Brain glioma microenvironment feature learning method combining super voxel and graph embedding

Abstract

The invention belongs to the technical field of brain glioma microenvironment, and particularly relates to a brain glioma microenvironment feature learning method combining supervoxels and graph embedding. According to the method, local features, invasion dynamics and a global space structure are fully fused, a multi-dimensional and comprehensive microenvironment representation is formed, a rich information basis is provided for accurate diagnosis and treatment, and the accuracy of diagnosis and treatment is improved through fine division of a tumor core area, an infiltration area and normal tissues. The method can more accurately describe the biological behavior of the tumor and the influence of the biological behavior on the surrounding tissue, is beneficial to the evaluation of invasiveness and prognosis, effectively captures the space, texture and morphological information in the image by using the super voxel graph embedding technology, can more intuitively reflect the complex microenvironment relation through the construction of the graph structure, improves the discrimination capability of the model, and improves the accuracy of the model. And dynamic optimization of the feature learning process is realized by introducing a learning compensation mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of glioma microenvironment, and particularly relates to a method for learning glioma microenvironment features combining supervoxels and graph embedding. Background Art

[0002] Glioma, as the main type of adult primary malignant brain tumors, its treatment and prognosis evaluation have always been research hotspots in the medical field. Gliomas are highly invasive and heterogeneous, and their microenvironment features have an important impact on tumor growth, invasion and metastasis. Therefore, in-depth understanding and accurate analysis of glioma microenvironment features are of great significance for improving treatment effects and patient prognosis.

[0003] Traditional methods for analyzing glioma microenvironment features mainly rely on pathological sections and microscopic observations. Although these methods can provide detailed cell and tissue structure information, they have limitations such as complex operations, long time consumption, and difficulty in comprehensively reflecting the dynamic changes of the tumor microenvironment. With the rapid development of medical imaging technologies, especially the wide application of non-invasive imaging technologies such as magnetic resonance imaging (MRI), new ideas and methods have been provided for the analysis of glioma microenvironment features.

[0004] In recent years, significant progress has been made in MRI-based brain tumor segmentation and feature extraction technologies. Among them, supervoxels, as three-dimensional voxel blocks composed of adjacent pixels, can provide richer local information than traditional pixel-level features, thus helping to more accurately capture the morphology and texture features of brain tumors. However, solely relying on supervoxel features for glioma microenvironment analysis still faces challenges, such as high feature dimensions, information redundancy, and difficulty in comprehensively reflecting the complex relationships of the tumor microenvironment.

[0005] Graph embedding technology, as an effective data dimensionality reduction and feature representation method, can map high-dimensional data to a low-dimensional space while preserving the main structure and feature information of the data. Applying graph embedding technology to glioma microenvironment feature learning can further explore the potential relationships between supervoxel features, reduce feature dimensions, and improve the robustness and interpretability of feature representation.

[0006] Currently, there have been studies attempting to apply supervoxels and graph embedding technologies to brain tumor segmentation and feature extraction, but the research on glioma microenvironment feature learning is not sufficient. Most existing methods only rely on voxel-level intensity information, ignoring local structure and morphological features, lacking effective embedding methods to integrate different modalities of image data, resulting in insufficient microenvironment feature expression ability, failing to fully consider the tumor microenvironment changes of different patients, and being difficult to provide precise personalized learning strategies. Summary of the Invention

[0007] The object of the present invention is to provide a method for learning glioma microenvironment features by combining supervoxels and graph embedding, so as to achieve efficient learning and optimization of glioma microenvironment features, and improve the accuracy and stability of image analysis.

[0008] The technical solution adopted by the present invention is specifically as follows:

[0009] A method for learning glioma microenvironment features by combining supervoxels and graph embedding, comprising:

[0010] Obtain the multi-modal three-dimensional medical image data of the brain of a glioma patient, extract the tumor core area, the infiltration area and the surrounding normal tissue area according to the multi-modal three-dimensional medical image data of the brain, and obtain the local characteristic information of the glioma microenvironment according to the tumor core area, the infiltration area and the surrounding normal tissue area;

[0011] Obtain the invasion information of the glioma according to the tumor core area and the infiltration area, and obtain the microenvironment compensation information according to the invasion information;

[0012] Obtain a set of supervoxel units according to the multi-modal three-dimensional medical image data of the brain. Each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information and spatial coordinate feature information, and obtain supervoxel graph embedding information according to the set of supervoxel units;

[0013] Obtain a graph embedding feature learning strategy according to the local characteristic information, the microenvironment compensation information and the supervoxel edge graph embedding information, and obtain the glioma microenvironment graph structure according to the graph embedding feature learning strategy;

[0014] Obtain learning compensation information according to the glioma microenvironment graph structure, obtain microenvironment learning compensation information according to the learning compensation information, the tumor core area and the infiltration area, and re-obtain the graph embedding feature learning strategy.

[0015] In a preferred solution, the step of obtaining the multi-modal three-dimensional medical image data of the brain of a glioma patient, extracting the tumor core area, the infiltration area and the surrounding normal tissue area according to the multi-modal three-dimensional medical image data of the brain, and obtaining the local characteristic information of the glioma microenvironment according to the tumor core area, the infiltration area and the surrounding normal tissue area includes:

[0016] Obtain the multi-modal three-dimensional medical image data of the brain of a glioma patient;

[0017] Extract the tumor core area, the infiltration area and the surrounding normal tissue area according to the multi-modal three-dimensional medical image data of the brain;

[0018] Obtain a corresponding plurality of tumor core vectors according to the tumor core area;

[0019] Obtain a corresponding plurality of infiltration vectors according to the infiltration area;

[0020] Obtain a plurality of corresponding surrounding normal tissue vectors according to the surrounding normal tissue regions;

[0021] Obtain the tumor core weight, the infiltration weight, and the surrounding normal tissue weight;

[0022] Obtain a local feature vector according to the plurality of tumor core vectors, the plurality of infiltration vectors, the plurality of surrounding normal tissue vectors, the tumor core weight, the infiltration weight, and the surrounding normal tissue weight, and label it as the local feature information of the glioma microenvironment.

[0023] In a preferred embodiment, the step of obtaining the tumor core weight, the infiltration weight, and the surrounding normal tissue weight includes:

[0024] Obtain the tumor core area of the tumor core region, the infiltration area of the infiltration region, and the surrounding normal tissue area of the surrounding normal tissue region;

[0025] Obtain the tumor core proportion, the infiltration proportion, and the surrounding normal tissue proportion respectively according to the tumor core area, the infiltration area, and the surrounding normal tissue area;

[0026] Obtain a proportion weight table, wherein the proportion weight table includes three weight sub-tables, namely a tumor core weight sub-table, an infiltration weight sub-table, and a surrounding normal tissue weight sub-table, and each weight sub-table includes corresponding multiple proportion intervals and the weights corresponding to each proportion interval;

[0027] Obtain the corresponding tumor core weight, infiltration weight, and surrounding normal tissue weight respectively from the proportion weight table according to the tumor core proportion, the infiltration proportion, and the surrounding normal tissue proportion.

[0028] In a preferred embodiment, the step of obtaining the tumor core area of the tumor core region, the infiltration area of the infiltration region, and the surrounding normal tissue area of the surrounding normal tissue region includes:

[0029] Obtain an overall image including the tumor core region, the infiltration region, and the surrounding normal tissue region from the brain multi-modal three-dimensional medical image data;

[0030] Construct a plane rectangular coordinate system in the overall image;

[0031] Obtain a plurality of corresponding edge contour inflection point coordinates of the tumor core region, the infiltration region, and the surrounding normal tissue region respectively according to the plane rectangular coordinate system;

[0032] Obtain the corresponding tumor core area, infiltration area, and surrounding normal tissue area respectively according to the plurality of corresponding edge contour inflection point coordinates.

[0033] In a preferred embodiment, the steps of obtaining the invasion information of glioma based on the tumor core region and the infiltration region and obtaining the microenvironment compensation information based on the invasion information include:

[0034] Obtain the tumor core change vector of the tumor core region;

[0035] Obtain the infiltration change vector of the infiltration region;

[0036] Obtain the invasion value based on the tumor core change vector and the infiltration change vector, and label it as invasion information;

[0037] Obtain a compensation table, where the compensation table includes multiple invasion interval values and the microenvironment compensation information corresponding to each invasion interval value;

[0038] Obtain the corresponding microenvironment compensation information from the compensation table according to the invasion value.

[0039] In a preferred embodiment, obtaining a set of supervoxel units based on brain multimodal three-dimensional medical image data, where each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information, and spatial coordinate feature information, and the steps of obtaining supervoxel map embedding information based on the set of supervoxel units include:

[0040] Obtain a set of supervoxel units based on brain multimodal three-dimensional medical image data, where each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information, and spatial coordinate feature information;

[0041] Obtain the corresponding morphological feature vector according to the morphological feature information;

[0042] Obtain the corresponding voxel intensity feature vector according to the voxel intensity feature information;

[0043] Obtain the corresponding texture feature vector according to the texture feature information;

[0044] Obtain the corresponding spatial coordinate feature vector according to the spatial coordinate feature information;

[0045] Obtain the graph embedding value according to the morphological feature vector, voxel intensity feature vector, texture feature vector, and spatial coordinate feature vector;

[0046] Obtain an embedding graph, where the embedding graph includes multiple graph embedding interval values and the supervoxel graph embedding information corresponding to each graph embedding interval value;

[0047] Obtain the corresponding supervoxel graph embedding information from the embedding graph according to the graph embedding value.

[0048] In a preferred embodiment, a graph embedding feature learning strategy is obtained based on local feature information, microenvironment compensation information, and supervoxel edge graph embedding information. The steps of obtaining the glioma microenvironment graph structure according to the graph embedding feature learning strategy include:

[0049] Obtain the corresponding local feature vector according to the local feature information;

[0050] Obtain the corresponding invasion value according to the microenvironment compensation information;

[0051] Obtain the corresponding graph embedding value according to the supervoxel edge graph embedding information;

[0052] Obtain the strategy value according to the local feature vector, invasion value, and graph embedding value;

[0053] Obtain a strategy learning table, where the strategy learning table includes multiple strategy interval values and the graph embedding feature learning strategies corresponding to each strategy interval value;

[0054] Obtain the corresponding graph embedding feature learning strategy from the strategy learning table according to the strategy value;

[0055] Obtain the corresponding supervoxel graph construction method according to the graph embedding feature learning strategy;

[0056] Construct the glioma microenvironment graph structure according to the supervoxel graph construction method and in combination with the brain multi-modal three-dimensional medical image data.

[0057] In a preferred embodiment, learning compensation information is obtained according to the glioma microenvironment graph structure, and microenvironment learning compensation information is obtained according to the learning compensation information, tumor core area, and infiltration area, and the steps of obtaining the graph embedding feature learning strategy are performed again, including:

[0058] Obtain the learning compensation information according to the glioma microenvironment graph structure;

[0059] Obtain the corresponding learning compensation value according to the learning compensation information;

[0060] Obtain the tumor core change vector of the tumor core area;

[0061] Obtain the infiltration change vector of the infiltration area;

[0062] Obtain the invasion compensation value according to the learning compensation value, tumor core area, and infiltration area;

[0063] Obtain a learning compensation table, where the learning compensation table includes multiple invasion compensation interval values and the microenvironment learning compensation information corresponding to each invasion compensation interval value;

[0064] Obtain the corresponding microenvironment learning compensation information from the learning compensation table according to the invasion compensation value;

[0065] Take the microenvironment learning compensation information as the microenvironment compensation information, and re-perform the acquisition of the graph embedding feature learning strategy.

[0066] In a preferred solution, the steps of obtaining the learning compensation information according to the glioma microenvironment graph structure include:

[0067] Obtain the glioma microenvironment vector according to the glioma microenvironment graph structure;

[0068] Obtain the actual glioma microenvironment vector of the glioma patient;

[0069] Obtain the microenvironment deviation value according to the glioma microenvironment vector and the actual glioma microenvironment vector;

[0070] Obtain the learning microenvironment compensation table, where the microenvironment learning compensation table includes multiple microenvironment deviation interval values and the learning compensation information corresponding to each microenvironment deviation interval value;

[0071] Obtain the corresponding learning compensation information from the microenvironment compensation table according to the microenvironment deviation value.

[0072] And, a glioma microenvironment feature learning terminal combining supervoxels and graph embedding includes:

[0073] One or more processors;

[0074] A storage device storing one or more programs thereon;

[0075] When the one or more programs are executed by the one or more processors, the one or more processors implement the glioma microenvironment feature learning method combining supervoxels and graph embedding.

[0076] The technical effects achieved by the present invention are:

[0077] In the present invention, by fully integrating local features, invasion dynamics, and global spatial structures, a multi-dimensional and comprehensive microenvironment representation is formed, providing a rich information basis for accurate diagnosis and treatment. Through the fine division of the tumor core area, infiltration area, and normal tissue, the biological behavior of the tumor and its impact on the surrounding tissues can be described more accurately, which helps to evaluate invasiveness and prognosis. By using the supervoxel graph embedding technology, the spatial, texture, and morphological information in the image is effectively captured. The construction of this graph structure can more intuitively reflect the complex microenvironment relationships and improve the discriminative ability of the model. By introducing a learning compensation mechanism, the dynamic optimization of the feature learning process is realized, enabling the model to better adapt to the heterogeneity among different patients and the complex changes in the tumor microenvironment, and enhancing the generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1It is the flowchart of the method provided by the present invention. Specific embodiments

[0079] To make the above objects, features and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given in conjunction with the accompanying drawings of the specification.

[0080] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0081] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that may be included in at least one implementation manner of the present invention. The "in a preferred embodiment" that appears in different places in this specification does not all refer to the same embodiment, nor is it a separate or alternative embodiment that excludes other embodiments.

[0082] Thirdly, the present invention is described in detail in conjunction with schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of explanation, the schematic diagrams are only examples and should not limit the scope of protection of the present invention here.

[0083] Please refer to the attached Figure 1 As shown, a method for learning glioma microenvironment features combining supervoxels and graph embedding is provided, including:

[0084] S1. Obtain the multi-modal three-dimensional medical image data of the brain of glioma patients, extract the tumor core area, infiltration area and surrounding normal tissue area according to the multi-modal three-dimensional medical image data of the brain, and obtain the local characteristic information of the glioma microenvironment according to the tumor core area, infiltration area and surrounding normal tissue area;

[0085] S2. Obtain the invasion information of glioma according to the tumor core area and infiltration area, and obtain the microenvironment compensation information according to the invasion information;

[0086] S3. Obtain a set of supervoxel units according to the multi-modal three-dimensional medical image data of the brain. Each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information and spatial coordinate feature information, and obtain supervoxel graph embedding information according to the set of supervoxel units;

[0087] S4. Obtain a graph embedding feature learning strategy according to the local characteristic information, microenvironment compensation information and supervoxel edge graph embedding information, and obtain the glioma microenvironment graph structure according to the graph embedding feature learning strategy;

[0088] S5. Obtain learning compensation information based on the glioma microenvironment map structure, obtain microenvironment learning compensation information based on the learning compensation information, the tumor core area, and the infiltration area, and re-perform the graph embedding feature learning strategy acquisition.

[0089] In the above steps S1 to S5, using the multi-modal three-dimensional medical image data of glioma patients, the tumor area is divided into the tumor core area, the infiltration area, and the surrounding normal tissue area through image segmentation technology. Thus, based on the local characteristic information extracted from these three areas, the invasion characteristics of the tumor are extracted according to the tumor core area and the infiltration area, and then microenvironment compensation information is generated. The three-dimensional medical image data is subdivided into multiple supervoxel units, and multi-dimensional features such as morphology, voxel intensity, texture, and spatial coordinates are extracted from each unit. Through graph embedding technology, the relationships and structural information of these supervoxel units are encoded into a graph structure. The extracted local characteristic information, microenvironment compensation information, and supervoxel graph embedding information are fused. By constructing a graph embedding feature learning strategy, a graph reflecting the overall structure of the glioma microenvironment is generated. Based on the preliminarily constructed glioma microenvironment graph structure, learning compensation information is further extracted, and then combined with the information of the tumor core area and the infiltration area, the graph embedding feature learning strategy is readjusted to form a closed-loop feedback of microenvironment learning compensation information, fully integrating local features, invasion dynamics, and global spatial structure, forming a multi-dimensional and comprehensive microenvironment representation, providing a rich information basis for accurate diagnosis and treatment. Through the fine division of the tumor core area, the infiltration area, and the normal tissue, the biological behavior of the tumor and its impact on the surrounding tissues can be described more accurately, which helps to evaluate invasiveness and prognosis. Using supervoxel graph embedding technology, the spatial, texture, and morphological information in the image is effectively captured. The construction of this graph structure can more intuitively reflect complex microenvironment relationships and improve the discriminative ability of the model. By introducing a learning compensation mechanism, the dynamic optimization of the feature learning process is realized, enabling the model to better adapt to the heterogeneity among different patients and the complex changes in the tumor microenvironment, and enhancing the generalization ability.

[0090] In a preferred embodiment, the steps of obtaining the multi-modal three-dimensional medical image data of the brain of a glioma patient and obtaining the local characteristic information of the glioma microenvironment according to the tumor core area, the infiltration area, and the surrounding normal tissue area include:

[0091] S101. Obtain the multi-modal three-dimensional medical image data of the brain of a glioma patient;

[0092] S102. Extract the tumor core area, the infiltration area, and the surrounding normal tissue area according to the multi-modal three-dimensional medical image data of the brain;

[0093] S103. Obtain a corresponding plurality of tumor core vectors according to the tumor core region;

[0094] S104. Obtain a corresponding plurality of infiltration vectors according to the infiltration region;

[0095] S105. Obtain a corresponding plurality of surrounding normal tissue vectors according to the surrounding normal tissue region;

[0096] S106. Obtain the tumor core weight, infiltration weight, and surrounding normal tissue weight;

[0097] S107. Obtain a local feature vector according to the plurality of tumor core vectors, the plurality of infiltration vectors, the plurality of surrounding normal tissue vectors, the tumor core weight, the infiltration weight, and the surrounding normal tissue weight, and label it as the local feature information of the glioma microenvironment.

[0098] In the above steps S101 to S107, multi-modal three-dimensional medical image data of the brain is obtained from glioma patients. These data usually include multiple imaging modalities such as MRI, CT, or PET. Using image segmentation technology, three key regions are extracted from the multi-modal three-dimensional image data: the tumor core region, the infiltration region, and the surrounding normal tissue region. This process can be achieved through deep learning algorithms (such as U-Net) or traditional image segmentation methods for automatic segmentation. A plurality of tumor core vectors representing the characteristics of the region are extracted from the tumor core region, a plurality of infiltration vectors are extracted from the infiltration region to reflect the state of tumor diffusion to the surrounding tissues, and corresponding vectors are extracted from the surrounding normal tissue region to reflect the normal information of its structure and function. For the vectors in different regions (tumor core, infiltration region, and normal tissue region), corresponding weights are assigned, and the extracted multiple vectors are fused according to the obtained weights to generate a comprehensive local feature vector. The calculation formula for the local feature vector is In the formula, Z represents the local feature vector, i represents the number of the plurality of tumor core vectors, i = 1, 2, 3... n, L i represents the i-th tumor core vector, h represents the number of the plurality of infiltration vectors, h = 1, 2, 3... m, J h represents the h-th infiltration vector, g represents the number of the plurality of surrounding normal tissue vectors, g = 1, 2, 3... k, W g represents the g-th surrounding normal tissue vector, α represents the tumor core weight, β represents the infiltration weight, It is expressed as the weight of the surrounding normal tissue and marked as the local characteristic information of the glioma microenvironment, which can capture information on various aspects such as the structure, function, and metabolism of glioma, improving the comprehensiveness of the microenvironment description. This enables the model to distinguish the pathological states of different regions in detail, thus more accurately reflecting the invasiveness of the tumor and the response of the surrounding environment. By assigning weights to the vectors of different regions, effective integration of information is achieved.

[0099] Example 1 is the first way of the steps to obtain the tumor core weight, infiltration weight, and the weight of the surrounding normal tissue.

[0100] Specifically, the steps to obtain the tumor core weight, infiltration weight, and the weight of the surrounding normal tissue include:

[0101] S1061a. Obtain the tumor core area of the tumor core region, the infiltration area of the infiltration region, and the area of the surrounding normal tissue of the surrounding normal tissue region;

[0102] S1062a. Obtain the tumor core ratio, infiltration ratio, and the ratio of the surrounding normal tissue according to the tumor core area, infiltration area, and the area of the surrounding normal tissue respectively;

[0103] S1063a. Obtain a ratio weight table, where the ratio weight table includes three weight sub - tables, namely the tumor core weight sub - table, the infiltration weight sub - table, and the weight sub - table of the surrounding normal tissue. Each weight sub - table includes corresponding multiple ratio intervals and the weights corresponding to each ratio interval;

[0104] S1064a. Obtain the corresponding tumor core weight, infiltration weight, and the weight of the surrounding normal tissue from the ratio weight table according to the tumor core ratio, infiltration ratio, and the ratio of the surrounding normal tissue respectively.

[0105] As in the above steps S1061a to S1064a, the areas of the tumor core area, the infiltration area and the surrounding normal tissue area are calculated respectively according to the segmented image data, and the area data of each area are used to calculate the proportion of each area (i.e., the tumor core proportion, the infiltration proportion and the surrounding normal tissue proportion), which is usually the area of the area divided by the total area of the overall area. A proportion weight table is pre-designed and constructed, which contains three weight sub-tables, corresponding to the tumor core, the infiltration area and the surrounding normal tissue, respectively. Multiple proportion intervals are set in each sub-table, and a corresponding weight value is defined for each interval. According to the calculated proportion of each area, the matching proportion interval is found from the corresponding proportion weight subtable, and the corresponding weight value is obtained, that is, the tumor core weight, infiltration weight and surrounding normal tissue weight are obtained respectively, which can objectively quantify the proportion of tumor core, infiltration area and normal tissue in the overall structure, avoid relying solely on subjective judgment, and thus enhance the reliability of data-driven analysis. Through the pre-constructed proportion weight table, expert knowledge and clinical experience can be integrated into the model, so that the weight assignment of different regions is more in line with the actual pathological characteristics, providing more meaningful input for subsequent microenvironment feature learning.

[0106] In a preferred embodiment, the step of obtaining the tumor core area of the tumor core region, the infiltration area of the infiltration area, and the surrounding normal tissue area of the surrounding normal tissue region comprises:

[0107] S1061a1, obtaining an overall image of a tumor core area, an infiltrated area, and surrounding normal tissue areas in multimodal three-dimensional medical image data of the brain;

[0108] S1061a2, constructing a plane rectangular coordinate system in the overall image;

[0109] S1061a3, respectively obtaining the coordinates of multiple edge contour inflection points corresponding to the tumor core area, the infiltration area, and the surrounding normal tissue area according to a plane rectangular coordinate system;

[0110] S1061a4. Obtain the corresponding tumor core area, infiltration area, and surrounding normal tissue area respectively according to the corresponding multiple edge contour inflection point coordinates.

[0111] In the above steps S1061a1 to S1061a4, an overall image including the tumor core region, the infiltration region, and the surrounding normal tissue region is extracted from the brain multi-modal three-dimensional medical image data to ensure that all target regions are presented in the same image, providing complete spatial information. A plane rectangular coordinate system is constructed on the overall image to provide a unified geometric reference for each pixel or voxel in the image. Usually, the origin of this coordinate system is the center of the image or a certain fixed point, and the position of each point is described by horizontal and vertical coordinates. Using image processing algorithms (such as edge detection, contour extraction, and inflection point detection algorithms), multiple inflection point coordinates on the edge contours of the tumor core region, the infiltration region, and the surrounding normal tissue region are respectively extracted. These inflection points are usually the key points where the shape of the edge curve changes significantly. According to the extracted multiple edge contour inflection point coordinates, geometric algorithms can be used to calculate the area of the boundary of each region. The corresponding area calculation formula is where S represents the corresponding area, f represents the number of the corresponding multiple edge contour inflection point coordinates, f = 1, 2, 3…t, X f represents the x-axis coordinate point of the f-th edge contour inflection point, X f+1 represents the x-axis coordinate point of the (f + 1)-th edge contour inflection point, Y f represents the y-axis coordinate point of the f-th edge contour inflection point, Y f+1 represents the y-axis coordinate point of the (f + 1)-th edge contour inflection point. When f takes the value of t, t + 1 represents 1. By constructing the coordinate system and extracting the edge inflection points, the boundary characteristics of each region can be accurately captured, making the area calculation more objective and accurate. This refined description helps to quantify the true morphology of the tumor and its surrounding regions. Using geometric algorithms to calculate the area avoids the subjective errors that may be brought by manual segmentation, ensuring the objectivity and repeatability of the data, and providing reliable quantitative indicators for clinical diagnosis and disease condition assessment.

[0112] In a preferred embodiment, the steps of obtaining the invasion information of glioma according to the tumor core region and the infiltration region, and obtaining the microenvironment compensation information according to the invasion information include:

[0113] S201. Obtain the tumor core change vector of the tumor core region;

[0114] S202. Obtain the infiltration change vector of the infiltration region;

[0115] S203. Obtain the invasion value according to the tumor core change vector and the infiltration change vector, and mark it as the invasion information;

[0116] S204. Obtain the compensation table, where the compensation table includes multiple invasion interval values and the microenvironment compensation information corresponding to each invasion interval value;

[0117] S205. Obtain the corresponding microenvironment compensation information from the compensation table according to the invasion value.

[0118] In the above steps S201 to S205, by extracting the internal dynamic change information from the image data of the tumor core area, such as changes in shape, density, texture or signal intensity, a vector reflecting the internal state change of the tumor core is constructed. Similarly, a vector describing the subtle changes in this area (such as the spread of tumor cells to the surrounding healthy tissues, blurred edges, etc.) is extracted from the infiltration area. The tumor core change vector and the infiltration change vector are fused to generate a comprehensive "invasion value". The calculation formula of the invasion value is In the formula, Q represents the invasion value, L change represents the tumor core change vector, J change represents the infiltration change vector. A compensation table is pre-constructed, which corresponds different ranges of invasion values (i.e., multiple invasion interval values) to the corresponding microenvironment compensation information. Specific compensation parameters are set within each interval, and these parameters are often obtained based on clinical experience or large-scale data analysis. According to the previously generated invasion value, the matching invasion interval is searched in the compensation table, and the corresponding microenvironment compensation information of this interval is obtained. By separately extracting the change vectors of the tumor core and the infiltration area and fusing them to generate the invasion value, the invasiveness of the tumor can be quantitatively evaluated, which helps to more accurately reflect the true situation of the spread of the lesion. Using the compensation table to associate the invasion value with the microenvironment compensation information can automatically adjust the microenvironment characteristics according to different tumor invasion levels, thereby compensating for the local information loss or imbalance problems caused by tumor invasion.

[0119] In a preferred embodiment, a supervoxel unit set is obtained according to the brain multi-modal three-dimensional medical image data. Each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information and spatial coordinate feature information. The steps of obtaining the supervoxel map embedding information according to the supervoxel unit set include:

[0120] S301. Obtain a supervoxel unit set according to the brain multi-modal three-dimensional medical image data, where each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information and spatial coordinate feature information;

[0121] S302. Obtain the corresponding morphological feature vector according to the morphological feature information;

[0122] S303. Obtain the corresponding voxel intensity feature vector according to the voxel intensity feature information;

[0123] S304. Obtain the corresponding texture feature vector according to the texture feature information;

[0124] S305. Obtain the corresponding spatial coordinate feature vector according to the spatial coordinate feature information;

[0125] S306. Obtain the graph embedding value according to the morphological feature vector, voxel intensity feature vector, texture feature vector and spatial coordinate feature vector;

[0126] S307. Obtain an embedding graph, where the embedding graph includes multiple graph embedding interval values and the supervoxel graph embedding information corresponding to each graph embedding interval value;

[0127] S308. Obtain the corresponding supervoxel graph embedding information from the embedding graph according to the graph embedding value.

[0128] In the above steps S301 to S308, using the brain multi-modal three-dimensional medical image data, through a pre-designed segmentation algorithm (such as a method based on clustering or region growing), the entire image is divided into several supervoxel units with consistent features. Each supervoxel unit contains information from different dimensions, including: morphological features, which describe geometric information such as structure, shape, and boundary; voxel intensity features, which reflect the signal intensity of different tissues or lesions; texture features, which capture local gray-scale distribution, roughness, patterns, etc.; spatial coordinate features, which determine the position and distribution in three-dimensional space. And according to the information of different dimensions, extract the morphological feature vector, voxel intensity feature vector, texture feature vector and spatial coordinate feature vector, fuse the four feature vectors (morphology, voxel intensity, texture, spatial coordinate) extracted from the supervoxel unit, and calculate the graph embedding value. The calculation formula of the graph embedding value is T = U·D·P·K, where T represents the graph embedding value, U represents the morphological feature vector, D represents the voxel intensity feature vector, P represents the texture feature vector, and K represents the spatial coordinate feature vector. Pre-construct an embedding graph, which contains multiple graph embedding interval values, and each interval value corresponds to a set of predefined supervoxel graph embedding information. This is equivalent to establishing a mapping mechanism. By discretizing the continuous graph embedding value into intervals and associating it with the corresponding embedding information, a lookup table is formed. According to the previously calculated supervoxel graph embedding value, find the corresponding interval from the embedding graph and obtain the predefined supervoxel graph embedding information within that interval. By extracting and fusing multi-dimensional features such as morphology, intensity, texture, and spatial position, the features of the supervoxel unit can be characterized more comprehensively, enabling subsequent graph embedding to capture more detailed and rich local structure information. Supervoxel segmentation helps to reduce the influence of image noise and discontinuity on feature extraction. Combining with a standardized graph embedding table can provide a consistent feature mapping between different data, thereby improving the robustness and accuracy of the overall analysis.

[0129] In a preferred embodiment, a graph embedding feature learning strategy is obtained according to local characteristic information, microenvironment compensation information, and supervoxel edge graph embedding information. The steps of obtaining the glioma microenvironment graph structure according to the graph embedding feature learning strategy include:

[0130] S401. Obtain the corresponding local characteristic vector according to the local characteristic information;

[0131] S402. Obtain the corresponding invasion value according to the microenvironment compensation information;

[0132] S403. Obtain the corresponding graph embedding value according to the supervoxel edge graph embedding information;

[0133] S404. Obtain the policy value according to the local characteristic vector, invasion value, and graph embedding value;

[0134] S405. Obtain a policy learning table, where the policy learning table includes multiple policy interval values and the graph embedding feature learning strategies corresponding to each policy interval value;

[0135] S406. Obtain the corresponding graph embedding feature learning strategy from the policy learning table according to the policy value;

[0136] S407. Obtain the corresponding supervoxel graph construction method according to the graph embedding feature learning strategy;

[0137] S408. Construct the glioma microenvironment graph structure according to the supervoxel graph construction method and in combination with the brain multi-modal three-dimensional medical image data.

[0138] In the above steps S401 to S408, using the local feature information extracted from the brain multi-modal three-dimensional medical images, the invasion value is obtained according to the microenvironment compensation information, and the corresponding graph embedding value is obtained according to the supervoxel edge map embedding information. The previously obtained local feature vector, invasion value, and graph embedding value are fused to calculate the strategy value. The calculation method of the strategy value is C = Z·Q·T, where C represents the strategy value, Z represents the local feature vector, Q represents the invasion value, and T represents the graph embedding value. A strategy learning table is pre-designed, which discretizes the strategy value into multiple intervals, and each interval corresponds to a specific graph embedding feature learning strategy. These strategies are usually determined based on clinical data or large-sample statistical analysis, reflecting the optimal feature extraction methods under different microenvironment states. According to the previously generated strategy value, the matching interval is searched in the strategy learning table, and the corresponding graph embedding feature learning strategy is extracted. According to the obtained graph embedding feature learning strategy, the corresponding supervoxel graph construction method is determined, which may include how to define the edges between supervoxels, how to set the connection weights, and how to use methods such as graph neural networks for feature transfer. Combining the determined supervoxel graph construction method with the brain multi-modal three-dimensional medical image data, a graph structure reflecting the glioma microenvironment is constructed. This graph structure uses each supervoxel unit as a node and uses edges to connect to express the relationship between them, forming a graph model that can effectively capture the complex interaction relationships between the tumor and its surrounding tissues. By integrating local features, invasion situations, and graph embedding information, and using the strategy learning table for adaptive strategy selection, the entire method can dynamically adjust the feature learning process to cope with the complex changes in different patients and different pathological states, integrating local details (such as morphological and texture features), global structures (through graph embedding values), and tumor invasion information, so that the finally constructed graph structure can comprehensively and accurately depict the glioma microenvironment.

[0139] In a preferred embodiment, learning compensation information is obtained according to the glioma microenvironment graph structure, and microenvironment learning compensation information is obtained according to the learning compensation information, the tumor core area, and the infiltration area, and the steps of obtaining the graph embedding feature learning strategy are re-performed, including:

[0140] S501. Obtain learning compensation information according to the glioma microenvironment graph structure;

[0141] S502. Obtain the corresponding learning compensation value according to the learning compensation information;

[0142] S503. Obtain the tumor core change vector of the tumor core area;

[0143] S504. Obtain the infiltration change vector of the infiltration area;

[0144] S505. Obtain the invasion compensation value according to the learning compensation value, the tumor core area, and the infiltration area;

[0145] S506. Obtain a learning compensation table, where the learning compensation table includes multiple invasion compensation interval values and the corresponding microenvironment learning compensation information for each invasion compensation interval value;

[0146] S507. Obtain the corresponding microenvironment learning compensation information from the learning compensation table according to the invasion compensation value;

[0147] S508. Use the microenvironment learning compensation information as the microenvironment compensation information and re - obtain the graph embedding feature learning strategy.

[0148] In the above steps S501 to S508, learning compensation information is extracted from the brain glioma microenvironment graph structure and quantified into a learning compensation value. Combining the change vectors of the tumor core area, the infiltration area, and the learning compensation value, the invasion compensation value is calculated. The calculation formula of the invasion compensation value is In the formula, E represents the invasion compensation value, B represents the learning compensation value, L change represents the tumor core change vector, J change represents the infiltration change vector. A learning compensation table is pre - constructed. This table stores multiple invasion compensation interval values and provides corresponding microenvironment learning compensation information for each interval. By querying the table, the learning compensation strategy applicable to the current invasion compensation value is determined, the best microenvironment learning compensation information is selected, and it is used as the microenvironment compensation information to optimize the subsequent graph embedding feature learning strategy, realizing adaptive learning and adjustment for the brain glioma microenvironment. Through the dynamic adjustment of the learning compensation information, the deviation in feature learning is reduced, the ability to finely depict the brain glioma microenvironment is improved. Combining real - time learning compensation, it can adapt to the invasion characteristics of different lesion areas, improve the generalization ability. By constructing the learning compensation table, a closed - loop optimization mechanism is formed, making the graph embedding feature learning strategy more accurate and enhancing the representation ability of the brain glioma microenvironment.

[0149] In a preferred embodiment, the steps of obtaining learning compensation information according to the brain glioma microenvironment graph structure include:

[0150] S5011. Obtain the brain glioma microenvironment vector according to the brain glioma microenvironment graph structure;

[0151] S5012. Obtain the actual brain glioma microenvironment vector of the brain glioma patient;

[0152] S5013. Obtain the microenvironment deviation value according to the brain glioma microenvironment vector and the actual brain glioma microenvironment vector;

[0153] S5014. Obtain the learning microenvironment compensation table, where the microenvironment learning compensation table includes multiple microenvironment deviation interval values and the corresponding learning compensation information for each microenvironment deviation interval value;

[0154] S5015. Obtain the corresponding learning compensation information from the microenvironment compensation table according to the microenvironment deviation value.

[0155] In the above steps S5011 to S5015, extract the glioma microenvironment vector from the glioma microenvironment map structure. This vector is used to characterize key information such as the spatial distribution and metabolic characteristics of the tumor. Obtain the actual glioma microenvironment vector of the patient, which is derived from real pathological data or imaging analysis results and reflects the current patient's tumor physiological state. Calculate the microenvironment deviation value. The calculation formula for the microenvironment deviation value is In the formula, N deviation represents the microenvironment deviation value, N now represents the glioma microenvironment vector, N reality represents the difference between the glioma actual microenvironment vector, that is, the glioma microenvironment vector and the actual microenvironment vector. This value is used to measure the deviation between the system prediction and the real situation. Pre-construct a microenvironment learning compensation table, which stores multiple microenvironment deviation interval values and assigns corresponding learning compensation information to each interval. By querying the compensation table, determine the learning compensation information applicable to the current microenvironment deviation value. By calculating the microenvironment deviation value, dynamically adjust the learning strategy, improve the ability to depict the glioma microenvironment characteristics, make the model more adaptable to individual patients, combine the learning compensation information, reduce the impact of noise data or abnormal features on the model, improve the robustness of the prediction result, be able to more accurately capture the glioma microenvironment characteristics, provide high-value data support for imaging analysis, tumor grading, treatment decision-making, etc., and enhance the application potential of artificial intelligence in tumor diagnosis.

[0156] Example 2. This example is a further improvement based on Example 1, and the difference from Example 1 is the different ways of obtaining the tumor core weight, infiltration weight, and surrounding normal tissue weight.

[0157] The steps of obtaining the tumor core weight, infiltration weight, and surrounding normal tissue weight include:

[0158] S106b1. Obtain the tumor core area of the tumor core region, the infiltration area of the infiltration region, and the surrounding normal tissue area of the surrounding normal tissue region;

[0159] S106b2. Define weight coefficients based on clinical priorities, namely the tumor core coefficient, infiltration coefficient, and surrounding normal tissue coefficient. Among them, the tumor core coefficient is greater than the infiltration coefficient, and the infiltration coefficient is greater than the surrounding normal tissue coefficient;

[0160] S106b3. Multiply the area of each region by its corresponding weight coefficient to obtain the weighted area;

[0161] S106b4. Assign tumor core weight, infiltration weight, and surrounding normal tissue weight according to the proportion of the weighted area to the total weighted area.

[0162] In the above steps S1061b1 to S1061b4, according to the segmented image data, calculate the areas of the tumor core region, infiltration region, and surrounding normal tissue region respectively (the area acquisition method of steps S1061a1 to S1061a4 can be adopted). According to clinical medical experience, clarify the importance differences of different regions for treatment or prognosis (for example, the tumor core region is usually more critical for treatment response), and directly assign weight coefficients to the three regions (for example, core coefficient = 0.6, infiltration coefficient = 0.3, surrounding coefficient = 0.1). The size of the coefficient reflects the clinical priority. Multiply the original area of each region by its weight coefficient to obtain the "weighted area". For example, if the core region has a large area and a high coefficient, its weighted area will be significantly higher than other regions. Through the formula weight = regional weighted area / total weighted area (core weighted area + infiltration weighted area + surrounding weighted area), convert the weighted area into a proportional weight, which not only retains the clinical priority but also balances the actual area differences of different regions. There is no need to pre-define a static "proportion interval table", and different clinical scenarios (such as different cancer types, different treatment goals) can be adapted by adjusting the weight coefficients. The weight coefficients directly reflect medical experience, avoiding underestimating key regions (such as a small core region with great harm) due to simply relying on area ratios, avoiding complex look-up table matching and interval judgment, and only requiring multiplication and normalization calculations, reducing the implementation complexity and the risk of human error. The weight coefficients can be dynamically adjusted based on expert consensus and clinical trial data, and the results have clear mathematical and medical interpretability.

[0163] And, a glioma microenvironment feature learning terminal combining supervoxels and graph embedding, comprising:

[0164] One or more processors;

[0165] A storage device storing one or more programs thereon;

[0166] When the one or more programs are executed by the one or more processors, the one or more processors implement the glioma microenvironment feature learning method combining supervoxels and graph embedding.

[0167] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention are implemented according to the conventional means in the art without special description and limitation.

Claims

1. A method for learning glioma microenvironment features by combining supervoxels and graph embedding, characterized in that Including: Obtain the multi-modal three-dimensional medical image data of the brain of a glioma patient, extract the tumor core area, infiltration area, and surrounding normal tissue area from the multi-modal three-dimensional medical image data of the brain, and obtain the local characteristic information of the glioma microenvironment according to the tumor core area, infiltration area, and surrounding normal tissue area; Obtain the invasion information of glioma according to the tumor core area and infiltration area, and obtain the microenvironment compensation information according to the invasion information; Obtain a supervoxel unit set according to the multi-modal three-dimensional medical image data of the brain. Each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information, and spatial coordinate feature information, and obtain supervoxel graph embedding information according to the supervoxel unit set; Obtain a graph embedding feature learning strategy according to the local characteristic information, microenvironment compensation information, and supervoxel edge graph embedding information, and obtain the glioma microenvironment graph structure according to the graph embedding feature learning strategy; Obtain learning compensation information according to the glioma microenvironment graph structure, obtain microenvironment learning compensation information according to the learning compensation information, tumor core area, and infiltration area, and re-obtain the graph embedding feature learning strategy.

2. The glioma microenvironment feature learning method combining supervoxels and graph embedding according to claim 1, wherein, The steps of obtaining the multi-modal three-dimensional medical image data of the brain of a glioma patient, extracting the tumor core area, infiltration area, and surrounding normal tissue area from the multi-modal three-dimensional medical image data of the brain, and obtaining the local characteristic information of the glioma microenvironment according to the tumor core area, infiltration area, and surrounding normal tissue area include: Obtain the multi-modal three-dimensional medical image data of the brain of a glioma patient; Extract the tumor core area, infiltration area, and surrounding normal tissue area from the multi-modal three-dimensional medical image data of the brain; Obtain a corresponding plurality of tumor core vectors according to the tumor core area; Obtain a corresponding plurality of infiltration vectors according to the infiltration area; Obtain a corresponding plurality of surrounding normal tissue vectors according to the surrounding normal tissue area; Obtain the tumor core weight, infiltration weight, and surrounding normal tissue weight; Obtain a local characteristic vector according to the plurality of tumor core vectors, plurality of infiltration vectors, plurality of surrounding normal tissue vectors, tumor core weight, infiltration weight, and surrounding normal tissue weight, and label it as the local characteristic information of the glioma microenvironment.

3. The glioma microenvironment feature learning method combining supervoxels and graph embedding according to claim 2, wherein The steps of obtaining the tumor core weight, infiltration weight, and surrounding normal tissue weight include: Obtain the tumor core area of the tumor core area, the infiltration area of the infiltration area, and the surrounding normal tissue area of the surrounding normal tissue area; Obtain the tumor core ratio, infiltration ratio, and surrounding normal tissue ratio according to the tumor core area, infiltration area, and surrounding normal tissue area respectively; Obtain a ratio weight table, where the ratio weight table includes three weight sub-tables, which are the tumor core weight sub-table, infiltration weight sub-table, and surrounding normal tissue weight sub-table respectively. Each weight sub-table includes corresponding multiple ratio intervals and the weights corresponding to each ratio interval; Obtain the corresponding tumor core weight, infiltration weight, and surrounding normal tissue weight from the ratio weight table according to the tumor core ratio, infiltration ratio, and surrounding normal tissue ratio respectively.

4. The method for learning glioma microenvironment features by combining supervoxels and graph embedding according to claim 3, wherein The steps for obtaining the tumor core area of the tumor core region, the infiltration area of the infiltration region, and the surrounding normal tissue area of the surrounding normal tissue region include: Obtain the overall image including the tumor core region, the infiltration region, and the surrounding normal tissue region from the brain multi-modal three-dimensional medical image data; Construct a plane rectangular coordinate system in the overall image; Obtain the corresponding multiple edge contour inflection point coordinates of the tumor core region, the infiltration region, and the surrounding normal tissue region respectively according to the plane rectangular coordinate system; Obtain the corresponding tumor core area, infiltration area, and surrounding normal tissue area respectively according to the corresponding multiple edge contour inflection point coordinates.

5. The method for learning glioma microenvironment features combining supervoxels and graph embedding according to claim 1, characterized in that The steps for obtaining the invasion information of glioma according to the tumor core region and the infiltration region, and obtaining the microenvironment compensation information according to the invasion information include: Obtain the tumor core change vector of the tumor core region; Obtain the infiltration change vector of the infiltration region; Obtain the invasion value according to the tumor core change vector and the infiltration change vector, and mark it as the invasion information; Obtain the compensation table, where the compensation table includes multiple invasion interval values and the corresponding microenvironment compensation information for each invasion interval value; Obtain the corresponding microenvironment compensation information from the compensation table according to the invasion value.

6. The method for learning glioma microenvironment features by combining supervoxels and graph embedding according to claim 1, wherein Obtain the supervoxel unit set according to the brain multi-modal three-dimensional medical image data, where each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information, and spatial coordinate feature information. The steps for obtaining the supervoxel graph embedding information according to the supervoxel unit set include: Obtain the supervoxel unit set according to the brain multi-modal three-dimensional medical image data, where each supervoxel unit includes morphological feature information, voxel intensity feature information, texture feature information, and spatial coordinate feature information; Obtain the corresponding morphological feature vector according to the morphological feature information; Obtain the corresponding voxel intensity feature vector according to the voxel intensity feature information; Obtain the corresponding texture feature vector according to the texture feature information; Obtain the corresponding spatial coordinate feature vector according to the spatial coordinate feature information; Obtain the graph embedding value according to the morphological feature vector, the voxel intensity feature vector, the texture feature vector, and the spatial coordinate feature vector; Obtain the embedding graph table, where the embedding graph table includes multiple graph embedding interval values and the corresponding supervoxel graph embedding information for each graph embedding interval value; Obtain the corresponding supervoxel graph embedding information from the embedding graph table according to the graph embedding value.

7. The glioma microenvironment feature learning method combining supervoxels and graph embedding according to claim 1, characterized in that The steps for obtaining the graph embedding feature learning strategy according to the local feature information, the microenvironment compensation information, and the supervoxel edge graph embedding information, and obtaining the glioma microenvironment graph structure according to the graph embedding feature learning strategy include: Obtain the corresponding local feature vector according to the local feature information; Obtain the corresponding invasion value according to the microenvironment compensation information; Obtain the corresponding graph embedding value according to the supervoxel edge graph embedding information; Obtain the strategy value according to the local feature vector, the invasion value, and the graph embedding value; Obtain the strategy learning table, where the strategy learning table includes multiple strategy interval values and the corresponding graph embedding feature learning strategy for each strategy interval value; Obtain the corresponding graph embedding feature learning strategy from the strategy learning table according to the strategy value; Obtain the corresponding supervoxel graph construction method according to the graph embedding feature learning strategy; Construct a glioma microenvironment map structure according to the supervoxel map construction method and in combination with brain multi-modal three-dimensional medical image data.

8. The glioma microenvironment feature learning method combining supervoxels and graph embedding according to claim 1, characterized in that Obtain learning compensation information according to the glioma microenvironment map structure, obtain microenvironment learning compensation information according to the learning compensation information, the tumor core area, and the infiltration area, and re-perform the steps of obtaining the graph embedding feature learning strategy, including: Obtain learning compensation information according to the glioma microenvironment map structure; Obtain the corresponding learning compensation value according to the learning compensation information; Obtain the tumor core change vector of the tumor core area; Obtain the infiltration change vector of the infiltration area; Obtain the invasion compensation value according to the learning compensation value, the tumor core area, and the infiltration area; Obtain a learning compensation table, where the learning compensation table includes multiple invasion compensation interval values and the microenvironment learning compensation information corresponding to each invasion compensation interval value; Obtain the corresponding microenvironment learning compensation information from the learning compensation table according to the invasion compensation value; Use the microenvironment learning compensation information as the microenvironment compensation information and re-perform the graph embedding feature learning strategy.

9. The glioma microenvironment feature learning method combining supervoxels and graph embedding according to claim 8, wherein The steps of obtaining learning compensation information according to the glioma microenvironment map structure include: Obtain the glioma microenvironment vector according to the glioma microenvironment map structure; Obtain the actual glioma microenvironment vector of the glioma patient; Obtain the microenvironment deviation value according to the glioma microenvironment vector and the actual glioma microenvironment vector; Obtain a learning microenvironment compensation table, where the microenvironment learning compensation table includes multiple microenvironment deviation interval values and the learning compensation information corresponding to each microenvironment deviation interval value; Obtain the corresponding learning compensation information from the microenvironment compensation table according to the microenvironment deviation value.

10. A glioma microenvironment feature learning terminal combining supervoxels and graph embedding, characterized in that Include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the glioma microenvironment feature learning method combining supervoxels and graph embedding according to any one of claims 1 to 9.

Images