Intelligent grading system for severity of gloomy integrated with image omics

Abstract

The invention relates to the field of medical image information, in particular to an intelligent grading system for the severity of a protruding eye integrated with imageomics. Comprising a data acquisition module used for acquiring clinical features and medical image features; the data processing module is used for determining dimension-reduced clinical features and dimension-reduced medical image features; constructing a fusion analysis matrix; calculating the standardization hindrance of each dimension-reduced medical image feature; determining clinical characteristics of the main components; constructing a stage fusion feature matrix, and calculating principal component standard hindrance; obtaining a final fusion feature matrix; and the data classification module is used for determining a reference vector according to the final fusion feature matrix, combining the dimensionality reduction medical image feature and the principal component clinical feature of each patient into a personal feature vector, and inputting the reference vector and the personal feature vector into a trained intelligent classification model. And outputting the classification level of each patient through the intelligent classification model. The method can improve the grading effect on the patients suffering from the eye protrusion.

Description

Technical Field

[0001] This invention relates to the field of medical imaging information, specifically to an intelligent grading system for the severity of exophthalmos that integrates radiomics. Background Technology

[0002] Thyroid eye disease is an autoimmune disorder associated with thyroid dysfunction (most commonly hyperthyroidism, especially Graves' disease). Its core mechanism involves the immune system producing antibodies (primarily TSH receptor antibodies) that attack the thyroid and orbital tissues, leading to autoimmune inflammation of the orbital tissues and muscles. This causes the eyes to bulge outwards, resulting in exophthalmos (protruding eye). Thyroid eye disease is mainly divided into active and inactive phases. The treatment methods and difficulties vary depending on the stage of the disease. Therefore, to effectively treat patients with thyroid eye disease, it is usually necessary to perform an eye examination and use radiomics to classify the severity of the exophthalmos.

[0003] Traditional radiomics refers to the process of converting standard medical images (such as CT, MRI, and PET) into high-dimensional, mineable quantitative data, and then analyzing them to support decision-making. Essentially, it is about "translating" images into data that can be statistically analyzed and modeled. In actual use, relying solely on extracting information from medical images cannot fully represent the entire nature of thyroid eye disease, and it also obscures the biological significance of the same texture feature changes in medical images. Therefore, existing technologies are considering integrating clinical data with radiomics to build powerful predictive models that enable intelligent grading of the severity of patients' exophthalmos.

[0004] In the process of data fusion, both clinical data and medical images contain a large number of features. Direct fusion can easily introduce redundant noise, causing the "curse of dimensionality", increasing the risk of model overfitting, and reducing the interpretability and robustness of the model. Summary of the Invention

[0005] This invention provides an intelligent grading system for exophthalmos severity that integrates radiomics to address existing problems.

[0006] The intelligent grading system for exophthalmos severity integrating radiomics of the present invention adopts the following technical solution: One embodiment of the present invention provides an intelligent grading system for exophthalmos severity that integrates radiomics, comprising: The data acquisition module is used to acquire the clinical and medical imaging characteristics of each patient. The data processing module is used to determine dimensionality-reduced clinical features and dimensionality-reduced medical imaging features from clinical features and medical imaging features for each patient. Based on the reduced-dimensional clinical features and reduced-dimensional medical image features, a fusion analysis matrix is ​​constructed for each reduced-dimensional medical image feature; based on the fusion analysis matrix for each reduced-dimensional medical image feature, the standardization barrier of each reduced-dimensional medical image feature is calculated. Based on the standardized barrier properties of each dimensionality-reduced medical image feature, principal component clinical features are determined from the dimensionality-reduced clinical features; based on the principal component clinical features and the dimensionality-reduced medical image features, a stage fusion feature matrix is ​​constructed, and the principal component standardized barrier properties of the stage fusion feature matrix are calculated. If the standard barrier of the principal component is less than the preset barrier threshold, the stage fusion feature matrix is ​​determined as the final fusion feature matrix; otherwise, the clinical features of the principal component are standardized, and the final fusion feature matrix is ​​constructed based on the standardized clinical features of the principal component. The data classification module is used to determine the baseline vector based on the final fused feature matrix, and to combine the dimensionality-reduced medical image features and principal component clinical features of each patient into a personal feature vector. The baseline vector and personal feature vector are then input into the trained intelligent grading model, which outputs the classification level of each patient.

[0007] Optionally, in the data processing module, for each patient, dimensionality-reduced clinical features and dimensionality-reduced medical imaging features are determined from clinical features and medical imaging features, specifically including: For each patient, a correlation standard feature matrix is ​​constructed based on clinical characteristics and medical imaging characteristics. Based on the correlation standard feature matrix of each patient, reference clinical characteristics and reference medical imaging characteristics are determined from the clinical characteristics and medical imaging characteristics, respectively. Calculate the first correlation degree of each reference clinical feature and the second correlation degree of each reference medical image feature. Based on the first preset correlation threshold and the second preset correlation threshold, determine the dimensionality-reduced clinical features and dimensionality-reduced medical image features from the reference clinical features and reference medical image features, respectively. The first preset correlation threshold and the second preset correlation threshold are in the range of (0,1). The first preset correlation threshold is related to the reference clinical features, and the second preset correlation threshold is related to the reference medical image features.

[0008] Optionally, in the data processing module, for each patient, a correlation standard feature matrix is ​​constructed based on clinical characteristics and medical imaging characteristics. Based on the correlation standard feature matrix for each patient, reference clinical features and reference medical imaging features are determined from the clinical characteristics and medical imaging characteristics, respectively, including: For each patient, each clinical feature is combined with a medical imaging feature to form a data pair, and each clinical feature data pair is used as an element in each row of the matrix to construct an initial feature matrix. The initial feature matrix is ​​then standardized to obtain an initial standard feature matrix. For each patient, the Pearson correlation coefficient between each clinical feature and the medical imaging feature is calculated. The elements in the initial standard feature matrix are then repositioned according to the Pearson correlation coefficient of each clinical feature to obtain the associated standard feature matrix. In the associated standard feature matrix, the number of rows of data pairs of clinical features is directly proportional to the size of the Pearson correlation coefficient of the clinical features. In the association standard feature matrix of each patient, clinical features with Pearson correlation coefficients greater than a preset first correlation threshold are identified as reference clinical features, and reference medical imaging features are obtained; the preset first correlation threshold is determined by the top 70% of clinical features.

[0009] Optionally, in the data processing module, the Pearson correlation coefficient between each clinical feature and the medical imaging feature is calculated, specifically including: Each medical image feature is used as an element in the data sequence to obtain the image feature sequence; Using the a-th clinical feature as an element in the data sequence, we obtain the clinical feature sequence of the a-th clinical feature, where the image feature sequence and the clinical feature sequence have the same length. Obtain the clinical feature sequence for each clinical feature; Calculate the Pearson correlation coefficient between the clinical feature sequence and the imaging feature sequence for each clinical feature to obtain the Pearson correlation coefficient for each clinical feature; The elements in the initial standard feature matrix are repositioned according to the Pearson correlation coefficient of each clinical feature to obtain the associated standard feature matrix.

[0010] Optionally, in the data processing module, the features of the reference medical image are obtained, specifically including: For each patient, each medical imaging feature and clinical feature is combined into a data pair, and each medical imaging feature data pair is used as an element in each row of the matrix to construct an image feature matrix. The image feature matrix is ​​then standardized to obtain the image standard feature matrix. For each patient, the Pearson correlation coefficient between each medical image feature and the clinical feature is calculated. The elements in the image feature matrix are then rearranged according to the Pearson correlation coefficient of each medical image feature to obtain the associated image feature matrix. In the associated image feature matrix, the number of rows of data pairs of medical image features is directly proportional to the size of the Pearson correlation coefficient of the medical image features. In the associated image feature matrix of each patient, the associated image features with a Pearson correlation coefficient greater than a preset second correlation threshold are determined as reference medical image features; the preset second correlation threshold is determined by the top 70% of medical image features.

[0011] Optionally, in the data processing module, the first correlation degree for each reference clinical feature and the second correlation degree for each reference medical imaging feature are calculated, specifically including: Obtain the number of patients including the b-th reference clinical feature, and obtain the row number of the data pair corresponding to the b-th reference clinical feature in the association standard feature matrix of the c-th patient; Obtain the row number of the corresponding data pair in the association standard feature matrix of each patient for the b-th reference clinical feature, and sum them to obtain the priority of the b-th reference clinical feature; The ratio of the number of patients with the b-th reference clinical feature to the priority of the b-th reference clinical feature is determined as the first degree of association of the b-th reference clinical feature. Obtain the first degree of association for each reference clinical feature; Obtain the number of patients including the d-th reference medical image feature, and obtain the row number of the data pair corresponding to the d-th reference medical image feature in the associated image feature matrix of the e-th patient; To obtain the priority of the d-th reference medical image feature, we obtain the row number of the corresponding data pair in the associated image feature matrix of each patient and sum them up. The ratio of the number of patients with the d-th reference medical image feature to the priority of the d-th reference medical image feature is determined as the second degree of association of the d-th reference medical image feature. Obtain the second degree of correlation for each reference medical image feature.

[0012] Optionally, in the data processing module, based on the reduced-dimensional clinical features and the reduced-dimensional medical image features, a fusion analysis matrix is ​​constructed corresponding to each reduced-dimensional medical image feature, specifically including: For the f-th dimensionless medical image feature, obtain the vector feature of the f-th dimensionless medical image feature in the g-th reference patient, where the dimensionless medical image feature of the reference patient includes the f-th dimensionless medical image feature. The vector features of the f-th dimensionality-reduced medical image feature in the g-th reference patient are combined with each dimensionality-reduced clinical feature to form a data pair. The data pair of the f-th dimensionality-reduced medical image feature in each reference patient is used as the element in each row of the matrix to construct the fusion analysis matrix corresponding to the f-th dimensionality-reduced medical image feature. Obtain the fusion analysis matrix corresponding to each dimension-reduced medical image feature.

[0013] Optionally, in the data processing module, based on the fusion analysis matrix corresponding to each dimensionality-reduced medical image feature, the standardization barrier of each dimensionality-reduced medical image feature is calculated, specifically including: The difference between the rows and columns in the fusion analysis matrix corresponding to the f-th dimensionless medical image feature is obtained to obtain the feature difference of the f-th dimensionless medical image feature. In the fusion analysis matrix corresponding to the f-th dimension-reduced medical image feature, the difference between each data pair and the preset empirical data pair is obtained, and the average of the differences is calculated to obtain the empirical difference value of each data pair in the fusion analysis matrix corresponding to the f-th dimension-reduced medical image feature. The preset empirical data pair is the range of the data pair composed of clinical features and medical image features. The empirical difference values ​​of the data pairs are averaged to obtain the empirical difference value of the f-th dimensionless medical image feature. The product of the feature difference of the f-th dimensionless medical image feature and the empirical difference value of the f-th dimensionless medical image feature is determined as the standardization barrier of the f-th dimensionless medical image feature. Obtain the standardized barrier properties of each dimension-reduced medical image feature.

[0014] Optionally, in the data processing module, based on the standardization barrier of each dimensionality-reduced medical image feature, principal component clinical features are determined from the dimensionality-reduced clinical features, specifically including: Dimensionally reduced medical image features with standardized barrier properties less than a preset barrier threshold are identified as first medical image features, and dimensionality reduced medical image features with standardized barrier properties greater than or equal to a preset barrier threshold are identified as second medical image features. When the number of first medical image features exceeds a preset threshold, the fusion analysis matrix of the second medical image features is standardized, and the standardization result is analyzed using principal component analysis to obtain principal component features. The dimensionality-reduced clinical features in the principal component features were identified as the principal component clinical features. When the number of first medical image features is less than or equal to a preset threshold, the fusion analysis matrix of each dimensionality-reduced medical image feature is standardized, and the standardization result is analyzed using principal component analysis to obtain principal component features. The dimensionality-reduced clinical features in the principal component features were identified as the principal component clinical features.

[0015] Optionally, in the data processing module, a stage fusion feature matrix is ​​constructed based on the principal component clinical features and the reduced-dimensional medical image features, and the principal component standard barrier property of the stage fusion feature matrix is ​​calculated, specifically including: The median of the clinical feature of the h-th principal component for each patient is obtained. Obtain the median value of the clinical characteristics of each principal component; The median of the i-th dimensionless medical image feature for each patient is calculated to obtain the median value of the i-th dimensionless medical image feature. Obtain the median value of each dimension-reduced medical image feature; The median value of each principal component clinical feature and the median value of each dimension-reduced medical image feature are combined in pairs to obtain data pairs. The data pairs are used as elements in the matrix to construct the stage fusion feature matrix. The mean of the clinical features of the h-th principal component is obtained by calculating the mean of the clinical features of the h-th principal component for each patient. Obtain the mean of the clinical features of each principal component, and calculate the mean to obtain the principal component mean of the stage fusion feature matrix; The standard deviation of the median value of each principal component's clinical feature is calculated to obtain the principal component representative value of the stage fusion feature matrix; The product of the principal component mean and the principal component representative value of the stage fusion feature matrix is ​​normalized, and the normalization result is determined as the principal component standard obstacle of the stage fusion feature matrix.

[0016] The beneficial effects of the technical solution of the present invention are: In this embodiment of the invention, before integrating clinical data and radiomics, preliminary dimensionality reduction is performed on both to reduce the high-dimensional pressure during subsequent data fusion. Then, the fusion obstacles caused by the unequal number of features during data fusion are analyzed, and the standardization process is intelligently completed to improve the rationality of principal component feature selection and the effectiveness of subsequent dimensionality reduction. This allows the grading results of thyroid eye disease patients in terms of exophthalmos severity to be combined with various influencing environments, thereby improving the practicality and rationality of the grading results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a structural diagram of an intelligent grading system for exophthalmos severity that integrates radiomics, provided in one embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the intelligent grading system for exophthalmos severity integrated with radiomics proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent grading system for exophthalmos severity integrated with radiomics provided by this invention.

[0022] This invention provides an intelligent grading system for exophthalmos severity that integrates radiomics. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates the structure of an intelligent grading system for exophthalmos severity integrating radiomics, according to an embodiment of the present invention, comprising the following modules: The data acquisition module 101 is used to acquire the clinical and medical imaging characteristics of each patient.

[0023] For example, feature extraction is performed on the clinical dataset and medical image dataset for each patient to obtain the clinical features and medical image features for each patient. The clinical dataset and medical image dataset for patients can be obtained by collecting clinical data (such as thyroid function status, disease course, smoking history, clinical activity score, etc.) and orbital CT or MRI medical images of patients with thyroid eye disease.

[0024] Feature extraction can be performed using pre-trained neural networks (such as convolutional neural networks CNNs) or traditional radiomics feature extraction tools (such as PyRadiomics) to extract initial clinical features and medical image features from clinical data and medical images, respectively. Each feature is represented as a numerical vector.

[0025] Data processing module 102 is used to determine dimensionality-reduced clinical features and dimensionality-reduced medical imaging features from clinical features and medical imaging features for each patient. Based on the reduced-dimensional clinical features and reduced-dimensional medical image features, a fusion analysis matrix is ​​constructed for each reduced-dimensional medical image feature; based on the fusion analysis matrix for each reduced-dimensional medical image feature, the standardization barrier of each reduced-dimensional medical image feature is calculated. Based on the standardized barrier properties of each dimensionality-reduced medical image feature, principal component clinical features are determined from the dimensionality-reduced clinical features; based on the principal component clinical features and the dimensionality-reduced medical image features, a stage fusion feature matrix is ​​constructed, and the principal component standardized barrier properties of the stage fusion feature matrix are calculated. If the standard barrier of the principal components is less than the preset barrier threshold, the stage fusion feature matrix is ​​determined as the final fusion feature matrix; otherwise, the clinical features of the principal components are standardized, and the final fusion feature matrix is ​​constructed based on the standardized clinical features of the principal components.

[0026] In this embodiment, the data processing module determines dimensionality-reduced clinical features and dimensionality-reduced medical imaging features for each patient from clinical features and medical imaging features, specifically including: For each patient, a correlation standard feature matrix is ​​constructed based on clinical characteristics and medical imaging characteristics. Based on the correlation standard feature matrix of each patient, reference clinical characteristics and reference medical imaging characteristics are determined from the clinical characteristics and medical imaging characteristics, respectively. Calculate the first degree of correlation for each reference clinical feature and the second degree of correlation for each reference medical image feature. Based on the first preset correlation threshold and the second preset correlation threshold, determine the dimensionality-reduced clinical features and dimensionality-reduced medical image features from the reference clinical features and reference medical image features, respectively.

[0027] In the data processing module, for each patient, a correlation standard feature matrix is ​​constructed based on clinical characteristics and medical imaging characteristics. Then, based on each patient's correlation standard feature matrix, reference clinical features and reference medical imaging features are determined from the clinical characteristics and medical imaging characteristics, respectively. Specifically, these include: For each patient, each clinical feature is combined with a medical imaging feature to form a data pair, and each clinical feature data pair is used as an element in each row of the matrix to construct an initial feature matrix. The initial feature matrix is ​​then standardized to obtain an initial standard feature matrix. For each patient, the Pearson correlation coefficient between each clinical feature and the medical imaging feature is calculated. The elements in the initial standard feature matrix are then repositioned according to the Pearson correlation coefficient of each clinical feature to obtain the associated standard feature matrix. In the associated standard feature matrix, the number of rows of data pairs of clinical features is directly proportional to the size of the Pearson correlation coefficient of the clinical features. In the association standard feature matrix of each patient, clinical features with Pearson correlation coefficients greater than a preset first correlation threshold are identified as reference clinical features, and reference medical image features are obtained.

[0028] In the data processing module, the Pearson correlation coefficient between each clinical feature and the medical imaging feature is calculated, specifically including: Each medical image feature is used as an element in the data sequence to obtain the image feature sequence; Using the a-th clinical feature as an element in the data sequence, we obtain the clinical feature sequence of the a-th clinical feature, where the image feature sequence and the clinical feature sequence have the same length. Obtain the clinical feature sequence for each clinical feature; Calculate the Pearson correlation coefficient between the clinical feature sequence and the imaging feature sequence for each clinical feature to obtain the Pearson correlation coefficient for each clinical feature; The elements in the initial standard feature matrix are repositioned according to the Pearson correlation coefficient of each clinical feature to obtain the associated standard feature matrix.

[0029] In the data processing module, reference medical image features are obtained, specifically including: For each patient, each medical imaging feature and clinical feature is combined into a data pair, and each medical imaging feature data pair is used as an element in each row of the matrix to construct an image feature matrix. The image feature matrix is ​​then standardized to obtain the image standard feature matrix. For each patient, the Pearson correlation coefficient between each medical image feature and the clinical feature is calculated. The elements in the image feature matrix are then rearranged according to the Pearson correlation coefficient of each medical image feature to obtain the associated image feature matrix. In the associated image feature matrix, the number of rows of data pairs of medical image features is directly proportional to the size of the Pearson correlation coefficient of the medical image features. For each patient, the associated image features with a Pearson correlation coefficient greater than a preset second correlation threshold are identified as reference medical image features in the associated image feature matrix.

[0030] In the data processing module, the first correlation degree for each reference clinical feature and the second correlation degree for each reference medical imaging feature are calculated, specifically including: Obtain the number of patients including the b-th reference clinical feature, and obtain the row number of the data pair corresponding to the b-th reference clinical feature in the association standard feature matrix of the c-th patient; Obtain the row number of the corresponding data pair in the association standard feature matrix of each patient for the b-th reference clinical feature, and sum them to obtain the priority of the b-th reference clinical feature; The ratio of the number of patients with the b-th reference clinical feature to the priority of the b-th reference clinical feature is determined as the first degree of association of the b-th reference clinical feature. Obtain the first degree of association for each reference clinical feature; Obtain the number of patients including the d-th reference medical image feature, and obtain the row number of the data pair corresponding to the d-th reference medical image feature in the associated image feature matrix of the e-th patient; To obtain the priority of the d-th reference medical image feature, we obtain the row number of the corresponding data pair in the associated image feature matrix of each patient and sum them up. The ratio of the number of patients with the d-th reference medical image feature to the priority of the d-th reference medical image feature is determined as the second degree of association of the d-th reference medical image feature. Obtain the second degree of correlation for each reference medical image feature.

[0031] In the data processing module, based on the reduced-dimensional clinical features and reduced-dimensional medical image features, a fusion analysis matrix is ​​constructed for each reduced-dimensional medical image feature, specifically including: For the f-th dimensionless medical image feature, obtain the vector feature of the f-th dimensionless medical image feature in the g-th reference patient, where the dimensionless medical image feature of the reference patient includes the f-th dimensionless medical image feature. The vector features of the f-th dimensionality-reduced medical image feature in the g-th reference patient are combined with each dimensionality-reduced clinical feature to form a data pair. The data pair of the f-th dimensionality-reduced medical image feature in each reference patient is used as the element in each row of the matrix to construct the fusion analysis matrix corresponding to the f-th dimensionality-reduced medical image feature. Obtain the fusion analysis matrix corresponding to each dimension-reduced medical image feature.

[0032] In the data processing module, based on the fusion analysis matrix corresponding to each dimensionality-reduced medical image feature, the standardization barrier of each dimensionality-reduced medical image feature is calculated, specifically including: The difference between the rows and columns in the fusion analysis matrix corresponding to the f-th dimensionless medical image feature is obtained to obtain the feature difference of the f-th dimensionless medical image feature. In the fusion analysis matrix corresponding to the f-th dimension-reduced medical image feature, the difference between each data pair and the preset empirical data pair is obtained, and the average of the differences is calculated to obtain the empirical difference value of each data pair in the fusion analysis matrix corresponding to the f-th dimension-reduced medical image feature. The preset empirical data pair is the range of the data pair composed of clinical features and medical image features. The empirical difference values ​​of the data pairs are averaged to obtain the empirical difference value of the f-th dimensionless medical image feature. The product of the feature difference of the f-th dimensionless medical image feature and the empirical difference value of the f-th dimensionless medical image feature is determined as the standardization barrier of the f-th dimensionless medical image feature. Obtain the standardized barrier properties of each dimension-reduced medical image feature.

[0033] In the data processing module, based on the standardization barrier of each dimensionality-reduced medical image feature, principal component clinical features are determined from the dimensionality-reduced clinical features, specifically including: Dimensionally reduced medical image features with standardized barrier properties less than a preset barrier threshold are identified as first medical image features, and dimensionality reduced medical image features with standardized barrier properties greater than or equal to a preset barrier threshold are identified as second medical image features. When the number of first medical image features exceeds a preset threshold, the fusion analysis matrix of the second medical image features is standardized, and the standardization result is analyzed using principal component analysis to obtain principal component features. The dimensionality-reduced clinical features in the principal component features were identified as the principal component clinical features. When the number of first medical image features is less than or equal to a preset threshold, the fusion analysis matrix of each dimensionality-reduced medical image feature is standardized, and the standardization result is analyzed using principal component analysis to obtain principal component features. The dimensionality-reduced clinical features in the principal component features were identified as the principal component clinical features.

[0034] In the data processing module, based on the principal component clinical features and the reduced-dimensional medical image features, a stage fusion feature matrix is ​​constructed, and the principal component standard obstacle of the stage fusion feature matrix is ​​calculated, specifically including: The median of the clinical feature of the h-th principal component for each patient is obtained. Obtain the median value of the clinical characteristics of each principal component; The median of the i-th dimensionless medical image feature for each patient is calculated to obtain the median value of the i-th dimensionless medical image feature. Obtain the median value of each dimension-reduced medical image feature; The median value of each principal component clinical feature and the median value of each dimension-reduced medical image feature are combined in pairs to obtain data pairs. The data pairs are used as elements in the matrix to construct the stage fusion feature matrix. The mean of the clinical features of the h-th principal component is obtained by calculating the mean of the clinical features of the h-th principal component for each patient. Obtain the mean of the clinical features of each principal component, and calculate the mean to obtain the principal component mean of the stage fusion feature matrix; The standard deviation of the median value of each principal component's clinical feature is calculated to obtain the principal component representative value of the stage fusion feature matrix; The product of the principal component mean and the principal component representative value of the stage fusion feature matrix is ​​normalized, and the normalization result is determined as the principal component standard obstacle of the stage fusion feature matrix.

[0035] For example, for each patient, each clinical feature is paired with each imaging feature to form a data pair (clinical feature value, imaging feature value). These data pairs are then filled into a matrix with clinical features as rows and imaging features as columns to construct an initial feature matrix, which is then standardized to obtain an initial standard feature matrix.

[0036] In this step, traditional feature screening is global and ignores individual differences. In reality, the causes of ophthalmopathy may differ among patients (e.g., patient A's exophthalmos is mainly caused by muscle hyperplasia, while patient B's is caused by fat hyperplasia). Therefore, the assessment of feature importance should also be personalized. Constructing an independent matrix for each patient is precisely to capture individualized patterns.

[0037] Organize the relationships between all feature pairs into a structured form. Standardizing this process eliminates differences in units and numerical ranges between different features, preparing for a fair calculation of correlations later. Without standardization, features with large numerical values ​​will unreasonably dominate the calculations.

[0038] For example, the patients in this embodiment may include three individuals (P1, P2, P3); clinical features include: C1: TSH receptor antibody level (unit: IU / L), C2: Clinical activity score (CAS, dimensionless); medical imaging features include: I1: ocular protrusion (unit: mm), I2: extraocular muscle volume (unit: cm). 3 ).

[0039] Following the order C1, C2, I1, I2, the corresponding values ​​for each patient can be P1 (5.1, 3, 18.5, 1.2), P2 (22.7, 6, 22.1, 1.8), and P3 (15.3, 4, 20.3, 1.5), respectively. Each element in the matrix can be a (clinical feature value, imaging feature value) data pair. An initial feature matrix M1 can be constructed using this data. Z-score normalization of the initial feature matrix yields the initial standard feature matrix M2.

[0040] For P1, calculate the Pearson correlation coefficient between the standard value sequence of C1 and the long sequence spliced ​​from the standard values ​​of all image features. For example, the calculation result is: the Pearson correlation coefficient γC1 = 0.85 for C1, and similarly, the Pearson correlation coefficient γC2 = 0.60 for C2.

[0041] Since γC1 is greater than γC2, the new order of rows in the initial standard characteristic matrix remains unchanged. Adjustments are actually needed based on the calculated Pearson correlation coefficient.

[0042] The Pearson correlation coefficient measures the linear correlation between changes in a clinical feature and changes in medical imaging features. It is a standard statistic for measuring this linear relationship, computationally efficient, and with clearly defined meaning. The elements in the initial standard feature matrix are rearranged according to the Pearson correlation coefficient for each clinical feature, placing the most important clinical features for the current patient at the top. This sorted matrix is ​​the "correlation standard feature matrix," which visually demonstrates the priority order of clinical features for that patient.

[0043] In this embodiment, radiomics is used to guide and interpret clinical data. The main focus is on analyzing which changes in clinical indicators are closely related to radiographic findings in specific patients. Therefore, clinical features are considered the object of analysis, while the overall radiographic features are treated as a reference system.

[0044] Because not every feature is meaningful for analysis, reference clinical features and reference medical imaging features are selected from clinical features and medical imaging features respectively by setting a first relevance threshold and a second relevance threshold. Furthermore, the first and second relevance thresholds are values ​​set according to actual needs or historical experience and can be modified as needed; no specific numerical limitations are imposed here. In a preferred embodiment, they can be 70% of the clinical features and 70% of the medical imaging features, respectively.

[0045] Optionally, the methods for obtaining reference medical imaging features can refer to the steps for obtaining reference clinical features, which will not be elaborated on here.

[0046] Taking 70% of clinical features and medical imaging features as an example, the top 70% of clinical features and medical imaging features in the association standard feature matrix are determined as reference clinical features and reference medical imaging features. For example, in the association matrix of each patient, the top 70% of clinical features are selected. Since there are only two features in this embodiment, the top 70% are the features ranked first: P1:C1, P2:C1, P3:C1. Therefore, the set of reference clinical features is {C1}, and the reference medical imaging features are determined similarly.

[0047] The correlation degree (including the first correlation degree of each reference clinical feature and the second correlation degree of each reference medical imaging feature) can be calculated as follows: in, Indicates the degree of correlation. Indicates the frequency of occurrence of the feature. Indicates the first The sorting value at the time of occurrence, that is, the row number. This represents the normalization function.

[0048] In the formula, It is used to measure the prevalence of a characteristic. The more frequently it occurs, the less likely it is to be an isolated phenomenon. Used to measure the excellence of this feature. The smaller the ranking value (e.g., 1st, 2nd), the higher its importance when it appears. Furthermore, the smaller the value, the higher the overall ranking of this feature. (Molecular) Encouraging universality, denominator Encouraging excellence. Therefore, a high Q value characteristic must be present in a large number of patients (high... ), and it always ranks very high (low) whenever it appears. This formula, in its feature selection criteria, aims to identify key features that are both universal and essential.

[0049] Optionally, the first preset association threshold and the second preset association threshold can be values ​​set according to actual needs or historical experience, and can be modified according to actual needs. There are no specific numerical restrictions here. The range of the first preset association threshold and the second preset association threshold is (0,1). In a preferred embodiment, they can be 0.3 and 0.4, respectively.

[0050] Taking C1 as an example, calculate its degree of association: N (number of occurrences) = 3 (it appears in the first row in P1, P2, and P3). The sum of the rankings equals 1(P1) + 1(P2) + 1(P3) = 3. Therefore, the Q value of C1 is QC1 = 3 / 3 = 1.0, which is greater than the first preset association threshold of 0.3. Thus, the dimensionality-reduced clinical feature = {C1}. The same operation is performed on the image features, assuming the final result is: dimensionality-reduced medical image features = {I1, I2} (i.e., both image features are preserved).

[0051] The first correlation degree of each reference clinical feature and the second correlation degree of each reference medical image feature are calculated, and the dimension-reduced clinical features and dimension-reduced medical image features are determined from the reference clinical features and reference medical image features according to the first preset correlation threshold and the second preset correlation threshold.

[0052] Before proceeding with subsequent feature analysis, it is necessary to assess the difficulty of feature association. Different feature combinations may yield poor results when directly merged due to differences in their data structures and scales. Therefore, based on the reduced-dimensional clinical features and reduced-dimensional medical imaging features, a fusion analysis matrix is ​​constructed for each reduced-dimensional medical imaging feature. The construction method is similar to the matrix construction method in the preceding steps, and will not be elaborated further here.

[0053] According to the above embodiments, the elements in the first row of the fusion analysis matrices F1 and F2 for each dimensionality-reduced medical image feature (since only one dimensionality-reduced clinical feature is retained above, the fusion analysis matrix is ​​a three-row, one-column matrix, with each row corresponding to each patient, and the fusion analysis matrix F1 of I1 is used as an example for analysis. The fusion analysis matrix of I2 is the fusion analysis matrix F2) can be (5.1, 18.5), the elements in the second row can be (22.7, 22.1), and the elements in the third row can be (15.3, 20.3). In each element, the first value represents the value of the current patient in the current dimensionality-reduced medical image feature, and the second value represents the value of the current patient in the current dimensionality-reduced clinical feature.

[0054] For example, the formula for calculating the standardization barrier of each dimension-reduced medical image feature can be: in, Indicating obstacles to standardization, This represents the difference between the row and the column. This represents the mean of the differences between each data pair and a preset empirical data pair.

[0055] In the formula, Used to measure the "lean" or "thin" nature of a data structure. The larger the value, the more patients there are than the number of features, the more stable the data, the more reliable the statistical results, and the lower the basic difficulty of fusion. Used to measure the difference in numerical range between different features. The larger the value, the greater the difference in dimensions and scale between features. Directly standardizing them together will result in severe distortion, and the technical difficulty of fusion is high. Combining the meanings of the two parameters, when A larger value indicates either a poor data structure with large scale differences, or a good data structure but huge scale differences. However, both of these cases require further processing.

[0056] The standardization barrier for each dimensionality-reduced medical image feature can be calculated using the above formula. When most fusion tasks are difficult (high R-value), it indicates that the data itself is challenging. In this case, strict criteria must be applied, requiring a clinical feature to be proven important in the fusion analysis of all image features (through PCA, principal component analysis) before it can be selected. This ensures that the selected features have strong universality and robustness. When most fusion tasks are simple (low R-value), it indicates that the data quality is high and easy to process. The risk here is not in the fusion itself, but in the potential omission of specialized features designed to solve "difficult cases" (high R-value tasks). Therefore, the system relaxes the criteria, focusing only on features proven effective in "high-difficulty" tasks, ensuring that key features for solving core problems are captured.

[0057] For example, the normalized barrier R of I1 can be calculated as: (Feature difference) equals the number of rows (3) - the number of columns (1) = 2. (Empirical variance): Calculate the range of all data pairs in the matrix. The range of C1 = 22.7 - 5.1 = 17.6; the range of I1 = 22.1 - 18.5 = 3.6. Take the average of the two. =(17.6+3.6) / 2=10.6. Therefore, R(I1)=ΔM×ΔZ =21.2. Similarly, R(I2) can be calculated as follows: =18.2.

[0058] After calculating the standardization barrier for each dimensionality-reduced medical image feature, not all image features face the same fusion difficulty. The strategy needs to be dynamically adjusted based on the overall severity of the environment. If more than half of the image features are easy to fuse (small R), the overall data environment is relatively favorable for data analysis. In this case, the main risk is not in the fusion itself, but in potentially overlooking specialized features designed for solving high-difficulty tasks. Therefore, only features selected by PCA from high-difficulty tasks (secondary medical image features) are trusted. These features are considered crucial for solving the core challenges. If more than half of the image features are difficult to fuse (large R), the overall data environment is relatively harsh. In this case, the system must adopt a universal strategy, requiring a clinical feature to be proven important by PCA in the fusion analysis of all image features (regardless of difficulty) before being selected. This ensures the extreme robustness of the final features. PCA (Principal Component Analysis) can identify the direction that best represents the variance of the original data in the new fused data space (clinical + image). The original clinical features corresponding to these directions (principal components) are naturally the core features that contribute the most and are the most representative in this fusion scenario.

[0059] The preset obstacle threshold and preset quantity threshold are values ​​set based on actual needs or historical experience, and can be modified according to actual needs; no specific numerical restrictions are imposed here. In a preferred embodiment, they can be 0.7 and 50% of the number of dimensionality-reduced medical image features, respectively.

[0060] For example, the number of image features with R < 0.5 is counted. In this embodiment, R(I1) = 21.2 and R(I2) = 18.2, both much greater than 0.5, so the number is 0. Since 0% < 50%, all fusion analysis matrices F1 and F2 are standardized, and PCA is performed. Assuming the PCA results show that C1 is the major contributor, the resulting principal component clinical feature is {C1}.

[0061] For example, the formula for calculating the principal component standard obstacle R1 of the stage fusion feature matrix can be: in, The principal component standard obstacle of the stage fusion feature matrix is ​​represented. This represents the principal component mean of the stage-fused feature matrix. This represents the principal component representative value of the stage fusion feature matrix.

[0062] After screening, the core feature set was obtained. Now, a baseline template that can represent the typical characteristic relationships of the entire population needs to be constructed. The median is used instead of the mean because the median is not sensitive to outliers, making the template more robust and less susceptible to the influence of individual extreme patient data. At this point, the calculation object is no longer the original data, but the newly constructed "template" matrix, i.e., the stage fusion feature matrix. The purpose of R1 is to check the consistency within this template.

[0063] The principal component mean represents the typical magnitude of these core clinical features in the original data. The representative value of the principal components (the standard deviation of the median) measures the volatility of a clinical feature's value across different imaging feature contexts. A standard deviation of 0 indicates extreme stability; a large standard deviation indicates that the feature's meaning is highly dependent on its paired imaging features and is unstable. A high R1 value means that the template contains a feature with a large and unstable value. Such a feature will become an unstable strong signal in the subsequent model, interfering with model learning. R1 is the metric used to detect such "inconsistent features."

[0064] When the principal component standard barrier is greater than or equal to a preset barrier threshold, the clinical features of the principal components are standardized, and the final fusion feature matrix is ​​constructed based on the standardized clinical features of the principal components. When inconsistencies within the template are detected (excessively high R1), the system does not directly use the problematic template. Instead, it traces back to the source and standardizes the original data of the problematic clinical features. This eliminates dimensional differences and instabilities at the source. Then, the template is reconstructed using clean data until the template quality meets the standard (R1 < threshold). This ensures that the final output fusion matrix is ​​highly pure and robust.

[0065] For example, in the construction phase, the fusion feature matrix M3 consists of rows: principal component clinical features (C1), columns: dimensionality-reduced medical imaging features (I1, I2), and elements: calculated median values ​​for all patients on each feature pair. (C1, I1): The numerical pairs are (5.1, 18.5), (22.7, 22.1), and (15.3, 20.3). The median value is [15.3, 20.3]. (C1, I2): The numerical pairs are (5.1, 1.2), (22.7, 1.8), and (15.3, 1.5). The median value is [15.3, 1.5].

[0066] The elements in the first row of the M3 matrix are [15.3, 20.3] and [15.3, 1.5], respectively. (Principal component mean) Calculate the mean of C1 in all patients. =(5.1+22.7+15.3) / 3=14.37. Since there's only one row, the principal component mean... =14.37. Principal component representative value: Take the median value of all clinical features in M3 from row C1, i.e., [15.3, 15.3], calculate its standard deviation, and get 0. R1=norm(14.37×0)=0. Since R1 is less than 0.7, it indicates that the matrix is ​​balanced and no further optimization is needed. And the final fused feature matrix is ​​the stage fused feature matrix obtained. When R1 is greater than or equal to 0.7, it indicates that the matrix is ​​unbalanced and the principal component clinical features need to be standardized. Construct the final fused feature matrix based on the standardized principal component clinical features, specifically including: The median of the j-th standardized principal component clinical feature for each patient is obtained. Obtain the median value of the clinical characteristics of each standardized principal component; The median of the kth dimensionless medical image feature for each patient is calculated to obtain the median value of the kth dimensionless medical image feature. Obtain the median value of each dimension-reduced medical image feature; The median value of each principal component clinical feature and the median value of each dimensionality-reduced medical image feature are combined in pairs to obtain data pairs. The data pairs are then used as elements in the matrix to construct the final fused feature matrix.

[0067] For example, the data used here is: Rows: Principal component clinical features (C1), Columns: Dimensionally reduced medical image features (I1, I2), and the elements in the first row of the M3 matrix are [15.3, 20.3] and [15.3, 1.5], respectively. (C1, I1): The numerical pairs are (5.1, 18.5), (22.7, 22.1), (15.3, 20.3). (C1, I2): The numerical pairs are (5.1, 1.2), (22.7, 1.8), (15.3, 1.5).

[0068] Z-score standardization was performed on column C1 [5.1, 22.7, 15.3]. The mean (μ) of C1 = (5.1 + 22.7 + 15.3) / 3 = 14.37, and the standard deviation (σ) ≈ 7.21. Standardization of C1 is as follows: C1P1 = (5.1 - 14.37) / 7.21 ≈ -1.29, C1P2 = (22.7 - 14.37) / 7.21 ≈ 1.16, C1P3 = (15.3 - 14.37) / 7.21 ≈ 0.13. The data for the dimensionality-reduced medical image features (I1, I2) remained unchanged. The medians were re-obtained: C1P1, C1P2, and C1P3 represent the standardized values ​​of the three numerical vectors in column C1.

[0069] For (C1, I1): the numerical pairs are: (-1.29, 18.5), (1.16, 22.1), (0.13, 20.3). The new medians are: Median of the C1 portion: taken from [-1.29, 1.16, 0.13], the result is 0.13. Median of the I1 portion: taken from [18.5, 22.1, 20.3], the result is 20.3. Therefore, the new data pair is [0.13, 20.3].

[0070] For (C1, I2): the numerical pairs are (-1.29, 1.2), (1.16, 1.8), (0.13, 1.5). The new median values ​​are: the median for the C1 portion remains 0.13. The median for the I2 portion is the median taken from [1.2, 1.8, 1.5], resulting in 1.5. Therefore, the new data pair is [0.13, 1.5].

[0071] Using the new median data pairs, construct the final fused feature matrix: The newly calculated median data pairs are then filled into the corresponding positions in the matrix to form the final fused feature matrix. Therefore, the first row of the final fused feature matrix contains elements [0.13, 20.3] and [0.13, 1.5]. This completes the construction of the final fused feature matrix.

[0072] The data classification module 103 is used to determine the baseline vector based on the final fusion feature matrix, and to combine the dimensionality-reduced medical image features and principal component clinical features of each patient into a personal feature vector. The baseline vector and personal feature vector are input into the trained intelligent grading model, and the intelligent grading model outputs the classification level of each patient.

[0073] In this embodiment, the data classification module determines the baseline vector based on the final fused feature matrix, specifically including: The final fused feature matrix is ​​expanded along its rows to obtain the baseline vector.

[0074] For example, the baseline vector contains the most stable and core feature relationship patterns extracted from all training data. It is a solidified, standardized knowledge base. The individual feature vector is the individualized information of the patient to be diagnosed, which is the specific value of the core features selected above. Concatenating the two and inputting them into the model is equivalent to providing the AI ​​model with two types of information.

[0075] This input method greatly enriches the model's decision context. The model no longer blindly guesses based solely on individual case data, nor does it rigidly apply group patterns. Guided by deep background knowledge, it makes judgments about individual situations, which significantly improves the accuracy, robustness, and interpretability of decisions.

[0076] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0077] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent grading system for exophthalmos severity integrating radiomics, characterized in that, include: The data acquisition module is used to acquire the clinical and medical imaging characteristics of each patient. The data processing module is used to determine dimensionality-reduced clinical features and dimensionality-reduced medical imaging features from clinical features and medical imaging features for each patient. Based on the reduced-dimensional clinical features and reduced-dimensional medical image features, a fusion analysis matrix is ​​constructed for each reduced-dimensional medical image feature; based on the fusion analysis matrix for each reduced-dimensional medical image feature, the standardization barrier of each reduced-dimensional medical image feature is calculated. Based on the standardized barrier properties of each dimensionality-reduced medical image feature, principal component clinical features are determined from the dimensionality-reduced clinical features; based on the principal component clinical features and the dimensionality-reduced medical image features, a stage fusion feature matrix is ​​constructed, and the principal component standardized barrier properties of the stage fusion feature matrix are calculated. If the standard barrier of the principal component is less than the preset barrier threshold, the stage fusion feature matrix is ​​determined as the final fusion feature matrix; otherwise, the clinical features of the principal component are standardized, and the final fusion feature matrix is ​​constructed based on the standardized clinical features of the principal component. The data classification module is used to determine the baseline vector based on the final fused feature matrix, and to combine the dimensionality-reduced medical image features and principal component clinical features of each patient into a personal feature vector. The baseline vector and personal feature vector are then input into the trained intelligent grading model, which outputs the classification level of each patient.

2. The intelligent grading system for exophthalmos severity integrating radiomics as described in claim 1, characterized in that, In the data processing module, for each patient, dimensionality-reduced clinical features and dimensionality-reduced medical imaging features are determined from clinical features and medical imaging features, specifically including: For each patient, a correlation standard feature matrix is ​​constructed based on clinical characteristics and medical imaging characteristics. Based on the correlation standard feature matrix of each patient, reference clinical characteristics and reference medical imaging characteristics are determined from the clinical characteristics and medical imaging characteristics, respectively. Calculate the first correlation degree of each reference clinical feature and the second correlation degree of each reference medical image feature. Based on the first preset correlation threshold and the second preset correlation threshold, determine the dimensionality-reduced clinical features and dimensionality-reduced medical image features from the reference clinical features and reference medical image features, respectively. The first preset correlation threshold and the second preset correlation threshold are in the range of (0,1). The first preset correlation threshold is related to the reference clinical features, and the second preset correlation threshold is related to the reference medical image features.

3. The intelligent grading system for exophthalmos severity integrating radiomics as described in claim 2, characterized in that, In the data processing module, for each patient, a correlation standard feature matrix is ​​constructed based on clinical characteristics and medical imaging characteristics. Based on this matrix, reference clinical features and reference medical imaging features are determined from the clinical characteristics and medical imaging characteristics, respectively. Specifically, this includes: For each patient, each clinical feature is combined with a medical imaging feature to form a data pair, and each clinical feature data pair is used as an element in each row of the matrix to construct an initial feature matrix. The initial feature matrix is ​​then standardized to obtain an initial standard feature matrix. For each patient, the Pearson correlation coefficient between each clinical feature and the medical imaging feature is calculated. The elements in the initial standard feature matrix are then repositioned according to the Pearson correlation coefficient of each clinical feature to obtain the associated standard feature matrix. In the associated standard feature matrix, the number of rows of data pairs of clinical features is directly proportional to the size of the Pearson correlation coefficient of the clinical features. In the association standard feature matrix of each patient, clinical features with Pearson correlation coefficients greater than a preset first correlation threshold are identified as reference clinical features, and reference medical imaging features are obtained; the preset first correlation threshold is determined by the top 70% of clinical features.

4. The intelligent grading system for exophthalmos severity integrating radiomics according to claim 3, characterized in that, The data processing module calculates the Pearson correlation coefficient between each clinical feature and the medical imaging feature, specifically including: Each medical image feature is used as an element in the data sequence to obtain the image feature sequence; Using the a-th clinical feature as an element in the data sequence, we obtain the clinical feature sequence of the a-th clinical feature, where the image feature sequence and the clinical feature sequence have the same length. Obtain the clinical feature sequence for each clinical feature; Calculate the Pearson correlation coefficient between the clinical feature sequence and the imaging feature sequence for each clinical feature to obtain the Pearson correlation coefficient for each clinical feature; The elements in the initial standard feature matrix are repositioned according to the Pearson correlation coefficient of each clinical feature to obtain the associated standard feature matrix.

5. The intelligent grading system for exophthalmos severity integrating radiomics according to claim 3, characterized in that, The data processing module acquires reference medical image features, specifically including: For each patient, each medical imaging feature and clinical feature is combined into a data pair, and each medical imaging feature data pair is used as an element in each row of the matrix to construct an image feature matrix. The image feature matrix is ​​then standardized to obtain the image standard feature matrix. For each patient, the Pearson correlation coefficient between each medical image feature and the clinical feature is calculated. The elements in the image feature matrix are then rearranged according to the Pearson correlation coefficient of each medical image feature to obtain the associated image feature matrix. In the associated image feature matrix, the number of rows of data pairs of medical image features is directly proportional to the size of the Pearson correlation coefficient of the medical image features. In the associated image feature matrix of each patient, the associated image features with a Pearson correlation coefficient greater than a preset second correlation threshold are determined as reference medical image features; the preset second correlation threshold is determined by the top 70% of medical image features.

6. The intelligent grading system for exophthalmos severity integrating radiomics according to claim 2, characterized in that, The data processing module calculates the first correlation degree for each reference clinical feature and the second correlation degree for each reference medical imaging feature, specifically including: Obtain the number of patients including the b-th reference clinical feature, and obtain the row number of the data pair corresponding to the b-th reference clinical feature in the association standard feature matrix of the c-th patient; Obtain the row number of the corresponding data pair in the association standard feature matrix of each patient for the b-th reference clinical feature, and sum them to obtain the priority of the b-th reference clinical feature; The ratio of the number of patients with the b-th reference clinical feature to the priority of the b-th reference clinical feature is determined as the first degree of association of the b-th reference clinical feature. Obtain the first degree of association for each reference clinical feature; Obtain the number of patients including the d-th reference medical image feature, and obtain the row number of the data pair corresponding to the d-th reference medical image feature in the associated image feature matrix of the e-th patient; To obtain the priority of the d-th reference medical image feature, we obtain the row number of the corresponding data pair in the associated image feature matrix of each patient and sum them up. The ratio of the number of patients with the d-th reference medical image feature to the priority of the d-th reference medical image feature is determined as the second degree of association of the d-th reference medical image feature. Obtain the second degree of correlation for each reference medical image feature.

7. The intelligent grading system for exophthalmos severity integrating radiomics according to claim 1, characterized in that, In the data processing module, a fusion analysis matrix is ​​constructed for each dimensionality-reduced medical image feature based on the dimensionality-reduced clinical features and dimensionality-reduced medical image features. Specifically, this includes: For the f-th dimensionless medical image feature, obtain the vector feature of the f-th dimensionless medical image feature in the g-th reference patient, where the dimensionless medical image feature of the reference patient includes the f-th dimensionless medical image feature. The vector features of the f-th dimensionality-reduced medical image feature in the g-th reference patient are combined with each dimensionality-reduced clinical feature to form a data pair. The data pair of the f-th dimensionality-reduced medical image feature in each reference patient is used as the element in each row of the matrix to construct the fusion analysis matrix corresponding to the f-th dimensionality-reduced medical image feature. Obtain the fusion analysis matrix corresponding to each dimension-reduced medical image feature.

8. The intelligent grading system for exophthalmos severity integrating radiomics according to claim 1, characterized in that, In the data processing module, based on the fusion analysis matrix corresponding to each dimensionality-reduced medical image feature, the standardization barrier of each dimensionality-reduced medical image feature is calculated, specifically including: The difference between the rows and columns in the fusion analysis matrix corresponding to the f-th dimensionless medical image feature is obtained to obtain the feature difference of the f-th dimensionless medical image feature. In the fusion analysis matrix corresponding to the f-th dimension-reduced medical image feature, the difference between each data pair and the preset empirical data pair is obtained, and the average of the differences is calculated to obtain the empirical difference value of each data pair in the fusion analysis matrix corresponding to the f-th dimension-reduced medical image feature. The preset empirical data pair is the range of the data pair composed of clinical features and medical image features. The empirical difference values ​​of the data pairs are averaged to obtain the empirical difference value of the f-th dimensionless medical image feature. The product of the feature difference of the f-th dimensionless medical image feature and the empirical difference value of the f-th dimensionless medical image feature is determined as the standardization barrier of the f-th dimensionless medical image feature. Obtain the standardized barrier properties of each dimension-reduced medical image feature.

9. The intelligent grading system for exophthalmos severity integrating radiomics according to claim 1, characterized in that, In the data processing module, based on the standardization barrier of each dimensionality-reduced medical image feature, principal component clinical features are determined from the dimensionality-reduced clinical features, specifically including: Dimensionally reduced medical image features with standardized barrier properties less than a preset barrier threshold are identified as first medical image features, and dimensionality reduced medical image features with standardized barrier properties greater than or equal to a preset barrier threshold are identified as second medical image features. When the number of first medical image features exceeds a preset threshold, the fusion analysis matrix of the second medical image features is standardized, and the standardization result is analyzed using principal component analysis to obtain principal component features. The dimensionality-reduced clinical features in the principal component features were identified as the principal component clinical features. When the number of first medical image features is less than or equal to a preset threshold, the fusion analysis matrix of each dimensionality-reduced medical image feature is standardized, and the standardization result is analyzed using principal component analysis to obtain principal component features. The dimensionality-reduced clinical features in the principal component features were identified as the principal component clinical features.

10. The intelligent grading system for exophthalmos severity integrating radiomics according to claim 1, characterized in that, In the data processing module, a stage fusion feature matrix is ​​constructed based on principal component clinical features and dimensionality-reduced medical imaging features, and the principal component standard obstacle of the stage fusion feature matrix is ​​calculated, specifically including: The median of the clinical feature of the h-th principal component for each patient is obtained. Obtain the median value of the clinical characteristics of each principal component; The median of the i-th dimensionless medical image feature for each patient is calculated to obtain the median value of the i-th dimensionless medical image feature. Obtain the median value of each dimension-reduced medical image feature; The median value of each principal component clinical feature and the median value of each dimension-reduced medical image feature are combined in pairs to obtain data pairs. The data pairs are used as elements in the matrix to construct the stage fusion feature matrix. The mean of the clinical features of the h-th principal component is obtained by calculating the mean of the clinical features of the h-th principal component for each patient. Obtain the mean of the clinical features of each principal component, and calculate the mean to obtain the principal component mean of the stage fusion feature matrix; The standard deviation of the median value of each principal component's clinical feature is calculated to obtain the principal component representative value of the stage fusion feature matrix; The product of the principal component mean and the principal component representative value of the stage fusion feature matrix is ​​normalized, and the normalization result is determined as the principal component standard obstacle of the stage fusion feature matrix.

Images