Breast cancer liver metastasis evaluation method and system based on imaging feature analysis

Abstract

The application discloses a breast cancer liver metastasis evaluation method and system based on imageomics feature analysis, which comprises the following steps: extracting an original imageomics feature set through collecting liver area multi-modal image data, constructing a feature time sequence evolution track after screening to obtain a calibration feature subset, generating a high-dimensional fusion feature matrix through feature fusion, combining a professional platform to complete tumor heterogeneity quantification grading, and finally realizing comprehensive evaluation of liver metastasis state and progress trend; the system comprises multiple functional units which are connected in series and interact bidirectionally, and realizes full-process connection from feature collection, screening, modeling, fusion to quantification evaluation and result output. The application comprehensively absorbs multi-dimensional static and dynamic feature information, avoids invasiveness through a non-invasive technical path, provides scientific support for clinical treatment scheme formulation, efficacy monitoring and prognosis judgment, and significantly enhances the precision, systematicness and clinical practicability of breast cancer liver metastasis evaluation.

Description

Technical Field

[0001] This invention relates to the field of breast cancer liver metastasis assessment technology, and in particular to a method and system for assessing breast cancer liver metastasis based on radiomics feature analysis. Background Technology

[0002] Breast cancer is a malignant tumor with a high incidence rate among women. Liver metastasis significantly increases treatment difficulty and affects patient prognosis. Therefore, accurately assessing the status of liver metastases, tracking their progression, and precisely understanding the intrinsic heterogeneity of the tumor are crucial prerequisites for clinical treatment planning and efficacy evaluation. Traditional assessment methods mainly rely on imaging morphology observation and invasive pathological examination, which struggle to quantify and capture the microscopic characteristics of tumors and their dynamic changes over time. Radiomics technology, by deeply mining high-dimensional features in multimodal images, provides a feasible direction for non-invasive and accurate assessment. Current clinical practice urgently needs a systematic solution that integrates multiple technologies to achieve full-process coverage from image feature extraction, screening, dynamic analysis to comprehensive assessment, solving core technical challenges such as efficient utilization of multi-dimensional features, effective capture of dynamic changes, and accurate quantification of heterogeneity.

[0003] Existing technologies suffer from two significant shortcomings: First, the feature processing and evaluation process lacks systematic integration. Most technologies only address a single step, failing to construct a complete technical chain from original feature acquisition, calibrated feature screening, dynamic feature modeling to deep feature fusion. This results in an inability to comprehensively integrate multi-dimensional information such as density distribution, texture structure, and morphological contours in images, as well as feature changes at different time points, leading to insufficient comprehensiveness and reliability of the evaluation results. Second, there is a lack of effective correlation between tumor heterogeneity assessment and metastatic lesion grading. Existing technologies have not established an organic link between heterogeneity quantitative analysis and metastatic lesion grading assessment. They have neither achieved accurate quantification and grading of heterogeneity through a professional platform nor effectively combined heterogeneity assessment results with dynamic and fused features. Consequently, they cannot accurately reflect the impact of heterogeneity on liver metastasis progression and cannot provide clinicians with assessment evidence that combines dynamic tracking capabilities with specificity. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a method and system for assessing liver metastasis of breast cancer based on radiomics feature analysis.

[0005] The technical solution adopted in this invention is a method for assessing liver metastasis of breast cancer based on radiomics feature analysis, comprising the following steps: S1, collecting multimodal image data of the liver region of breast cancer patients through a tumor image heterogeneity analysis platform, and extracting an original radiomics feature set including density distribution, texture structure, morphological contour, and spatial correlation dimensions; S2, performing feature screening on the original radiomics feature set based on a liver metastasis grading radiomics network, removing redundant and noisy features, and obtaining a calibrated radiomics feature subset related to liver metastasis of breast cancer; S3, inputting the calibrated radiomics feature subset into a time-series dynamic... The system employs a dynamic radiomics prediction model to construct a temporal evolution trajectory of features and explore the fluctuation patterns and correlations of features at different time points. In step S4, a gradient-enhanced image feature classification algorithm is used to fuse and enhance the dimensions of the temporal evolution trajectory and the calibrated feature subset, generating a high-dimensional fused feature matrix. In step S5, a tumor image heterogeneity analysis platform is used to perform heterogeneity quantification analysis on the high-dimensional fused feature matrix, dividing the tumor heterogeneity level intervals. Finally, combining the grading results and heterogeneity level intervals output by the liver metastasis grading radiomics network, a comprehensive assessment of the status and progression trend of breast cancer liver metastasis is completed.

[0006] Furthermore, the grading model expression of the liver metastasis grading radiomics network is as follows: ,in, The grading coefficient for liver metastases. These are the network weight parameters. For the first A calibrated image omics feature value For the first The hierarchical influence factors of each characteristic, For feature space correlation function, The entropy value is the characteristic distribution. For the feature clustering parameter, For the first One heterogeneous related feature, For the first The weighting coefficients of each heterogeneous feature. These represent the number of calibrated features and the number of heterogeneous features, respectively.

[0007] Furthermore, the prediction expression of the temporal dynamic image omics prediction model is: ,in, These are predicted values ​​for time-series features. Adjusting parameters for the model, For a moment The calibration characteristic value, For a moment The characteristic attenuation coefficient, For the first The characteristic changes in each time interval For the first Weighting factors for each time interval, This is the initial time point. As the termination point, This represents the total number of time intervals.

[0008] Furthermore, the fusion expression of the gradient boosting image feature classification algorithm is as follows: ,in, The output value of the high-dimensional fusion feature matrix. For the first The output of the gradient boosting tree For the first The confidence coefficients of each gradient boosting tree. For the first The dimensionality enhancement coefficient of each feature. For the first One feature value to be fused, For the first The fusion modulator of each feature To increase the number of gradient boosting trees, The total dimension of the features to be fused.

[0009] Furthermore, the quantitative model expression of the tumor image heterogeneity analysis platform is as follows: ,in, This is a value quantified for heterogeneity. For the first Feature values ​​of a local region The mean of the global features. For the first The weight of each local region This is the heterogeneity amplification factor. For the first A structural heterogeneity characteristic For the first The influence weight of each structural heterogeneity feature To divide the local area into quantities, This represents the number of types of structural heterogeneity features.

[0010] Furthermore, the comprehensive model expression for assessing breast cancer liver metastasis based on radiomics feature analysis is as follows: ,in, This is a comprehensive assessment value for breast cancer liver metastasis. These are the weighting coefficients for the classification results. These are the weighting coefficients for the time series prediction results. The joint weighting coefficients for integrating features and heterogeneity. The grading coefficient for liver metastases. These are predicted values ​​for time-series features. The output value of the high-dimensional fusion feature matrix. This is the value for heterogeneity quantification.

[0011] Further, S3 includes the following sub-steps: S31, based on the calibrated radiomics feature subset extracted by the tumor image heterogeneity analysis platform, arrange the feature data corresponding to each examination time point in time sequence, establish a feature time series database, and clarify the timestamp and acquisition conditions corresponding to each feature data; S32, segment the feature data in the time series database according to a preset time interval, calculate the change amplitude and rate of change of the feature data in each segment, and determine the time nodes and duration of feature fluctuations; S33, input the segmented feature data and fluctuation parameters into the time series dynamic radiomics prediction model, and construct a dynamic trajectory curve of feature evolution over time through the time series correlation calculation module inside the model to capture the change pattern of features at different stages; S34, based on the dynamic trajectory curve, mine the correlation strength and dependency relationship of feature data at adjacent time nodes to form a feature time series correlation matrix, providing time series dimension support for feature fusion.

[0012] Further, step S4 includes the following sub-steps: S41, obtaining the feature temporal correlation matrix output by S3 and the calibration image omics feature subset obtained by S2, and performing dimensional unification processing on the two types of data to ensure that the data structure meets the input requirements of the gradient boosting image feature classification algorithm; S42, initializing the number of decision trees, learning rate, and depth parameters of the gradient boosting image feature classification algorithm, and inputting the dimension-unified feature data into the algorithm model for the first round of feature fusion calculation; S43, based on the first round of fusion results, adjusting the weight coefficients and fusion ratios of each feature through the gradient iteration optimization module of the algorithm, performing multiple rounds of iterative fusion calculation, and gradually improving the recognition and discrimination of the fused features; S44, when the number of iterations reaches a preset threshold or the stability of the fusion result meets the set conditions, stopping the iteration, and outputting the final high-dimensional fusion feature matrix, which includes comprehensive information of temporal features and calibration features.

[0013] Further, step S5 includes the following sub-steps: S51, inputting the high-dimensional fusion feature matrix output from S4 into the tumor image heterogeneity analysis platform, and using the platform's region division module to divide the feature space corresponding to the tumor image into multiple non-overlapping local feature regions; S52, for each local feature region, calculating the distribution variance, mean deviation, and spatial clustering heterogeneity-related parameters of the feature data within the region, and establishing a heterogeneity parameter set for each region; S53, using the platform's parameter fusion module to perform weighted fusion of the heterogeneity parameter sets of each local region to obtain the global heterogeneity quantification value of the tumor image, with the weight coefficients determined based on the feature importance of each region; S54, based on the magnitude of the global heterogeneity quantification value and referring to a preset grading standard, dividing the tumor image heterogeneity into different level intervals, with each interval corresponding to a fixed range of heterogeneity degree.

[0014] A radiomics-based system for assessing breast cancer liver metastases is developed. This system, applied to a radiomics-based approach for assessing breast cancer liver metastases, includes: a radiomics feature acquisition and preprocessing unit, a calibration feature screening and optimization unit, a temporal feature dynamic modeling unit, a high-dimensional feature fusion calculation unit, a tumor heterogeneity quantitative grading unit, and a comprehensive assessment result output unit. These units are sequentially connected and allow for bidirectional data interaction. The radiomics feature acquisition and preprocessing unit collects multimodal imaging data of the liver region from breast cancer patients and extracts the raw radiomics feature set, which is then transmitted to the calibration feature screening and optimization unit. The calibration feature screening and optimization unit filters the raw feature set using a liver metastasis grading radiomics network and outputs calibration radiomics feature subsets. The data is collected by the temporal feature dynamic modeling unit; the temporal feature dynamic modeling unit constructs the feature temporal evolution trajectory based on the temporal dynamic image omics prediction model, and sends the trajectory data to the high-dimensional feature fusion calculation unit; the high-dimensional feature fusion calculation unit uses the gradient boosting image feature classification algorithm to fuse the trajectory data with the calibrated feature subset, generates a high-dimensional fusion feature matrix, and transmits it to the tumor heterogeneity quantification grading unit; the tumor heterogeneity quantification grading unit performs quantitative analysis on the high-dimensional fusion feature matrix through the tumor image heterogeneity analysis platform, divides the heterogeneity level intervals, and feeds them back to the comprehensive evaluation result output unit; the comprehensive evaluation result output unit combines the liver metastasis grading results with the heterogeneity level intervals to complete the assessment of the liver metastasis status and progression trend of breast cancer and outputs the assessment results.

[0015] Beneficial Effects: This invention proposes a method and system for assessing liver metastasis in breast cancer based on radiomics feature analysis. Through a multi-stage, coherent process, it achieves complete coverage from the extraction of original image features, the screening of calibrated features, dynamic feature modeling, to the fusion of high-dimensional features. It comprehensively integrates multi-dimensional static features such as density distribution, texture structure, and morphological contour with dynamic change information at different time points, breaking the limitations of single-stage technology applications and significantly improving the comprehensiveness and reliability of the assessment results. Addressing the lack of effective correlation between tumor heterogeneity assessment and metastatic lesion grading, it achieves precise quantification and grading of tumor heterogeneity through a specialized analysis platform. It deeply integrates the heterogeneity assessment results with dynamic and fused features, clearly reflecting the impact of heterogeneity on liver metastasis progression, providing clinical assessment evidence that combines dynamic tracking capabilities with targeted analysis. Meanwhile, this method and system utilize non-invasive radiomics technology, avoiding the trauma caused by invasive testing. Through in-depth mining and comprehensive analysis of multi-dimensional features, it achieves accurate assessment of the status and progression trend of liver metastases, providing scientific support for treatment plan formulation, efficacy monitoring and prognosis, and significantly improving the accuracy, systematicness and practicality of clinical assessment of breast cancer liver metastases. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0017] Figure 2 This is a flowchart of method step S3 of the present invention;

[0018] Figure 3 This is a flowchart of method step S4 of the present invention;

[0019] Figure 4 This is a flowchart of step S5 of the method of the present invention;

[0020] Figure 5 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, the method for assessing breast cancer liver metastasis based on radiomics feature analysis includes the following steps:

[0023] S1. Multimodal image data of the liver region of breast cancer patients were collected through a tumor image heterogeneity analysis platform, and the original radiomics feature set including density distribution, texture structure, morphological contour and spatial correlation dimensions was extracted.

[0024] Specifically, in step S1, multimodal image data of the liver region of breast cancer patients are collected through a tumor image heterogeneity analysis platform. During the collection process, the image scanning parameters are strictly controlled. The slice thickness of CT images is set to 0.625 to 1.25 mm, the tube voltage is maintained at 120 kV, the tube current is adjusted to 200 to 300 mA / s, and the scanning range covers the entire liver region and the surrounding 2 cm of tissue. For MRI images, T1-weighted, T2-weighted, and enhanced sequences are used, with a repetition time of 500 to 800 ms, an echo time of 10 to 30 ms, and a matrix size of 256×256 to 512×512 to ensure that the image resolution meets the feature extraction requirements. After data acquisition, the platform automatically extracts the original image omics feature set, which includes four dimensions: density distribution, texture structure, morphological contour, and spatial correlation. The density distribution dimension includes 32 features such as pixel gray-level mean, median, and standard deviation; the texture structure dimension includes 48 features derived from the gray-level co-occurrence matrix and gray-level run-length matrix; the morphological contour dimension includes 16 features such as tumor maximum diameter, surface area, volume, and sphericity; and the spatial correlation dimension includes 24 features such as adjacent pixel correlation and regional contrast. In total, there are 120 original features. All feature extraction processes are completed in the platform's built-in feature extraction module. Through batch processing algorithms, synchronous feature extraction of multimodal image data is achieved, ensuring the consistency and completeness of feature extraction and providing comprehensive and high-quality raw data support for subsequent feature selection and model analysis.

[0025] S2, based on the graded radiomics network of liver metastases, feature screening of the original radiomics feature set was performed to remove redundant and noisy features, resulting in a subset of calibrated radiomics features related to liver metastases of breast cancer.

[0026] Specifically, step S2 uses a liver metastasis grading radiomics network to screen features from the original radiomics feature set. The network has a built-in three-layer feature screening module. The first layer uses a correlation analysis algorithm to calculate the Pearson correlation coefficient between any two features, setting a correlation coefficient threshold of 0.85, and removing redundant features with correlation coefficients higher than this threshold, retaining about 85 features after preliminary screening. The second layer uses an analysis of variance algorithm to calculate the variance of each feature, setting a variance threshold of 0.01, and removing low-discrimination features with variance values ​​lower than this threshold, retaining about 60 features after secondary screening. The third layer uses the network's built-in importance assessment module to score the contribution of features to the diagnosis of breast cancer liver metastasis, setting a contribution threshold of 0.6, and screening out labeled features with contributions higher than this threshold, finally obtaining a subset of labeled radiomics features related to breast cancer liver metastasis, with the number of features controlled between 35 and 40. Throughout the screening process, the network learning rate was set to 0.001, the number of iterations was 500, and the batch size was 32. The screening parameters were continuously optimized using the gradient descent algorithm to ensure that the selected calibration features included key information in various dimensions such as density distribution, texture structure, morphological contour, and spatial correlation, while effectively eliminating redundant and noisy features, reducing the computational complexity of subsequent models, improving evaluation efficiency and accuracy, and laying a high-quality data foundation for subsequent time series modeling and feature fusion.

[0027] S3 inputs the calibrated radiomics feature subset into the time-series dynamic radiomics prediction model to construct the feature time-series evolution trajectory and explore the feature fluctuation patterns and correlation patterns at different time nodes.

[0028] Specifically, step S3 inputs the calibrated radiomics feature subset into the time-series dynamic radiomics prediction model. The model first labels the calibrated feature subset with a time dimension, dividing it into five time stages based on the patient's imaging examination time nodes: baseline, 1 month, 3 months, 6 months, and 12 months post-treatment. Each time stage corresponds to a set of calibrated feature data. The model has a built-in time-series trajectory construction module, which uses a time-series interpolation algorithm to supplement the missing feature values ​​between each time stage. The interpolation interval is set to 7 days to ensure the continuity of the feature time-series data. Subsequently, the time-series feature data is segmented using a sliding window algorithm. The window size is set to 3 consecutive time nodes, and the step size is 1 time node. The rate of change, fluctuation amplitude, and trend slope of the feature data within each window are calculated to explore the feature fluctuation patterns at different time nodes. Meanwhile, the model analyzes the interaction relationship of each calibration feature at different time stages through association rule mining algorithm, sets the support threshold to 0.3 and the confidence threshold to 0.7, and selects feature combinations with significant correlation to construct a feature time-series evolution trajectory map. This map includes the time change curve of each feature and the heat map of the correlation strength between features, clearly presenting the dynamic change process and mutual influence pattern of the calibration features over time, providing rich time-series dimension information for subsequent feature fusion, and improving the dynamism and foresight of the evaluation results.

[0029] S4. Gradient boosting image feature classification algorithm is used to perform feature fusion and dimensionality boosting on temporal evolution trajectory and calibration feature subset to generate high-dimensional fusion feature matrix.

[0030] Specifically, step S4 utilizes a gradient boosting image feature classification algorithm to perform feature fusion and dimensionality enhancement on the temporal evolution trajectory and the calibrated feature subset. The algorithm first standardizes the two types of data, mapping parameters such as the feature change rate and fluctuation amplitude of the temporal evolution trajectory to the 0-1 range for each feature value of the calibrated feature subset, ensuring data scale consistency. The algorithm sets the number of decision trees to 100, with a maximum depth of 8 layers per tree, a learning rate of 0.05, and a minimum number of sample splits of 20. A gradient boosting model is constructed through multiple iterations. The first iteration builds an initial decision tree based on the calibrated feature subset. Subsequent iterations correct the prediction error of the previous iteration by introducing temporal evolution trajectory features, gradually improving the model's fitting accuracy. During feature fusion, the algorithm employs a weighted fusion strategy, setting the weight coefficient of the labeled feature subset to 0.6 and the weight coefficient of the time-series evolution trajectory features to 0.4. This weight allocation highlights the importance of different types of features. Simultaneously, the algorithm generates new higher-order features through a feature interaction module. Each higher-order feature is obtained from 2 to 3 basic features through multiplication, summation, and other operations, ultimately generating a high-dimensional fusion feature matrix with a dimension of 256. Throughout the fusion process, the algorithm verifies model stability through a cross-validation mechanism, using 5-fold cross-validation to ensure that the high-dimensional fusion feature matrix includes both the basic information of the labeled features and integrates the dynamic changes of the time-series features, resulting in stronger discriminative power and generalization performance, providing high-quality feature input for subsequent heterogeneity analysis.

[0031] S5 uses a tumor imaging heterogeneity analysis platform to perform heterogeneity quantification analysis on the high-dimensional fusion feature matrix and divide the tumor heterogeneity level intervals.

[0032] Specifically, step S5 uses a tumor image heterogeneity analysis platform to perform heterogeneity quantification analysis on the high-dimensional fusion feature matrix. The platform first divides the image region corresponding to the high-dimensional fusion feature matrix into 10×10 grid units, totaling 100 local regions, each corresponding to a set of high-dimensional fusion feature data. The platform has a built-in heterogeneity quantification module, using three core indicators for quantification: first, the feature dispersion within a region, calculating the standard deviation of feature values ​​within each local region to reflect the uniformity of features within that region; second, the feature difference between regions, calculating the Euclidean distance between feature values ​​of any two adjacent local regions to reflect the degree of differentiation between features in different regions; and third, the global feature entropy value, calculating the information entropy based on the feature distribution of all local regions to reflect the degree of disorder in the overall feature distribution. During the quantification process, the weight of the feature dispersion within a region is set to 0.3, the weight of the feature difference between regions is set to 0.4, and the weight of the global feature entropy value is set to 0.3. The tumor heterogeneity quantification value for each patient is obtained through weighted summation, with the quantification value ranging from 0 to 10. Subsequently, the platform categorized tumor heterogeneity into gradations based on quantitative values: 0-3 for low heterogeneity, 3-6 for medium heterogeneity, and 6-10 for high heterogeneity, each with a clearly defined degree of heterogeneity. Throughout the analysis, the platform employed a parallel computing architecture with 8 threads to ensure high efficiency. Through multi-dimensional quantitative indicators and scientific grading, the platform accurately presented tumor heterogeneity characteristics, providing crucial heterogeneity references for subsequent comprehensive assessments.

[0033] S6 combines the grading results and heterogeneity level intervals output by the liver metastasis grading radiomics network to complete a comprehensive assessment of the status and progression trend of breast cancer liver metastases.

[0034] Specifically, step S6 combines the grading results output by the liver metastasis grading radiomics network with the heterogeneity grading intervals to complete a comprehensive assessment of the status and progression trend of breast cancer liver metastases. First, the grading results output by the liver metastasis grading radiomics network are obtained. These results are divided into four levels (Level 1, Level 2, Level 3, and Level 4) based on the analysis of labeled feature subsets, corresponding to different stages of liver metastasis development: Level 1 for early metastasis, Level 2 for intermediate metastasis, Level 3 for advanced metastasis, and Level 4 for late metastasis. Then, a correlation assessment matrix is ​​established between the grading results and the heterogeneity grading intervals. The matrix horizontally represents the four grading levels and vertically represents the three heterogeneity intervals, with each intersection corresponding to a specific comprehensive assessment criterion. During the assessment, the patient's corresponding grading level and heterogeneity interval are first determined. Then, a comprehensive judgment is made based on the correlation assessment matrix, and the temporal evolution trend of the features output by the temporal dynamic radiomics prediction model is combined to predict the direction of liver metastasis progression in the next 6 to 12 months. For example, patients with a grade II classification and low heterogeneity are assessed as having stable metastatic disease and a low risk of progression; patients with a grade III classification and high heterogeneity are assessed as having rapidly progressing metastatic disease and a high risk of progression. The entire assessment process employs a weighted scoring mechanism, with the grade classification weighted at 0.5, the heterogeneity grade range weighted at 0.3, and the temporal evolution trend weighted at 0.2. A comprehensive assessment score is obtained through weighted calculation, and based on the score, three assessment levels—low risk, intermediate risk, and high risk—are assigned. The final output is a comprehensive assessment report including metastatic disease status, degree of heterogeneity, and progression risk, providing a comprehensive and accurate reference for clinical treatment decisions.

[0035] Preferably, the grading model expression of the liver metastasis grading radiomics network is as follows: ,in, The grading coefficient for liver metastases. These are the network weight parameters. For the first A calibrated image omics feature value For the first The hierarchical influence factors of each characteristic, For feature space correlation function, The entropy value is the characteristic distribution. For the feature clustering parameter, For the first One heterogeneous related feature, For the first The weighting coefficients of each heterogeneous feature. These represent the number of calibrated features and the number of heterogeneous features, respectively.

[0036] Specifically, the grading model of the liver metastasis radiomics network is based on the correlation analysis between the imaging features and grading of breast cancer liver metastases. First, statistical methods are used to determine the linear correlation between labeled radiomics features and metastasis grading. Then, a feature space correlation function is introduced to quantify the nonlinear effects between features. Finally, a logarithmic transformation is used to balance the weighting of heterogeneous features, forming a multi-factor comprehensive grading expression. The model is established because metastasis grading is directly influenced by labeled features and indirectly influenced by heterogeneous features. A dual-weight coefficient is needed to adjust the contribution of these two types of factors. The linear superposition term of the labeled features reflects the basic grading basis, while the combination of the feature space correlation function and the logarithmic term of the heterogeneous features reflects the corrective effect of complex correlations on grading. The parameter values ​​were validated using extensive clinical data. The primary weight in the network weight parameters was set to 0.6, the secondary weight to 0.4, the feature grading influence factor ranging from 0.1 to 0.9 (assigned based on feature importance), the feature distribution entropy ranging from 1 to 5, the feature clustering parameter ranging from 0.2 to 0.8, the heterogeneity feature weight coefficient ranging from 0.3 to 0.7, the number of calibrated features set to 35 to 40, and the number of heterogeneous features set to 20 to 25. During implementation, the selected calibrated and heterogeneous features were first input into the model. The weight parameters were optimized using a gradient descent algorithm with 500 iterations, a learning rate of 0.001, and a batch size of 32. The grading coefficients, ranging from 0 to 10, were calculated, corresponding to the four grading levels of metastatic lesions. The significance of this model lies in achieving accurate grading of metastatic lesions through the comprehensive quantification of multi-dimensional features, providing a core grading basis for subsequent comprehensive evaluation, and solving the problem of insufficient accuracy in grading based on a single feature.

[0037] Preferably, the prediction expression of the time-series dynamic imagemics prediction model is: The prediction expression of the time-series dynamic imagemics prediction model is: ,in, These are predicted values ​​for time-series features. Adjusting parameters for the model, For a moment The calibration characteristic value, For a moment The characteristic attenuation coefficient, For the first The characteristic changes in each time interval For the first Weighting factors for each time interval, This is the initial time point. As the termination point, This represents the total number of time intervals.

[0038] Specifically, the predictive expression of the time-series dynamic radiomics prediction model is based on time series analysis theory. First, it integrates feature values ​​and decay coefficients from different time points through integral operations to reflect the cumulative effect of features over time. Then, it introduces a linear superposition term with feature changes and time interval weights to capture the dynamic fluctuations of features, ultimately forming an integral-form prediction model. The model is based on the fact that the feature changes of breast cancer liver metastases have both continuity and decay characteristics. It needs to consider both the immediate effect of features at individual time points and the cumulative effect of feature changes. The integral term describes the smooth evolution of features over time, while the linear superposition term corrects for the impact of feature mutations. The parameters were calibrated using clinical follow-up data. The primary adjustment parameter was set to 0.7, the secondary adjustment parameter to 0.3, and the feature decay coefficient ranged from 0.8 to 0.95, gradually decreasing over time. Feature change was calculated using the difference in feature values ​​between adjacent time points. The time interval weighting factor ranged from 0.1 to 0.5, increasing with the time interval. The initial time point was set to the patient's first examination, and the termination time point was set to the last follow-up. The total number of time intervals was determined based on the number of follow-ups, typically 4 to 5. During implementation, the calibrated feature subset was first sorted by time point, missing values ​​were supplemented, and then feature change and decay coefficients were calculated. These were then substituted into the model for integration, with an integration step size of 1 day. The predicted value was obtained through a numerical integration algorithm. The significance of this model lies in accurately capturing the temporal evolution of features, predicting the progression trend of metastatic lesions in advance, providing a prospective basis for clinical treatment adjustments, and overcoming the deficiency of traditional static assessments in reflecting dynamic changes.

[0039] Preferably, the fusion expression of the gradient boosting image feature classification algorithm is: ,in, The output value of the high-dimensional fusion feature matrix. For the first The output of the gradient boosting tree For the first The confidence coefficients of each gradient boosting tree. For the first The dimensionality enhancement coefficient of each feature. For the first One feature value to be fused, For the first The fusion modulator of each feature To increase the number of gradient boosting trees, The total dimension of the features to be fused.

[0040] Specifically, the fusion expression of the gradient boosting image feature classification algorithm is based on the gradient boosting tree principle. First, it integrates the outputs and confidence coefficients of multiple gradient boosting trees through a product operation, reflecting the ensemble effect of the base model. Then, it introduces the arctangent function to perform a nonlinear transformation on the product of the feature values ​​and the fusion adjustment factor, enhancing the feature discriminative power. Finally, it forms the fused output value through linear superposition. The model is based on the principle that the fusion of calibration features and temporal features needs to balance the stability of the base model with the nonlinear correlation of features. Gradient boosting tree ensemble can improve the accuracy of base predictions, and the nonlinear transformation of the arctangent function can strengthen feature differences. The combination of the two achieves deep feature fusion. Parameter values ​​were determined through cross-validation. The number of gradient boosting trees was set to 100, with the confidence coefficient of each tree ranging from 0.8 to 0.98, assigned based on the prediction accuracy of the trees. The feature dimension boosting coefficient ranged from 0.5 to 1.2, the fusion adjustment factor ranged from 0.4 to 0.9, and the total dimension of the features to be fused was set to 256. The number of gradient boosting trees was adjusted according to the feature dimension. In the implementation process, the gradient boosting tree parameters are first initialized, and the feature data with unified dimensions is input into the algorithm. A decision tree is constructed in each iteration, the prediction error is calculated, and the feature weights are adjusted. The number of iterations is set to 100, the learning rate is 0.05, and the maximum depth of the decision tree is 8 layers. Finally, the fused output value is obtained through product and linear superposition operations. The significance of this model lies in achieving efficient fusion of statically labeled features and dynamic temporal features, improving the discriminative power and generalization performance of features, providing a high-quality, high-dimensional feature matrix for heterogeneity analysis, and solving the problems of insufficient feature dimensionality and inadequate fusion.

[0041] Preferably, the quantitative model expression of the tumor image heterogeneity analysis platform is: ,in, This is a value quantified for heterogeneity. For the first Feature values ​​of a local region The mean of the global features. For the first The weight of each local region This is the heterogeneity amplification factor. For the first A structural heterogeneity characteristic For the first The influence weight of each structural heterogeneity feature To divide the local area into quantities, This represents the number of types of structural heterogeneity features.

[0042] Specifically, the quantitative model expression of the tumor image heterogeneity analysis platform is based on statistical heterogeneity theory. First, it quantifies the difference between local region features and the global mean through variance calculation, reflecting intra-regional heterogeneity. Then, an exponential function is introduced to amplify the influence of structural heterogeneity features. Finally, a weighted summation is used to obtain the global heterogeneity quantification value. The model is based on the premise that tumor heterogeneity includes two levels: intra-regional homogeneity and inter-regional variability. The variance term describes the degree of feature dispersion within a region, while the exponential term reinforces the role of structural heterogeneity. The combination of these two terms achieves comprehensive quantification of heterogeneity. Parameter values ​​are determined through statistical analysis of image data. The weight of local region features ranges from 0.2 to 0.6, assigned according to the importance of features within the region. The heterogeneity amplification coefficient is set to 0.5, and the influence weight of structural heterogeneity features ranges from 0.3 to 0.8. The number of local region divisions is set to 100, using a 10×10 grid. The number of structural heterogeneity feature types is set to 15 to 20, including dimensions such as texture and morphology. In the implementation process, the image regions corresponding to the high-dimensional fusion feature matrix are first gridded, and the feature values ​​and global mean of each region are calculated to obtain the variance value. Then, structural heterogeneity features are extracted and weighted sums are calculated. These are substituted into the model, and quantitative values ​​are obtained through exponential operations and weighted summation. The significance of this model lies in achieving accurate quantification and hierarchical classification of tumor heterogeneity, clearly presenting the degree of differentiation within the tumor, providing key heterogeneity references for comprehensive assessment, and solving the problems of qualitative and vague traditional heterogeneity assessment.

[0043] The preferred expression for the comprehensive model for assessing breast cancer liver metastasis based on radiomics feature analysis is as follows: ,in, This is a comprehensive assessment value for breast cancer liver metastasis. These are the weighting coefficients for the classification results. These are the weighting coefficients for the time series prediction results. The joint weighting coefficients for integrating features and heterogeneity. The grading coefficient for liver metastases. These are predicted values ​​for time-series features. The output value of the high-dimensional fusion feature matrix. This is the value for heterogeneity quantification.

[0044] Specifically, the comprehensive assessment model for breast cancer liver metastasis is based on multi-indicator comprehensive evaluation theory. It integrates grading coefficients, time-series predicted values, fusion feature output values, and heterogeneity quantification values ​​through linear weighted summation. Each indicator is assigned a corresponding weight coefficient to reflect its contribution to the comprehensive assessment. The model is established based on the fact that breast cancer liver metastasis assessment requires comprehensive consideration of four core factors: metastatic lesion grading, temporal progression trend, feature fusion information, and degree of heterogeneity. These factors are interrelated and complementary, and the linear weighting form can intuitively reflect the importance proportion of each factor, ensuring the objectivity and comprehensiveness of the assessment results. Parameter values ​​were determined through clinical efficacy data validation and analytic hierarchy process (AHP). The weight coefficient for grading results was set to 0.4, the weight coefficient for time-series predicted results was set to 0.25, and the combined weight coefficient for fusion features and heterogeneity was set to 0.35. The grading coefficient ranges from 0 to 10, the time-series predicted value ranges from 5 to 15, the fusion feature output value ranges from 0 to 20, and the heterogeneity quantification value ranges from 0 to 10. During implementation, the grading coefficients, time-series predicted values, fusion feature output values, and heterogeneity quantification values ​​output by each preceding model are first obtained. After standardization, these are input into the model, and a weighted sum is used to obtain a comprehensive evaluation value, ranging from 0 to 20, corresponding to low, medium, and high risk levels. The significance of this model lies in integrating multi-dimensional evaluation indicators to achieve a comprehensive determination of the status and progression trend of breast cancer liver metastasis, providing a scientific and accurate comprehensive reference for the formulation of clinical treatment plans, and overcoming the shortcomings of the one-sidedness of single-indicator evaluation.

[0045] Preferred, such as Figure 2 As shown, S3 includes the following sub-steps: S31, based on the calibrated radiomics feature subset extracted by the tumor image heterogeneity analysis platform, the feature data corresponding to each examination time point are arranged in time series order to establish a feature time series database, and the timestamps and acquisition conditions corresponding to each feature data are clarified; S32, the feature data in the time series database are segmented according to a preset time interval, the change amplitude and change rate of the feature data in each segment are calculated, and the time nodes and duration of feature fluctuations are determined; S33, the segmented feature data and fluctuation parameters are input into the time series dynamic radiomics prediction model, and the dynamic trajectory curve of feature evolution over time is constructed through the time series correlation calculation module inside the model to capture the change law of features at different stages; S34, based on the dynamic trajectory curve, the correlation strength and dependency relationship of feature data at adjacent time nodes are mined to form a feature time series correlation matrix, providing time series dimension support for feature fusion.

[0046] Specifically, step S3 achieves precise construction of the feature time-series evolution trajectory through four sub-steps. S31 first relies on the calibrated radiomics feature subset extracted by the tumor imaging heterogeneity analysis platform, sorting the feature data corresponding to each examination time point according to the chronological order of the patient's imaging examinations, and establishing a feature time-series database including feature name, feature value, acquisition timestamp, and scanning parameters. The timestamp is accurate to the hour, and the scanning parameters record key indicators such as tube voltage, tube current, and slice thickness to ensure data traceability and repeatability. S32 segments the feature data in the time-series database according to a preset 30-day fixed time interval, calculating the absolute change amplitude and daily average change rate of each feature data within each segment. By setting a change amplitude threshold of 0.2 and a change rate threshold of 0.005, the key time nodes and duration of feature fluctuations are identified, with periodic statistics accurate to the nearest whole number. S33 inputs the segmented feature data and fluctuation parameters into the time-series dynamic image omics prediction model. The time-series correlation calculation module inside the model adopts the sliding window algorithm, with the window size set to 3 consecutive time nodes and the step size to 1 time node. Through multiple rounds of iteration, a dynamic trajectory curve of feature evolution over time is constructed, with the curve sampling interval set to 1 day to ensure the continuity and smoothness of the trajectory. S34, based on the dynamic trajectory curve, uses the correlation coefficient analysis method to calculate the correlation strength of feature data at adjacent time nodes, sets the correlation strength threshold to 0.7, filters out strongly correlated feature pairs and forms a feature time-series correlation matrix. The matrix dimension is consistent with the number of calibrated features, providing comprehensive time-series dimensional support for feature fusion in the subsequent step S4. This step-by-step system, through progressive processing, realizes in-depth mining and pattern capture of time-series features, improving the dynamism and accuracy of subsequent evaluation.

[0047] Preferred, such as Figure 3 As shown, step S4 includes the following sub-steps: S41, obtaining the feature temporal correlation matrix output by S3 and the calibration image omics feature subset obtained by S2, and performing dimensional unification processing on the two types of data to ensure that the data structure meets the input requirements of the gradient boosting image feature classification algorithm; S42, initializing the number of decision trees, learning rate, and depth parameters of the gradient boosting image feature classification algorithm, and inputting the dimension-unified feature data into the algorithm model for the first round of feature fusion calculation; S43, based on the first round of fusion results, adjusting the weight coefficients and fusion ratios of each feature through the gradient iteration optimization module of the algorithm, performing multiple rounds of iterative fusion calculation, and gradually improving the recognition and discrimination of the fused features; S44, when the number of iterations reaches a preset threshold or the stability of the fusion result meets the set conditions, stopping the iteration, and outputting the final high-dimensional fusion feature matrix, which includes comprehensive information of temporal features and calibration features.

[0048] Specifically, step S4 constructs the high-dimensional fusion feature matrix through four sub-steps. S41 first obtains the feature temporal correlation matrix output in step S3 and the calibrated radiomics feature subset obtained in step S2. Data normalization is used to map the numerical range of the two types of data to the interval 0 to 1. The element values ​​of the feature temporal correlation matrix are normalized column-wise, and the calibrated radiomics feature subset is normalized by feature term, ensuring that the data structure meets the input requirements of the gradient boosting image feature classification algorithm. During normalization, the precision is set to retain four decimal places. S42 initializes the key parameters of the gradient boosting image feature classification algorithm: the number of decision trees is set to 100, the learning rate is set to 0.05, the maximum depth of the decision trees is 8 layers, the minimum number of sample splits is 20, and the minimum number of leaf nodes is 10. The unified feature data is then divided into training and validation sets in an 8:2 ratio and input into the algorithm model for the first round of feature fusion. S43: Based on the first round of fusion results, the prediction error is calculated through the gradient iteration optimization module of the algorithm. The gradient descent method is used to adjust the weight coefficients and fusion ratios of each feature. The step size for adjusting the weight coefficients is set to 0.01. The fusion ratio is dynamically optimized according to the initial ratio of 0.6 for the calibration features and 0.4 for the temporal features. 100 rounds of iterative fusion calculations are performed. Intermediate results are output every 10 rounds of iterations and the stability is verified. S44: The iteration number threshold is set to 100 times and the stability threshold is set to 0.001. When the number of iterations reaches the threshold or the fluctuation amplitude of the intermediate results of 5 consecutive rounds is less than the stability threshold, the iteration is stopped. A high-dimensional fusion feature matrix with a dimension of 256 is output. This matrix includes the dynamic change information of the temporal features and the basic information of the calibration features. Through step-by-step parameter optimization and iterative fusion, the recognition and discrimination of the fusion features are significantly improved, providing high-quality input for the heterogeneity analysis in step S5.

[0049] Preferred, such as Figure 4 As shown, step S5 includes the following sub-steps: S51, inputting the high-dimensional fusion feature matrix output from S4 into the tumor image heterogeneity analysis platform, and dividing the feature space corresponding to the tumor image into multiple non-overlapping local feature regions through the platform's region division module; S52, for each local feature region, calculating the distribution variance, mean deviation, and spatial clustering heterogeneity-related parameters of the feature data within the region, and establishing a heterogeneity parameter set for each region; S53, through the platform's parameter fusion module, performing weighted fusion on the heterogeneity parameter sets of each local region to obtain the global heterogeneity quantification value of the tumor image, with the weight coefficient determined based on the feature importance of each region; S54, according to the magnitude of the global heterogeneity quantification value, referring to the preset grading standard, dividing the tumor image heterogeneity into different level intervals, with each interval corresponding to a fixed range of heterogeneity degree.

[0050] Specifically, step S5 performs quantitative grading of tumor heterogeneity. In S51, the high-dimensional fusion feature matrix output from step S4 is input into the tumor image heterogeneity analysis platform. The platform's region division module uses an adaptive grid division algorithm to divide the feature space corresponding to the tumor image into 100 non-overlapping local feature regions with a grid density of 10×10. The boundary of each region is adaptively adjusted according to the tumor contour to ensure that the region division conforms to the actual morphology of the tumor. In S52, for each local feature region, three heterogeneity-related parameters are calculated: distribution variance, mean deviation, and spatial clustering. The distribution variance is calculated using the sample variance formula; the mean deviation is the absolute difference between the mean of the feature data within the region and the global mean; and the spatial clustering is calculated as the ratio of the number of clusters of feature values ​​within the region to the area of ​​the region. All parameters are accurate to six decimal places. In S53, the process continues... The platform's parameter fusion module performs weighted fusion of heterogeneity parameter sets for each local region according to a weighting ratio of 0.3 for distribution variance, 0.4 for mean deviation, and 0.3 for spatial clustering, obtaining a heterogeneity sub-quantification value for each local region. Then, regional weights are assigned according to the proportion of tumor tissue in each local region, with regional weights ranging from 0.005 to 0.02. The weighted summation yields the global heterogeneity quantification value of the tumor image. S54 presets a heterogeneity grading standard: a global heterogeneity quantification value of 0 to 3.5 is the low heterogeneity interval, 3.5 to 7 is the medium heterogeneity interval, and 7 to 10 is the high heterogeneity interval. Each interval corresponds to a clear description of the degree of heterogeneity and a defined range of quantification values. Through step-by-step regional division, parameter calculation, fusion quantification, and grading, accurate quantitative assessment of tumor heterogeneity is achieved, providing a key basis for the comprehensive assessment in step S6.

[0051] like Figure 5As shown, a breast cancer liver metastasis assessment system based on radiomics feature analysis is applied to a breast cancer liver metastasis assessment method based on radiomics feature analysis. The system includes: a radiomics feature acquisition and preprocessing unit, a calibration feature screening and optimization unit, a temporal feature dynamic modeling unit, a high-dimensional feature fusion calculation unit, a tumor heterogeneity quantitative grading unit, and a comprehensive assessment result output unit. These units are sequentially connected and allow for bidirectional data interaction. The radiomics feature acquisition and preprocessing unit collects multimodal imaging data of the liver region from breast cancer patients and extracts the original radiomics feature set, which is then transmitted to the calibration feature screening and optimization unit. The calibration feature screening and optimization unit filters the original feature set through a liver metastasis grading radiomics network and outputs the calibration radiomics feature set. The system first collects a subset of data and sends it to the temporal feature dynamic modeling unit. The temporal feature dynamic modeling unit constructs a feature temporal evolution trajectory based on a temporal dynamic image omics prediction model and sends the trajectory data to the high-dimensional feature fusion calculation unit. The high-dimensional feature fusion calculation unit uses a gradient-enhanced image feature classification algorithm to fuse the trajectory data with the calibrated feature subset, generating a high-dimensional fusion feature matrix, which is then transmitted to the tumor heterogeneity quantification and grading unit. The tumor heterogeneity quantification and grading unit performs quantitative analysis of the high-dimensional fusion feature matrix through a tumor image heterogeneity analysis platform, divides the heterogeneity level intervals, and feeds them back to the comprehensive evaluation result output unit. The comprehensive evaluation result output unit combines the liver metastasis grading results with the heterogeneity level intervals to complete the assessment of the status and progression trend of breast cancer liver metastasis and outputs the assessment results.

[0052] A method and system for assessing liver metastasis in breast cancer based on radiomics feature analysis addresses the lack of systematic integration in feature processing and assessment. It achieves full-process coverage from raw feature extraction, calibration feature screening, dynamic feature modeling to high-dimensional feature fusion through a multi-stage progressive operation. It comprehensively incorporates multi-dimensional static information such as density distribution, texture structure, and morphological contours from images, as well as dynamic change data at different time points. This breaks through the fragmented limitations of single-stage technology applications and completely solves the problem of traditional methods' difficulty in integrating multi-dimensional and multi-temporal features, making the assessment results more comprehensive and reliable. Addressing the lack of effective correlation between tumor heterogeneity assessment and metastatic lesion grading, it uses a professional analysis platform to accurately quantify and classify tumor heterogeneity, and deeply integrates heterogeneity assessment results with dynamic and fused features. This clearly presents the impact mechanism of heterogeneity on liver metastasis progression, making up for the lack of connection between the two in traditional techniques and providing clinical assessment evidence with both dynamic tracking value and targeted application.

[0053] This method and system utilize a non-invasive radiomics approach, avoiding the physical trauma and risks to patients caused by invasive testing, thus improving the safety and acceptability of clinical applications. Through in-depth mining and comprehensive analysis of multi-dimensional features, it achieves accurate judgment of the status and progression trend of liver metastases, providing scientific and specific support for treatment plan formulation, dynamic monitoring of efficacy, and prognostic risk assessment. The construction of a complete technical system enables efficient data flow and deep integration at each stage, significantly improving the systematicness and standardization of the assessment process and effectively reducing the errors that may occur in a single technical step. By incorporating dynamic change capture and heterogeneity quantification into the comprehensive assessment, the assessment results are more in line with the development pattern of tumor diseases, significantly improving the accuracy, practicality, and clinical translational value of clinical assessment of breast cancer liver metastases.

[0054] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0055] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing liver metastasis of breast cancer based on radiomics feature analysis, characterized in that, Includes the following steps: S1. Multimodal image data of the liver region of breast cancer patients were collected through a tumor image heterogeneity analysis platform, and the original radiomics feature set including density distribution, texture structure, morphological contour and spatial correlation dimensions was extracted. S2, based on the graded radiomics network of liver metastases, feature screening of the original radiomics feature set was performed to remove redundant and noisy features, resulting in a subset of calibrated radiomics features related to liver metastases of breast cancer. S3 inputs the calibrated radiomics feature subset into the time-series dynamic radiomics prediction model to construct the feature time-series evolution trajectory and explore the feature fluctuation patterns and correlation patterns at different time nodes. S4. Gradient boosting image feature classification algorithm is used to perform feature fusion and dimensionality boosting on temporal evolution trajectory and calibration feature subset to generate high-dimensional fusion feature matrix. S5 uses a tumor imaging heterogeneity analysis platform to perform heterogeneity quantification analysis on the high-dimensional fusion feature matrix and divide the tumor heterogeneity level intervals. S6, combining the grading results and heterogeneity grade intervals output by the liver metastasis grading radiomics network, completes a comprehensive assessment of the status and progression trend of breast cancer liver metastasis. The grading model expression of the liver metastasis grading radiomics network is as follows: ,in, The grading coefficient for liver metastases. These are the network weight parameters. For the first A calibrated image omics feature value For the first The hierarchical influence factors of each characteristic, For feature space correlation function, The entropy value is the characteristic distribution. For the feature clustering parameter, For the first One heterogeneous related feature, For the first The weighting coefficients of each heterogeneous feature. These represent the number of calibrated features and the number of heterogeneous features, respectively. The prediction expression of the time-series dynamic image omics prediction model is: ,in, These are predicted values ​​for time-series features. , Adjusting parameters for the model, For a moment The calibration characteristic value, For a moment The characteristic attenuation coefficient, For the first The characteristic changes in each time interval For the first Weighting factors for each time interval, This is the initial time point. As the termination point, This represents the total number of time intervals. The fusion expression of the gradient boosting image feature classification algorithm is as follows: ,in, The output value of the high-dimensional fusion feature matrix. For the first The output of the gradient boosting tree For the first The confidence coefficients of each gradient boosting tree. For the first The dimensionality enhancement coefficient of each feature. For the first One feature value to be fused, For the first The fusion modulator of each feature To increase the number of gradient boosting trees, The total dimension of the features to be fused; The quantitative model expression for the tumor image heterogeneity analysis platform is as follows: ,in, This is a value quantified for heterogeneity. For the first Feature values ​​of a local region The mean of the global features. For the first The weight of each local region This is the heterogeneity amplification factor. For the first A structural heterogeneity characteristic For the first The influence weight of each structural heterogeneity feature To divide the local area into quantities, This represents the number of types of structural heterogeneity features.

2. The method for assessing breast cancer liver metastasis based on radiomics feature analysis according to claim 1, characterized in that, The comprehensive model expression for assessing breast cancer liver metastasis based on radiomics feature analysis is as follows: ,in, This is a comprehensive assessment value for breast cancer liver metastasis. These are the weighting coefficients for the classification results. These are the weighting coefficients for the time series prediction results. The joint weighting coefficients for integrating features and heterogeneity. The grading coefficient for liver metastases. These are predicted values ​​for time-series features. The output value of the high-dimensional fusion feature matrix. This is the value for heterogeneity quantification.

3. The method for assessing breast cancer liver metastasis based on radiomics feature analysis according to claim 1, characterized in that, S3 includes the following sub-steps: S31, based on the calibrated radiomics feature subset extracted by the tumor image heterogeneity analysis platform, arrange the feature data corresponding to each examination time point in time sequence, establish a feature time series database, and clarify the timestamp and acquisition conditions corresponding to each feature data; S32, segment the feature data in the time series database according to a preset time interval, calculate the change amplitude and change rate of the feature data in each segment, and determine the time nodes and duration of feature fluctuations; S33, input the segmented feature data and fluctuation parameters into the time series dynamic radiomics prediction model, and construct a dynamic trajectory curve of feature evolution over time through the time series correlation calculation module inside the model to capture the change pattern of features at different stages; S34, based on the dynamic trajectory curve, mine the correlation strength and dependency relationship of feature data at adjacent time nodes to form a feature time series correlation matrix, providing time series dimension support for feature fusion.

4. The method for assessing breast cancer liver metastasis based on radiomics feature analysis according to claim 1, characterized in that, S4 includes the following sub-steps: S41, obtaining the feature temporal correlation matrix output by S3 and the calibration image omics feature subset obtained by S2, and performing dimensional unification processing on the two types of data to ensure that the data structure meets the input requirements of the gradient boosting image feature classification algorithm; S42, initializing the number of decision trees, learning rate, and depth parameters of the gradient boosting image feature classification algorithm, and inputting the dimension-unified feature data into the algorithm model for the first round of feature fusion calculation; S43, based on the first round of fusion results, adjusting the weight coefficients and fusion ratios of each feature through the gradient iteration optimization module of the algorithm, performing multiple rounds of iterative fusion calculation, and gradually improving the recognition and discrimination of the fused features; S44, when the number of iterations reaches a preset threshold or the stability of the fusion result meets the set conditions, stopping the iteration, and outputting the final high-dimensional fusion feature matrix, which includes comprehensive information of temporal features and calibration features.

5. The method for assessing breast cancer liver metastasis based on radiomics feature analysis according to claim 1, characterized in that, S5 includes the following sub-steps: S51, inputting the high-dimensional fusion feature matrix output from S4 into the tumor image heterogeneity analysis platform, and using the platform's region division module to divide the feature space corresponding to the tumor image into multiple non-overlapping local feature regions; S52, for each local feature region, calculating the distribution variance, mean deviation, and spatial clustering heterogeneity-related parameters of the feature data within the region, and establishing a heterogeneity parameter set for each region; S53, using the platform's parameter fusion module to perform weighted fusion of the heterogeneity parameter sets of each local region to obtain the global heterogeneity quantification value of the tumor image, with the weight coefficients determined based on the feature importance of each region; S54, based on the magnitude of the global heterogeneity quantification value and referring to a preset grading standard, dividing the tumor image heterogeneity into different level intervals, with each interval corresponding to a fixed range of heterogeneity degree.

6. A breast cancer liver metastasis assessment system based on radiomics feature analysis, characterized in that, This system is applied to the breast cancer liver metastasis assessment method based on radiomics feature analysis as described in claim 1, comprising: a radiomics feature acquisition and preprocessing unit, a calibration feature screening and optimization unit, a time-series feature dynamic modeling unit, a high-dimensional feature fusion calculation unit, a tumor heterogeneity quantitative grading unit, and a comprehensive assessment result output unit. These units are sequentially connected and interact bidirectionally. The radiomics feature acquisition and preprocessing unit acquires multimodal image data of the liver region from breast cancer patients and extracts the original radiomics feature set, transmitting the original feature set to the calibration feature screening and optimization unit. The calibration feature screening and optimization unit filters the original feature set through a liver metastasis grading radiomics network and outputs a calibration radiomics feature subset to the time-series feature dynamic modeling unit. The system consists of a dynamic modeling unit and a temporal feature dynamic modeling unit. The dynamic modeling unit constructs a feature temporal evolution trajectory based on a temporal dynamic image omics prediction model and sends the trajectory data to the high-dimensional feature fusion calculation unit. The high-dimensional feature fusion calculation unit uses a gradient boosting image feature classification algorithm to fuse the trajectory data with a calibrated feature subset, generating a high-dimensional fusion feature matrix and transmitting it to the tumor heterogeneity quantification and grading unit. The tumor heterogeneity quantification and grading unit performs quantitative analysis on the high-dimensional fusion feature matrix through a tumor image heterogeneity analysis platform, divides the heterogeneity level intervals, and feeds them back to the comprehensive evaluation result output unit. The comprehensive evaluation result output unit combines the liver metastasis grading results with the heterogeneity level intervals to complete the assessment of the status and progression trend of breast cancer liver metastasis and outputs the assessment results.

Images