron number. Each neuron can be represented as... (1≤ ≤ ), Represents neurons The dendritic structure can be viewed as a dendritic tree composed of multiple dendritic branches. Let be variables located in the cell body used for dendritic information integration. dendrites In depth Dendritic branches at the location, among which, Variables for storing and calculating matrices, This indicates the communication rules.

[0027] (2) Indicates from depth To the depth The set of dendritic connections of dendritic branches (sparse connections).

[0028] (3) It is a set of synapses.

[0029] (4) and These correspond to the input neuron and the output neuron, respectively.

[0030] Specifically, the preset communication rules This includes rules for dendrite calculation and rules for structural plasticity: a) Dendrite Calculation Rules

[0031] Dendrite computation rules utilize learning functions Perform linear or nonlinear calculations. and Representing variables respectively and At any moment The value of . Indicates dendritic branching At any moment The output value. If the dendrite rule is at time... If enabled, dendritic branches Output value Along dendrite connection Passed to variable If dendritic branches If the depth is 1, the output value is passed to the cell body. .

[0032] b) Plasticity rule

[0033] Plasticity rules dynamically regulate the connection of dendritic branches through structural plasticity. This represents the dendrite weight. and Representing variables respectively and At any moment The value of . Indicates dendritic branching At any moment The output value. If the plasticity rule is at time... If enabled, dendritic connections Will be deleted, output value Connect along the newly established dendritic branches Passed to variable The weights of deleted dendritic connections are reset to zero, while reconstructed dendritic connections are assigned weights in the range [0,1] and can be adjusted subsequently.

[0034] The LDN membrane system is controlled by a global clock synchronization. In each time step, each dendritic branch executes at most one rule, and the rules in all neurons are executed in parallel. If a dendritic branch has multiple applicable rules in a given time step, the system will select one of them to apply probabilistically. Figure 2 Taking numerical matrix computation as an example, this paper compares and illustrates the difference between the traditional SN membrane system, which uses neurons for computation, and the system in this paper, which uses dendrites for computation. Figure 2 As shown in (a), for the matrix Traditional SN membrane systems require Includes The neurons that generate each pulse are used for representation and computation. For example... Figure 2 As shown in (b), in this system, each dendritic branch... A corresponding receptive field is used to calculate the matrix. Therefore, for the same numerical matrix, this system only requires a single neuron to complete the computational tasks required by the entire neuronal network in a traditional SN membrane system.

[0035] In the system In the middle, if the same dendritic branch If both dendritic computation rules and plasticity rules are included, then the dendritic computation rules are executed before the plasticity rules. For rules of the same type, the system randomly selects one to execute based on a uniform probability distribution. After a rule is executed, the variables within the branch are updated, and this update may trigger subsequent computation steps. Neuron At any moment The state is denoted as dendritic branches The state is denoted as Branch At any moment The result obtained from the calculation is defined as its output value, denoted as . Similarly, neurons At any moment The output value is Dendritic tree The computation is performed in a hierarchical manner, and the results are integrated into neurons. Within the cell body. The state update of this hierarchical calculation follows formula (1): (1) in, Dendritic branch At any moment The output value, For variables At any moment The value of .

[0036] If at time Dendritic connections exist Then dendritic branches The output value will be executed after the rule is executed. Passed to dendritic branches variables ,at this time ;otherwise .

[0037] If dendritic branches At any moment Its output value Transmission to cell body variables Then the neuron Triggered. The cell state update formula is shown in equation (2): (2) in, Dendritic branch At any moment The output value, For variables At any moment The value of .

[0038] Hypothetical neuron For neurons The postsynaptic neuron, then the neuron Through synapses Its output value Transmitted to neurons dendrites If dendritic branches Received Then the state update of this branch is as shown in equation (3): (3) in, For neurons At any moment The output value, For variables At any moment The value of .

[0039] Figure 3 (a) shows a dendritic tree structure with a depth of 4 consisting of 10 dendritic branches. Figure 3(b) reveals the complex hierarchical computation process within the dendritic tree: each branch contains a computation rule, where branch 8 executes the plasticity rule, and the remaining branches execute the dendritic computation rule. The computation process follows a strict sequential dependency: any dendritic branch only initiates computation after all its predecessor nodes have completed their computations. In the initial stage, terminal branches 2, 3, 5, 8, and 10 are identified as candidate nodes, with branches 2, 5, 8, and 10 being randomly selected for execution. By executing the plasticity rule, branch 8 undergoes structural reorganization, disconnecting from branch 7 and establishing a new connection with branch 6. Subsequent computation stages proceed as follows... Sequential progression: In the second stage, branches 3, 6, and 9 are designated as candidate nodes and the dendritic calculation rules are executed; in the third stage, branches 1 and 7 are selected and activated; in the final stage, branch 4 is activated, completing the entire calculation process.

[0040] Inspired by the diverse dendritic morphologies and plasticity of neurons in biological neural systems, this invention designs four learnable neurons with variable dendritic structures for survival prediction tasks. These neurons can achieve attention-based feature extraction at both global and local scales, such as... Figure 4 As shown, variable dendritic morphology enables neurons to possess multiple computational functions. Therefore, neurons with different dendritic morphologies can perform modular complex operations and can be combined into models with multiple functional structures. Within these neurons, dendritic branches achieve adaptive adjustments in structure and function through dendritic computation rules and plasticity rules. This dendritic-level processing enables a single neuron to possess computational capabilities similar to multilayer networks. Table 1 details the specific application of the aforementioned rules in the dendritic branches of each neuron.

[0041] Table 1. Detailed rules for four exemplary learnable neurons

[0042] The four types of neurons are: 1) Learnable neurons with convolution operation capabilities and a dendritic tree depth of 2. ( Figure 4 a) Among them, dendritic branches and Multiple sets of convolutional kernels were computed in synergistic calculation using dendritic computation rules and plasticity rules, and the results were obtained in the cell body. Integrate in; 2) Neurons that implement channel attention calculation using a dendritic tree of depth 3. ( Figure 4 b), where the branches , and The dendrite computation rules are executed, and operations such as convolution, max pooling, and average pooling are performed, ultimately at the cell body. Generate channel attention weights; 3) Neurons that perform spatial attention computation through a dendritic structure of depth 3 ( Figure 4 c), branch , and Max pooling, average pooling, concatenation, and convolution operations are performed using dendritic computation rules, thereby enabling cell-body operations. Spatial attention weights are formed in the middle; 4) Neurons using dendritic trees of depth 3 for feature fusion ( Figure 4 d), where branches and Attention feature maps that fuse global and local information are generated through dendritic computation and plasticity rules. The initial feature representation is then enhanced through weighted attention computation, ultimately resulting in a representation in the cell body. It integrates global and local features internally.

[0043] Specifically, the step of selecting the multiple image patches with the highest information content from the numerical matrix based on information entropy includes: First, the information entropy value of the WSI image region is calculated, and a preset number of regions with the highest information entropy value are selected as the image blocks.

[0044] Furthermore, in this embodiment, the input neuron Encoding of whole-slice digital images (WSI) based on RGB color mode, and encoding of color images using digital sequences. Input neurons. The R, G, and B channels of the WSI (or image block) are read in parallel, and the intensity values ​​of each channel (range 0-255) are converted into 8-bit binary sequences with values ​​ranging from 00000000 to 11111111. Figure 5 Taking the red channel as an example, this demonstrates a pixel in a full slice image of size W×H. The RGB encoding method.

[0045] M WSIs, represented by a numerical sequence, are input to the input neuron. The terminal dendritic branches, where M=4 represents the number of images processed in parallel. In the input neuron In this model, each dendritic branch independently processes one WSI, and the number of branches matches the number of input images. Considering that WSIs have different feature representations at different resolutions, the LDN membrane system is based on a "pyramid" image architecture and uses image information entropy for image patch sampling. Specifically, the terminal dendritic branch first executes the dendritic calculation rule to obtain n resolution levels of the WSI, as shown in Equation (4); subsequently, the downstream dendritic branch performs image patch sampling at these different resolution levels.

[0046] (4) in, The numerical matrix representing WSI, For functions that perform multi-resolution sampling, For this image Each resolution level, here =17.

[0047] Input neurons The output consists of the κ image patches with the highest information content from each WSI, selected by information entropy filtering, as shown below. These filtered image patches will be directly input into the encoder. Decoder system.

[0048]

[0049] in, (5) In the formula, Indicates the number of occurrences of adjacent coded element pairs. Given the joint probability, and Both are 8-bit binary sequences; W and H represent the width and height of the image, respectively.

[0050] Specifically, the dendritic morphology neural computational model also includes a self-supervised learning architecture, which comprises an encoder and a decoder. For example... Figure 6 As shown, the self-supervised learning module consists of two main components: the encoder. and decoder These two components, through image reconstruction tasks, are able to work in parallel from... Global and local features were extracted from the WSI (Weighted Injection Sample) and corresponding slides of the patients. The aim was to comprehensively understand the tumor tissue and its surrounding environment, while simultaneously capturing finer cellular and nuclear details. Encoder It has four built-in sub-modules , , , They are specifically responsible for extracting feature information closely related to image reconstruction. The decoder... It is equipped with six neurons These neurons work together to reconstruct the image. To focus more on texture and color features in the channel and spatial dimensions that are closely related to image reconstruction, each module of the encoder cleverly incorporates the CBAM attention mechanism. This mechanism relies on neurons... , , as well as The synergistic effect of the encoder. With decoder Information exchange between them is done through neurons. It was accomplished through this process.

[0051] encoder The calculation process starts from the sub-module neurons Initiation begins. In the initial phase, neurons... dendritic branches It will perform masking operations according to specific dendrite rules (6). Simultaneously, it will... Slicing was performed using WSI, with slice values ​​of 20, 50, 100, 200, and 500.

[0052] (6) in, This represents the numerical matrix of WSI, while This represents the numerical matrix of the slice. and These are binary vectors used to mask WSI and slices, respectively. (Symbol) This indicates that element-wise multiplication is performed. Furthermore, the mask's scale is determined by a parameter. It was set up that way.

[0053] Two types of mask images, namely sets and They were input into the dendritic branches respectively. and Convolution operations are performed using dendritic rules (7) to extract global and local features.

[0054] (7) in, and These represent the masked WSI and the set of slices, respectively. and These are the learnable parameters in the convolution operation.

[0055] neurons The extracted features will then be and Transmitted to neurons respectively and Neuron Its responsibility is to generate channel attention weights. Then it is sent to the neuron Next, the neuron According to the dendrite rule (8), it will be achieved through matrix multiplication. The calculated weights are added to the feature matrix. and On each of the channels.

[0056] (8) Subsequently, the channel feature matrix and Entered into neurons and Processing occurs within this process. Similarly, neurons... Responsible for generating spatial attention weights Subsequently, in neurons In this process, these weights are obtained through matrix multiplication. The channel feature matrices from WSI and slices were applied respectively. and At each spatial location, this process follows the dendritic rule (9).

[0057] (9) submodule The calculation process and Similar, but different, neurons The masking operation was eliminated during the calculation, and a downsampling operation was used instead. The calculation results are respectively integrated into the neurons of the decoder. and In addition, it is worth mentioning the submodules. neurons Responsible for processing the global-local depth features extracted by the encoder and Transmitted to neurons For image reconstruction purposes. Meanwhile, and It will also be used for survival prediction in the next phase.

[0058] neurons Each dendritic branch of the layer performs convolution operations in sequence, and strictly follows the dendritic rule (10) throughout the process.

[0059] (10) in, This represents the features extracted from WSI, while This represents the features extracted from the slice. and These are the learnable parameters in the model.

[0060] From neurons Reconstructed images were obtained from the cell bodies. and These reconstructed images are then transmitted to the neurons. And apply dendrite rules (11) to supervise the reconstruction of the image and the original image. and Consistency between them. This calculation process generates the following mean squared error (MSE) loss value and transmits it to the cell body. : (11) Subsequently, the LDN membrane system in neurons The stochastic gradient descent algorithm is used for optimization. The objective function is calculated through backpropagation. Relative to all model parameters The gradient is calculated, and the parameters are updated according to rule (12): (12) in, For the number of iterations, For learning rate, This represents the mean square error loss.

[0061] Specifically, the feature fusion neuron generates the fused deep features according to the following rules. Firstly, based on: (13) Generate attention feature map Furthermore, based on: (14) Feature weighting and concatenation are performed; among them, and These are the global features and local features, respectively. Represents matrix multiplication. This represents element-wise multiplication. This indicates a splicing operation.

[0062] This process utilizes the WSI extracted by the encoder of the self-supervised learning module and the deep features of the sliced ​​image for survival prediction. Specifically, neurons... It will utilize dendritic branches to execute specific dendritic rules (13), cleverly fusing WSI with the features of the sliced ​​image to generate an attention feature map. To more comprehensively capture and represent the original features, dendritic branches further utilize this attention feature map. And according to the dendrite rule (14), the features are... and We perform weighted processing to obtain richer feature representations.

[0063] Specifically, the calculation rule for the output neuron is as follows: Preliminary predicted values ​​were obtained. ,in, The fused depth features are referred to as such. and These are learnable parameters; Output neuron Survival predictions will be made based on the already fused features. At the same time, the dendritic branch will use dendritic rules (15) to output the specific results of the survival prediction. Then, the first layer of the dendritic branch will further apply dendritic rules (16) to effectively supervise the entire prediction process.

[0064] (15) in, Indicates origin from neurons The fusion characteristics and These are learnable parameters.

[0065] (16) Here, This refers to the total number of patients. It is the patient's OS time predicted by the model, and This represents the patient's actual OS time.

[0066] Based on the preliminary predicted values Compared to actual lifespan According to the loss function Calculate the predicted loss, where the predicted total survival value is... ,in, For the number of patients.

[0067] To evaluate the performance of the LDN membrane system in predicting the survival of GBM patients, this invention used histopathological Western blots (WSIs) of 110 patients from the Cancer Genome Atlas GBM cohort. For each WSI, the 20, 50, 100, 200, and 500 image patches with the highest information content were selected based on information entropy. Each image patch was uniformly adjusted to 256×256 pixels before processing.

[0068] The LDN membrane system implemented in this invention is built on the PyTorch framework and utilizes the parallel processing capabilities of an NVIDIA Tesla V100 GPU equipped with 32GB of video memory. During optimization, a stochastic gradient descent algorithm is employed, with a weight decay factor of 0.001 and a learning rate (lr) of 0.005. In the experiments, the batch size is fixed at 16, and the number of training epochs is 80. To comprehensively evaluate system performance, a five-fold cross-validation study was conducted based on 110 WSIs. In each fold, the WSIs are allocated as follows: 88 for training, 7 for validation, and 15 for testing.

[0069] To evaluate the performance of the overall survival prediction task, the root mean square error (RMSE), mean absolute error (MAE), and consistency index (C) were used. The index is used as an evaluation metric. Lower RMSE and MAE values ​​indicate better prediction results, while higher C... The index indicates better consistency between the prediction and the actual results. After five-fold cross-validation experiments, the RMSE, MAE, and C of the LDN membrane system... The index values ​​were 206.4±29.1, 178.5±32.2, and 0.727, respectively, which were significantly better than the previous methods.

[0070] Figure 7 This demonstrates the impact of the CBAM attention mechanism on global-local feature extraction within a self-supervised learning module. To evaluate this impact, four WSI images and image patches were selected, such as... Figure 7 (a) and Figure 7 As shown in (d), feature extraction heatmaps are displayed when the attention mechanism is enabled and disabled in the self-supervised learning module.

[0071] Example 2 like Figure 8 As shown, this embodiment provides a survival prediction system for glioblastoma patients, including: The image acquisition and encoding unit is used to acquire digital whole-slice images of glioblastoma patients and encode them into a numerical matrix; An image patch filtering unit is used to filter out the multiple image patches with the highest information content from the numerical matrix based on information entropy; The feature extraction unit is used to input the image patch into the dendritic morphology neural computing model for processing. The model contains at least one learnable dendritic neuron. Each neuron has a multi-layered dendritic tree structure. Each dendritic branch is an independent computing unit. By executing preset communication rules, it extracts global and local features from the image patch. The feature fusion unit is used to fuse the global features and local features through the feature fusion neurons in the model to generate fused deep features; The survival prediction unit is used to input the fused deep features into the output neuron and calculate the predicted total survival value of the patient.

[0072] Example 3 Embodiment 3 of the present invention provides an electronic device.

[0073] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the glioblastoma patient survival prediction method as described in Embodiment 1 of the present invention.

[0074] The detailed steps are the same as those for the glioblastoma patient survival prediction method provided in Example 1, and will not be repeated here.

[0075] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0076] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the glioblastoma patient survival prediction method as described in Embodiment 1 of the present invention.

[0077] The detailed steps are the same as those for the glioblastoma patient survival prediction method provided in Example 1, and will not be repeated here.

[0078] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0079] A computer program product includes software code, wherein the program in the software code performs the steps in the method for predicting the survival of glioblastoma patients as described in Embodiment 1 of the present invention.

[0080] The detailed steps are the same as those for the glioblastoma patient survival prediction method provided in Example 1, and will not be repeated here.

[0081] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for predicting the survival of patients with glioblastoma, characterized in that, include: Digital whole-section images of glioblastoma patients were acquired and encoded into numerical matrices. Based on information entropy, select the image patches with the highest information content from the numerical matrix; The image patch is input into a dendritic morphology neural computation model for processing. The model contains at least one learnable dendritic neuron. Each neuron has a multi-layered dendritic tree structure. Each dendritic branch serves as an independent computational unit. By executing preset communication rules, global and local features are extracted from the image patch. The global and local features are fused using the feature fusion neurons in the model to generate fused deep features. The fused deep features are input into the output neuron to calculate the predicted overall survival value of the patient.

2. The method as described in claim 1, characterized in that, The preset communication rules include dendrite calculation rules and structural plasticity rules: The dendrite calculation rule is as follows: ,in, For dendritic branching variables, It is a learnable linear or nonlinear function; The structural plasticity rule is as follows: ,in, The weights are values ​​in the range [0,1] and are used to dynamically adjust dendritic connections.

3. The method as described in claim 2, characterized in that, The step of selecting the multiple image patches with the highest information content from the numerical matrix based on information entropy includes: The information entropy value of the image region is calculated according to the dendrite calculation rule, and a preset number of regions with the highest information entropy value are selected as the image blocks.

4. The method as described in claim 1, characterized in that, The feature fusion neuron generates fused deep features according to the following rules. First, according to Generate attention feature map ; and then according to Feature weighting and concatenation are performed; among them, and These are the global features and local features, respectively. Represents matrix multiplication. This represents element-wise multiplication. This indicates a splicing operation.

5. The method as described in claim 1, characterized in that, The dendritic morphology neural computation model also includes a self-supervised learning architecture, which comprises an encoder and a decoder. The encoder performs multi-level downsampling, feature extraction, and attention enhancement on the image patches, while the decoder reconstructs the image based on the depth features output by the encoder and minimizes the mean square error loss between the reconstructed image and the original image. To optimize model parameters, model parameters according to To be updated, among which, This is the learning rate.

6. The method as described in claim 1, characterized in that, The calculation rule for the output neuron is as follows: Preliminary predicted values ​​were obtained. ,in, The fused depth features are referred to as such. and These are learnable parameters; Based on the preliminary predicted values Compared to actual lifespan According to the loss function Calculate the predicted loss, where the predicted total survival value is... ,in, For the number of patients.

7. A survival prediction system for glioblastoma patients, characterized in that, include: The image acquisition and encoding unit is used to acquire digital whole-slice images of glioblastoma patients and encode them into a numerical matrix; An image patch filtering unit is used to filter out the multiple image patches with the highest information content from the numerical matrix based on information entropy; The feature extraction unit is used to input the image patch into the dendritic morphology neural computing model for processing. The model contains at least one learnable dendritic neuron. Each neuron has a multi-layered dendritic tree structure. Each dendritic branch is an independent computing unit. By executing preset communication rules, it extracts global and local features from the image patch. The feature fusion unit is used to fuse the global features and local features through the feature fusion neurons in the model to generate fused deep features; The survival prediction unit is used to input the fused deep features into the output neuron and calculate the predicted total survival value of the patient.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for predicting the survival of glioblastoma patients as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for predicting the survival of glioblastoma patients as described in any one of claims 1 to 6.

10. A computer program product, comprising software code, characterized in that, The program in the software code performs the steps of the method for predicting the survival of glioblastoma patients as described in any one of claims 1 to 6.

Images