Construction method of brain tumor distinguishing model

Abstract

The invention relates to the technical field of medical data processing, and particularly provides a brain tumor distinguishing model construction method, which is characterized in that a radiomics model is constructed based on a brain-tumor junction region (BTI), and the model can effectively distinguish glioblastoma (GBM) and isolated brain metastases (SBM); the GBM and the SBM have significant overlap in subjective image features, the brain tumor distinguishing model based on radiomics constructed by the invention effectively makes up for uncertainty possibly caused by subjective judgment, and the BTI region-based logic regression model provided by the invention has high prediction capability and clinical practicability. An important auxiliary tool is provided for clinical diagnosis, and the diagnosis accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and particularly to a method for constructing a brain tumor differentiation model. Background Art

[0002] Glioblastoma multiforme (GBM) and brain metastases (BM) are the most common malignant brain tumors in adults, and they have significant clinical and pathological differences. GBM is the most common and highly invasive primary malignant brain tumor in adults, with extremely poor prognosis. In contrast, BM is a secondary tumor formed by hematogenous dissemination of malignant tumors from other sites, and its treatment strategy and prognosis are closely related to the type and stage of the primary tumor. Although some cases of BM can be clearly diagnosed through a previous history of malignancy or multiple lesions, approximately 30% of BM patients present with brain metastases as the initial manifestation, and approximately 50% of them are solitary brain metastases (SBM). In addition, GBM and SBM often exhibit overlapping features on conventional MRI images, such as similar enhancement patterns and high signal intensity of the surrounding edema shown on T2-weighted images (T2WI), which poses a major challenge to accurate differentiation.

[0003] Given the different treatment strategies for GBM and SBM, their differential diagnosis has important clinical significance. Therefore, developing a non-invasive diagnostic method that can accurately distinguish GBM and SBM has considerable clinical value.

[0004] In recent years, radiomics has emerged as a new quantitative analysis method, which extracts high-throughput quantitative features from medical images and provides a new technical means for accurate tumor diagnosis. Radiomics can overcome the subjectivity of traditional diagnostic methods, discover microscopic radiomics features that are difficult to identify with the naked eye, and thus improve the objectivity and accuracy of diagnosis. Currently, the application of radiomics in brain tumor research mainly focuses on the feature analysis of the entire tumor region or specific sub-regions (such as tumor enhancement, necrosis, or the surrounding edema zone), and certain research results have been achieved.

[0005] The brain tumor interface (BTI), as the transitional region between tumor tissue and normal brain tissue, has unique pathophysiological significance. This region not only contains information on tumor cell infiltration but also reflects the interaction between the tumor and the surrounding brain tissue, such as angiogenesis, inflammation, and extracellular matrix remodeling. The present invention focuses on this transitional region and aims to construct a stable and efficient differential diagnosis model based on the radiomics features of the 10mm BTI region to provide support for clinical decision-making. Summary of the Invention

[0006] The purpose of the present invention is to overcome the drawbacks and deficiencies of the prior art, and to provide a method for constructing a brain tumor differentiation model. The technical solution adopted by the present invention is as follows:

[0007] A method for constructing a brain tumor differentiation model, comprising the following steps:

[0008] S1. Obtain the radiomics information and diagnostic information of at least two brain tumors to be differentiated;

[0009] S2. Perform image preprocessing on the radiomics information;

[0010] S3. Outline the ROI region in the preprocessed image and perform brain tumor interface (BTI) region segmentation;

[0011] S4. Extract the radiomics features within the BTI region and perform hierarchical screening on the obtained radiomics features within the BTI region, including: screening for robust features, screening for features with statistical significance from the robust features, and screening for representative features from the features with statistical significance;

[0012] S5. Input the several features obtained after hierarchical screening into a logistic regression model for training.

[0013] Preferably, in step S1, the screening conditions for the radiomics information used to construct this brain tumor differentiation model include: the types of brain tumors are confirmed by pathology; there is no history of brain lesion surgery, radiotherapy, or other treatments; there is no mixed tumor type within the lesion; the exclusion conditions include: poor image quality; the maximum diameter of the lesion is less than 2 cm; multiple intracranial lesions.

[0014] As one of the embodiments, in step S1, obtain the radiomics information and diagnostic information of pathologically confirmed brain metastases and glioblastoma multiforme.

[0015] Preferably, the image preprocessing in step S2 includes: format conversion, resampling to achieve an isotropic voxel size of 1×1×1 mm 3 and image correction.

[0016] Preferably, in step S3, the specific steps for outlining the ROI region include: on the T1-weighted contrast-enhanced image, layer by layer depict the ROI, and finally synthesize a three-dimensional volume of interest (VOI) covering the entire tumor lesion.

[0017] Preferably, in step S3, the BTI region segmentation is an automatic segmentation of the brain tumor interface based on Python code. The specific operation includes: expanding the entire tumor edge outward by 5 mm and then contracting it inward by 5 mm to generate a BTI ROI with a diameter of 10 mm.

[0018] Preferably, in step S4, use PyRadiomics to extract features from the TBI region according to the Image Biomarker Standardization Initiative, including first-order features, shape features, texture features, and wavelet features.

[0019] Further, in step S4, the radiomics features extracted within the BTI region are subjected to ICC test to obtain robust features; the robust features are subjected to Z-score normalization, followed by hypothesis testing to screen out features with statistical significance; based on the features with statistical significance, feature selection is performed through Pearson correlation coefficient and minimum redundancy maximum correlation algorithm to retain the most representative features.

[0020] Preferably, in step S4, after obtaining the most representative features, Lasso regression is used to further optimize the feature set, and finally several key features are selected.

[0021] Preferably, several features obtained after step-by-step screening are used as a training set to input into a logistic regression model for training, and its accuracy is evaluated using a test set with the same screening and processing criteria as the training set.

[0022] The beneficial effects of the present invention are as follows: The present invention constructs a radiomics model based on the brain-tumor interface region, which can effectively distinguish glioblastoma multiforme (GBM) and solitary brain metastasis (SBM); there are significant overlaps in the subjective imaging features between GBM and SBM, which may lead to misjudgment by doctors. The brain tumor differentiation model based on radiomics constructed by the present invention effectively makes up for this deficiency, provides an important auxiliary tool for clinical diagnosis, and significantly improves the diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, obtaining other drawings based on these drawings still belongs to the scope of the present invention.

[0024] Figure 1 is the entire usage process and internal decision-making process of the tumor discrimination model;

[0025] Figure 2 is the feature weight map finally included in the modeling;

[0026] Figure 3 in which, the ROC curve graphs of the model in the training set (a) and the test set (b);

[0027] Figure 4 is the calibration curve graph of the model on the training set (a) and the test set (b)

[0028] Figure 5 is the decision curve graph of the model on the training set (a) and the test set (b)

[0029] Figure 6 Confusion matrix diagrams of the model on the training set (a) and the test set (b).

[0030] Figure 7 Among them, (a) is the distribution diagram of feature importance visualized by SHAP, (b) is the sample decision diagram visualized by SHAP, and (c) is the visualization presentation of the first sample in SHAP visualization. Detailed implementation manners

[0031] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the implementation cases and the accompanying drawings.

[0032] The present invention provides a method for constructing a brain tumor differentiation model. Before introducing the embodiments, the material sources of the present invention and various data acquisition and processing methods used in the embodiments are first introduced.

[0033] I. Material sources and collection methods

[0034] 1. Data sources and research objects

[0035] The present invention retrospectively analyzed the data in the databases of three medical centers. The inclusion criteria are as follows: (1) Brain metastases or glioblastoma multiforme (GBM) confirmed by pathology; (2) No history of brain lesion surgery, radiotherapy or other treatments; (3) No mixed tumor types within the lesion. The exclusion criteria include: (1) Poor image quality; (2) The maximum diameter of the lesion is less than 2 cm; (3) Multiple intracranial lesions. The total number of patients finally included in the analysis is 432.

[0036] Among them, 187 patients (96 GBM and 91 SBM) were included in Center 1; 106 patients (68 GBM and 38 SBM) were included in Center 2; 139 patients (62 GBM and 77 SBM) were included in Center 3. The present invention uses the data sets of Center 1 and Center 2 as the training set, and the data of Center 3 as the test set. Ethical approval has been approved by the ethics committee, and patient informed consent has been waived.

[0037] II. Image processing methods

[0038] Preprocessing

[0039] 1. Format conversion: The T1WI enhanced sequence data of eligible patients are exported from the Picture Archiving and Communication System (PACS) in the Digital Imaging and Communications in Medicine (DICOM) format, and then converted to the Neuroimaging Informatics Technology Initiative (NIfTI) format using MRIcroGL software (https: / / www.nitrc.org / projects / mricrogl / ), and the patient information is anonymized.

[0040] 2. Resampling: Use Python (https: / / www.python.org / ) code to resample all sample files to achieve an isotropic voxel size of 1×1×1 mm 3 for each voxel.

[0041] 3. Image correction: Implement N4 bias correction through a Python script to correct the intensity inhomogeneity (i.e., bias field) in MRI images. Image normalization eliminates variations caused by different imaging devices and scanning parameters, ensuring data consistency. Discretization converts continuous pixel intensity values into a finite number of discrete levels, improving the repeatability and stability of feature extraction.

[0042] Region of Interest (ROI) Segmentation

[0043] 1. Tumor segmentation:

[0044] After image preprocessing, import the T1-weighted contrast-enhanced MRI images into the ITK-SNAP software (version 4.2.0; http: / / www.itksnap.org / pmwiki / pmwiki.php) to delineate the tumor region (ROI). The specific steps are as follows: On the T1-weighted contrast-enhanced image, manually delineate the ROI layer by layer, and finally synthesize a three-dimensional volume of interest (VOI) that covers the entire tumor lesion. To evaluate the stability of radiological features, inter-observer and intra-observer reproducibility analyses were performed. Specifically, 20 images were randomly selected, and observer one with 2 years of MRI diagnosis experience and observer two with 6 years of MRI diagnosis experience independently completed the lesion segmentation. The radiological features extracted were used to evaluate inter-observer reproducibility. For the remaining cases, observer one performed the lesion segmentation to ensure data consistency.

[0045] 2. BTI region segmentation: Automatic segmentation of the brain tumor interface (BTI) based on Python code. By expanding the entire tumor margin outward by 5 mm and then contracting it inward by 5 mm, an automatic BTI ROI with a diameter of 10 mm is generated. After processing, import the ROI and the corresponding image into the ITK-SNAP software, and manually erase the parts covering irrelevant regions (such as the skull) to ensure segmentation accuracy. The same automatic segmentation and manual correction procedures were also applied to 20 samples used for consistency testing to verify the reliability and robustness of the method.

[0046] III. Radiomics Feature Extraction

[0047] After the image preprocessing, PyRadiomics was used to extract radiomics features from the 10 mm TBI region according to the Image Biomarker Standardization Initiative (IBSI) (31). A total of 833 features were extracted, including 18 first-order features, 14 shape features, 73 texture features, and 728 wavelet features. The detailed information of these features can be found in the supplementary material.

[0048] IV. Consistency Test

[0049] Twenty samples were randomly selected, covering different categories and feature distributions, to ensure that the test results were widely representative. For each radiomics feature, the intraclass correlation coefficient (ICC) was used for the consistency test. ICC is a commonly used statistical index to evaluate the consistency of measurement values between different evaluators or measurements, and is applicable to evaluating the reliability of continuous variables. To ensure that the selected radiomics features had high stability, only the features with an ICC coefficient greater than 0.8 were retained.

[0050] V. Feature Selection

[0051] A multi-step feature selection strategy was adopted to process and screen the radiomics features. First, the dataset was divided into a training set and a test set according to the predetermined grouping information. To eliminate the dimensional differences between different features, Z-score normalization was applied to all features. Feature selection included three stages: First, hypothesis testing was performed, and a normality test (Shapiro-Wilk test) was conducted for each feature. Features that met the normal distribution criteria were evaluated using an independent samples t-test, and those that did not were evaluated using the Mann-Whitney U test. Only the features with statistical significance (p < 0.05) were retained. Next, the Person correlation threshold was set to 0.9, and the top 20 features were selected according to the maximum correlation and minimum redundancy (mRMR) algorithm. Finally, the least absolute shrinkage and selection operator (LASSO) method was used to further optimize the most representative features for constructing the radiomics model and their corresponding calculation formulas.

[0052] VI. Model Construction

[0053] After comprehensive feature selection, an independent model was constructed for each feature set. Nine different algorithms were used in this invention for model construction, including logistic regression, naive Bayes, support vector machine, random forest, extremely randomized trees, light gradient boosting, gradient boosting decision tree, adaptive boosting, and multi-layer perceptron. The training and evaluation of all models were completed using the scikit-learn library in Python, and five-fold cross-validation was used to evaluate the performance.

[0054] VII. Radiologist Diagnosis Model

[0055] To comprehensively evaluate the performance of the radiomics model, this study constructed diagnostic models based on radiologists with different levels of experience on the test dataset and compared these models with the radiomics model. Three radiologists (with 2 years, 6 years, and 18 years of diagnostic experience respectively) independently diagnosed all test samples under double-blind conditions (i.e., without knowing the patient's clinical information and pathological results). By comparing the diagnostic results with radiologists of different years of experience, this study aimed to verify the diagnostic effectiveness and reliability of the radiomics model in actual clinical applications and evaluate its value in assisting diagnosis at different experience levels.

[0056] VIII. Model Performance Evaluation

[0057] The robustness of the model performance was evaluated through five-fold cross-validation. The performance of each model was quantified by multiple metrics, including the area under the ROC curve (AUC), accuracy (ACC), sensitivity (Sen), specificity (Spe), F1-score (F1), recall, positive predictive value (PPV), and negative predictive value (NPV). To further compare the performance differences between models, the DeLong test was performed using MedCalc software to evaluate the statistical significant differences in AUC values between models. All analyses used a p-value < 0.05 as the threshold for statistical significance.

[0058] IX. SHAP Visualization

[0059] SHAP (Shapley Additive exPlanations) is an interpretable machine learning tool based on the Shapley value in cooperative game theory. It provides an accurate and intuitive method to explain complex machine learning models. Its core principle is to regard each feature as a "player" in the cooperative game and quantify the contribution of each feature to the model prediction result through the Shapley value. The unique advantage of SHAP is that it can not only evaluate the importance of features but also clearly reveal the specific direction (positive or negative) and influence degree of features on the prediction result. This fine-grained interpretability ability is crucial for enhancing the interpretability and credibility of the model.

[0060] X. Statistical Analysis

[0061] All statistical analyses were performed using R and Python packages. When analyzing the differences between two sets of clinical data, appropriate statistical methods were selected according to the characteristics of the data distribution. For continuous variables that follow a normal distribution, an independent samples t-test was used, and the results were expressed as mean ± standard deviation. For continuous variables that do not follow a normal distribution, non-parametric tests such as the Mann-Whitney U test were adopted, and the results were described by the median (interquartile range). Categorical variables were compared between groups using the chi-square test. To compare the areas under the ROC curves (AUC) between different models, the DeLong test was used, and a p-value < 0.05 was used as the criterion for statistical significance. Calibration curves were plotted to evaluate the consistency between predicted probabilities and actual probabilities, and the Hosmer-Lemeshow test was used to evaluate the model fit, with a p-value > 0.05 indicating a good model fit. In addition, decision curve analysis (DCA) was performed to evaluate the clinical net benefit of the model at different threshold probabilities, thereby quantifying its value in clinical applications.

[0062] Example

[0063] 1. Clinical and magnetic resonance imaging characteristics

[0064] A total of 432 brain tumor patients from three centers were included in this example, including 226 patients with glioblastoma multiforme (GBM) and 206 patients with solitary brain metastases (SBM). Table 1 provides the detailed distribution of demographic information and clinical characteristics in the training set and test set. In the training set, the age of patients in the GBM group was significantly lower than that in the SBM group (P = 0.005), and there were significant differences in tumor location between the two groups (P < 0.001), but there was no significant difference in gender distribution (P > 0.05). In the test set, there were also significant differences in tumor location between the GBM group and the SBM group (P < 0.001), while no significant differences were observed in age and gender distribution (P > 0.05).

[0065] Table 1. Comparison of general clinical information of patients in two datasets

[0066]

[0067] 2. Selection of radiomics features

[0068] In this embodiment, a total of 833 radiomics features were extracted from a 10-mm region at the junction of brain tumors and normal brain tissues. To ensure the robustness and reproducibility of the selected features, an intraclass correlation coefficient (ICC) analysis was performed on the features extracted from the regions of interest (ROIs) independently delineated by two doctors, and the ICC threshold was set at 0.8. After the ICC test, 828 robust features were retained within the ROI. This screening process effectively eliminated the features with large observer variability, ensuring the reliability and stability of the features used in subsequent analyses.

[0069] After the ICC test, the 828 retained radiomics features were first subjected to Z-score normalization to eliminate the dimensional differences between the features. Subsequently, a hypothesis test was conducted to further screen out 463 features with statistical significance. On this basis, feature selection was performed through the Pearson correlation coefficient and the minimum redundancy maximum relevance (mRMR) algorithm, and finally 20 of the most representative features were retained. Finally, Lasso regression was used to further optimize the feature set, and 10 key features were ultimately selected. Figure 2 The weight distribution of these 10 selected features is shown, intuitively reflecting the contribution of each feature to the model prediction result. The calculation formula is as follows:

[0070] Rad_socre = 0.5597

[0071] + 0.0356 × BTI_original_shape_MajorAxisLength

[0072] + 0.0743 × BTI_original_glszm_SizeZoneNonUniformity

[0073] - 0.0330 × BTI_wavelet_LLL_glszm_SmallAreaLowGrayLevelEmphasis

[0074] - 0.0018 × BTI_original_gldm_LowGrayLevelEmphasis

[0075] + 0.1397 × BTI_wavelet_HHH_gldm_DependenceNonUniformity

[0076] - 0.0276 × BTI_wavelet_LLL_glcm_MaximumProbability

[0077] +0.0120×BTI_wavelet_HHH_glszm_SizeZoneNonUniformityNormalized

[0078] +0.0019×BTI_wavelet_LHL_glszm_SizeZoneNonUniformity

[0079] -0.0045×BTI_wavelet_HHH_ngtdm_Strength

[0080] -0.0607×BTI_original_shape_Elongation

[0081] 3. Model construction

[0082] After feature selection, nine machine learning algorithms were adopted, including Logistic Regression (LR), Naive Bayes (NB), Support Vector Machine (SVM), Random Forest (RF), Extra Trees, XGBoost, Gradient Boosting Decision Tree (GBDT), AdaBoost, and Multi-Layer Perceptron (MLP), to construct multiple models. Through comparative analysis, the optimal model was selected. Among the models constructed based on the 10-mm brain tumor and normal brain tissue interface (BTI), the Logistic Regression model was determined to be the best model. The AUC value of this model on the test set was 0.808, and its ROC curve is as Figure 3 shown. The Hosmer-Lemeshow (HL) test showed that the p-values of both the training set and the test set were greater than 0.05, indicating good model fitting. The calibration curve of the model is as Figure 4 shown, and the results of the Decision Curve Analysis (DCA) are as Figure 5 shown. In addition, the confusion matrix of the model is as Figure 6 shown, further illustrating its performance in the classification task. These results indicate that the Logistic Regression model based on the BTI region has high predictive ability and clinical practicability, and can effectively distinguish different types of brain tumors.

[0083] Control group

[0084] Radiologist diagnosis model

[0085] The present invention invited three radiologists with different clinical experiences to participate in constructing a clinical diagnosis model. Their diagnostic experiences were 2 years (Doctor 1), 6 years (Doctor 2), and 18 years (Doctor 3) respectively. Each doctor independently completed the discrimination task for the test set samples and established the corresponding diagnosis model. Through the systematic evaluation of various indicators of the model, the results showed that generally, the diagnosis model of the doctor with the most experience (Doctor 3) performed the best, with an AUC value of 0.789 and an accuracy rate of 0.784; the performance of the doctor with medium experience (Doctor 2) was the second, with an AUC value of 0.740 and an accuracy rate of 0.741; while the model performance of the doctor with the least experience (Doctor 1) was relatively low, with an AUC value of 0.699 and an accuracy rate of 0.698.

[0086] Comparison of Diagnostic Methods

[0087] The present invention established a radiomics model based on the 10-mm brain-tumor interface to distinguish glioblastoma multiforme (GBM) from solitary brain metastases (SBM). To evaluate the differences between this radiomics model and traditional manual diagnosis, three doctors with different diagnostic experiences (Doctor 1 with 2 years of experience; Doctor 2 with 6 years of experience; Doctor 3 with 18 years of experience) were invited to manually judge the test data. The results showed that on the test set, the AUC value of Model TBI was 0.808. The DeLong test showed that the performance of Model TBI was significantly better than that of Doctor 1 (p < 0.05). Although the AUC value of Model TBI was higher than that of Doctor 2 and Doctor 3 on the test set, the difference was not statistically significant. In addition, in the comparison between doctors, the AUC value of Doctor 3 was significantly higher than that of the less experienced Doctor 1 (p < 0.05), but the difference between Doctor 3 and Doctor 2 was not statistically significant. Similarly, the AUC value of Doctor 2 was higher than that of Doctor 1, but the difference was also not statistically significant.

[0088] SHAP Visualization

[0089] The present invention used the SHAP method to perform a visualization analysis on the prediction results of Model TBI . SHAP values can quantify the contribution of each feature to the model output, thus helping to understand the decision-making process of the model ( Figure 7 ).

[0090] In summary, the present invention constructs a radiomics model based on the 10-mm brain-tumor junction region, which can effectively distinguish glioblastoma multiforme (GBM) from solitary brain metastases (SBM). Although there are significant overlaps in subjective imaging features between GBM and SBM, the radiomics model constructed in this study effectively makes up for this deficiency, provides an important auxiliary tool for clinical diagnosis, and significantly improves the diagnostic accuracy.

[0091] The foregoing disclosure is only a preferred embodiment of the present invention, and of course it cannot be used to limit the scope of rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope covered by the present invention.

Claims

1. A method for constructing a brain tumor differentiation model, characterized in that, It includes the following steps: S1. Obtain the radiomics information and diagnostic information of at least two types of brain tumors to be distinguished; S2. Perform image preprocessing on the radiomics information; S3. Outline the ROI region in the preprocessed image and perform brain tumor interface (BTI) region segmentation; S4. Extract the radiomics features within the BTI region, and perform step-by-step screening on the obtained radiomics features within the BTI region, including: screening for robust features, screening for features with statistical significance from the robust features, and screening for representative features from the features with statistical significance; S5. Input the several features obtained after step-by-step screening into a logistic regression model for training.

2. The method for constructing a brain tumor differentiation model according to claim 1, wherein, In step S1, the screening conditions for the radiomics information used to construct this brain tumor discrimination model include: the types of brain tumors are confirmed by pathology; there is no history of brain lesion surgery, radiotherapy or other treatments; there is no mixed tumor type within the lesion; The exclusion conditions include: poor image quality; the maximum diameter of the lesion is less than 2 cm; multiple intracranial lesions.

3. The method for constructing a brain tumor differentiation model according to claim 1, wherein, In step S2, the image preprocessing includes format conversion, resampling to achieve an isotropic voxel size of 1×1×1 mm 3 , and image correction.

4. The construction method of a brain tumor differentiation model according to claim 1, characterized in that In step S3, the specific steps for outlining the ROI region include: on the T1-weighted contrast-enhanced image, depict the ROI layer by layer, and finally synthesize a three-dimensional volume of interest (VOI) covering the entire tumor lesion.

5. The construction method of a brain tumor differentiation model according to claim 1, characterized in that, In step S3, the BTI region segmentation is an automatic segmentation of the brain tumor interface based on Python code. The specific operation includes: expanding the entire tumor edge outward by 5 mm and then contracting it inward by 5 mm to generate a BTI ROI with a diameter of 10 mm.

6. The construction method of a brain tumor differentiation model according to claim 1, characterized in that In step S4, use PyRadiomics to extract features from the TBI region according to the Image Biomarker Standardization Initiative, including first-order features, shape features, texture features, and wavelet features.

7. The construction method of a brain tumor differentiation model according to claim 1, wherein, In step S4, the radiomics features extracted within the BTI region are subjected to ICC test to obtain robust features; Perform Z-score standardization on the robust features, and then perform hypothesis testing to screen out features with statistical significance; Based on the features with statistical significance, perform feature selection through the Pearson correlation coefficient and the minimum redundancy maximum correlation algorithm to retain the most representative features.

8. The construction method of a brain tumor differentiation model according to claim 7, characterized in that, In step S4, after obtaining the most representative features, use Lasso regression to further optimize the feature set, and finally select several key features.

9. The construction method of a brain tumor differentiation model according to claim 1, characterized in that Use the several features obtained after step-by-step screening as the training set to input into a logistic regression model for training, and evaluate its accuracy using a test set with the same screening and processing standards as the training set.

10. A method for constructing a brain tumor differentiation model according to any one of claims 1-9, characterized in that, In step S1, obtain the radiomics information and diagnostic information of pathologically confirmed brain metastases and glioblastoma.

Images