Tumor heterogeneity analysis system based on multi-modal characteristics

Abstract

The invention relates to the technical field of machine learning, and discloses a tumor heterogeneity analysis system based on multi-modal features, which comprises a data acquisition module, a feature preprocessing module, a feature fusion module, a clustering analysis module, a heterogeneity analysis module and a result visualization module. In the clustering analysis, a double-approximation graph structure combining a core approximation graph and a global approximation graph is constructed by using a self-adaptive neighborhood ratio and a mutual reachable distance, and in-cluster branch analysis is realized by combining eccentricity weighted hierarchical clustering, so that the robustness and the recognition precision under the conditions of variable density and complex structures are improved; in heterogeneity analysis, a feature selection kernel SVM model is introduced, redundant feature interference is suppressed, key feature contribution degrees are highlighted, and a clear discrimination boundary and a feature evaluation result are obtained; according to the method, robustness, precision and interpretability are considered under the condition of multi-modal data, and an efficient and reliable technical means is provided for tumor heterogeneity analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and particularly relates to a tumor heterogeneity analysis system based on multi-modal features. BACKGROUND

[0002] As a highly complex disease, tumor has significant heterogeneity, which is manifested in the differences in molecular characteristics, imaging features and clinical indicators between different patients, different parts of the same patient, and even the same tumor sample. Such heterogeneity directly affects the reliability of the analysis results of tumor characteristics. In order to more comprehensively depict tumor heterogeneity, current researches tend to combine medical image data, genomic data and clinical indicator data to conduct comprehensive analysis of multi-modal features. However, the existing tumor heterogeneity analysis methods still have the following shortcomings: firstly, in the process of multi-modal data fusion, the dimensions, distribution densities and noise levels of different modalities are quite different. Traditional methods often rely on fixed distance metrics in clustering analysis, which leads to incorrect division under the condition of variable density and complex structure data, thereby reducing the accuracy of subpopulation identification. Secondly, tumor heterogeneity usually manifests as a bifurcated structure of potential subpopulations. Existing methods lack fine-grained hierarchical modeling when mining fine-grained branch relationships, and are difficult to reveal the potential typing characteristics of tumor subpopulations. In addition, when using machine learning methods to analyze tumor heterogeneity, the existence of redundant or irrelevant features can mask the contribution of key features, reducing the explainability and accuracy of the analysis results. Although traditional support vector machines can perform nonlinear classification, they lack effective feature selection mechanisms and are difficult to accurately identify the core factors that determine tumor subpopulation differentiation in high-dimensional multi-modal feature space, which easily causes the discrimination boundary to be blurred. SUMMARY

[0003] The present application aims to overcome the problems of insufficient clustering robustness, limited subpopulation identification accuracy and unclear key feature contribution evaluation in existing tumor heterogeneity analysis methods, and proposes a tumor heterogeneity analysis system based on multi-modal features to improve the stability, accuracy and interpretability of analysis under multi-modal data fusion. The innovation lies in the proposed analysis method combining double approximate graph-eccentricity hierarchical clustering and feature selection kernel SVM model: in the clustering stage, by introducing adaptive neighborhood ratio and mutual reachable distance, a double approximate graph structure combining core approximate graph and global approximate graph is constructed, and the eccentricity weighted hierarchical clustering mechanism is used to realize the fine-grained analysis of intra-cluster branches, thereby significantly improving the clustering robustness under variable density and neck structure data conditions, and more accurately revealing the tumor subpopulation and its internal potential bifurcation structure; in the heterogeneity analysis stage, by embedding binary feature selection variables in the kernel function and combining the nonlinear mapping of polynomial kernel and Gaussian kernel, the interference of redundant features is effectively suppressed, the contribution of key features to subpopulation differentiation is highlighted, and then clear discrimination boundary and feature contribution evaluation results are output, enhancing the interpretability and accuracy of the analysis results.

[0004] The present application provides a tumor heterogeneity analysis system based on multi-modal features, which comprises: a data acquisition module, a feature processing module, a feature extraction and fusion module, a clustering analysis module, a heterogeneity analysis module and a result visualization and reporting module

[0005] The data acquisition module acquires multi-modal data of tumor patients, including medical image data, genomic data and clinical indicator data.

[0006] The feature processing module performs missing value filling, normalization, noise removal and feature alignment processing on the multi-modal data to generate a multi-modal feature matrix.

[0007] The feature extraction and fusion module extracts image features, gene features and clinical features from the multi-modal feature matrix and performs feature fusion to generate a multi-modal fusion feature vector.

[0008] The clustering analysis module analyzes the multi-modal fusion feature vector by a double approximate graph-eccentricity hierarchical clustering algorithm, identifies potential tumor subpopulations, and generates sample clustering labels and confidence results; the sample clustering labels include tumor subpopulation labels and intra-cluster branch labels.

[0009] The heterogeneity analysis module processes the sample clustering labels and the multi-modal fusion feature vector by a feature selection kernel SVM heterogeneity analysis model to obtain tumor heterogeneity analysis results.

[0010] The result visualization and reporting module generates a typing distribution graph, a feature contribution graph and a typing report from the tumor heterogeneity analysis results to generate an analysis report.

[0011] Further, the process of generating sample clustering labels by using a double approximation graph-eccentricity hierarchical clustering algorithm specifically includes the following steps:

[0012] Step S1: Calculate the kernel distance and mutual reachable distance of the multi-modal fusion feature vector, and construct a cluster approximation graph according to the kernel distance and the mutual reachable distance. In the process of constructing the cluster approximation graph, an adaptive neighborhood ratio is introduced to optimize the edge construction mechanism and structure representation ability of the cluster approximation graph, thereby improving the clustering robustness and accuracy under the condition of variable density and neck structure data, and obtaining a double approximation graph; the double approximation graph includes a core approximation graph and a global approximation graph;

[0013] Step S2: Perform single-link hierarchical clustering using the double approximation graph to generate a clustering hierarchical tree describing the hierarchical relationship between samples; set a minimum cluster size parameter to condense the clustering hierarchical tree, thereby removing sub-clusters with insufficient size and compressing redundant branches, obtaining an initial cluster set and cluster membership probability;

[0014] Step S3: For each cluster in the initial cluster set, calculate the weighted centroid of the cluster respectively; and define an eccentricity function based on the distance between the sample points in the initial cluster set and the weighted centroid; calculate the eccentricity weight based on the eccentricity function, and assign the eccentricity weight to the edges of the double approximation graph to obtain a weighted double approximation graph; based on the weighted double approximation graph, a single-link hierarchical clustering method is used to construct an eccentricity hierarchical tree; preset a minimum branch size parameter and a branch duration threshold to condense the eccentricity hierarchical tree and extract an intra-cluster branch set;

[0015] Step S4: For the intra-cluster branch set, determine the branch root of each branch; and calculate the geodesic distance from each sample point in the intra-cluster branch set to the corresponding branch root based on the weighted double approximation graph; define a similarity function according to the geodesic distance, calculate the similarity, and convert the similarity into branch membership probability by introducing a temperature parameter softmax;

[0016] Step S5: Fuse the cluster membership probability and the branch membership probability to generate sample clustering labels and confidence results; the sample clustering labels include tumor subpopulation labels and intra-cluster branch labels.

[0017] Further, the process of obtaining tumor heterogeneity analysis results by using a feature selection kernel SVM heterogeneity analysis model specifically includes the following steps:

[0018] Step B1: Establish a sample labeling data set according to the sample clustering labels and the multi-modal fusion feature vector;

[0019] Step B2: introducing an embedded feature selection mechanism, constructing a binary selection variable-nonlinear kernel function, performing feature selection on the sample labeled data set, and obtaining a feature subset after selection; the binary selection variable-nonlinear kernel function includes a feature selection polynomial kernel and a feature selection Gaussian kernel;

[0020] Step B3: constructing a kernel SVM dual optimization model with feature selection constraints, training based on the feature subset after selection, solving the optimal solution of the kernel SVM dual optimization model with feature selection constraints, and obtaining an optimal Lagrange multiplier parameter set;

[0021] Step B4: using the optimal Lagrange multiplier parameter set to construct a discriminant function and calculate the proximity value of each sample relative to the discriminant hyperplane; wherein the proximity value is used to quantify the distribution characteristics and differences of the samples in the high-dimensional kernel space, so as to realize fine differentiation and heterogeneity description of the potential subgroups of tumor samples;

[0022] Step B5: combining the sample clustering label and the proximity value to obtain tumor heterogeneity analysis results, including the discriminant boundary of the potential subgroups, the contribution evaluation of the key features, and the typing information.

[0023] By using the above scheme, the application has the following beneficial effects:

[0024] Firstly, by introducing adaptive neighborhood ratio and mutual reachable distance, the application constructs a double approximation graph structure combining core approximation graph and global approximation graph, and realizes accurate modeling of complex topological relationship in multi-modal fusion data; this mechanism effectively solves the problem of false edges and false segmentation of existing methods under variable density and neck structure data conditions, improves the robustness and stability of the clustering process, and enables the application to accurately identify potential subgroups in a multi-source, multi-dimensional tumor data environment; through this technical scheme, the universality and reliability of the application in complex sample distribution conditions are further enhanced.

[0025] Secondly, the application combines an eccentricity weighted hierarchical clustering mechanism to analyze the intra-cluster branches in a fine-grained manner, and realizes high-precision revelation of the potential bifurcation structure within the tumor subgroups; this improvement not only improves the accuracy of subgroup identification, but also solves the problem of insufficient heterogeneity description of traditional methods, enabling the application to more clearly show the hierarchical relationship within the tumor population; through this mechanism, the application enhances the refinement and scientificity of subgroup division in a complex tumor heterogeneity environment, providing researchers with more valuable clustering results.

[0026] Finally, the application proposes a feature selection kernel SVM model in the heterogeneity analysis stage, realizes effective suppression of redundant features and highlighting of key features by embedding binary feature selection variables in the kernel function and combining a polynomial kernel with a Gaussian kernel, and improves the clarity of the discriminant boundary; the design solves the problems of unclear key feature contribution and insufficient interpretability in existing methods, so that the application can output a feature contribution degree evaluation result with credibility in a multi-modal feature space; through the improvement, the application enhances the explanation ability of tumor heterogeneity features, and provides stable, accurate and interpretable support for subsequent analysis and application. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A module schematic diagram of a tumor heterogeneity analysis system based on multi-modal features provided by the application;

[0028] Figure 2 A breast cancer patient typing distribution diagram proposed in embodiment six;

[0029] Figure 3 A key feature contribution degree diagram in embodiment six.

[0030] Figure 2 In the figure, there are three parts, namely Luminal A, HER2 positive and Basal-like;

[0031] Figure 3 In the figure, the X-axis label: feature name (including HER2 expression level, ER expression level, Ki-67 proliferation index, image texture energy value, clinical stage); Y-axis: contribution degree (0-0.5). DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application; based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0033] Embodiment one, according to Figure 1 The application provides a tumor heterogeneity analysis system based on multi-modal features, which comprises a data acquisition module, a feature processing module, a feature extraction and fusion module, a clustering analysis module, a heterogeneity analysis module and a result visualization and reporting module.

[0034] The data acquisition module acquires multi-modal data of tumor patients, including medical image data, genomic data and clinical index data.

[0035] The feature processing module performs missing value filling, normalization, noise removal and feature alignment processing on the multi-modal data to generate a multi-modal feature matrix;

[0036] The feature extraction and fusion module extracts image features, gene features and clinical features from the multi-modal feature matrix, and performs feature fusion to generate a multi-modal fusion feature vector;

[0037] The clustering analysis module analyzes the multi-modal fusion feature vector by a double-approximation graph-eccentricity hierarchical clustering algorithm to identify potential tumor subgroups and generate sample clustering labels and confidence results; the sample clustering labels include tumor subgroup labels and intra-cluster branch labels; the double-approximation graph-eccentricity hierarchical clustering algorithm is constructed by introducing an adaptive neighborhood ratio and a mutual reachable distance to establish a double-approximation graph structure, and combining eccentricity weighting and hierarchical clustering mechanisms;

[0038] The heterogeneity analysis module processes the sample clustering labels and the multi-modal fusion feature vector by a feature selection kernel SVM heterogeneity analysis model to obtain tumor heterogeneity analysis results; the tumor heterogeneity analysis results include discriminant boundaries of potential subgroups, contribution degree evaluation of key features and typing information;

[0039] The result visualization and report module generates typing distribution graphs, feature contribution degree graphs and typing reports based on the tumor heterogeneity analysis results to generate an analysis report.

[0040] In this embodiment, a double-approximation graph-eccentricity hierarchical clustering algorithm is used to generate sample clustering labels, and the process specifically includes the following steps:

[0041] Step S1: Calculate the kernel distance and mutual reachable distance of the multi-modal fusion feature vector, and construct a cluster approximation graph based on the kernel distance and mutual reachable distance; in the cluster approximation graph construction process, an adaptive neighborhood ratio is introduced to optimize the edge construction mechanism and structure representation ability of the cluster approximation graph, thereby improving the clustering robustness and accuracy under variable density and neck structure data conditions, and obtaining a double-approximation graph; the double-approximation graph includes a core approximation graph and a global approximation graph; the core approximation graph is used to describe the skeleton structure of the cluster based on local kernel distance constraints to improve the calculation efficiency; the global approximation graph is used to introduce an additional edge set under the constraint of the maximum edge weight of the minimum spanning tree to enhance the representation ability of the cluster internal structure; through the double-approximation graph mechanism, the accuracy of cluster internal structure description and the clustering efficiency are balanced, and the formula is as follows:

[0042] Adaptive neighborhood ratio formula:

[0043] , ;

[0044] Wherein, represents a sample point in the dataset, i.e., a single data point corresponding to the multi-modal fusion feature vector; represents a sample point ; represents a sample point ; represents the Euclidean distance between the sample point and the nearest neighbor sample ; represents the local average neighborhood scale of the sample point ; represents the number of nearest neighbors used for neighborhood calculation, controlling the range of the local neighborhood; represents the local average neighborhood distance of the th nearest neighbor of

[0045] ; By introducing this ratio, the existence of an edge no longer depends only on the absolute distance, but also incorporates a local density adaptive constraint, thereby effectively avoiding false connections or incorrect segmentation problems under conditions of variable density data and neck structure data.

[0046] Core approximation graph edge set formula:

[0047] ;

[0048] wherein, represents two sample points in the global dataset, derived from multi-modal fusion feature vectors, represents the mutual reachable distance of the sample point , used to characterize the proximity of point pairs under local density constraints; represents the local kernel scale parameter of the sample point , calculated from the kernel distance, used to limit the edge connection range; represents the maximum value of the local scales of the two points, serving as a constraint threshold; represents an indicator function, taking the value 1 if the edge satisfies the adaptive neighborhood ratio constraint, otherwise 0; represents the edge set in the core approximation graph, i.e., the point pairs that satisfy the density constraint and AND filtering;

[0049] Explanation: The core graph forms the skeleton structure of the cluster through mutual reachable distance and local kernel scale constraints, and introduces adaptive neighborhood ratio (AND) filtering to remove noise and abnormal edges.

[0050] Global approximation graph edge set formula:

[0051] ;

[0052] wherein, represents the cluster The maximum weight of all edges in the minimum spanning tree (MST) is used to control the longest edge allowed in the global graph; represents the set of edges in the global approximate graph;

[0053] Note: Under the constraints of the MST skeleton, the global graph introduces an additional set of edges in a controlled manner to enhance the internal structure representation of the cluster; AND also acts as an adaptive filter to avoid incorrect edge connections across density regions or neck regions;

[0054] Step S2: Perform single-link hierarchical clustering using a biproximate graph to generate a clustering tree that describes the hierarchical relationships between samples. Set a minimum cluster size parameter to condense the clustering tree, thereby removing subclusters of insufficient size and compressing redundant branches to obtain an initial set of clusters and cluster membership probabilities. This clustering does not require a preset number of clusters and can adaptively identify heterogeneous sample groups and effectively eliminate noise points.

[0055] Step S3: For each cluster in the initial cluster set, the weighted centroid of the cluster is calculated; an eccentricity function is defined based on the distance between the sample points within the cluster and the weighted centroid in the initial cluster set; eccentricity weights are calculated based on the eccentricity function, and eccentricity weights are assigned to the edges of the bi-approximate graph to obtain a weighted bi-approximate graph; based on the weighted bi-approximate graph, an eccentricity hierarchical tree is constructed using a single-link hierarchical clustering method; a minimum branch size parameter and a branch persistence threshold are preset, and the eccentricity hierarchical tree is condensed to extract the intra-cluster branch set to reveal the potential tumor subpopulation bifurcation structure. The formula used is as follows:

[0056] The eccentricity formula is as follows:

[0057] ;

[0058] in, represents a single cluster in the initial set of clusters, Represents a cluster The first sample points; express The weighted average vector of all member samples obtained by weight is the weighted centroid of the cluster; represents a distance metric; Indicates eccentricity;

[0059] Step S4: For the branch set within the cluster, determine the branch root of each branch; and calculate the geodesic distance from each sample point in the cluster to the corresponding branch root based on the weighted bi-approximate graph; define a similarity function based on the geodesic distance, calculate the similarity, and convert the similarity into branch membership probability by introducing the softmax of the temperature parameter, so as to realize the probabilistic distribution of samples among multiple branches. The formula used is as follows:

[0060] ;

[0061] wherein, denotes a specific branch within a cluster denotes a specific branch within a cluster denotes a sample point belongs to a branch soft membership probability, denotes a temperature parameter, denotes a specific branch within a cluster denotes a specific branch within a cluster traverses i.e. covers all branches within the cluster; denotes a sample point denotes a specific branch within a cluster denotes a specific branch within a cluster denotes a similarity function value of a sample point denotes a similarity function value of a sample point denotes a specific branch within a cluster denotes a specific branch within a cluster denotes a similarity function value of a sample point

[0062] Step S5: fuse the cluster membership probability and the branch membership probability to generate a sample clustering label and a confidence result; the sample clustering label comprises a tumor subpopulation label and a branch label within a cluster.

[0063] In Example Three, which is based on Example One, the process of generating a sample clustering label specifically comprises the following steps:

[0064] Step R1: calculate the kernel distance and the mutual reachable distance of the multi-modal fusion feature vector, and construct a cluster approximation graph according to the kernel distance and the mutual reachable distance; the cluster approximation graph comprises a core approximation graph and a global approximation graph;

[0065] Step R2: perform single-link hierarchical clustering using the cluster approximation graph to generate a clustering hierarchical tree describing the hierarchical relationship between samples; set a minimum cluster size parameter, condense the clustering hierarchical tree, thereby removing sub-clusters with insufficient size and compressing redundant branches, to obtain an initial cluster set and a cluster membership probability;

[0066] Step R3: for each cluster in the initial cluster set, calculate the weighted centroid of the cluster respectively; and define a eccentricity function based on the distance between the sample points within the cluster in the initial cluster set and the weighted centroid; calculate the eccentricity weight based on the eccentricity function, assign the eccentricity weight to the edges of the cluster approximation graph, obtain the weighted cluster approximation graph; based on the weighted cluster approximation graph, construct an eccentricity hierarchical tree using single-link hierarchical clustering; preset a minimum branch size parameter and an eccentricity threshold, condense the eccentricity hierarchical tree, extract a cluster branch set to reveal the potential tumor subpopulation bifurcation structure;

[0067] Step R4: determining branch roots of each branch in the intra-cluster branch set, and calculating geodesic distances from each sample point in the intra-cluster branch set to the corresponding branch root based on the weighted approximate graph of the cluster; defining a similarity function according to the geodesic distances, calculating the similarity, and converting the similarity into branch membership probability by introducing a temperature parameter of softmax, so as to realize the probabilistic distribution of the sample between multiple branches;

[0068] Step R5: fusing the cluster membership probability and the branch membership probability to generate a sample clustering label and a confidence result; the sample clustering label comprises a tumor subpopulation label and an intra-cluster branch label.

[0069] In this embodiment, a feature selection kernel SVM heterogeneity analysis model is used to obtain a tumor heterogeneity analysis result, and the process specifically includes the following steps:

[0070] Step B1: establishing a sample annotation data set according to a sample clustering label and a multi-modal fusion feature vector;

[0071] Step B2: introducing an embedded feature selection mechanism to construct a binary selection variable-nonlinear kernel function, performing feature screening on the sample annotation data set to obtain a screened feature subset; and realizing sparse control of the input feature dimension; the binary selection variable-nonlinear kernel function includes a feature selection polynomial kernel and a feature selection Gaussian kernel, and the used formula is as follows:

[0072] Binary selection variable-nonlinear kernel function:

[0073] ;

[0074] wherein, represents a binary feature selection variable, represents an element-by-element product, represents row and column indexes of a kernel matrix, represents a feature vector of a sample and a sample ; represents a kernel function, which defines the similarity between two samples in a hidden space; represents a similarity under the binary selection variable-nonlinear kernel function;

[0075] Explanation: only the selected features will enter the kernel function calculation, and the unselected features have no contribution to the result;

[0076] Feature selection polynomial kernel:

[0077] ;

[0078] wherein, ​denotes the total number of features (dimensionality), denotes the index of feature dimension; denotes the scaling parameter, denotes the offset parameter, denotes the polynomial order; denotes the sample the value of the sample in the th dimension, denotes the sample the value of the sample denotes the th component of the binary feature selection variable;

[0079] Feature selection Gaussian kernel:

[0080] ;

[0081] where, denotes the weighted Euclidean distance square of two samples in the selected feature dimension;

[0082] Step B3: Construct the kernel SVM dual optimization model with feature selection constraints, and train based on the screened feature subset, solve the optimal solution of the kernel SVM dual optimization model with feature selection constraints, obtain the optimal Lagrange multiplier parameter set, the formula used is as follows:

[0083] Kernel SVM dual optimization model with feature selection constraints:

[0084] , ;

[0085] where, denotes the kernel matrix, denotes the total number of samples in the training set, denotes the th element of the kernel matrix; denotes the class label of the th and th sample, respectively;

[0086] , ;

[0087] where, denotes the Lagrange multiplier vector, denotes the transpose symbol, denotes the all-1 row vector, denotes the label row vector, denotes the optimization variable index (Lagrange multiplier), denotes the th training sample of the Lagrange multiplier; represents a penalty factor; represents the abbreviation of subject to, in the case of satisfying the following constraint condition; represents a label consistency constraint, represents a range constraint of the Lagrange multiplier;

[0088] Step B4: using the optimal Lagrange multiplier parameter set, constructing a discriminant function, and calculating the proximity value of each sample relative to the discriminant hyperplane; wherein the proximity value is used to quantify the distribution characteristics and differences of the samples in the high-dimensional kernel space, so as to realize the fine distinction and heterogeneity description of the potential subgroups of tumor samples;

[0089] Step B5: obtaining tumor heterogeneity analysis results by combining sample clustering labels and proximity values, wherein the tumor heterogeneity analysis results include the discriminant boundary of the potential subgroups, the contribution degree evaluation of the key features, and the typing information.

[0090] In example five, based on example two, in this example, the process of obtaining tumor heterogeneity analysis results specifically includes the following steps:

[0091] Step E1: establishing a sample labeled data set according to the sample clustering labels and the multi-modal fusion feature vectors;

[0092] Step E2: performing feature processing on the sample labeled data set, including normalization and screening, to form a screened feature subset;

[0093] Step E3: constructing a dual optimization model of a conventional kernel SVM, and training based on the screened feature subset to solve the optimal solution of the dual optimization model of the kernel SVM, and obtaining an optimal Lagrange multiplier parameter set;

[0094] Step E4: using the optimal Lagrange multiplier parameter set, constructing a discriminant function, and calculating the proximity value of each sample relative to the discriminant hyperplane; wherein the proximity value is used to quantify the distribution characteristics and differences of the samples in the high-dimensional kernel space, so as to realize the fine distinction and heterogeneity description of the potential subgroups of tumor samples;

[0095] Step E5: obtaining tumor heterogeneity analysis results by combining sample clustering labels and proximity values, wherein the tumor heterogeneity analysis results include the discriminant boundary of the potential subgroups, the contribution degree evaluation of the key features, and the typing information.

[0096] In example six, based on example five, in this example, according to Figure 2 , Figure 3 ,

[0097] a heterogeneity analysis module configured to process the sample cluster labels and the multi-modal fusion feature vectors by the feature selection kernel SVM heterogeneity analysis model to obtain a tumor heterogeneity analysis result, wherein the tumor heterogeneity analysis result comprises a discrimination boundary of a potential subpopulation, a contribution degree evaluation of a key feature, and typing information;

[0098] The tumor heterogeneity analysis results of 50 breast cancer patients were analyzed.

[0099] Discrimination boundary of potential subpopulation:

[0100] The system identified three main tumor subpopulations (Cluster A, Cluster B, and Cluster C);

[0101] Cluster A: HER2-positive type, the boundary is determined by the expression level of the HER2 gene (> 2.0 times the baseline) and the imaging mass enhancement mode;

[0102] Cluster B: Luminal A type, the boundary is defined by ER positivity (> 70%) + low Ki-67 (< 15%);

[0103] Cluster C: Basal-like type, the boundary is determined by triple-negative (ER- / PR- / HER2-) + high heterogeneity index of image texture;

[0104] Contribution degree evaluation of key features (Top-1):

[0105] 1. HER2 expression level - contribution degree 0.42;

[0106] 2. ER expression level - contribution degree 0.33;

[0107] 3. Ki-67 proliferation index - contribution degree 0.27;

[0108] 4. Image texture energy value - contribution degree 0.21;

[0109] Summary of typing information:

[0110] Cluster A (HER2-positive type): 15 cases, accounting for 30%;

[0111] Cluster B (Luminal A type): 22 cases, accounting for 44%;

[0112] Cluster C (Basal-like type): 13 cases, accounting for 26%;

[0113] The result visualization and reporting module generates a typing distribution graph, a feature contribution degree graph, and a typing report based on the tumor heterogeneity analysis result to generate an analysis report.

[0114] Generate cancer patient typing distribution chart: including three parts, namely Luminal A, HER2 positive and Basal-like;

[0115] Generate key feature contribution chart:

[0116] X-axis label: feature name (including HER2 expression level, ER expression level, Ki-67 proliferation index, image texture energy value, clinical stage);

[0117] Y-axis: contribution (0-0.5);

[0118] Analysis report:

[0119] Luminal A type (Cluster B, 22 cases, 44%): the subpopulation is characterized by high expression of ER (average positive rate 78%), low Ki-67 proliferation index (mean 12%), mainly distributed in I-II stage patients, and has good prognosis;

[0120] HER2 positive type (Cluster A, 15 cases, 30%): the subpopulation is mainly characterized by HER2 gene amplification (average expression value 2.4 times), and the image shows obvious mass enhancement;

[0121] Basal-like type (Cluster C, 13 cases, 26%): the subpopulation is a typical triple negative (ER- / PR- / HER2-), with high Ki-67 proliferation index (mean 35%), and the image texture shows high heterogeneity signal, indicating poor prognosis.

[0122] The above describes the present application and its embodiments, which is not limited, and the drawings shown are only one of the embodiments of the present application, and the actual structure is not limited thereto; in general, if a person skilled in the art is inspired, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical scheme should belong to the protection scope of the present application.

Claims

1. A tumor heterogeneity analysis system based on multimodal features, characterized by: The system includes feature extraction and fusion module, cluster analysis module, and heterogeneity analysis module; Feature extraction and fusion module, obtaining multimodal fusion feature vectors; The cluster analysis module uses the bi-approximation graph-eccentricity hierarchical clustering algorithm to analyze the multimodal fusion feature vector and generate sample cluster labels; The heterogeneity analysis module uses the feature selection kernel SVM heterogeneity analysis model to process sample cluster labels and multimodal fusion feature vectors to obtain tumor heterogeneity analysis results.

2. The tumor heterogeneity analysis system based on multimodal features according to claim 1, characterized in that: The process of generating sample cluster labels using the bi-proximity graph-eccentricity hierarchical clustering algorithm includes the following steps: Step S1: Calculate the kernel distance and mutual reachability distance of the multimodal fusion feature vector, and construct a cluster approximation graph based on the kernel distance and mutual reachability distance. In the process of cluster approximation graph construction, introduce an adaptive neighborhood ratio to optimize the edge construction mechanism and structural representation ability of the cluster approximation graph, and obtain a dual approximation graph; Step S2: Use the bi-proximity graph to perform single-link hierarchical clustering to generate a clustering hierarchy tree; set the minimum cluster size parameter, condense the clustering hierarchy tree, and obtain the initial cluster set and cluster membership probability; Step S3: For each cluster in the initial cluster set, the weighted centroid of the cluster is calculated respectively; and an eccentricity function is defined based on the distance between the sample points within the cluster in the initial cluster set and the weighted centroid of the cluster; eccentricity weights are calculated based on the eccentricity function, and eccentricity weights are assigned to the edges of the bi-approximate graph to obtain a weighted bi-approximate graph; based on the weighted bi-approximate graph, an eccentricity hierarchical tree is constructed using a single-link hierarchical clustering method; a minimum branch size parameter and a branch persistence threshold are preset, and the eccentricity hierarchical tree is condensed to extract the intra-cluster branch set; Step S4: Determine the branch root for the branch set within the cluster; and calculate the geodesic distance from the sample points within the cluster to the branch root based on the weighted bi-approximation graph; calculate the similarity based on the geodesic distance, and convert the similarity into branch soft membership probability through softmax with temperature parameter; Step S5: Fuse the cluster membership probability with the branch soft membership probability to generate a sample cluster label.

3. The tumor heterogeneity analysis system based on multimodal features according to claim 2, characterized in that: The dual approximation graph includes the core approximation graph and the global approximation graph.

4. The tumor heterogeneity analysis system based on multimodal features according to claim 2, characterized in that: Sample cluster labels include tumor subgroup labels and intra-cluster branch labels.

5. The tumor heterogeneity analysis system based on multimodal features according to claim 1, characterized in that: The process of obtaining tumor heterogeneity analysis results using the feature selection kernel SVM heterogeneity analysis model specifically includes the following steps: Step B1: Create a sample annotation dataset based on sample cluster labels and multimodal fusion feature vectors; Step B2: Introduce an embedded feature selection mechanism, construct a binary selection variable-nonlinear kernel function, perform feature screening on the sample annotation dataset, and obtain the filtered feature subset; Step B3: Construct a kernel SVM dual optimization model with feature selection constraints, and train it based on the filtered feature subset to solve the optimal solution of the kernel SVM dual optimization model with feature selection constraints and obtain the optimal Lagrange multiplier parameter set; Step B4: Using the optimal Lagrange multiplier parameter set, construct a discriminant function and calculate the proximity value; Step B5: Combine the sample cluster labels and proximity values ​​to obtain tumor heterogeneity analysis results.

6. The tumor heterogeneity analysis system based on multimodal features according to claim 5, characterized in that: Binary selection variable-nonlinear kernel functions include feature selection polynomial kernel and feature selection Gaussian kernel.

Images