Breast cancer recurrence risk prediction method based on G2 and G3 grading patients

Abstract

A breast cancer recurrence risk prediction method based on G2 and G3 grading patients comprises the following steps: step 1, collecting data of breast cancer patients, and labeling the data to obtain an integrated data set with class labels and an integrated data set without class labels; 2, constructing a breast cancer recurrence prediction model suitable for clinical G2 and G3 grading; step 3, based on a Self-training semi-supervised learning principle, obtaining a semi-supervised learning integrated prediction model; step 4, calculating an SHAP value of each feature in the data set to obtain an importance sequence of the features to a prediction result, and establishing a semi-supervised integrated prediction model based on the SHAP; and step 5, collecting MRI image omics characteristics and clinical data of breast cancer patients of real clinical G2 and G3 grades, inputting the MRI image omics characteristics and clinical data into the semi-supervised integrated prediction model based on the SHAP, and analyzing and predicting the recurrence risk of the breast cancer patients by using the model parameters obtained in the step 3. The method can be used for breast cancer recurrence prediction, and the accuracy of recurrence risk prediction is improved.

Description

Technical Field

[0001] The invention belongs to the field of breast cancer recurrence risk prediction, and specifically is a breast cancer recurrence risk prediction method based on G2 and G3 graded patients. Background Art

[0002] Breast cancer is one of the most common malignant tumors in women worldwide. Its increasing incidence and mortality have become a major challenge in the field of public health. Although significant progress has been made in the field of breast cancer risk assessment, especially in the prediction and treatment of HER2-negative breast cancer, there is still a large gap in the risk assessment research for breast cancer patients with clinical grades of G2 and G3. The recurrence risk assessment of these high-grade tumors is crucial to guide clinical treatment decisions, but there is currently a lack of accurate prediction tools and sufficient clinical data support.

[0003] HER2-negative breast cancers account for the majority of all breast cancers, and G2 and G3 grade tumors generally indicate a higher malignant potential and risk of recurrence. However, existing risk assessment models often do not adequately take into account this important factor, tumor grade, or their application in these specific subtypes has not been fully validated. Some models may perform well in low-grade tumors but lack accuracy in G2 and G3 grade tumors.

[0004] In addition, current risk assessment tools often rely on limited clinical and pathological parameters, without making full use of emerging biomarkers, gene expression data and imaging genomics features. The application of these new biomarkers and advanced technologies is expected to improve the accuracy of predicting the risk of recurrence in patients with G2 and G3 breast cancer. In order to fill this gap, it is of great significance to develop and verify new prediction models to integrate more comprehensive biological information and advanced data analysis methods. At the same time, enhancing the interpretability of the prediction results, thereby providing more accurate risk assessment and personalized treatment recommendations for these high-risk patients, is also a technical problem that the present invention aims to solve. Summary of the invention

[0005] The present invention provides a method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grades, so as to solve the defects in the prior art.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grading, comprising the following steps:

[0008] Step 1: Collect MRI imaging features and clinical data of breast cancer patients with clinical G2 and G3 grades, perform data feature engineering on the collected data, construct the processed data into an integrated data set of clinical features combined with MRI imaging features, and label the data to obtain an integrated data set with class labels and an integrated data set without class labels;

[0009] Step 2: Build a breast cancer recurrence prediction model suitable for clinical G2 and G3 grades, use the integrated data set with class labels obtained in step 1 as a training set to train the model and obtain the trained model parameters;

[0010] Step 3: Based on the principle of self-training semi-supervised learning, the unlabeled integrated data set obtained in step 1 is used as a test set to evaluate the model trained in step 2 and adjust the parameters according to the evaluation results to obtain a semi-supervised learning integrated prediction model;

[0011] Step 4: Based on the semi-supervised learning ensemble prediction model obtained in step 3, calculate the SHAP value of each feature in the data set, obtain the importance ranking of the features in the semi-supervised ensemble prediction model to the prediction results, analyze the contribution of the features to the model prediction results, and establish a semi-supervised ensemble prediction model based on SHAP;

[0012] Step 5: Collect MRI imaging features and clinical data of real clinical G2 and G3 grade breast cancer patients, input them into the SHAP-based semi-supervised ensemble prediction model, and use the model parameters obtained in step 3 to analyze and predict the recurrence risk of breast cancer patients.

[0013] As described above, in the method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grades, in the step 1, when collecting the MRI imaging features of G2 and G3 breast cancer patients, Pyradiomics is used to extract imaging features from the patients' MRI images, and the MRI imaging feature data covers the tumor conditions of G2 and G3 breast cancer patients with different degrees of disease.

[0014] In the method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grades as described above, the MRI imaging feature data include: partial tumor volume (SER), T1-weighted image washed tissue volume (SER), tumor-free area (Euler median), and tumor lesion signal intensity.

[0015] As described above, a method for predicting the risk of breast cancer recurrence based on patients with G2 and G3 grades,

[0016] The integrated data set with class labels in step 1 is:

[0017] a. Data containing clear label information, wherein the label information indicates the category or result to which it belongs;

[0018] b. An integrated dataset containing clinical feature data and MRI radiomics feature data;

[0019] c. Data used to train and test machine learning models to ensure that the models can learn the predictive patterns of the data;

[0020] The integrated dataset without class labels in step 1 is:

[0021] a. Data that does not contain clear label information, and the category or result is unknown;

[0022] b. An integrated dataset containing clinical feature data and MRI radiomics feature data;

[0023] c. Used in the optimization step of semi-supervised learning to give pseudo labels through model prediction to enhance the generalization ability of the model.

[0024] In the method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 stages as described above, when dividing the integrated data set with class labels and the integrated data set without class labels in the step one, the label of patients who have no recurrence for more than five years after chemotherapy is recorded as 1, and the label of patients who have recurred for more than five years after chemotherapy is recorded as 0, both of which are used as data with class labels; the average recurrence time of breast cancer patients with G2 and G3 stages after chemotherapy in the step one is concentrated in 2-5 years, and the label of patients who have no recurrence within two years after chemotherapy is recorded as 3, as data without class labels.

[0025] In the above-mentioned method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grades, the extreme gradient boosting (XGBoost) algorithm is used to construct a machine learning model in step 2.

[0026] In the above-mentioned method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grades, the specific operation of using the unlabeled integrated data set obtained in step 1 as a test set to evaluate the model trained in step 2 is:

[0027] Step 1: Input the unlabeled integrated data set into the model, and assign pseudo label 1 to the patient data predicted by the model as positive and with a probability greater than the mean probability of positive prediction of the current test set, i.e., recurrence;

[0028] Step 2: Incorporate pseudo-labeled patient data into the labeled dataset to train a new integrated model and adjust parameters;

[0029] Step 3: Repeat steps 1 and 2 until there is no positive label in the unlabeled data predicted by the model or the positive label prediction result is not greater than the mean prediction probability of the entire class in this round of test set. The remaining labels are assigned pseudo labels of 0, i.e., no recurrence, and the parameters are adjusted to finally obtain the semi-supervised learning integrated prediction model.

[0030] In the above-mentioned method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grades, the specific operation of obtaining the importance ranking of features to the prediction results in step 4 is: according to the SHAP average absolute value of the model, the SHAP value of each feature is calculated, and the importance is ranked according to the average absolute SHAP value.

[0031] As described above, a method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grades, the features include partial tumor volume (SER), T1 weighted image washout tissue volume (SER), tumor-free area (Euler median), progesterone receptor, tumor lesion signal intensity, tumor stage, chemotherapy method, lymph node stage, menopause, lymph node lesions, and estrogen receptor; the top 5 features in terms of feature importance are used as predictive features for predicting the recurrence risk of breast cancer; according to the importance ranking of the average absolute SHAP value, the top 5 features in terms of importance are: partial tumor volume (SER), T1 weighted image washout tissue volume (SER), tumor-free area (Euler median), progesterone receptor, and tumor lesion signal intensity.

[0032] The advantages of the present invention are as follows: the present invention establishes an integrated data set with class labels and an integrated data set without class labels based on clinical characteristic data and MRI impact omics feature data of breast cancer patients with clinical G2 and G3 grades, further expands the available data of recurrent patients, improves model performance for better prediction, and solves the problem that the existing G2 and G3 breast cancer recurrence lacks accurate prediction tools and sufficient clinical data support; at the same time, the present invention combines the clinical characteristics and MRI imaging data of patients, uses the XGBoost algorithm to construct a preliminary recurrence risk prediction model, and optimizes the model using the Self-training semi-supervised learning principle to enhance the generalization ability and prediction performance of the model; moreover, the present invention also performs interpretability analysis through the SHAP method, provides clinicians with a transparent explanation of the model prediction, and brings more accurate treatment plans to breast cancer patients; the present invention creatively uses the SHAP-based semi-supervised learning model in combination with MRI imaging omics features for G2 and G3 breast cancer recurrence prediction, which not only improves the accuracy of recurrence risk prediction, but also provides the interpretability of the model through the SHAP value. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0034] Figure 1 is a flow chart diagram of the present invention;

[0035] Figure 2 is a flow chart of model construction training of the present invention;

[0036] Figure 3 It is an important characteristic SHAP diagram of the model of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] like Figure 1 A method for predicting the recurrence risk of breast cancer based on patients with G2 and G3 grading is shown, comprising the following steps:

[0039] Step 1: Collect MRI imaging features and clinical data of breast cancer patients with clinical G2 and G3 grades, perform data feature engineering on the collected data, construct the processed data into an integrated data set of clinical features combined with MRI imaging features, and label the data to obtain an integrated data set with class labels and an integrated data set without class labels;

[0040] Step 2: Build a breast cancer recurrence prediction model suitable for clinical G2 and G3 grades, use the integrated data set with class labels obtained in step 1 as a training set to train the model and obtain the trained model parameters;

[0041] Step 3: Based on the principle of self-training semi-supervised learning, the unlabeled integrated data set obtained in step 1 is used as a test set to evaluate the model trained in step 2 and adjust the parameters according to the evaluation results to obtain a semi-supervised learning integrated prediction model;

[0042] Step 4: Based on the semi-supervised learning ensemble prediction model obtained in step 3, calculate the SHAP value of each feature in the data set, obtain the importance ranking of the features in the semi-supervised ensemble prediction model to the prediction results, analyze the contribution of the features to the model prediction results, and establish a semi-supervised ensemble prediction model based on SHAP;

[0043] Step 5: Collect MRI imaging features and clinical data of real clinical G2 and G3 grade breast cancer patients, input them into the SHAP-based semi-supervised ensemble prediction model, and use the model parameters obtained in step 3 to analyze and predict the recurrence risk of breast cancer patients.

[0044] Specifically, in step 1 described in this embodiment, when collecting the MRI imaging features of G2 and G3 breast cancer patients, Pyradiomics is used to extract imaging features from the patients' MRI images. The MRI imaging feature data covers the tumor conditions of G2 and G3 breast cancer patients with different degrees of disease.

[0045] Specifically, the MRI imaging feature data described in this embodiment include: partial tumor volume (SER), T1-weighted image washed tissue volume (SER), tumor-free area (Euler median), and tumor lesion signal intensity.

[0046] More specifically, the integrated data set with class labels in step 1 of this embodiment as described above refers to:

[0047] a. Data containing clear label information, wherein the label information indicates the category or result to which it belongs;

[0048] b. An integrated dataset containing clinical feature data and MRI radiomics feature data;

[0049] c. Data used to train and test machine learning models to ensure that the models can learn the predictive patterns of the data;

[0050] The integrated dataset without class labels in step 1 is:

[0051] a. Data that does not contain clear label information, and the category or result is unknown;

[0052] b. An integrated dataset containing clinical feature data and MRI radiomics feature data;

[0053] c. Used in the optimization step of semi-supervised learning to give pseudo labels through model prediction to enhance the generalization ability of the model.

[0054] More specifically, in step one of this embodiment, when dividing the integrated data set with class labels and the integrated data set without class labels, the label of patients who have no recurrence for more than five years after chemotherapy is recorded as 1, and the label of patients who have recurred for more than five years after chemotherapy is recorded as 0, both of which are used as data with class labels; the average recurrence time of breast cancer patients in G2 and G3 stages after chemotherapy in step one is concentrated in 2-5 years, and the label of patients who have no recurrence within two years after chemotherapy is recorded as 3 as data without class labels.

[0055] More specifically, in step 2 described in this embodiment, the Extreme Gradient Boosting (XGBoost) algorithm is used to build a machine learning model.

[0056] Furthermore, in step 3 of this embodiment, the specific operation of using the unlabeled integrated data set obtained in step 1 as a test set to evaluate the model trained in step 2 is:

[0057] Step 1: Input the unlabeled integrated data set into the model, and assign pseudo label 1 to the patient data predicted by the model as positive and with a probability greater than the mean probability of positive prediction of the current test set, i.e., recurrence;

[0058] Step 2: Incorporate pseudo-labeled patient data into the labeled dataset to train a new integrated model and adjust parameters;

[0059] Step 3: Repeat steps 1 and 2 until there is no positive label in the unlabeled data predicted by the model or the positive label prediction result is not greater than the mean prediction probability of the entire class in this round of test set. The remaining labels are assigned pseudo labels of 0, i.e., no recurrence, and the parameters are adjusted to finally obtain the semi-supervised learning integrated prediction model.

[0060] Furthermore, the specific operation of obtaining the importance ranking of features to the prediction results in step 4 described in this embodiment is: according to the SHAP mean absolute value of the model, the SHAP value of each feature is calculated, and the importance of the average absolute SHAP value is ranked.

[0061] Furthermore, the features described in this embodiment include partial tumor volume (SER), T1 weighted image washout tissue volume (SER), tumor-free area (Euler median), progesterone receptor, tumor lesion signal intensity, tumor stage, chemotherapy method, lymph node stage, menopause, lymph node lesions, and estrogen receptor; the top 5 features in terms of feature importance are used as predictive features for predicting the risk of breast cancer recurrence; according to the importance ranking of the average absolute SHAP value, the top 5 features in terms of importance are: partial tumor volume (SER), T1 weighted image washout tissue volume (SER), tumor-free area (Euler median), progesterone receptor, and tumor lesion signal intensity.

[0062] Model Validation

[0063] like Figure 2 As shown, the model construction and evaluation verification of the present invention include the following operations:

[0064] Step 1: Collect MRI imaging features and clinical data of 173 G2 and G3 breast cancer patients, perform data feature engineering on the collected data to obtain 11 most critical and important related factors, construct the processed data into an integrated data set of clinical features combined with MRI imaging features, and label the data to obtain 127 integrated data sets with class labels and 46 integrated data sets without class labels;

[0065] Step 2: Build a breast cancer recurrence prediction model suitable for clinical G2 and G3 grades, use the integrated data set of 127 cases with class labels obtained in step 1 as the training set to train the model and obtain the trained model parameters;

[0066] Step 3: Based on the principle of self-training semi-supervised learning, the 46 unlabeled integrated data sets obtained in step 1 are used as the test set to evaluate the model trained in step 2 and adjust the parameters according to the evaluation results to obtain a semi-supervised learning integrated prediction model;

[0067] Step 4: Based on the semi-supervised learning ensemble prediction model obtained in step 3, calculate the SHAP value of each feature in the data set, and get the importance ranking of the 11 important features in the semi-supervised ensemble prediction model to the prediction results (such as Figure 3 As shown in Figure 2, the contribution of the features to the model prediction results was analyzed, and 5 predictive features that affect the recurrence of breast cancer with clinical G2 and G3 grades were obtained;

[0068] Specifically, in step 1 described in this embodiment, when collecting MRI imaging features of G2 and G3 breast cancer patients, Pyradiomics is used to extract imaging features of the patient's MRI images, and the Boruta algorithm is used to calculate the importance score of the imaging features of each MRI image to determine 4 key MRI image imaging features as predictors, including partial tumor volume (SER), T1-weighted image washed tissue volume (SER), tumor-free area (Euler median), and tumor lesion signal intensity.

[0069] More specifically, in step 1 described in this embodiment, when collecting the clinical characteristics of G2 and G3 breast cancer patients, the chi-square test was used to evaluate the independence between different categories for discrete data, and the Mann-Whitney U test was used for continuous data, and 7 clinical characteristics significantly associated with the risk of breast cancer recurrence were identified, including progesterone receptor, tumor stage, chemotherapy method, lymph node stage, menopause, lymph node lesions, and estrogen receptor.

[0070] More specifically, in step one described in this embodiment, when dividing the integrated data set with class labels and the integrated data set without class labels, the label of patients who have no recurrence for more than five years after chemotherapy is recorded as 1, and the label of patients who have recurred for more than five years after chemotherapy is recorded as 0, both of which are regarded as classified labeled data. The average recurrence time after chemotherapy is concentrated in 2-5 years, and the label of patients who have no recurrence within two years after chemotherapy is recorded as 3 as unclassified labeled data.

[0071] More specifically, in step 2 described in this embodiment, the extreme gradient boosting (XGBoost) algorithm is used to construct a machine learning model, and the AUC of the integrated model obtained by verification of the test set is 0.8284, the accuracy is 80.77%, the sensitivity is 76.92%, and the specificity is 84.62%.

[0072] To be more specific, in step three described in this embodiment, an optimized semi-supervised learning ensemble model is used. After verification using the test set, the AUC of the model is 0.9211, the Accuracy is 88.57%, the Sensitivity is 81.25%, and the Specificity is 94.74%. Compared with before optimization, the evaluation indicators are improved by 11.19%, 5.94%, 5.63%, and 23.17%, respectively. The semi-supervised learning optimization effect of the self-training principle is obvious.

[0073] More specifically, the top five features ranked by feature importance described in this embodiment are used as predictive features for predicting the risk of breast cancer recurrence, namely, partial tumor volume (SER), T1-weighted image washout tissue volume (SER), tumor-free area (Euler median), progesterone receptor, and tumor lesion signal intensity.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the risk of breast cancer recurrence based on G2 and G3 graded patients, characterized in that: The steps include: Step 1: Collect MRI imaging features and clinical data of breast cancer patients with clinical G2 and G3 grades, perform data feature engineering on the collected data, construct the processed data into an integrated data set of clinical features combined with MRI imaging features, and label the data to obtain an integrated data set with class labels and an integrated data set without class labels; Step 2: Build a breast cancer recurrence prediction model suitable for clinical G2 and G3 grades, use the integrated data set with class labels obtained in step 1 as a training set to train the model and obtain the trained model parameters; Step 3: Based on the principle of self-training semi-supervised learning, the unlabeled integrated data set obtained in step 1 is used as a test set to evaluate the model trained in step 2 and adjust the parameters according to the evaluation results to obtain a semi-supervised learning integrated prediction model; Step 4: Based on the semi-supervised learning ensemble prediction model obtained in step 3, calculate the SHAP value of each feature in the data set, obtain the importance ranking of the features in the semi-supervised ensemble prediction model to the prediction results, analyze the contribution of the features to the model prediction results, and establish a semi-supervised ensemble prediction model based on SHAP; Step 5: Collect MRI imaging features and clinical data of real clinical G2 and G3 grade breast cancer patients, input them into the SHAP-based semi-supervised ensemble prediction model, and use the model parameters obtained in step 3 to analyze and predict the recurrence risk of breast cancer patients.

2. A method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 1, characterized in that: In the step 1, when collecting the MRI imaging features of G2 and G3 breast cancer patients, Pyradiomics is used to extract imaging features from the patients' MRI images. The MRI imaging feature data covers the tumor conditions of G2 and G3 breast cancer patients with different degrees of disease.

3. A method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 2, characterized in that: The MRI imaging feature data include: partial tumor volume, T1-weighted image washed tissue volume, tumor-free area, and tumor lesion signal intensity.

4. A method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 1, characterized in that: In the step 1, when dividing the integrated data set with class labels and the integrated data set without class labels, the label of patients who have no recurrence for more than five years after chemotherapy is recorded as 1, and the label of patients who have recurred for more than five years after chemotherapy is recorded as 0, both of which are regarded as data with class labels.

5. A method for predicting the risk of breast cancer recurrence based on G2 and G3 grading in patients according to claim 1, characterized in that: The integrated data set with class labels in step 1 is: a. Data containing clear label information, wherein the label information indicates the category or result to which it belongs; b. An integrated dataset containing clinical feature data and MRI radiomics feature data; c. Data used to train and test machine learning models to ensure that the models can learn the predictive patterns of the data; The integrated dataset without class labels in step 1 is: a. Data that does not contain clear label information, and the category or result is unknown; b. An integrated dataset containing clinical feature data and MRI radiomics feature data; c. Used in the optimization step of semi-supervised learning to give pseudo labels through model prediction to enhance the generalization ability of the model.

6. A method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 1, characterized in that: In the step 1, when dividing the integrated data set with class labels and the integrated data set without class labels, the label of patients who have no recurrence for more than five years after chemotherapy is recorded as 1, and the label of patients who have recurred for more than five years after chemotherapy is recorded as 0, both of which are used as data with class labels; the average recurrence time of breast cancer patients in G2 and G3 stages after chemotherapy in the step 1 is concentrated in 2-5 years, and the label of patients who have no recurrence within two years after chemotherapy is recorded as 3 as data without class labels.

7. A method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 1, characterized in that: In the step 2, the extreme gradient boosting algorithm is used to build a machine learning model.

8. The method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 1, characterized in that: In the step 3, the specific operation of using the unlabeled integrated data set obtained in step 1 as a test set to evaluate the model trained in step 2 is: Step 1: Input the unlabeled integrated data set into the model, and assign pseudo label 1 to the patient data predicted by the model as positive and with a probability greater than the mean probability of positive prediction of the current test set, i.e., recurrence; Step 2: Incorporate pseudo-labeled patient data into the labeled dataset to train a new integrated model and adjust parameters; Step 3: Repeat steps 1 and 2 until there is no positive label in the unlabeled data predicted by the model or the positive label prediction result is not greater than the mean prediction probability of the entire class in this round of test set. The remaining labels are assigned pseudo labels of 0, i.e., no recurrence, and the parameters are adjusted to finally obtain the semi-supervised learning integrated prediction model.

9. The method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 1, characterized in that: The specific operation of obtaining the importance ranking of features to the prediction results in the step 4 is: according to the SHAP average absolute value of the model, the SHAP value of each feature is calculated, and the importance ranking is based on the average absolute SHAP value.

10. A method for predicting breast cancer recurrence risk based on G2 and G3 graded patients according to claim 9, characterized in that: The features include partial tumor volume, T1-weighted image washout tissue volume, tumor-free area, progesterone receptor, tumor lesion signal intensity, tumor stage, chemotherapy method, lymph node stage, menopause, lymph node lesions, and estrogen receptor; the top 5 features in terms of feature importance are used as predictive features for predicting the risk of breast cancer recurrence; according to the importance ranking of the average absolute SHAP value, the top 5 features in terms of importance are: partial tumor volume, T1-weighted image washout tissue volume, tumor-free area, progesterone receptor, and tumor lesion signal intensity.

Images