Breast cancer prognosis real-time evaluation system and method fusing multi-modal image and co-disease network

Abstract

The invention discloses a breast cancer prognosis real-time evaluation system and method fusing a multi-modal image and a co-disease network, and relates to the technical field of breast cancer prognosis evaluation. According to the system, a comprehensive feature matrix fusing images, genes and clinical features is constructed by acquiring a mammary gland medical image, extracting focus features and combining gene expression information of a focus area and co-disease data of a patient, feature weighting is carried out based on an attention mechanism, and finally a prognosis risk score is output by utilizing a multi-layer perceptron model. And risk grading and intervention suggestion generation are realized. According to the system, a co-disease network and a gene interaction network are introduced for modeling, relevance expression among multi-source information is enhanced, and pathological features and systematic health status of patients can be reflected comprehensively. The intelligent level of prognosis evaluation can be improved, and good clinical popularization value is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of breast cancer prognosis assessment, and in particular to a real-time breast cancer prognosis assessment system and method integrating multimodal imaging and comorbidity networks. Background Art

[0002] Accurately assessing breast cancer prognosis is crucial for developing personalized treatment plans and optimizing medical resource allocation. Traditional prognostic assessment methods rely primarily on histopathological classification, TNM staging, and the expression levels of several biomarkers (such as HER2, ER, and PR). However, these methods suffer from limitations in timeliness, subjectivity, and inadequate responsiveness to individual patient differences, making them inadequate for precision medicine.

[0003] In recent years, with the rapid development of medical imaging technology, molecular diagnostic methods, and artificial intelligence algorithms, tumor prognosis assessment methods based on multimodal data fusion have gradually emerged. Among them, the joint modeling of imaging omics features and molecular omics features has been shown to improve the accuracy of prognostic judgment to a certain extent. However, existing technologies generally have the following shortcomings: First, they fail to fully explore and integrate the potential interactions between imaging, genetic, and clinical comorbidity data; second, they lack systematic modeling of disease comorbidity network structures, and cannot fully reflect the impact of the patient's overall health status on the risk of tumor progression; third, most fusion strategies use static weights or linear combinations, failing to dynamically perceive the importance of different features in specific scenarios, resulting in poor individual adaptability and robustness of the model.

[0004] Therefore, there is an urgent need for a comprehensive evaluation framework that can integrate multimodal imaging information, molecular expression characteristics, and disease comorbidity networks to improve the real-time, accuracy, and personalization of breast cancer prognosis assessment. Summary of the Invention

[0005] In response to the limitations of existing methods such as insufficient feature dimensions, shallow information fusion, poor accuracy and adaptability of evaluation results, the present invention proposes a real-time breast cancer prognosis assessment system and method that integrates multimodal imaging and comorbidity networks. In the breast cancer prognosis assessment, multimodal medical imaging, molecular genetic information and comorbidity clinical data are integrated to achieve real-time, accurate and personalized judgment of patients' prognostic risk.

[0006] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0007] A real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network, comprising:

[0008] The image preprocessing module is used to obtain the patient's breast medical images and perform normalization, denoising and contrast enhancement to generate a basic image dataset;

[0009] a lesion analysis module, configured to extract lesion contours based on the basic image dataset and calculate a contour integrity score; if the contour integrity score exceeds a preset threshold, analyze the lesion enhancement pattern in combination with the dynamic contrast-enhanced magnetic resonance imaging sequence, output preliminary lesion localization information, and extract lesion image features; otherwise, return to the basic image dataset and adjust filtering parameters before reprocessing;

[0010] A gene feature extraction module is used to extract gene expression data of the corresponding area based on the preliminary lesion location information, construct a gene interaction network, identify core gene nodes with high variability and high connectivity, and extract expression feature outputs of the core gene nodes;

[0011] The comorbidity feature extraction module is used to obtain the patient's comorbidity information, map it to the comorbidity network model, and extract comorbidity feature parameters related to breast cancer prognosis;

[0012] A feature fusion processing module is used to fuse the lesion imaging features, the expression features of the core gene nodes and the comorbidity feature parameters to construct a comprehensive feature matrix, and to perform a weighted evaluation of the importance of each feature in the comprehensive feature matrix based on an attention mechanism to generate a weighted feature vector;

[0013] a risk assessment module, configured to input the weighted feature vector into a pre-trained multi-layer perceptron model and output a prognostic risk score for breast cancer;

[0014] The result generation module is used to compare the prognostic risk score with a set risk threshold. If the prognostic risk score is greater than the risk threshold, the patient's risk level is determined to be a high-risk group; otherwise, it is determined to be a low-risk group, and personalized intervention recommendations are generated.

[0015] Preferably, the image preprocessing module includes:

[0016] A resolution normalization unit, configured to uniformly adjust breast medical images from multiple imaging devices or imaging modalities to a predetermined spatial resolution and grayscale intensity range, and to perform time frame registration and signal intensity normalization on a magnetic resonance dynamic contrast-enhanced image sequence;

[0017] A noise suppression unit is used to perform a median filter operation on the breast medical image after resolution normalization based on spatial neighborhood statistical information, so as to suppress speckle noise and random interference while maintaining the edge structure;

[0018] A contrast enhancement unit is used to perform regional brightness redistribution processing on the breast medical image after noise suppression processing based on the image grayscale distribution characteristics, so as to enhance the grayscale contrast between the lesion and the adjacent tissue;

[0019] The breast medical images processed by the contrast enhancement unit constitute a basic image data set, which is used as input for the subsequent lesion analysis module.

[0020] Preferably, the lesion analysis module specifically includes:

[0021] An image scoring unit, configured to calculate an image quality score for each breast medical image in the basic image dataset, wherein the image quality score is determined based on a combined function of cell nucleus density, image clarity, and background area ratio;

[0022] An image screening unit, configured to select a breast medical image with the highest image quality score as a representative breast image;

[0023] a contour extraction unit, configured to perform an edge detection operation on the representative breast image to extract a morphological contour of the lesion;

[0024] A contour scoring unit, configured to calculate a contour integrity score C based on the lesion morphological contour s , the calculation formula is:

[0025] C s =α1C closure +α2C continuity +α3(1-C artifact );

[0026] Where C closure Indicates the closure of the contour, C continuity Indicates edge continuity, C artifact represents the artifact interference intensity; α1, α2, and α3 are weight coefficients, and satisfy α1+α2+α3=1;

[0027] The dynamic enhancement analysis unit is used to calculate the enhancement rate of the lesion based on the magnetic resonance dynamic contrast enhancement image sequence corresponding to the representative breast image when the contour integrity score is greater than the preset threshold. The calculation formula is:

[0028]

[0029] Where, E(t) is the enhancement rate at time t; S(t) is the signal intensity of the lesion area at time t; S(0) is the baseline signal intensity;

[0030] The combined determination unit is used to determine the location of the lesion based on the combined rule of the contour integrity score and the enhancement rate according to the following logic function:

[0031]

[0032] Where L(x,y) is a binary function, which indicates whether the image coordinate position (x,y) belongs to the lesion area, 1 for lesion and 0 for non-lesion; E threshold is the set enhancement rate threshold;

[0033] When L(x,y) is 1, the preliminary lesion localization information is output; otherwise, a filter parameter adjustment instruction is sent to the image preprocessing module to regenerate the basic image data set to repeat the above lesion localization process;

[0034] An image feature extraction unit is used to extract lesion image features from representative breast images based on the preliminary lesion positioning information, including grayscale distribution parameters, texture detail index, lesion area, lesion perimeter, lesion boundary curvature and lesion enhancement rate.

[0035] Preferably, the gene feature extraction module includes:

[0036] A gene expression data extraction unit is used to extract gene expression data corresponding to the lesion area from the gene expression database based on the preliminary lesion location information to form a gene expression data set;

[0037] A gene interaction network construction unit is used to construct an undirected weighted graph model G=(V, E, W) based on the gene expression data set, where V represents a set of gene nodes, E represents a set of connection relationships between genes, and W represents a set of connection edge weights of expression correlations;

[0038] Gene attribute evaluation unit, used to calculate the gene interaction strength of each gene pair, the connectivity score of each gene node and the expression variability, in order to characterize the network connectivity characteristics and expression stability of each gene;

[0039] A core gene identification unit is used to identify gene nodes that meet the conditions from the gene node set as core gene nodes based on the rule that the gene interaction strength, connectivity score and expression variability all exceed preset thresholds, thereby forming a core gene node set;

[0040] A gene correlation analysis unit is used to perform correlation analysis on each gene in the core gene node set and the lesion imaging feature, and calculate the average correlation score of each core gene node;

[0041] A core distribution structure analysis unit is used to analyze the network concentration of core gene nodes based on the topological distribution of the core gene node set in an undirected weighted graph model; when the network concentration is higher than a preset threshold, the core gene node set is divided into several gene clusters based on a graph clustering algorithm, and the topological feature vectors of each gene cluster are classified to extract the stable structural characteristics of each gene cluster. The topological feature vectors include the average connectivity within the cluster, the edge weight variance, and the node betweenness centrality extreme difference; based on the stable structural characteristics, topological distribution, and DisGeNET disease-gene association database, disease label association information is generated for the gene cluster to which each core gene node belongs;

[0042] The expression feature output unit is used to extract the expression value, gene interaction strength, connectivity score, expression variability, average correlation score, and disease label association information of each core gene node in the core gene node set as expression feature output.

[0043] Preferably, the gene interaction strength S of each gene pair is calculated based on the common neighbor edge weight ij , the formula is:

[0044]

[0045] Where w ik 、w jk They represent the edge weights between gene i and gene k, and between gene j and gene k, respectively, and K is the total number of gene nodes;

[0046] Based on the gene interaction strength S ij Calculate the connectivity score D of each gene node i , the formula is:

[0047] Calculate the expression variability V of each gene node based on the gene expression data set i , the formula is: in is the expression variance of gene i, μ i is the mean expression value of gene i;

[0048] The formula for calculating the average correlation score of each core gene node is:

[0049]

[0050] Where R i represents the average correlation score of gene i; M is the number of dimensions of lesion imaging features; (g i ,f m ) represents the expression value g of gene i i and the image feature f of the mth lesionm The correlation coefficient between .

[0051] Preferably, the comorbidity feature extraction module includes:

[0052] A comorbidity data extraction unit is used to extract a list of confirmed diseases and corresponding visit times, diagnosis codes, treatment records, and clinical status data from the patient's electronic medical record or health file;

[0053] A comorbidity network mapping unit is used to map each disease diagnosed by the patient to a pre-established comorbidity network model, and obtain the location nodes and connection relationships of the patient's confirmed disease in the comorbidity network;

[0054] The comorbidity network model is an undirected weighted network model constructed based on population epidemiological statistical data. A node in the network represents a disease, and an edge weight represents the co-occurrence probability between diseases. The undirected weighted network model includes network versions at different time stages to reflect the co-occurrence relationship of diseases in different periods.

[0055] A comorbidity parameter extraction unit is used to calculate comorbidity characteristic parameters related to breast cancer prognosis based on the node distribution of the patient in the comorbidity network, including: the number of comorbidity nodes, the weight value of the edge connecting to the breast cancer node, and the set of major comorbidity nodes whose median centrality or betweenness centrality of the comorbidity nodes is higher than a set threshold;

[0056] The comorbidity status updating unit is used to select the disease information corresponding to the current evaluation time point based on the patient's diagnosis and treatment information including multiple time points, and update the above-mentioned comorbidity characteristic parameters to reflect the patient's current comorbidity status.

[0057] Preferably, the feature fusion processing module includes:

[0058] A feature matrix construction unit is used to standardize and integrate the lesion imaging features, the expression features of the core gene nodes, and the comorbidity feature parameter set to construct a comprehensive feature matrix, where each column corresponds to a standardized structural feature;

[0059] The feature importance evaluation unit is used to calculate the weight of the feature dimension of the comprehensive feature matrix based on the trainable attention network to generate a weighted feature vector X = (ω1x1,...,ω d x d ), where x d represents the d-th standardized feature, ω d Represents feature x d Corresponding contribution weight;

[0060] The weighted feature output unit is used to output the weighted feature vector X as input to the risk prediction module to perform the breast cancer prognosis risk scoring task.

[0061] Preferably, the risk assessment module includes:

[0062] A structural relevance scoring unit is used to score the average relevance score R i and connectivity score D i , construct the structural relevance score R score , the formula is: Where G is the number of core gene nodes;

[0063] The risk score modeling unit is used to combine the weighted feature vector with the structural correlation score value R score Input them into the pre-trained multi-layer perceptron model and output the prognostic risk score y for breast cancer risk ∈[0,1];

[0064] The multilayer perceptron model includes at least three layers of fully connected neurons. During the training process, a binary cross entropy loss function or a mean square error loss function is used, and the Adam optimization algorithm is used for parameter update. The learning rate is set to η = 0.001;

[0065] The training process introduces an early stopping mechanism. When the performance indicators of the model do not improve within eight consecutive training cycles on the validation set, the training is terminated and the model parameter configuration is returned to the optimal performance state.

[0066] A real-time breast cancer prognosis assessment method integrating multimodal imaging and comorbidity networks, including:

[0067] Obtain the patient's original medical images, perform normalization, denoising, and contrast enhancement, and generate a basic image dataset;

[0068] Extracting lesion contours based on the basic image dataset and calculating a contour integrity score; if the contour integrity score exceeds a preset threshold, analyzing the lesion enhancement pattern in combination with a dynamic contrast-enhanced magnetic resonance imaging sequence, outputting preliminary lesion localization information, and extracting lesion image features; otherwise, returning to the basic image dataset, adjusting filtering parameters, and reprocessing;

[0069] Extracting gene expression data of the corresponding region based on the preliminary lesion location information, constructing a gene interaction network, identifying core gene nodes with high variability and high connectivity, and extracting expression feature outputs of the core gene nodes;

[0070] Obtain the patient's comorbidity information, map it to the comorbidity network model, and extract comorbidity characteristic parameters related to breast cancer prognosis;

[0071] The lesion imaging features, the expression features of the core gene nodes, and the comorbidity feature parameters are integrated to construct a comprehensive feature matrix, and the importance of each feature in the comprehensive feature matrix is ​​weightedly evaluated based on the attention mechanism to generate a weighted feature vector;

[0072] Inputting the weighted feature vector into a pre-trained multi-layer perceptron model to output a prognostic risk score for breast cancer;

[0073] The risk score is compared with the set risk threshold. If the prognostic risk score is greater than the risk threshold, the patient's risk level is determined to be a high-risk group, otherwise it is determined to be a low-risk group, and personalized intervention recommendations are generated.

[0074] The beneficial effects of the present invention are: based on the patient's imaging features, the integration of gene expression information and individual comorbidity characteristics can more comprehensively characterize the heterogeneity of tumors and the disease evolution process, thereby significantly improving the accuracy and credibility of the prognostic evaluation results. The introduction of disease comorbidity network and gene interaction network models to achieve topological structure modeling between features is helpful to mine key nodes and paths, improve the feature expression ability from the association level, and strengthen the model's understanding and reasoning ability of complex medical data; based on the attention mechanism, the importance of multi-source features is adaptively weighted, and the model decision strategy can be flexibly adjusted according to different patient feature patterns, thereby improving the generalization ability and individual adaptability of the evaluation system; support for risk level division based on preset risk thresholds, and automatic generation of intervention recommendations, providing clinicians with an operational decision-making support basis, and improving the foresight and initiative of disease management. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0076] in:

[0077] Figure 1 It is a system structure block diagram of the present invention;

[0078] Figure 2 This is a schematic structural diagram of a lesion analysis module according to an embodiment of the present invention;

[0079] Figure 3 This is a schematic diagram of the structure of a gene feature extraction module in an embodiment of the present invention;

[0080] Figure 4 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0081] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of the present invention.

[0082] like Figure 1 As shown, an embodiment of the present invention provides a real-time breast cancer prognosis assessment system that integrates multimodal imaging and comorbidity networks, including an image preprocessing module, a lesion analysis module, a gene feature extraction module, a comorbidity feature extraction module, a feature fusion processing module, a risk assessment module, and a result generation module.

[0083] 1. Image preprocessing module

[0084] The image preprocessing module is used to obtain the patient's breast medical images, and perform standardization, denoising and contrast enhancement to generate a basic image dataset.

[0085] In a preferred embodiment of the present invention, the image preprocessing module is designed to perform unified standardization, noise suppression, and contrast enhancement on the patient's original breast medical image data to construct a basic image dataset for downstream lesion analysis, thereby improving the system's compatibility and robustness with multi-modal, multi-source heterogeneous medical images. It includes:

[0086] The resolution normalization unit is used to process breast medical images from multiple imaging devices or imaging modalities (such as ultrasound, mammography, MRI, etc.). The normalization process includes:

[0087] Unified spatial resolution: resample the input image to a preset pixel size (e.g., 0.5 mm × 0.5 mm per pixel) to ensure that all images are spatially comparable.

[0088] Grayscale intensity range normalization: Perform linear stretching or normalization functions (such as Min-Max normalization) on image pixel values ​​to unify them to a fixed grayscale range (such as 0–255) to improve input consistency between multimodal images;

[0089] Dynamic image sequence registration processing (for MRIDCE): Execute rigid or non-rigid image registration algorithms (such as mutual information or gradient difference based) on each time frame of the dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) image sequence;

[0090] After the registration is completed, the signal intensity of each time frame is further normalized to offset the impact of individual differences and device drift on the enhancement curve.

[0091] To improve image quality, a noise suppression unit performs an adaptive median filter on normalized breast medical images based on spatial neighborhood statistics. For each pixel, a k×k neighborhood is selected and the median of the neighboring pixels is used to replace the current pixel value. An edge-preserving strategy is introduced for edge regions to avoid blurring of edge structures. This filtering technique demonstrates excellent denoising capabilities for images with speckle noise and random interference, such as those found in MRI and ultrasound.

[0092] The contrast enhancement unit further improves the distinguishability of the lesion area and surrounding tissue in the image, including:

[0093] Grayscale distribution feature extraction: statistical image histogram, grayscale mean and standard deviation and other indicators;

[0094] Regional brightness redistribution processing: histogram equalization, local adaptive contrast enhancement (such as CLAHE) or regional contrast stretching methods can be used to locally enhance the lesion area and preserve the stability of normal tissue structure;

[0095] The processing results can significantly improve the display of tumor boundaries, morphology, and structural details. The processed images constitute the basic imaging data set as the input of the subsequent lesion analysis module.

[0096] 2. Lesion Analysis Module

[0097] The lesion analysis module is used to extract lesion contours based on the basic image dataset and calculate the contour integrity score. The contour closure and continuity are evaluated by comparing the degree of fit between the detected edges and the ideal tumor shape. If the contour integrity score exceeds the preset threshold, the lesion enhancement pattern is analyzed in combination with the magnetic resonance dynamic contrast-enhanced image sequence, and the preliminary lesion positioning information is output and the lesion image features are extracted. Otherwise, the basic image dataset is returned to adjust the filtering parameters and reprocessed.

[0098] In one preferred embodiment, the lesion analysis module identifies the most representative breast image from the basic image data set, extracts the breast lesion contour and evaluates its structural integrity, decides whether to introduce a dynamic enhancement sequence for further analysis based on the evaluation results, and finally outputs preliminary lesion location information and lesion image features. Figure 2 As shown, the lesion analysis module includes:

[0099] The image scoring unit calculates the image quality score for each breast medical image in the basic image dataset. The scoring criteria include but are not limited to:

[0100] The density of cell nuclei in the image D nuc (Based on local texture analysis and edge kernel extraction);

[0101] Image clarity clarity (using Laplace operator or frequency domain energy analysis);

[0102] Background area ratio B ratio (Calculate the ratio of background area to breast tissue);

[0103] By constructing a weighted scoring function Q = β1D nuc +β2S clarity +β3(1-B ratio ) to obtain the comprehensive quality score of each image, where β1, β2, and β3 are weight coefficients;

[0104] The image screening unit selects the breast image with the highest image quality score from all scored images as the representative breast image for subsequent lesion analysis. This screening mechanism ensures that downstream processing is based on high-quality images, improving the stability and reliability of the results.

[0105] The contour extraction unit performs edge detection and morphological processing operations on representative breast images to extract the morphological contours of the lesions. Commonly used algorithms include but are not limited to: Canny edge detection, Sobel / Laplacian operators, region growing and morphological closing operations.

[0106] The extraction result is a set of continuous contour paths

[0107] The contour scoring unit calculates the contour integrity score C based on the lesion morphological contour s , to avoid misjudgment caused by using only a single length ratio, the calculation formula is:

[0108] C s =α1C closure +α2C continuity +α3(1-C artifact );

[0109] Where C closure Indicates the closure of the contour. Based on the connectivity of the first and last points of the contour and the topological closure index, it detects whether a complete closed path is formed. The value range is [0,1]; C continuity Indicates edge continuity, which is used to evaluate the consistency of gradient direction and the number of break points between contour edge points; C artifact It represents the artifact interference intensity and is used to analyze the statistical variance and abnormal enhancement pattern of signal changes in the contour area to identify motion artifacts or non-lesion interference; α1, α2, and α3 are weight coefficients, and they satisfy α1+α2+α3=1;

[0110] When the contour integrity score is greater than a preset threshold (e.g., 0.75), the contour structure is considered stable and valuable for subsequent analysis. The dynamic enhancement analysis unit activates the dynamic enhancement analysis path: a dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) sequence corresponding to a representative breast image is acquired, and the signal intensity enhancement rate curve of the target area is calculated based on the frame registration results. The calculation formula is:

[0111]

[0112] Where, E(t) is the enhancement rate at time t; S(t) is the signal intensity of the lesion area at time t; S(0) is the baseline signal intensity;

[0113] The joint judgment unit determines the location of the lesion based on the joint rule of the contour integrity score and the enhancement rate according to the following logical function:

[0114]

[0115] Where L(x,y) is a binary function, which indicates whether the image coordinate position (x,y) belongs to the lesion area, 1 for lesion and 0 for non-lesion; E threshold is the set enhancement rate threshold;

[0116] When L(x,y) is 1, the preliminary lesion localization information (such as the lesion center point and region boundary) is output. Otherwise, a filter parameter adjustment instruction is sent to the image preprocessing module to regenerate the basic image data set to repeat the above lesion localization process.

[0117] The image feature extraction unit is used to extract lesion image features from representative breast images based on the preliminary lesion positioning information, including grayscale distribution parameters (average grayscale, grayscale entropy, etc.), texture detail indicators (GLCM, LBP, etc.), lesion area, lesion perimeter, lesion boundary curvature, and lesion enhancement rate.

[0118] 3. Gene feature extraction module

[0119] The gene feature extraction module is used to extract the gene expression data of the corresponding area based on the preliminary localization information of the lesion, construct a gene interaction network, identify core gene nodes with high variability and high connectivity, and extract the expression feature output of the core gene nodes.

[0120] like Figure 3 As shown, in a preferred embodiment of the present invention, the gene feature extraction module includes:

[0121] The gene expression data extraction unit extracts gene expression data corresponding to the lesion area from the gene expression database according to the spatial area provided by the preliminary lesion location information to form a gene expression data set, and the steps include:

[0122] Using image annotations (such as tumor masks) and pathological zoning matching tables, extract mRNA expression data of relevant tissue or cell samples in the region. Expression data can be formatted as FPKM, TPM, or counts, and after normalization, output as a gene expression data set.

[0123] The gene interaction network construction unit constructs an undirected weighted graph model G = (V, E, W) based on the gene expression data set, where V represents the set of gene nodes, E represents the set of gene connection relationships, and W represents the set of connection edge weights of expression correlations. Furthermore, protein-protein interaction relationships provided by public databases (such as STRING and KEGG) can be optionally introduced to assign biological pathway support weights to edge weights.

[0124] The gene attribute evaluation unit is used to calculate the gene interaction strength of each gene pair, the connectivity score of each gene node, and the expression variability to characterize the network connectivity characteristics and expression stability of each gene;

[0125] Calculate the gene interaction strength S of each gene pair based on the common neighbor edge weight ij , the formula is:

[0126]

[0127] Where w ik 、w jk They represent the edge weights between gene i and gene k, and between gene j and gene k, respectively, and K is the total number of gene nodes;

[0128] Based on the gene interaction strength S ij Calculate the connectivity score D of each gene node i , the formula is:

[0129] Calculate the expression variability V of each gene node based on the gene expression data set i , the formula is: in is the expression variance of gene i, μ i is the mean expression value of gene i;

[0130] Based on the attribute evaluation, the core gene identification unit identifies the gene nodes that meet the conditions as core gene nodes from the gene node set according to the rule that the gene interaction strength, connectivity score and expression variability all exceed the preset threshold, and constitutes the core gene node set;

[0131] The gene correlation analysis unit performs correlation analysis on each gene in the core gene node set and the lesion imaging characteristics, and calculates the average correlation score of each core gene node. The formula is:

[0132]

[0133] Where R i represents the average correlation score of gene i; M is the number of dimensions of lesion imaging features; (g i ,f m ) represents the expression value g of gene i i and the image feature f of the mth lesion m The correlation coefficient between .

[0134] The core distribution structure analysis unit analyzes the network concentration of core gene nodes (such as average neighbor distance and network density) based on the topological distribution of the core gene node set in the undirected weighted graph model. When the network concentration is higher than the preset threshold, the core gene node set is divided into several gene clusters based on a graph clustering algorithm (such as Louvain and K-means on embedding). The topological feature vector of each gene cluster is classified. The topological feature vector includes the average connectivity within the cluster, the variance of the edge weight, and the extreme difference of the node betweenness centrality. The stable structural characteristics of each gene cluster are extracted. When the average connectivity within the cluster, the variance of the edge weight, and the extreme difference of the node betweenness centrality are all lower than the preset threshold, it is determined to be a stable structural feature. Based on the stable structural characteristics, topological distribution, and the DisGeNET disease-gene association database, the disease label association information (such as HER2 pathway and immune escape) of the gene cluster to which each core gene node belongs is generated.

[0135] The expression feature output unit extracts the expression value, gene interaction strength, connectivity score, expression variability, average correlation score, and disease label association information of each core gene node in the core gene node set as the expression feature output.

[0136] 4. Comorbidity feature extraction module

[0137] The comorbidity feature extraction module is used to obtain the patient's comorbidity information, map it to the comorbidity network model, and extract comorbidity feature parameters related to breast cancer prognosis.

[0138] In this embodiment of the present invention, the comorbidity feature extraction module performs structured modeling on the patient's past and current medical conditions to quantify the propagation location of comorbidity information in the disease network and its potential impact on breast cancer prognosis. This includes:

[0139] The comorbidity data extraction unit extracts the list of diagnosed diseases (such as diabetes, hypertension, heart failure, osteoporosis, etc.), the corresponding visit time (the initial diagnosis time and the most recent visit time for each disease), the diagnosis code (standard ICD-10 or ICD-11 code for cross-system universal mapping), treatment record (whether to receive continuous drug treatment, surgical intervention, complication situation, etc.), and clinical status data (disease activity status (active period / remission period), control stability, medication compliance, etc.) from the patient's electronic medical record (EMR) or health record. The extraction results are recorded in a structured form for comorbidity network modeling input;

[0140] The comorbidity network mapping unit maps the list of diagnosed diseases of the patient to the pre-established comorbidity network model as follows:

[0141] Each diagnosed disease exists as a node in the comorbidity network;

[0142] A one-to-one correspondence is established between the standardized diagnosis code and the node library in the comorbidity network;

[0143] After mapping is completed, the positioning node subgraph structure of the patient in the comorbidity network is extracted, and the following attributes are recorded:

[0144] Node position (node ID);

[0145] Connection edges (i.e., co-occurrence edges) between other disease nodes;

[0146] The weight value of the connection edge reflects the influence strength of the disease suffered by the target patient on the breast cancer node.

[0147] Preferably, the system supports the introduction of patient age, gender, race, and other demographic factors for risk adjustment to more accurately complete the mapping.

[0148] The comorbidity network model is a directed weighted network model based on population epidemiological statistical data, with nodes representing a disease and edge weights representing the co-occurrence probability between diseases. The directed weighted network model includes multiple versions corresponding to different periods or age groups (such as <45 years old, 45-65 years old, >65 years old) to reflect the time evolution characteristics of disease relationships. The model can select the current version through a timestamp to ensure alignment with the current time point of the patient;

[0149] The comorbidity parameter extraction unit calculates comorbidity characteristic parameters related to breast cancer prognosis based on the patient's node distribution in the comorbidity network. These parameters include: the number of comorbidity nodes (i.e., the total number of comorbidities a patient has, used to measure comorbidity burden), the weight of the edges connecting to the breast cancer node, and the set of major comorbidity nodes whose degree centrality or betweenness centrality of the comorbidity nodes is higher than a set threshold. For example, if "diabetes" has a high betweenness centrality to the breast cancer node in the network and the patient is in an uncontrolled state, then the disease is marked as a major comorbidity node factor.

[0150] Because changes in comorbidities are time-sensitive, the comorbidity status update unit selects the disease information corresponding to the current assessment time point based on the patient's diagnosis and treatment information covering multiple time points, and updates the above comorbidity characteristic parameters to ensure alignment with the current risk assessment time point:

[0151] Input: EMR data at multiple time points (e.g., stored quarterly or annually);

[0152] Filtering strategy: select the data closest to the current forecast evaluation date;

[0153] Dynamic update indicators:

[0154] Whether a comorbidity has been newly diagnosed / relieved;

[0155] Whether the main comorbidity nodes are updated;

[0156] Whether there is structural shift in the network position (such as changes in edge weights);

[0157] Output updated comorbidity characteristic parameters for dynamic individualized modeling.

[0158] 5. Feature fusion processing module

[0159] The feature fusion processing module is used to fuse lesion imaging features, expression features of core gene nodes and comorbidity feature parameters to construct a comprehensive feature matrix. It also performs a weighted evaluation of the importance of each feature in the comprehensive feature matrix based on the attention mechanism to generate a weighted feature vector.

[0160] In one embodiment of the present invention, the feature fusion processing module performs unified fusion processing on the individualized features of patients from multiple modal sources to construct a high-dimensional structured input vector for breast cancer prognostic risk modeling. It includes:

[0161] The feature matrix construction unit integrates the lesion imaging features, the expression features of the core gene nodes, and the comorbidity feature parameter set in a standardized manner (e.g., z-score normalization, min-max normalization) to construct a comprehensive feature matrix. Each column corresponds to a standardized structural feature, indicating the standard value under that feature dimension, and each row represents a feature source channel (e.g., imaging / gene / comorbidity).

[0162] The feature importance evaluation unit is used to calculate the weight of the feature dimension of the comprehensive feature matrix based on the trainable attention network to generate a weighted feature vector X = (ω1x1,...,ω d x d ), where x d represents the d-th standardized feature, ω d Represents feature x d Corresponding contribution weight;

[0163] The weighting coefficients are obtained through multi-layer attention sub-network training, which can adopt the Transformer structure or lightweight self-attention mechanism; the attention mechanism can be obtained through joint training or independent training, where the loss function can be an error indicator for predicting risk, such as cross entropy or mean square error. The output dimension of the attention mechanism is consistent with the dimension of the comprehensive feature matrix, and supports weight sparsity control to improve the interpretability of the model. The weighting coefficients are multiplied dimension by dimension with the standardized features in the comprehensive feature matrix to generate the final weighted feature vector. In some embodiments, a feature importance threshold can be set, and the importance scores below the threshold can be compressed (such as set to 50% of the original value) to enhance the contribution stability of highly correlated features.

[0164] The weighted feature output unit outputs the weighted feature vector X, which is used as the input of the risk prediction module to perform the breast cancer prognosis risk scoring task.

[0165] 6. Risk Assessment Module

[0166] The risk assessment module is used to input the weighted feature vector into a pre-trained multi-layer perceptron model and output the prognostic risk score of breast cancer.

[0167] In this embodiment of the present invention, the risk assessment module is used to receive the weighted feature vector output by the feature fusion processing module, combine it with the structural correlation score, and use a pre-trained multi-layer perceptron neural network model to quantitatively predict the prognostic risk of breast cancer patients. It includes:

[0168] The structural relevance scoring unit is based on the average relevance score R i and connectivity score D i , construct the structural relevance score R score , the formula is: Where G is the number of core gene nodes; the structural correlation score is used to reflect the cross-modal structural consistency between the extracted features, which helps to enhance the model's perception of the quality of feature distribution.

[0169] The risk score modeling unit combines the weighted feature vector with the structural correlation score value R score Input them into the pre-trained multi-layer perceptron model and output the prognostic risk score y for breast cancer risk ∈[0,1];

[0170] The multilayer perceptron model includes at least three layers of fully connected neurons (e.g., 128-64-32), each layer followed by a ReLU activation function; during training, a binary cross entropy loss function or a mean square error loss function is used, and the Adam optimization algorithm is used for parameter update, with a learning rate set to η = 0.001;

[0171] The training process introduces an early stopping mechanism. When the performance indicators of the model do not improve within eight consecutive training cycles on the validation set, the training is terminated and the model parameter configuration is returned to the optimal performance state.

[0172] 7. Result generation module

[0173] The result generation module is used to compare the prognostic risk score with the set risk threshold. If the prognostic risk score is greater than the risk threshold, the patient's risk level is determined to be high-risk, with a high probability of breast cancer recurrence or progression during the follow-up period. Otherwise, the patient is determined to be low-risk, with a relatively controllable risk level. The risk level can be optionally divided into two levels (high / low) or three levels (high / medium / low), and a dynamic threshold adjustment strategy is supported. The specific value can be determined based on clinical experience or the optimal compromise point under the ROC curve.

[0174] After completing the risk level determination, the system matches the patient's built-in intervention knowledge base based on the patient's individual characteristics (including imaging indicators, gene expression characteristics, comorbidity parameters, etc.) and outputs corresponding precise intervention plan recommendations.

[0175] The matching process includes:

[0176] Input features: including but not limited to the following information: lesion imaging characteristics (such as unclear boundaries, enhancement speed); high-risk gene expression patterns (such as high expression of BRCA1 and TP53); high-central comorbidity information (such as diabetes combined with cardiovascular disease);

[0177] The system's preset intervention suggestion knowledge base includes:

[0178] Intervention pathways for high-risk patients: such as increased follow-up frequency (once every 3 months) and multimodal imaging review;

[0179] Drug recommendations: Matching targeted drugs (e.g., trastuzumab is recommended for HER2+), adjuvant hormone therapy;

[0180] Lifestyle guidance: Provide lifestyle management advice on comorbidity control (such as controlling blood sugar / blood pressure);

[0181] Output: Textual or structured personalized intervention recommendation reports, which can be pushed to the electronic medical record system, doctor terminal or patient mobile terminal as auxiliary decision support data.

[0182] The result generation module can be configured as a rule-based matching, table-based recommendation, or a deep learning-driven intervention strategy matching model (such as a multi-task learning framework), supporting decision-making after review by clinicians.

[0183] like Figure 4 FIG. 1 is another embodiment of the present invention, which provides a real-time breast cancer prognosis assessment method integrating multimodal imaging and comorbidity network, including:

[0184] Obtain the patient's original medical images, perform normalization, denoising, and contrast enhancement, and generate a basic image dataset;

[0185] Extracting lesion contours based on the basic image dataset and calculating a contour integrity score; if the contour integrity score exceeds a preset threshold, analyzing the lesion enhancement pattern in combination with a dynamic contrast-enhanced magnetic resonance imaging sequence, outputting preliminary lesion localization information, and extracting lesion image features; otherwise, returning to the basic image dataset, adjusting filtering parameters, and reprocessing;

[0186] Extracting gene expression data of the corresponding region based on the preliminary lesion location information, constructing a gene interaction network, identifying core gene nodes with high variability and high connectivity, and extracting expression feature outputs of the core gene nodes;

[0187] Obtain the patient's comorbidity information, map it to the comorbidity network model, and extract comorbidity characteristic parameters related to breast cancer prognosis;

[0188] The lesion imaging features, the expression features of the core gene nodes, and the comorbidity feature parameters are integrated to construct a comprehensive feature matrix, and the importance of each feature in the comprehensive feature matrix is ​​weightedly evaluated based on the attention mechanism to generate a weighted feature vector;

[0189] Inputting the weighted feature vector into a pre-trained multi-layer perceptron model to output a prognostic risk score for breast cancer;

[0190] The risk score is compared with the set risk threshold. If the prognostic risk score is greater than the risk threshold, the patient's risk level is determined to be a high-risk group, otherwise it is determined to be a low-risk group, and personalized intervention recommendations are generated.

[0191] In summary, the present invention constructs an individualized real-time prognosis assessment system for breast cancer by integrating breast multimodal medical images, gene expression characteristics and comorbidity network information, realizes standardized preprocessing and efficient integration of multi-source medical data, and improves the accuracy of lesion identification and feature extraction; utilizes image features and gene interaction networks to mine core biomarkers, and enhances the ability to characterize tumor biological behavior; introduces comorbidity network modeling to model the patient's systemic disease association, and strengthens the overall assessment of prognostic risk; dynamically weights multimodal features based on the attention mechanism to improve the model's predictive robustness in different individuals; realizes prognostic risk grading and personalized intervention recommendations through deep learning models, and enhances the clinical guidance value of the assessment results.

[0192] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily conceive of various modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network, characterized by: The system includes: an image preprocessing module for acquiring a patient's breast medical image and performing standardization, denoising, and contrast enhancement to generate a basic image data set; a lesion analysis module, configured to extract lesion contours based on the basic image dataset and calculate a contour integrity score; if the contour integrity score exceeds a preset threshold, analyze the lesion enhancement pattern in combination with the dynamic contrast-enhanced magnetic resonance imaging sequence, output preliminary lesion localization information, and extract lesion image features; otherwise, return to the basic image dataset and adjust filtering parameters before reprocessing; A gene feature extraction module is used to extract gene expression data of the corresponding area based on the preliminary lesion location information, construct a gene interaction network, identify core gene nodes with high variability and high connectivity, and extract expression feature outputs of the core gene nodes; The comorbidity feature extraction module is used to obtain the patient's comorbidity information, map it to the comorbidity network model, and extract comorbidity feature parameters related to breast cancer prognosis; A feature fusion processing module is used to fuse the lesion imaging features, the expression features of the core gene nodes and the comorbidity feature parameters to construct a comprehensive feature matrix, and to perform a weighted evaluation of the importance of each feature in the comprehensive feature matrix based on an attention mechanism to generate a weighted feature vector; a risk assessment module, configured to input the weighted feature vector into a pre-trained multi-layer perceptron model and output a prognostic risk score for breast cancer; The result generation module is used to compare the prognostic risk score with a set risk threshold. If the prognostic risk score is greater than the risk threshold, the patient's risk level is determined to be a high-risk group; otherwise, it is determined to be a low-risk group, and personalized intervention recommendations are generated.

2. The real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network according to claim 1 is characterized in that: The image preprocessing module includes: A resolution normalization unit, configured to uniformly adjust breast medical images from multiple imaging devices or imaging modalities to a predetermined spatial resolution and grayscale intensity range, and to perform time frame registration and signal intensity normalization on a magnetic resonance dynamic contrast-enhanced image sequence; A noise suppression unit is used to perform a median filter operation on the breast medical image after resolution normalization based on spatial neighborhood statistical information, so as to suppress speckle noise and random interference while maintaining the edge structure; A contrast enhancement unit is used to perform regional brightness redistribution processing on the breast medical image after noise suppression processing based on the image grayscale distribution characteristics, so as to enhance the grayscale contrast between the lesion and the adjacent tissue; The breast medical images processed by the contrast enhancement unit constitute a basic image data set, which is used as input for the subsequent lesion analysis module.

3. The real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network according to claim 2 is characterized in that: The lesion analysis module specifically includes: An image scoring unit, configured to calculate an image quality score for each breast medical image in the basic image dataset, wherein the image quality score is determined based on a combined function of cell nucleus density, image clarity, and background area ratio; An image screening unit, configured to select a breast medical image with the highest image quality score as a representative breast image; a contour extraction unit, configured to perform an edge detection operation on the representative breast image to extract a morphological contour of the lesion; A contour scoring unit, configured to calculate a contour integrity score C based on the lesion morphological contour s , the calculation formula is: C s =α1C closure +α2C continuity +α3(1-C artifact ); Where C closure Indicates the closure of the contour, C continuity Indicates edge continuity, C artifact represents the artifact interference intensity; α1, α2, and α3 are weight coefficients, and satisfy α1+α2+α3=1; The dynamic enhancement analysis unit is used to calculate the enhancement rate of the lesion based on the magnetic resonance dynamic contrast enhancement image sequence corresponding to the representative breast image when the contour integrity score is greater than the preset threshold. The calculation formula is: Where E(t) is the enhancement rate at time t; S(t) is the signal intensity of the lesion area at time t; S(0) is the baseline signal intensity; the joint judgment unit is used to judge the lesion location based on the joint rule of contour integrity score and enhancement rate according to the following logic function: Where L(x,y) is a binary function, which indicates whether the image coordinate position (x,y) belongs to the lesion area, 1 for lesion and 0 for non-lesion; E threshold is the set enhancement rate threshold; When L(x,y) is 1, the preliminary lesion localization information is output; otherwise, a filter parameter adjustment instruction is sent to the image preprocessing module to regenerate the basic image data set to repeat the above lesion localization process; An image feature extraction unit is used to extract lesion image features from representative breast images based on the preliminary lesion positioning information, including grayscale distribution parameters, texture detail index, lesion area, lesion perimeter, lesion boundary curvature and lesion enhancement rate.

4. The real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network according to claim 3 is characterized in that: The gene feature extraction module includes: A gene expression data extraction unit is used to extract gene expression data corresponding to the lesion area from the gene expression database based on the preliminary lesion location information to form a gene expression data set; A gene interaction network construction unit is used to construct an undirected weighted graph model G=(V, E, W) based on the gene expression data set, where V represents a set of gene nodes, E represents a set of connection relationships between genes, and W represents a set of connection edge weights of expression correlations; Gene attribute evaluation unit, used to calculate the gene interaction strength of each gene pair, the connectivity score of each gene node and the expression variability, in order to characterize the network connectivity characteristics and expression stability of each gene; A core gene identification unit is used to identify gene nodes that meet the conditions from the gene node set as core gene nodes based on the rule that the gene interaction strength, connectivity score and expression variability all exceed preset thresholds, thereby forming a core gene node set; A gene correlation analysis unit is used to perform correlation analysis on each gene in the core gene node set and the lesion imaging feature, and calculate the average correlation score of each core gene node; A core distribution structure analysis unit is used to analyze the network concentration of core gene nodes based on the topological distribution of the core gene node set in an undirected weighted graph model; when the network concentration is higher than a preset threshold, the core gene node set is divided into several gene clusters based on a graph clustering algorithm, and the topological feature vectors of each gene cluster are classified to extract the stable structural characteristics of each gene cluster. The topological feature vectors include the average connectivity within the cluster, the edge weight variance, and the node betweenness centrality extreme difference; based on the stable structural characteristics, topological distribution, and DisGeNET disease-gene association database, disease label association information is generated for the gene cluster to which each core gene node belongs; The expression feature output unit is used to extract the expression value, gene interaction strength, connectivity score, expression variability, average correlation score, and disease label association information of each core gene node in the core gene node set as expression feature output.

5. The real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network according to claim 4 is characterized in that: Calculate the gene interaction strength S of each gene pair based on the common neighbor edge weight ij , the formula is: Where w ik 、w jk They represent the edge weights between gene i and gene k, and between gene j and gene k, respectively, and K is the total number of gene nodes; Based on the gene interaction strength S ij Calculate the connectivity score D of each gene node i , the formula is: Calculate the expression variability V of each gene node based on the gene expression data set i , the formula is: in is the expression variance of gene i, μ i is the mean expression value of gene i; The formula for calculating the average correlation score of each core gene node is: Where R i represents the average correlation score of gene i; M is the number of dimensions of lesion imaging features; (g i ,f m ) represents the expression value g of gene i i and the image feature f of the mth lesion m The correlation coefficient between .

6. The real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network according to claim 1, characterized in that: The comorbidity feature extraction module includes: A comorbidity data extraction unit is used to extract a list of confirmed diseases and corresponding visit times, diagnosis codes, treatment records, and clinical status data from the patient's electronic medical record or health file; A comorbidity network mapping unit is used to map each disease diagnosed by the patient to a pre-established comorbidity network model, and obtain the location nodes and connection relationships of the patient's confirmed disease in the comorbidity network; The comorbidity network model is an undirected weighted network model constructed based on population epidemiological statistical data. A node in the network represents a disease, and an edge weight represents the co-occurrence probability between diseases. The undirected weighted network model includes network versions at different time stages to reflect the co-occurrence relationship of diseases in different periods. A comorbidity parameter extraction unit is used to calculate comorbidity characteristic parameters related to breast cancer prognosis based on the node distribution of the patient in the comorbidity network, including: the number of comorbidity nodes, the weight value of the edge connecting to the breast cancer node, and the set of major comorbidity nodes whose median centrality or betweenness centrality of the comorbidity nodes is higher than a set threshold; The comorbidity status updating unit is used to select the disease information corresponding to the current evaluation time point based on the patient's diagnosis and treatment information including multiple time points, and update the above-mentioned comorbidity characteristic parameters to reflect the patient's current comorbidity status.

7. The real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network according to claim 1, characterized in that: The feature fusion processing module includes: The feature matrix construction unit is used to standardize and integrate the lesion imaging features, the expression features of the core gene nodes and the comorbidity feature parameter set to construct a comprehensive feature matrix, in which each column corresponds to a standardized structured feature; the feature importance evaluation unit is used to calculate the weight of the feature dimension of the comprehensive feature matrix based on a trainable attention network to generate a weighted feature vector X = (ω1x1,...,ω d x d ), where x d represents the d-th standardized feature, ω d Represents feature x d Corresponding contribution weight; The weighted feature output unit is used to output the weighted feature vector X as input to the risk prediction module to perform the breast cancer prognosis risk scoring task.

8. The real-time breast cancer prognosis assessment system integrating multimodal imaging and comorbidity network according to claim 5, characterized in that: The risk assessment module includes: A structural relevance scoring unit is used to score the average relevance score R i and connectivity score D i , construct the structural relevance score R score , the formula is: Where G is the number of core gene nodes; The risk score modeling unit is used to combine the weighted feature vector with the structural correlation score value R score Input them into the pre-trained multi-layer perceptron model and output the prognostic risk score y for breast cancer risk ∈[0,1]; The multilayer perceptron model includes at least three layers of fully connected neurons. During the training process, a binary cross entropy loss function or a mean square error loss function is used, and the Adam optimization algorithm is used for parameter update. The learning rate is set to η = 0.001; The training process introduces an early stopping mechanism. When the performance indicators of the model do not improve within eight consecutive training cycles on the validation set, the training is terminated and the model parameter configuration is returned to the optimal performance state.

9. A method for real-time prognosis assessment of breast cancer by integrating multimodal imaging and comorbidity networks, applied to a real-time prognosis assessment system for breast cancer by integrating multimodal imaging and comorbidity networks as claimed in any one of claims 1 to 8, characterized in that: The method comprises: Obtain the patient's original medical images, perform normalization, denoising, and contrast enhancement, and generate a basic image dataset; Extracting lesion contours based on the basic image dataset and calculating a contour integrity score; if the contour integrity score exceeds a preset threshold, analyzing the lesion enhancement pattern in combination with a dynamic contrast-enhanced magnetic resonance imaging sequence, outputting preliminary lesion localization information, and extracting lesion image features; otherwise, returning to the basic image dataset, adjusting filtering parameters, and reprocessing; Extracting gene expression data of the corresponding region based on the preliminary lesion location information, constructing a gene interaction network, identifying core gene nodes with high variability and high connectivity, and extracting expression feature outputs of the core gene nodes; Obtain the patient's comorbidity information, map it to the comorbidity network model, and extract comorbidity characteristic parameters related to breast cancer prognosis; The lesion imaging features, the expression features of the core gene nodes, and the comorbidity feature parameters are integrated to construct a comprehensive feature matrix. The importance of each feature in the comprehensive feature matrix is ​​weighted and evaluated based on the attention mechanism to generate a weighted feature vector. The weighted feature vector is input into a pre-trained multi-layer perceptron model to output a prognostic risk score for breast cancer. The risk score is compared with the set risk threshold. If the prognostic risk score is greater than the risk threshold, the patient's risk level is determined to be a high-risk group, otherwise it is determined to be a low-risk group, and personalized intervention recommendations are generated.

Images