Construction method of breast cancer axillary lymph node metastasis prediction semi-supervised model

Abstract

The invention provides a construction method of a breast cancer axillary lymph node metastasis prediction semi-supervised model, and belongs to the field of medical modeling. Comprising the steps that S1, a certain chest enhancement CT sample size is fed, and a part of a labeled queue and a part of a non-labeled queue are included; s2, creating a first machine learning model according to the samples with the label queues; s3, using a first machine learning model to predict label-free lymph nodes in the label-free queue, and pasting pseudo labels on the label-free lymph nodes, including low-risk lymph nodes and high-risk lymph nodes; s4, creating a second machine learning model by using the queues with the labels and the pseudo labels as samples; and S5, combining every two models, and creating a third machine learning model for clinically predicting whether the breast cancer has armpit lymph node metastasis or not. The model can noninvasively predict the armpit lymph node tiredness condition of a breast cancer patient, and is beneficial to screening high-risk patients, determining a proper armpit treatment scheme and avoiding unnecessary armpit operations and complications.

Description

Technical Field

[0001] The present invention belongs to the field of breast cancer metastasis prediction models in medical modeling, and specifically relates to a method for constructing a semi-supervised model for predicting axillary lymph node metastasis of breast cancer. Background Art

[0002] Axillary lymph node metastasis of breast cancer is an independent prognostic factor for breast cancer and also determines the patient's treatment method. Clinical axillary lymph node evaluation is an important part of breast cancer diagnosis and treatment.

[0003] Accurate assessment of axillary lymph nodes (ALNs) is essential for determining the optimal staging and treatment of breast cancer patients. Currently, clinical practice still relies on invasive biopsy and axillary surgery to evaluate the axillary lymph node staging of breast cancer patients. As one of the effective axillary assessment methods, sentinel lymph node biopsy (SLNB) is often an important basis for determining whether neoadjuvant therapy (NAT) is needed. Although sentinel lymph node biopsy (SLNB) can reduce the complications associated with axillary lymph node dissection (ALND), this surgery still has a certain false negative rate. Currently, the accuracy of ALNs imaging methods commonly used in clinical practice is not sufficient to achieve accurate clinical staging of ALNs.

[0004] The emergence of radiomics has brought convenience to medical diagnosis and treatment. Currently, there are more and more radiomics studies on breast cancer. However, before the inventor studied axillary lymph node metastasis, all studies were based on breast lesions to predict axillary lymph node metastasis, and there was no prediction of axillary lymph node metastasis based on the lymph nodes themselves. The inventor previously applied for patent CN202310560075.4 to provide a prediction model for axillary lymph node metastasis in patients with breast cancer, including three-dimensional reconstruction of two-dimensional lung enhanced CT films, establishing an axillary lymph node atlas, selecting all axillary lymph nodes as ROI areas, and selecting more than 5 radiomics features from all radiomics features of each axillary lymph node to distinguish whether the breast cancer has axillary lymph node metastasis; and using a random forest machine learning prediction model to construct the prediction model. The model established by this invention can non-invasively predict the axillary lymph node involvement of breast cancer patients. The patient's clinical pathological characteristics are not included in the model, and the image cutting in the model is not based on breast tumors but on axillary lymph nodes. The model described in this invention is used to determine the appropriate axillary treatment plan, thereby avoiding unnecessary axillary surgery and complications, and helping to carry out more precise surgical and adjuvant treatment modes for breast cancer.

[0005] However, this machine learning model is a fully supervised prediction model, and the breadth of its clinical application is still limited; therefore, the art needs a semi-supervised model for predicting axillary lymph node metastasis of breast cancer and a method for constructing it. Summary of the invention

[0006] Therefore, the present invention provides a method for constructing a semi-supervised model for predicting axillary lymph node metastasis of breast cancer, the method comprising the following steps:

[0007] S1: A certain amount of chest enhanced CT samples are fed, including some labeled cohorts and some unlabeled cohorts. The labeled cohort is known to have axillary lymph node metastasis, and the lymph nodes of these patients are marked as metastatic lymph nodes and non-metastatic lymph nodes. The unlabeled cohort is unknown to have axillary lymph node metastasis, and the lymph nodes of these patients are marked as unlabeled lymph nodes.

[0008] S2: extracting and screening a number of CT imaging features from the samples in the labeled cohort, and training and creating a first machine learning model;

[0009] S3: predicting the unlabeled lymph nodes using the first machine learning model, and labeling the unlabeled lymph nodes with pseudo labels, wherein the pseudo labels include low-risk lymph nodes and high-risk lymph nodes;

[0010] S4: using the metastatic lymph nodes, non-metastatic lymph nodes, low-risk lymph nodes and high-risk lymph nodes as samples, extracting and screening multiple CT imaging features therein, and training and creating a second machine learning model;

[0011] S5: Merge the non-metastatic lymph nodes and low-risk lymph nodes in the second machine learning model to obtain lymph nodes with no metastasis risk, and merge the metastatic lymph nodes and high-risk lymph nodes therein to obtain lymph nodes with metastasis risk, to create a third machine learning model; that is, the third machine learning model only includes two categories of lymph nodes with no metastasis risk and lymph nodes with metastasis risk, and the third machine learning model is used to clinically predict whether breast cancer has axillary lymph node metastasis.

[0012] In a specific embodiment, the plurality of CT imaging features include 10 to 20 CT imaging features, and preferably include 14 CT imaging features.

[0013] Specifically, the 14 CT radiomics features include elongation, maximum 3D diameter, surface area to volume ratio, sphericity, 10% of eigenvalues, minimum eigenvalue, root mean square error, maximum correlation coefficient, correlation, inverse difference normalization, joint distribution of large correlations with higher gray values, regional entropy, low gray region emphasis, and small area low gray region emphasis.

[0014] In a specific embodiment, in step S3, the first machine learning model marks the lymph nodes whose risk index of axillary lymph node metastasis of breast cancer is less than or equal to 50% and greater than 0 as low-risk lymph nodes, and marks the lymph nodes whose risk index of axillary lymph node metastasis of breast cancer is greater than 50% and less than 100% as high-risk lymph nodes.

[0015] In a specific embodiment, in step S4, the multiple CT imaging features include 60 to 90 CT imaging features, preferably include 70 to 82 CT imaging features, and more preferably include 77 CT imaging features.

[0016] In a specific embodiment, in step S5, lymph nodes with a risk index of axillary lymph node metastasis of breast cancer less than or equal to 50% are classified as lymph nodes with no metastasis risk, and lymph nodes with a risk index of axillary lymph node metastasis of breast cancer greater than 50% are classified as lymph nodes with metastasis risk.

[0017] In a specific embodiment, the first machine learning model is any one of the following 11 models: logistic regression LR, naive Bayes, support vector machine SVM, k-nearest neighbor, random forest RF, extreme random tree ExtraTrees, extreme gradient boosting XGBoost, lightweight gradient boosting machine LightGBM, gradient boosting, adaptive boosting, and multilayer perceptron MLP model; the second machine learning model and the third machine learning model are any one of the following 7 models: logistic regression LR, support vector machine SVM, random forest RF, extreme random tree ExtraTrees, extreme gradient boosting XGBoost, lightweight gradient boosting machine LightGBM, and multilayer perceptron MLP model; preferably, the first machine learning model, the second machine learning model, and the third machine learning model are all MLP models, i.e., multilayer perceptron models.

[0018] The present invention has at least the following beneficial effects:

[0019] The machine learning models established in the prior art are all fully supervised models, that is, it is known whether all the samples used for modeling have been transferred or not; while the second machine learning model and the third machine learning model established in the present invention are both semi-supervised models, that is, some are samples of which it is known whether they have been transferred, and some are samples of which it is unknown whether they have been transferred. The samples of which it is known whether they have been transferred are classified as "labeled queues", while the samples of which it is unknown whether they have been transferred are classified as "unlabeled queues". The semi-supervised prediction model provided by the present invention can be trained using a combination of labeled data and unlabeled data; this method not only helps to discover potential patterns or structures by using unlabeled data, but also effectively utilizes existing labeled data, thereby improving the overall performance of the model.

[0020] The present invention finally divides the lymph nodes into two categories in the third machine learning model, and the sensitivity of the model application is improved compared with the first machine learning model. The second machine learning model in the present invention belongs to a multi-classifier, while the third machine learning model belongs to a binary classifier. When the third machine learning model of the present invention is used to predict axillary lymph node metastasis in breast cancer patients, it is more suitable for clinical application than the second machine learning model.

[0021] In general, the present invention establishes a semi-supervised model for predicting axillary lymph node metastasis of breast cancer based on enhanced chest CT. This model can be used to non-invasively predict the axillary lymph node involvement of breast cancer patients, which is helpful for screening high-risk patients and determining appropriate axillary treatment plans, thereby avoiding unnecessary axillary surgery and complications, and contributing to the development of more accurate surgical and adjuvant treatment modes for breast cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The figure is a flow chart of the overall construction method of the present invention.

[0023] Figure 2 Schematic diagram of delineating axillary lymph nodes on CT images.

[0024] FIG3 is an evaluation of the performance of the first machine learning model.

[0025] Figure 4 Evaluate the performance of the second machine learning model MLP.

[0026] Figure 5 The distribution of samples with different labels in the two-dimensional space predicted by the classifier, that is, the prediction effect of the second machine learning model.

[0027] Figure 6 The distribution of samples with different labels in the two-dimensional space predicted by the classifier is the prediction effect of the third machine learning model. DETAILED DESCRIPTION

[0028] like Figure 1As shown, the present invention included a total of 494 female breast cancer patients, divided into three cohorts: the total cohort (i.e., including the labeled cohort and the unlabeled cohort), the labeled cohort and the unlabeled cohort. The median age of the patients in the total cohort was 52 years old, of which 94.7% of the patients had invasive ductal carcinoma (IDC), and 68.6% of the patients had a tumor size grade of T2. The patients were molecularly typed according to hormone receptor status, HER2 status, and Ki-67 status. The results showed that 65 patients (13.2%) were luminal-A type, 218 patients (44.1%) were luminal-B type, 123 patients (24.9%) were HER2-positive type, and 88 patients (17.8%) were triple-negative type. All patients underwent axillary lymph node dissection (ALND) surgery, and ALNs pathological staging was performed based on the pathological examination results of the axillary lymph nodes (ALNs). According to whether there were ALNs labels, the 494 patients were divided into a labeled cohort of 214 patients and an unlabeled cohort of 280 patients, as shown in Table 1. Table 1 shows the baseline characteristics of the patients in the total cohort, labeled cohort, and unlabeled cohort.

[0029] Figure 2 This is a schematic diagram of outlining axillary lymph nodes on a CT image; the leftmost one is a chest enhanced CT image, the right one is a schematic diagram of the axillary lymph nodes on a chest enhanced CT film, the right one is a multi-slice CT image, and the rightmost one is a three-dimensional lymph node status.

[0030] In Table 1, grade III in the histological grading is the most serious, T3 in the pathological tumor size staging is the most serious, N3 in the pathological lymph node staging is the most serious, ER status represents the expression status of estrogen receptor, PR status represents the expression status of progesterone receptor, HER-2 status represents the expression status of HER-2 receptor, Ki-67 ≥ 14 is a high-risk factor, and molecular typing of triple-negative type is a high-risk factor.

[0031] Using postoperative pathological examination as the gold standard, the labeled cohort was divided into a metastatic group (n=53) and a non-metastatic group (n=161). The metastatic group contained a total of 597 ALNs, while the non-metastatic group contained 1172 ALNs. The two groups did not show significant differences in age, pathological type, histological grade, ER status, PR status, HER-2 status, Ki-67, and molecular typing. However, there was a statistical difference in T stage between the two groups (P<0.001), as shown in Table 2. Table 2 shows the baseline characteristics of patients with axillary lymph node metastasis and non-metastasis of breast cancer in the labeled cohort.

[0032] Table 1

[0033]

[0034] Table 2

[0035]

[0036] exist Figure 1 In the 280 unlabeled cohorts shown in Table 1, all patients were ALNs patients, of which N1 patients accounted for the highest proportion (61.8%). A total of 2422 ALNs were delineated on the CT images of these patients. Since these lymph nodes cannot be matched one-to-one with the pathological examination results, accurate labels cannot be obtained.

[0037] In order to further solve the problem of labeling 2422 unlabeled lymph nodes, we first established a supervised model based on 1769 labeled ALNs from 214 patients, namely the first machine learning model. After feature screening, 14 features were selected from the 107 radiomic features of 1769 labeled ALNs and included in the model, as shown in Table 3. Table 3 shows the categories and specific names of the 14 radiomic features included in the first machine learning model.

[0038] Table 3

[0039]

[0040] We applied 11 supervised machine learning (ML) classification algorithms to build models for the 14 radiomics features, as shown in Table 4. Table 4 compares the AUC values ​​and prediction effect indicators of the 11 supervised models, with 95% confidence intervals in brackets.

[0041] In Table 4, the higher the AUC, sensitivity, specificity, accuracy and recall, the better. From the results in Table 4, these 11 models are all good, because any model with an AUC greater than 0.8 is considered a good model; but overall, the MLP model is the best, and the logistic regression model is the second best.

[0042] As can be seen from Table 4, in the cross-validation set, the prediction model built using the multilayer perceptron algorithm performed best, with a sensitivity of 0.908 and a specificity of 0.923. The ROC curve shows that the AUC of the validation set is as high as 0.958, indicating that the first machine learning model has excellent classification performance ( Figure 3A-B ).

[0043] Table 4

[0044]

[0045] Figure 3 is an evaluation of the performance of the first machine learning model. Among them, Figure AB: ROC curve and AUC percentage of the supervised model; CD: confusion matrix of the MLP model training set and test set; EF: decision curve analysis of the MLP model training set and test set; GH: radiomics nomination map score of each lymph node. The dark bar represents the score of non-metastatic ALN, and the light bar represents the score of metastatic ALN.

[0046] Further evaluation of the multilayer perceptron model revealed a large number of true positives and true negatives in the confusion matrix (Figure 3C-D). Decision curve analysis (DCA) showed that the net benefit curve of the model was superior to the full treatment curve and the no treatment curve, further validating the usefulness of the model in clinical decision making (Figure 3E-F). In addition, the radiomics nomination map score for each ALN showed that the model had strong discriminatory power (Figure 3G-H). Given the excellent diagnostic performance of the supervised model, we subsequently used the model to predict 2422 unlabeled ALNs, and a predicted probability of metastasis was generated for each ALN.

[0047] After obtaining the predicted probability of unlabeled lymph nodes, we assigned labels to the 4191 lymph nodes enrolled in the study in turn. In the labeled cohort, 1172 ALNs from 161 patients were determined to be free of metastasis by pathological examination results, and their predicted metastasis probability was 0, so they were defined as the "non-metastatic lymph nodes" label. For the 597 ALNs from 53 patients, pathological examination results showed that all of them had metastases, so they were defined as the "metastatic lymph nodes" label. In the unlabeled cohort, after prediction by the MLP model, 1479 ALNs with a metastasis probability of 0-50% were considered to be possible non-metastatic lymph nodes, defined as the "low-risk lymph nodes" label; on the contrary, 943 ALNs with a metastasis probability of 50-100% were identified as possible metastatic lymph nodes, defined as the "high-risk lymph nodes" label, see for details. Figure 1 .

[0048] By outlining CT images of 4191 ALNs, we extracted a total of 1561 radiomic features. After feature screening, 77 radiomic features were finally selected to construct a semi-supervised multi-classifier, namely Figure 1 The second machine learning model shown. Seven machine learning algorithms were applied to establish this multi-classifier. Table 4 shows the diagnostic capabilities of these seven different models in the training set. Among the seven models, LR, SVM, Random Forest, Extra Trees, XGBoost, LightGBM and MLP, the AUC (macro / micro) values ​​of the training cohort were 0.780 / 0.793, 0.869 / 0.880, 0.785 / 0.816, 0.724 / 0.743, 0.842 / 0.858, 0.835 / 0.826 and 0.950 / 0.953, respectively. Figure 4 Evaluate the performance of the second machine learning model MLP. Figure 4 AB: ROC curve and AUC percentage of MLP semi-supervised multi-classifier; Figure 4 CD: Confusion matrix of the training and test sets of the MLP semi-supervised multi-classifier. Figure 4 A and Figure 4 B has good effects, but Figure 4 The effect of A training set is better than Figure 4 The effect of test set B is better. Figure 4 C and Figure 4 In D, the horizontal axis is the predicted data of the second machine learning model, and the vertical axis is the real data. Figure 4 This shows that the prediction effect of the second machine learning model is very good, for example, see Figure 4 The data of 520 and 941 on the diagonal in C. And from Figure 4 C and Figure 4 D It can be seen that if the non-metastatic lymph nodes and low-risk lymph nodes are combined to form lymph nodes without metastasis risk, and the metastatic lymph nodes and high-risk lymph nodes are combined to form lymph nodes with metastasis risk, the prediction effect of the obtained third machine learning model will be better. In the present invention, the semi-supervised multi-classifier is the second machine learning model, and the semi-supervised binary classifier is the third machine learning model. Both the second machine learning model and the third machine learning model are new models in the field.

[0049] Comparison of the diagnostic performance of these seven multi-classifiers showed that the MLP model performed better than the other models, with AUCs of 0.924, 0.949, 0.964, and 0.963 for non-metastatic lymph nodes, low-risk lymph nodes, high-risk lymph nodes, and metastatic lymph nodes, respectively, and sensitivities of 0.818, 0.903, 0.943, and 0.894, respectively, and specificities of 0.860, 0.860, 0.864, and 0.902, respectively (Table 5). The confusion matrices of the seven models applied in the training and test cohorts showed that the MLP model achieved the highest prediction accuracy ( Figure 4 CD). Table 5 is a comparison of the AUC values ​​and prediction effect indicators of the semi-supervised multi-classifier, i.e., the second machine learning model.

[0050] Next, we reduced the radiomic features to two dimensions based on the feature and label information to observe the distribution of samples with different labels in the two-dimensional space. It is worth noting that the "non-metastatic lymph node" label predicted by the semi-supervised multi-classifier is very close to the "low-risk lymph node" label in the two-dimensional space, while the "high-risk lymph node" label is also relatively close to the "metastatic lymph node" label on the two-dimensional plane, as shown in Figure 2. Figure 5 As shown. Therefore, we further optimized the semi-supervised multi-classifier and created a semi-supervised binary classifier, i.e., the third machine learning model, to better support subsequent clinical applications. The semi-supervised binary classifier, i.e., the third machine learning model, classifies the "non-metastatic lymph nodes" and "low-risk lymph nodes" predicted by the multi-classifier into one category, collectively referred to as "lymph nodes without metastasis risk", while "high-risk lymph nodes" and "lymph nodes with metastasis" are classified into another category, collectively referred to as "lymph nodes with metastasis risk". Figure 5-6The distribution of samples with different labels in the two-dimensional space predicted by the classifier. Figure 5 is the prediction effect of the second machine learning model, Figure 6 This is the prediction effect of the third machine learning model.

[0051] Table 5

[0052]

[0053] Table 6

[0054]

[0055] To test the diagnostic performance of the semi-supervised binary classifier, the third machine learning model, we retrospectively collected data from 212 patients who underwent axillary lymph node (ALN) biopsy in the past year. These patients were from the Second Xiangya Hospital of Central South University (144 cases), the First People's Hospital of Chenzhou (31 cases), the First People's Hospital of Changde (30 cases), and the First People's Hospital of Xiangtan (7 cases). All patients were positive in clinical evaluation, and a biopsy (CNB or FNA) was performed on each suspected ALN. The results showed that 165 biopsies were positive (pN+ (b) ), 47 cases were negative (pN- (b) We validated this multicenter cohort using a semi-supervised binary classifier, the third machine learning model mentioned above, which accurately predicted 165 cases of cN+ / pN+ (b) 163 cases of ALN, with a sensitivity of 98.79%; at the same time, 41 of 47 cN+ / pN-(b) ALN were correctly identified as non-metastatic, with a specificity of 87.23% and a correct classification rate (CCR) of 96.23%. These results show the effectiveness of the third machine learning model in screening high-risk patients (see Tables 6-7).

[0056] Table 7

[0057]

[0058] After evaluating the axillary lymph node biopsy cohort with a semi-supervised binary classifier, we found that it showed excellent sensitivity (98.79%) in identifying high-risk axillary metastatic lymph nodes, which provides the possibility of timely detection for high-risk patients who are often missed in clinical practice. To further verify this finding, we collected data from 450 patients with clinically negative evaluation (cN0) and performed a patient-level assessment (see Table 8). In this retrospective cohort of 450 cN0 patients, 342 patients underwent sentinel lymph node biopsy (SLNB), and the results showed that the sentinel lymph nodes were free of metastasis (pN0 (sn)In addition, 108 patients underwent SLNB first, and their frozen pathology results showed metastasis in the sentinel lymph nodes (pN+ (sn) ), and subsequently underwent axillary lymph node dissection (ALND).

[0059] In 342 patients with cN0 / pN0 (sn) Among the patients, the semi-supervised binary classifier accurately identified 307 as non-metastatic with a specificity of 89.77%. (sn) Among the patients, the semi-supervised binary classifier detected 99 patients with metastasis, with a sensitivity of 91.67%. Therefore, the correct classification rate (CCR) of the semi-supervised binary classifier for cN0 patients was 90.22%, further confirming its potential in clinical application (see Table 9).

[0060] Table 8

[0061]

[0062]

[0063] Table 9

[0064]

[0065] The present invention has at least the following characteristics:

[0066] First, the latest axillary lymph node metastasis (ALNM) prediction methods of other inventors are all based on the primary lesion of breast cancer, and few studies predict ALNM based on lymph node lesions themselves. Considering that the image of the lymph node itself can better reflect the pathological information of ALN, the present invention conducts radiomics research on a single ALN, accurately segments and evaluates each ALN, so as to achieve a comprehensive and accurate judgment.

[0067] Secondly, the present invention has developed a semi-supervised classifier based on ALNs CT images, namely the second machine learning model and the third machine learning model, which is easy to operate, objective and economical, not interfered by subjective factors, and has good diagnostic effect. This classifier provides a new method for clinicians to judge ALNM, and is expected to play an important role in guiding ALN surgical decision-making and promoting the development of accurate diagnosis and treatment of the axilla.

[0068] Furthermore, the present invention is the first semi-supervised multi-classifier for axillary lymph nodes of breast cancer, i.e., the second machine learning model, which realizes the prediction of multiple labels and improves the prediction efficiency. In addition, the number of samples trained on the model in the present invention is huge, which improves the accuracy of the model. The semi-supervised second machine learning model and the third machine learning model have the ability of self-learning, which makes the model have the potential for continuous improvement.

[0069] In addition, the present invention has undergone multi-center external validation, specifically including the Second Xiangya Hospital of Central South University, the First People's Hospital of Xiangtan City, the First People's Hospital of Chenzhou City, and the First People's Hospital of Changde City; including patient-level evaluation and lymph node-level validation, all of which proved that the diagnostic efficacy of the model is very high and is conducive to clinical application.

[0070] In general, the present invention provides a method for constructing a semi-supervised model for predicting axillary lymph node metastasis of breast cancer, which belongs to the field of medical modeling. It includes: S1: feeding a certain amount of chest enhanced CT samples, including part of the labeled cohort and part of the unlabeled cohort; S2: creating a first machine learning model based on the samples of the labeled cohort; S3: using the first machine learning model to predict the unlabeled lymph nodes, and attaching pseudo labels to them, including low-risk lymph nodes and high-risk lymph nodes; S4: using the labeled cohort and the pseudo labels as samples to create a second machine learning model; S5: merging two by two to create a third machine learning model for clinical prediction of whether breast cancer has axillary lymph node metastasis. The model of the present invention can non-invasively predict the axillary lymph node involvement of breast cancer patients, which is helpful to screen high-risk patients, determine appropriate axillary treatment plans, and avoid unnecessary axillary surgery and complications.

[0071] The breast cancer axillary lymph node metastasis prediction model provided by the present invention is ingeniously conceived. First, pseudo labels are generated in a supervised manner. Secondly, a semi-supervised multi-classifier is established according to the prediction probability using a machine learning algorithm, that is, the second machine learning model. Finally, it is optimized into a semi-supervised binary classifier, that is, the third machine learning model. Compared with previous prediction models, the present invention has stronger interpretability and compatibility.

[0072] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a semi-supervised model for predicting axillary lymph node metastasis of breast cancer, characterized in that: The method comprises the following steps: S1: A certain amount of chest enhanced CT samples are fed, including some labeled cohorts and some unlabeled cohorts. The labeled cohort is known to have axillary lymph node metastasis, and the lymph nodes of these patients are marked as metastatic lymph nodes and non-metastatic lymph nodes. The unlabeled cohort is unknown to have axillary lymph node metastasis, and the lymph nodes of these patients are marked as unlabeled lymph nodes. S2: extracting and screening a number of CT imaging features from the samples in the labeled cohort, and training and creating a first machine learning model; S3: predicting the unlabeled lymph nodes using the first machine learning model, and labeling the unlabeled lymph nodes with pseudo labels, wherein the pseudo labels include low-risk lymph nodes and high-risk lymph nodes; S4: using the metastatic lymph nodes, non-metastatic lymph nodes, low-risk lymph nodes and high-risk lymph nodes as samples, extracting and screening multiple CT imaging features therein, and training and creating a second machine learning model; S5: Merge the non-metastatic lymph nodes and low-risk lymph nodes in the second machine learning model to obtain lymph nodes with no metastasis risk, and merge the metastatic lymph nodes and high-risk lymph nodes therein to obtain lymph nodes with metastasis risk, to create a third machine learning model; that is, the third machine learning model only includes two categories of lymph nodes with no metastasis risk and lymph nodes with metastasis risk, and the third machine learning model is used to clinically predict whether breast cancer has axillary lymph node metastasis.

2. The construction method according to claim 1, characterized in that: In step S2, the plurality of CT imaging features include 10 to 20 CT imaging features, and preferably include 14 CT imaging features.

3. The construction method according to claim 1, characterized in that: In step S3, the first machine learning model marks the lymph nodes whose risk index of axillary lymph node metastasis of breast cancer is less than or equal to 50% and greater than 0 as low-risk lymph nodes, and marks the lymph nodes whose risk index of axillary lymph node metastasis of breast cancer is greater than 50% and less than 100% as high-risk lymph nodes.

4. The construction method according to claim 1, characterized in that: In step S4, the plurality of CT imaging genomics features include 60 to 90 CT imaging genomics features, preferably include 70 to 82 CT imaging genomics features, and more preferably include 77 CT imaging genomics features.

5. The construction method according to claim 1, characterized in that: In step S5, lymph nodes with a risk index of breast cancer axillary lymph node metastasis less than or equal to 50% are classified as lymph nodes without metastasis risk, and lymph nodes with a risk index of breast cancer axillary lymph node metastasis greater than 50% are classified as lymph nodes with metastasis risk.

6. The construction method according to any one of claims 1 to 5, characterized in that: The first machine learning model is any one of the following 11 models: logistic regression LR, naive Bayes, support vector machine SVM, k-nearest neighbor, random forest RF, extreme random tree ExtraTrees, extreme gradient boosting XGBoost, lightweight gradient boosting machine LightGBM, gradient boosting, adaptive boosting, and multilayer perceptron MLP model; the second machine learning model and the third machine learning model are any one of the following 7 models: logistic regression LR, support vector machine SVM, random forest RF, extreme random tree ExtraTrees, extreme gradient boosting XGBoost, lightweight gradient boosting machine LightGBM, and multilayer perceptron MLP model; preferably, the first machine learning model, the second machine learning model, and the third machine learning model are all MLP models, i.e., multilayer perceptron models.

Images