Infiltrative lung adenocarcinoma high-level component prediction method based on machine learning

By combining radiomics and clinical radiological features into a logistic regression model, the reliability and applicability of predicting high-grade components in invasive lung adenocarcinoma were addressed, enabling more accurate individualized assessment and treatment planning.

CN118941564AInactive Publication Date: 2024-11-12JINAN PEOPLES HOSPITAL
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411426224.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-14
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing technologies have limitations in predicting high-grade pathological subtypes of lung adenocarcinoma, especially high-grade components of invasive lung adenocarcinoma, in terms of reliability and applicability, particularly in early-stage lung cancer.

Method used

By acquiring CT image data and three-dimensional volumes of interest of lesions from patients with invasive lung adenocarcinoma, radiomics features were extracted and combined with clinical radiological features. A logistic regression algorithm was used to build a radiomics score, and the SHAP algorithm was used to perform feature importance analysis to predict high-level components.

Benefits of technology

It improves the accuracy and reliability of predicting high-grade components of invasive lung adenocarcinoma, provides more accurate individualized assessments, and helps physicians develop more effective treatment plans.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941564A_ABST
    Figure CN118941564A_ABST
Patent Text Reader

Abstract

The invention discloses an invasive lung adenocarcinoma high-level component prediction method based on machine learning. The method comprises the following steps: acquiring original CT image data of an invasive lung adenocarcinoma patient, a focus three-dimensional volume of interest (VOI) and corresponding clinical radiology characteristics; the original CT image data are preprocessed, and image omics features are extracted and screened in combination with the focus VOI; constructing a radiomics score through the screened radiomics features; performing feature screening on clinical radiology features by using statistical analysis; and in combination with radiomics scoring and screened clinical radiology characteristics, modeling is carried out by using a logistic regression algorithm to realize prediction of high-level components of the invasive lung adenocarcinoma. According to the method, image omics and clinical radiology characteristics of an invasive lung adenocarcinoma patient are extracted and analyzed, an image omics score is constructed based on image omics characteristics, a logistic regression model is established, an SHAP model is used for characteristic importance analysis, and prediction of high-level components of the invasive lung adenocarcinoma is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of medical image data processing, and in particular to a method for predicting high-level components of invasive lung adenocarcinoma based on machine learning. Background Art

[0002] Lung adenocarcinoma has become the most common subtype of non-small cell lung cancer. Common subtypes of invasive non-mucinous adenocarcinoma of the lung include five histopathological subtypes: lepidic predominant adenocarcinoma, acinar predominant adenocarcinoma, papillary predominant adenocarcinoma, micropapillary predominant adenocarcinoma, and solid predominant adenocarcinoma with mucin secretion. The pathological subtypes of lung adenocarcinoma HGP (high grade patterns) mainly include micropapillary predominant adenocarcinoma, solid predominant adenocarcinoma with mucin secretion, and complex glandular patterns. When high-grade pathological subtypes of lung adenocarcinoma exist as the main component (>5%), they have the potential for rapid metastasis, and patients have a poor prognosis and are prone to recurrence. In addition, recent studies have found that when high-grade pathological subtypes appear as minor components (components <5%), they will also have a significant negative impact on the patient's overall survival rate. Therefore, preoperative determination of the presence of HGP is of great significance for doctors to formulate more aggressive surgical plans and other adjuvant treatments.

[0003] SHAP is an algorithm based on cooperative game theory. It quantifies the contribution of each feature to model prediction by calculating the Shapley value, providing a transparent explanation for complex black box models and used to explain the output of machine learning models. Radiomics is an interdisciplinary field method that combines medical imaging and machine learning. It aims to use large-scale medical imaging data and related clinical information to discover and quantify disease-related imaging features through image analysis and pattern recognition technology to improve clinical problems such as early diagnosis of diseases, treatment selection, and prognosis assessment. Radiomics extracts and analyzes quantitative features in images, such as morphological features, texture features, functional features, etc., and associates them with clinical outcomes to provide more accurate and individualized assessment and prediction of diseases.

[0004] There are some studies in the prior art that have used radiomics methods to predict HGP in lung adenocarcinoma. However, these studies are generally limited to early lung cancer with ground glass opacity, which limits the reliability and applicability of the results. Summary of the invention

[0005] In order to solve the above technical problems, this application proposes the following technical solutions: In a first aspect, the present application embodiment provides a method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning, comprising: Obtain the original CT image data, lesion three-dimensional volume of interest (VOI) and corresponding clinical radiological features of patients with invasive lung adenocarcinoma; Preprocessing the original CT image data and extracting and screening imaging omics features in combination with lesion VOI; The radiomics score was constructed by screening the radiomics features; The clinical radiological features were screened using statistical analysis; Combining the radiomics score and the screened clinical radiological features, a logistic regression algorithm was used to model the prediction of high-grade components in invasive lung adenocarcinoma.

[0006] In a possible implementation, preprocessing the original CT image data and extracting and screening radiomics features in combination with lesion VOI includes: Resample the original CT image and set the pixel spacing of the image in three anatomical directions; Extracting radiomics features from the resampled CT images and lesion VOIs, wherein the image category data in the radiomics features include preprocessed images, Gaussian-Laplacian filtered images, and wavelet images, and the feature category data include shape features, histogram features, texture features, grayscale run matrices, grayscale size area matrices, grayscale dependency matrices, and neighborhood grayscale difference matrices; All data are randomly divided into training sets and test sets while keeping the ratio of positive and negative samples unchanged; For the data in the training set, Z-scores are standardized, and the formula is: ,in is the standardized feature, is the average value of the feature, is the standard deviation of the feature; The t-test was used to detect the significant difference between each feature and whether the lesion had a high-grade component, and the radiomics features with P < 0.05 were screened out. Then, the LASSO algorithm was used to calculate the weighted ranking of the features to select the optimal features in the training set.

[0007] In a possible implementation, the use of the LASSO algorithm to calculate the weighted ranking of features to select the optimal features in the training set includes: LASSO introduces regularization terms into the regression model to automatically select the optimal features and perform weighted ranking on them. The formula is: Where n is the number of samples; p is the number of features; is the jth feature of the i-th sample; is the true label value of the i-th sample; is the coefficient of feature j; is the regularization parameter, which controls the degree of shrinkage of the coefficients.

[0008] In a possible implementation, constructing a radiomics score using the screened radiomics features includes: According to the screened radiomics features, the R-Score was constructed by logistic regression method. The calculation formula of R-Score was: R-Score = -2.0379 + 0.2249 * boxsigmaimage_glszm_SizeZoneNonUniformityNormalized + 0.0396 * log_glszm_log_sigma_2_mm_3D_SizeZoneNonUniformi ty +1.0020 * log_glszm_log_sigma_2_mm_3D_SmallAreaHighGrayLevelEmphasis + 0.3679*wavelet_glcm_wavelet_HHH_ClusterShade + 0.6478* laplaciansharpening_ glrlm_LongRunHighGrayLevelEmphasis+0.4029*specklenoise_gldm_DependenceVariance+0.3354*recursivegaussian_gldm_DependenceVariance+0.1831*shotnoise_firstorder_InterquartileRange-0.2873 * shotnoise_glcm_InverseVariance-0.2305 *shotnoise_glcm_Idn; Among them: boxsigmaimage_glszm_SizeZoneNonUniformityNormalized, log_glszm_log_sigma_2_mm_3D_SizeZoneNonUniformity, log_glszm_log_sigma_2_mm_ 3D_SmallAreaHighGrayLevelEmphasis, wavelet_glcm_wavelet_HHH_ClusterShade, laplaciansharpening_glrlm_LongRunHighGrayLevelEmphasis, specklenoise_gldm_DependenceVariance, recursivegaussian_gldm_DependenceVariance, shotnoise _firstorder_InterquartileRange, shotnoise_glcm_InverseVariance and shotnoise_glcm_Idn are all imaging features screened out in the training set data.

[0009] In a possible implementation, the step of using statistical analysis to perform feature screening on the clinical radiological features includes: The clinical radiological characteristics in the training set data were placed into continuous variables, and the t-test was used when they met the normal distribution, otherwise the Mann-Whitney U test was used; For categorical variables, Chi-squared test was used when the distribution was relatively uniform, otherwise Fisher's exact test was used, and the features with significant differences were retained.

[0010] In a possible implementation, the combination of the radiomics score and the screened clinical radiological features, using a logistic regression algorithm to perform modeling to achieve prediction of high-grade components of invasive lung adenocarcinoma, includes: The model is built based on logistic regression. Based on the input independent variables, the probability of the output being a certain category is estimated. The formula is: Among them, P is the probability of predicting the target event, that is, the probability that the lesion has a high-grade component; , , …, is the input feature variable; , , …, It is the estimated parameter of the model, which indicates the influence weight of each variable on the prediction result; The training set data of R-Score and clinical radiological features with significant differences were input into the model for fitting, and the regression coefficient corresponding to each feature was calculated. , that is, to find a set of optimal coefficients so that the model can predict the probability of the target event to the greatest extent based on the input features; The preset threshold is set according to the predicted probability P for classification. When P ≥ the preset threshold, the prediction is positive, that is, the lesion has high-level components; when P < the preset threshold, the prediction is negative, that is, the lesion has no high-level components.

[0011] In a possible implementation, the calculation formula of the predicted probability P is as follows: .

[0012] In a possible implementation, it also includes: using the SHAP algorithm to perform feature importance analysis and explain the prediction results of the model.

[0013] In a possible implementation, the use of the SHAP algorithm to perform feature importance analysis and explain the prediction results of the model includes: using the SHAP model to analyze the prediction results of the logistic regression model, explaining the contribution of the feature to the prediction results by calculating the Shapley value of the feature, and for each feature The Shapley value measures the importance of a feature by calculating its marginal contribution to the model prediction result when the feature is added to all possible feature combinations. The formula is: Among them, S is a feature subset that does not include feature i; N is the set of all features; is the model output based only on the feature subset S; is the model output after adding feature i to subset S; represents the number of features in the subset S; It is a feature The Shapley value of the characteristic The marginal contribution to the forecast result.

[0014] In the present application example, by extracting and analyzing the radiomics and clinical radiological features of patients with invasive lung adenocarcinoma, a radiomics score based on the radiomics features was constructed, and a logistic regression model was established together with the clinical radiological features, and the SHAP model was used to perform feature importance analysis to explain the prediction results of the model. The experimental results show that the constructed logistic regression model has a good effect in predicting the high-grade components of invasive lung adenocarcinoma. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a process flow of a method for predicting high-level components of invasive lung adenocarcinoma based on machine learning provided in an embodiment of the present application; Figure 2 Schematic diagram of ROC curve comparison of different models provided in the embodiments of the present application, A is a comparison diagram of ROC curves of training set data on the imaging omics model, clinical model and fusion model, B is a comparison diagram of ROC curves of test set data on the imaging omics model, clinical model and fusion model; Figure 3 A bar chart showing the contribution of each feature to the prediction of high-grade components of invasive lung adenocarcinoma using SHAP analysis of the model using R-Score and screened clinical radiological features provided in the embodiment of the present application; Figure 4 A schematic scatter plot of the model provided in the embodiment of the present application using R-Score and screened clinical radiological features using SHAP to analyze the contribution of each feature to the prediction of high-grade components of invasive lung adenocarcinoma. DETAILED DESCRIPTION

[0016] The present solution is described below in conjunction with the accompanying drawings and specific implementation methods.

[0017] See also Figure 1 The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning provided in this embodiment includes: S101, obtaining original CT image data of patients with invasive lung adenocarcinoma, lesion three-dimensional volume of interest VOI and corresponding clinical radiological features.

[0018] The clinicoradiological characteristics of the patients included age, sex, smoking history, location, size, shape, density, spiculation, lobulation, cavitation, air bronchogram, pleural indentation, major histological type, histological subtype, T1 stage, and the presence or absence of STAS.

[0019] In this example, a total of 288 patients with invasive lung adenocarcinoma confirmed by postoperative pathology were included in the CT images, including 64 cases with high-grade components and 224 cases without high-grade components. The collected clinical radiological characteristics included age, gender, smoking history, location, size, shape, density, spicule sign, lobulation sign, cavitation sign, air bronchogram sign, pleural indentation sign, major histological type, histological subtype, T1 stage, and presence of STAS.

[0020] S102, preprocessing the original CT image data and extracting and screening radiomics features in combination with lesion VOI.

[0021] In this embodiment, the original CT image is resampled. Since the pixel spacing of the original image is usually inconsistent, the image needs to be resampled. The nearest neighbor interpolation method is used to calculate new pixel values ​​according to the new pixel spacing, thereby generating image data with a spacing of 1.0 mm in each direction to ensure uniformity of spatial resolution.

[0022] Extract radiomics features from preprocessed images and lesion VOIs, where the image categories of the extracted features include preprocessed images, LoG (Laplacian of Gaussian) filtered images, and wavelet images, and the feature categories include shape features, histogram features, and texture features (GLCM (Gray Level Co-occurence Matrix), GLRLM (Gray Level Run Length Matrix), GLSZM (GrayLevel Size Zone Matrix), GLDM (Gray Level Difference Matrix), and NGTDM (Neigbouring Gray Tone Difference Matrix); All data were randomly divided into training and test sets while keeping the ratio of positive and negative samples unchanged. The Z-scores of the imaging features in the training set data were standardized. The specific calculation formula is: ,in is the standardized feature, is the average value of the feature, is the standard deviation of the feature.

[0023] For the radiomic features in the training set data, the t-test was first used to evaluate the significant difference between each feature and whether there was a high-level component in the lesion, and the radiomic features with a P value less than 0.05 were selected to proceed to the next step. Then, the LASSO algorithm was used to calculate the weighted ranking of features to select the optimal features in the training set. LASSO automatically selects the optimal features and performs weighted ranking on them by introducing a regularization term in the regression model. The formula is: Where n is the number of samples; p is the number of features; is the jth feature of the i-th sample; is the true label value of the i-th sample; is the coefficient of feature j; is the regularization parameter, which controls the degree of shrinkage of the coefficients ( ≥0).

[0024] The key to LASSO lies in its regularization term , this regularization term has an effect on all coefficients Penalties are added to force some unimportant feature coefficients to shrink to 0, thereby playing the role of feature selection.

[0025] S103, constructing a radiomics score using the screened radiomics features.

[0026] Specifically, based on the screened radiomics features, the R-Score was constructed by logistic regression method, and the calculation formula of R-Score was: R-Score =-2.0379+ 0.2249 * boxsigmaimage_glszm_SizeZoneNonUniformityNormalized+0.0396 * log_glszm_log_sigma_2_mm_3D_SizeZoneNonUniformity +1.0020 * log_glszm_log_sigma_2_mm_3D_SmallAreaHighGrayLevelEmphasis+0.3679*wavelet_glcm_wavelet_HHH_ClusterShade+0.6478* laplaciansharpening_ glrlm_LongRunHighGrayLevelEmphasis+0.4029*specklenoise_gldm_DependenceVariance+0.3354*recursivegaussian_gldm_DependenceVariance+0.1831*shotnoise_firstorder_InterquartileRange-0.2873 * shotnoise_glcm_InverseVariance-0.2305 *shotnoise_glcm_Idn; Among them: boxsigmaimage_glszm_SizeZoneNonUniformityNormalized, log_glszm_log_sigma_2_mm_3D_SizeZoneNonUniformity, log_glszm_log_sigma_2_mm_ 3D_SmallAreaHighGrayLevelEmphasis, wavelet_glcm_wavelet_HHH_ClusterShade, laplaciansharpening_glrlm_LongRunHighGrayLevelEmphasis, specklenoise_gldm_DependenceVariance, recursivegaussian_gldm_DependenceVariance, shotnoise _firstorder_InterquartileRange, shotnoise_glcm_InverseVariance and shotnoise_glcm_Idn are all imaging features screened out in the training set data.

[0027] S104, performing feature screening on the clinical radiological features using statistical analysis.

[0028] For the clinical radiological features in the training set data, for continuous variables, t-test was used when they met the normal distribution, otherwise Mann-Whitney U test was used; for categorical variables, Chi-squared test was used when their distribution was relatively uniform, otherwise Fisher's exact test was used, and features with significant differences were retained.

[0029] S105, combining the radiomics score and the screened clinical radiological features, a logistic regression algorithm was used to perform modeling to predict the high-grade component of invasive lung adenocarcinoma.

[0030] The basic idea of ​​building a model based on logistic regression is to estimate the probability of the output being a certain category based on the input independent variables. The formula is: Among them, P is the probability of predicting the target event, that is, the probability that the lesion has a high-grade component; , , …, is the input feature variable; , , …, It is the estimated parameter of the model, which indicates the influence weight of each variable on the prediction result.

[0031] The training set data of R-Score and clinical radiological features with significant differences were input into the model for fitting, and the regression coefficient corresponding to each feature was calculated. , that is, find a set of optimal coefficients so that the model can predict the probability of the target event to the greatest extent based on the input features. The output of the logistic regression model is the probability P of the event occurring. The calculation formula of P is as follows: A threshold (here P is set to 0.5) is set according to the predicted probability P for classification: when P is greater than or equal to 0.5, the prediction is positive, that is, the lesion has high-level components; when P is less than 0.5, the prediction is negative, that is, the lesion has no high-level components.

[0032] The model was constructed based on logistic regression, and the R-Score and clinical radiological features with significant differences were input into the model. When calculating the results of the test set, the R-Score formula obtained from the training set features was directly used in this embodiment to calculate the R-Score of the test set, combined with the screened clinical radiological features, and input into the logistic regression model, and 0.5 was used as the cutoff threshold to obtain the prediction results. The comparison of the prediction effects of different models is shown in Table 1, and the corresponding ROC curves are shown in Figure 2 As shown in the figure. A is the ROC curve comparison of the training set data on the radiomics model, clinical model and fusion model; B is the ROC curve comparison of the test set data on the radiomics model, clinical model and fusion model. In contrast, the model combining R-Score and screened clinical radiological features has the best comprehensive performance on the test set, with the highest AUC (Area Under Curve) and a more balanced performance.

[0033] This embodiment also uses the SHAP algorithm to perform feature importance analysis and explain the prediction results of the model. Specifically, the SHAP model is used to analyze the prediction results of the logistic regression model, and the contribution of the feature to the prediction results is explained by calculating the Shapley value of the feature. For each feature The Shapley value measures the importance of a feature by calculating its marginal contribution to the model prediction result when the feature is added to all possible feature combinations. It is based on the following ideas: Among them, S is a feature subset that does not include feature i; N is the set of all features; is the model output based only on the feature subset S; is the model output after adding feature i to subset S; represents the number of features in the subset S; It is a feature The Shapley value of the characteristic The marginal contribution to the forecast result.

[0034] The larger the Shapley value of a feature, the greater its contribution to the model's prediction results. Figure 3 The results showed that the feature with the greatest impact on predicting the high-grade component of invasive lung adenocarcinoma was R-Score, followed by lobulation, density, tumor stage, pleural indentation, and spiculation. Figure 4 The results showed that patients with invasive lung adenocarcinoma with high R-Score, significant lobulation sign, high density, high tumor stage, and insignificant pleural indentation and spiculation sign were more likely to have high-grade components.

[0035] In summary, the present embodiment of the present invention uses radiomics technology to extract a large number of features from CT images and corresponding three-dimensional VOIs of patients with invasive lung adenocarcinoma, constructs a radiomics model based on logistic regression, and constructs R-Score through logistic regression. At the same time, clinical radiological features are introduced to establish a clinical model based on logistic regression, and then the final logistic regression prediction model is constructed by combining R-Score and clinical radiological features. Finally, based on the best performing combined model, the SHAP algorithm is used to perform feature importance analysis, and the results show that R-Score contributes most significantly to the prediction.

[0036] In the embodiments of the present application, "at least one" refers to one or more, and "plurality" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. Among them, A and B can be singular or plural. The character " / " generally indicates that the previous and subsequent associated objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c can be represented by: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0037] The above is only a specific implementation of the present application. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. The protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A machine learning-based method for predicting high-grade components in invasive lung adenocarcinoma. It is characterized in that include: Obtain the original CT image data, lesion three-dimensional volume of interest (VOI) and corresponding clinical radiological features of patients with invasive lung adenocarcinoma; Preprocessing the original CT image data and extracting and screening imaging omics features in combination with lesion VOI; The radiomics score was constructed by screening the radiomics features; The clinical radiological features were screened using statistical analysis; Combining the radiomics score and the screened clinical radiological features, a logistic regression algorithm was used to model the prediction of high-grade components in invasive lung adenocarcinoma.

2. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 1, It is characterized in that The preprocessing of the original CT image data and extraction and screening of imaging omics features in combination with the lesion VOI include: Resample the original CT image and set the pixel spacing of the image in three anatomical directions; Extracting radiomics features from the resampled CT images and lesion VOIs, wherein the image category data in the radiomics features include preprocessed images, Gaussian-Laplacian filtered images, and wavelet images, and the feature category data include shape features, histogram features, texture features, grayscale run matrices, grayscale size area matrices, grayscale dependency matrices, and neighborhood grayscale difference matrices; All data are randomly divided into training sets and test sets while keeping the ratio of positive and negative samples unchanged; For the data in the training set, Z-scores are standardized, and the formula is: ,in is the standardized feature, is the average value of the feature, is the standard deviation of the feature; The t-test was used to detect whether there was a significant difference between each feature and whether the lesion had a high-grade component, and the radiomics features with a P value less than 0.05 were screened out. Then, the LASSO algorithm was used to calculate the weighted ranking of the features to select the optimal features in the training set.

3. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 2, It is characterized in that The method of using the LASSO algorithm to calculate the weighted ranking of features to select the optimal features in the training set includes: LASSO introduces regularization terms into the regression model to automatically select the optimal features and perform weighted ranking on them. The formula is: Where n is the number of samples; p is the number of features; is the jth feature of the i-th sample; is the true label value of the i-th sample; is the coefficient of feature j; is the regularization parameter, which controls the degree of shrinkage of the coefficients.

4. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 1, It is characterized in that The radiomics score is constructed by screening the radiomics features, including: According to the screened radiomics features, the R-Score was constructed by logistic regression method. The calculation formula of R-Score was: R-Score = -2.0379 + 0.2249 * boxsigmaimage_glszm_SizeZoneNonUniformityNormalized + 0.0396 * log_glszm_log_sigma_2_mm_3D_SizeZoneNonUniformi ty +1.0020 * log_glszm_log_sigma_2_mm_3D_SmallAreaHighGrayLevelEmphasis + 0.3679 *wavelet_glcm_wavelet_HHH_ClusterShade + 0.6478 * laplaciansharpenin g_glrlm_LongRunHighGrayLevelEmphasis + 0.4029 * specklenoise_gldm_Dependen ceVariance+0.3354 * recursivegaussian_gldm_DependenceVariance+ 0.1831 * shotnoise_firstorder_InterquartileRange-0.2873 * shotnoise_glcm_InverseVariance-0.2305* shotnoise_glcm_Idn; Among them: boxsigmaimage_glszm_SizeZoneNonUniformityNormalized, log_glszm_ log_sigma_2_mm_3D_SizeZoneNonUniformity, log_glszm_log_sigma_2_mm_ 3D_SmallAreaHighGrayLevelEmphasis, wavelet_glcm_wavelet_HHH_ClusterShade, laplaciansharpening_glrlm_LongRunHighGrayLevelEmphasis, specklenoise_gldm_DependenceVariance, recursivegaussian_gldm_DependenceVariance, shotnoise _firstorder_InterquartileRange, shotnoise_glcm_InverseVariance and shotnoise_glcm_Idn are all imaging features screened out in the training set data.

5. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 1, It is characterized in that The method of using statistical analysis to screen the clinical radiological features comprises: The clinical radiological characteristics in the training set data were placed into continuous variables, and the t-test was used when they met the normal distribution, otherwise the Mann-Whitney U test was used; For categorical variables, Chi-squared test was used when the distribution was relatively uniform, otherwise Fisher's exact test was used, and the features with significant differences were retained.

6. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 1, It is characterized in that The combination of the radiomics score and the screened clinical radiological features, using a logistic regression algorithm to model and predict the high-grade component of invasive lung adenocarcinoma, includes: The model is built based on logistic regression. Based on the input independent variables, the probability of the output being a certain category is estimated. The formula is: Among them, P is the probability of predicting the target event, that is, the probability that the lesion has a high-grade component; , , …, is the input feature variable; , , …, It is the estimated parameter of the model, which indicates the influence weight of each variable on the prediction result; The training set data of R-Score and clinical radiological features with significant differences were input into the model for fitting, and the regression coefficient corresponding to each feature was calculated. , that is, to find a set of optimal coefficients so that the model can predict the probability of the target event to the greatest extent based on the input features; The preset threshold is set according to the predicted probability P for classification. When P is greater than or equal to the preset threshold, the prediction is positive, that is, the lesion has high-level components; when P is less than the preset threshold, the prediction is negative, that is, the lesion has no high-level components.

7. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 6, It is characterized in that The calculation formula for the predicted probability P is as follows: 。 8. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 1, It is characterized in that Also includes: The SHAP algorithm is used to perform feature importance analysis and explain the prediction results of the model.

9. The method for predicting high-grade components of invasive lung adenocarcinoma based on machine learning according to claim 8, It is characterized in that The SHAP algorithm is used to perform feature importance analysis and explain the prediction results of the model, including: using the SHAP model to analyze the prediction results of the logistic regression model, explaining the contribution of the feature to the prediction results by calculating the Shapley value of the feature, and for each feature The Shapley value measures the importance of a feature by calculating its marginal contribution to the model prediction results when the feature is added to all possible feature combinations. The formula is: Among them, S is a feature subset that does not include feature i; N is the set of all features; is the model output based only on the feature subset S; is the model output after adding feature i to subset S; represents the number of features in subset S; It is a feature The Shapley value of the characteristic The marginal contribution to the forecast result.

Citation Information

Patent Citations

  • Auxiliary assessment method for prognosis of nasopharynx cancer based on enhanced MRI radiomics

    CN111657945A

  • Lung cancer prognosis auxiliary evaluation method and system based on CT radiomics

    CN112951406A

  • Medical image classification method and equipment combining radiomics features and clinical features

    CN116229176A

  • Pneumocystis and aspergillus pulmonary infection identification system based on radiomics

    CN116630719A

  • Method, device and equipment for predicting curative effect of advanced breast cancer treatment based on CT (Computed Tomography) radiomics

    CN117115615A