A lymph node metastasis prediction model for breast cancer patients without incorporating clinicopathological features

Abstract

The application provides a breast cancer patient lymph node metastasis prediction model without clinical pathological characteristics, comprising the following steps: after three-dimensional reconstruction of two-dimensional lung enhanced CT films, an axillary lymph node atlas is established, all axillary lymph nodes in the atlas are selected as ROI regions, and more than 5 combined image features of each axillary lymph node are selected to distinguish whether breast cancer has axillary lymph node metastasis; and a logistic regression machine learning prediction model is used to construct the breast cancer patient axillary lymph node metastasis prediction model. The model established by the application can non-invasively predict whether breast cancer has axillary lymph node metastasis, the clinical pathological characteristics of the patient are not included in the model, and the image cutting in the model is not based on breast tumors, but based on axillary lymph nodes; the model is used to determine a suitable axillary treatment scheme, thereby avoiding unnecessary axillary surgery and complications, and helping to carry out more accurate surgery and adjuvant therapy mode of breast cancer.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of breast cancer prediction models, in particular to a lymph node metastasis prediction model for breast cancer patients without clinical pathological features and a construction method thereof. BACKGROUND

[0002] Axillary lymph node metastasis of breast cancer is an independent prognostic factor of breast cancer, and determines the treatment mode of patients. Clinical axillary lymph node evaluation is an important part of breast cancer diagnosis and treatment.

[0003] Currently, the clinical evaluation of axillary lymph nodes mainly relies on clinical palpation, B-ultrasound, CT and other two-dimensional imaging methods, which have low sensitivity and specificity, low diagnostic efficiency, and the positioning of lymph nodes is difficult, with low detection rate. Although sentinel lymph node biopsy can accurately evaluate axillary lymph nodes, it is an invasive operation, which increases the risk of a series of complications such as upper limb edema, surgical site pain and wound infection. The current non-invasive evaluation methods used in clinical practice are not satisfactory.

[0004] The emergence of imageomics brings convenience to medical diagnosis and treatment. Currently, there are more and more studies on imageomics of breast cancer. However, the current research is based on breast lesions to predict axillary lymph node metastasis. Due to the difficulty in positioning and low detection rate of imaging lymph nodes, there is currently no research based on lymph nodes themselves to predict axillary lymph node metastasis.

[0005] For example, Chinese patent application CN202010957962 provides a breast cancer patient axillary lymph node metastasis prediction model and a construction method thereof. The invention uses artificial intelligence machine learning algorithm, based on magnetic resonance image data and clinical feature data of breast cancer patients, to establish an artificial intelligence prediction model for axillary lymph node metastasis of breast cancer patients. The prediction model has the advantages of precision, simplicity, non-invasiveness, etc., and can effectively evaluate the preoperative axillary lymph node metastasis of breast cancer patients, which helps to assist clinical diagnosis and treatment decision-making of breast cancer, reduces unnecessary axillary lymph node dissection surgery of patients, reduces the occurrence of surgical complications, improves the quality of life of patients, has higher prediction efficiency and clinical benefits, and has important guiding significance for guiding clinical treatment strategies, strengthening clinical treatment intervention and subsequent individualized follow-up. However, this patent application uses magnetic resonance image (MRI) data, and the model of this patent includes clinical pathological features, and the image cutting of this patent is based on breast tumors, so this model and method have many defects; for example, the MRI detection range is limited and cannot detect all axillary lymph nodes.

[0006] Chinese patent application CN202110463515 discloses a method for predicting axillary lymph node metastasis based on breast-specific PET imaging. The method extracts MAMMIPET imaging features of breast cancer primary lesions, fuses primary lesion immunohistochemical pathology and related clinical factors, and constructs a model for predicting breast cancer lymph node metastasis risk. It realizes accurate and non-invasive prediction of breast cancer patient pre-treatment lymph node staging, supplements the low specificity of conventional imaging technology, and provides more specific reference for doctors to accurately stage and treat breast cancer patients, improving treatment success rate. However, this patent application uses PET image data, and the image cutting of this patent is also based on breast tumors. Therefore, this method also has many defects.

[0007] Chinese patent application CN202110271312 provides a method and related device for detecting the metastasis status of axillary lymph nodes of breast cancer. The method includes obtaining B-mode ultrasound images and shear wave elastography images of the axillary lymph node part of the patient; inputting the B-mode ultrasound images and shear wave elastography images into a deep learning model for feature extraction, feature fusion, and classification prediction to obtain the detection result of the metastasis status of the axillary lymph nodes of the patient's breast cancer. This application solves the technical problem of poor prediction performance of the machine learning model due to the small amount of data samples in the existing technology for detecting the metastasis status of axillary lymph nodes of breast cancer through machine learning methods. However, this patent application uses ultrasound image data, and the interpretation of ultrasound images relies on the subjective evaluation of radiologists, which leads to differences between different observers. Therefore, this detection method and related device also have obvious drawbacks.

[0008] Chinese patent application CN202110712181 provides a method for constructing a lymph node metastasis prediction model for breast cancer patients based on imaging. The method includes collecting magnetic resonance image data and clinical feature data of the patient, extracting image features based on the magnetic resonance image data, using a random forest algorithm to screen the image features to obtain a number of key image features, establishing an image feature prediction model based on the key image features using a support vector machine algorithm, screening key clinical features by performing single-factor analysis on the clinical feature data, and establishing a clinical feature prediction model based on the key clinical features using a support vector machine algorithm. The support vector machine algorithm is used to establish a comprehensive lymph node metastasis prediction model based on the key image features and key clinical features. The embodiments of this application use a random forest algorithm and a support vector machine algorithm to establish a model, which can establish a prediction model based on the structural risk minimization principle, avoid over-learning problems, and make the constructed prediction model more stable and accurate. However, this patent application uses magnetic resonance image data, and the image cutting of this patent is also based on breast tumors. Therefore, this model and method still have many defects.

[0009] Chinese patent application CN201810557069 provides a breast cancer axillary lymph node metastasis state discrimination method based on LASSO regression, which comprises the following steps: step 1, segmenting the collected breast cancer X-ray image of the patient to determine multiple lesion site images where the tumor is located; step 2, dividing the multiple lesion site images into a training set and a test set, extracting high-dimensional features from all the lesion site images, and constructing a breast lesion image feature database; step 3, using the LASSO regression algorithm to construct a breast cancer axillary lymph node metastasis identification model according to the breast lesion image feature data in the training set and the patient's pathological information; step 4, identifying the lesion site images in the test set according to the breast cancer axillary lymph node metastasis identification model to determine the breast cancer axillary lymph node metastasis state of the patient. The application can further mine data meaningful for identification in the image, and improve the automatic identification level based on X-ray images. However, the patent application uses X-ray image data, and the image cutting is based on breast tumors. The visualization of axillary lymph nodes by digital mammography is limited, and only 50% of patients can see axillary I-level lymph nodes in routine breast X-ray examination. Deep I-level or II-level lymph nodes are generally difficult to see. Therefore, the accuracy of the breast cancer axillary lymph node metastasis state discrimination method is greatly limited.

[0010] Therefore, there is still a need in the art for a new breast cancer patient lymph node metastasis prediction model without incorporating clinical pathological features and a construction method thereof. SUMMARY

[0011] Therefore, the present application provides a breast cancer patient lymph node metastasis prediction model without incorporating clinical pathological features, which is used to non-invasively predict whether breast cancer has axillary lymph node metastasis. The model comprises the following steps: after three-dimensional reconstruction of two-dimensional lung enhanced CT images, an axillary lymph node atlas is established, all axillary lymph nodes in the atlas are selected as ROI regions, and more than 5 image features of all image features of each axillary lymph node are combined to distinguish whether the breast cancer has axillary lymph node metastasis, and a logistic regression machine learning prediction model is used to construct the breast cancer patient axillary lymph node metastasis prediction model.

[0012] In the present application, each axillary lymph node has 107 image features.

[0013] The present application uses three-dimensional reconstruction technology to identify and reconstruct the tiny axillary lymph node lesions on two-dimensional CT images, and performs spatial positioning to extract the image features of the axillary lymph nodes, and establishes a machine learning model to predict the metastasis state of the axillary lymph nodes of breast cancer patients.

[0014] In a specific embodiment, more than 9 of the following 22 image features of axillary lymph nodes are used in the model; the 22 image features include: elongation, flatness, minimum axis length, maximum 2D diameter, sphericity, value of 10% of eigenvalues, quartile distance, kurtosis, mean, minimum eigenvalue, gray value range, root mean square error, skewness, cluster prominence, cluster shadow, correlation, differential entropy, correlation information measure 2, joint energy, joint distribution of large correlation with higher gray value, gray non-uniform normalization, intensity.

[0015] In a specific embodiment, all of the 22 image features are used.

[0016] In the present application, more than 9 of the 22 image features are used to distinguish whether there is axillary lymph node metastasis, which is relatively accurate, and if all of the 22 image features are used, the effect is best, specifically, the area under the curve (AUC) can reach more than 0.93.

[0017] In a specific embodiment, the three-dimensional reconstruction uses Vitaworks software.

[0018] In a specific embodiment, the features of the model are all based on image features, and do not include clinical and pathological features.

[0019] The present application also provides a method for constructing a breast cancer lymph node metastasis prediction model without including clinical and pathological features as described above, comprising the following steps:

[0020] 1) Data collection: collect and process two-dimensional lung high-resolution thin-slice enhanced CT image data and its clinical and pathological feature data of patients;

[0021] 2) Construction of breast cancer axillary lymph node three-dimensional atlas: automatically identify, segment and extract axillary lymph node information in two-dimensional CT images using algorithms to construct a three-dimensional axillary lymph node atlas;

[0022] 3) Establish an image feature prediction model: screen the key image features related to the status of breast cancer axillary lymph nodes in the CT axillary lymph node image data of breast cancer patients through single factor variance analysis, correlation analysis and Lasso regression, and establish a corresponding machine learning prediction model according to the logistic regression algorithm.

[0023] In a specific embodiment, between step 2) and step 3), the method further comprises using the Segmentation module in the 3D-Slicer software to draw the ROI of the multiple layers of the lymph nodes of the enrolled patient, and using the Radiomics image module in the 3D-Slicer software to extract features of the ROI mask in the CT original image.

[0024] In a specific embodiment, in step 3), after the prediction model is established, the method further comprises evaluating the performance of the prediction model by using the method of nested cross-validation, and the indicators representing the prediction performance include the area under the curve, specificity, sensitivity, false positive rate and false negative rate.

[0025] Compared with the prior art, the present application has at least the following beneficial effects:

[0026] 1. The present application is based on lung high-resolution thin-layer enhanced CT image data, which is accurate, covers all axillary lymph node locations, and is routine preoperative examination for patients, which is economical and convenient. Moreover, the lesions drawn by the present application are all lymph nodes, which can better reflect the morphological characteristics of metastatic lymph nodes, and the diagnostic efficiency is higher. In addition, compared with color ultrasound, lung high-resolution thin-layer enhanced CT has higher resolution and more accurate examination, which can capture smaller structural changes, thereby better judging whether the lymph nodes are invaded by tumors, and can simultaneously compare the size, shape and number of bilateral axillary lymph nodes on the image, reducing subjective judgment errors and making the examination more objective and comprehensive.

[0027] 2. The present application establishes multiple machine learning models, and selects the optimal prediction model by comparing the diagnostic efficiency, which has high diagnostic efficiency. The multiple machine learning models established by the present application are Logistic Regression (LR), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Naive Bayes (NB), and Classification Trees (CART).

[0028] 3. The prediction model of the present application only contains CT image features, and does not contain clinical and pathological features, which means that the prediction model does not need to consider clinical and pathological factors, so that the prediction model has a wider clinical application range and is easier to realize clinical application.

[0029] 4、In the previous imaging research, the positioning of lymph nodes almost completely depends on the subjective judgment of radiologists, and it is often difficult to accurately find the specific position of each axillary lymph node, so most studies are based on breast lesions to predict axillary lymph node metastasis, ignoring the lesion changes of lymph nodes themselves, which reduces the accuracy of lymph node prediction, the present application realizes the accurate positioning of axillary lymph nodes in imaging based on three-dimensional visualization technology, so that the positioning process is no longer affected by the subjective reading level of radiologists, thereby greatly reducing the artificial error and improving the objectivity and accuracy of positioning. In general, the present application predicts lymph nodes based on the lymph nodes themselves rather than through breast lesions, and the results will be more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is a schematic diagram of the development process of the prediction model, and sequentially includes CT three-dimensional reconstruction, drawing of ROI region, feature extraction and screening, and construction of model schematic diagram.

[0031] Figure 2 It is the ROC curve of the logistic regression prediction model in the present application.

[0032] Figure 3 It is the PR curve of the logistic regression prediction model in the present application. DETAILED DESCRIPTION

[0033] The lymph nodes in the present application specifically refer to axillary lymph nodes. Those skilled in the art can know that breast lesions are relatively large, generally about 2-5 cm, which are easy for doctors to identify with the naked eye. However, lymph nodes are relatively small, generally only about 0.5-1 cm, and doctors can also see some lymph nodes with the naked eye, but it is generally difficult to find all the lymph nodes on the CT film with the naked eye. Therefore, the present application clearly knows how many lymph nodes a patient has after positioning by three-dimensional reconstruction of lung CT, and each lymph node corresponds to the CT image of the patient. Some patients have several lymph nodes on the lung CT film, and some patients have dozens of lymph nodes on the lung CT film.

[0034] The present application adopts three-dimensional reconstruction technology to identify and reconstruct the micro-lesions on the two-dimensional CT image, and performs spatial positioning, extracts the imaging features of the lymph nodes, establishes a machine learning model, and predicts the metastasis status of axillary lymph nodes of breast cancer.

[0035] The present application retrospectively collects the data of breast cancer patients who have completed preoperative lung high-resolution thin-slice enhanced CT examination and subsequent surgery in the Department of Breast Surgery of Xiangya Second Hospital of Central South University.

[0036] The case entry criteria are as follows: (1) receiving breast cancer surgery in the Second Xiangya Hospital of Central South University and being confirmed as invasive breast cancer by postoperative pathology; (2) receiving ipsilateral axillary lymph node dissection in the Second Xiangya Hospital of Central South University; (3) completing lung high-resolution thin-slice enhanced CT examination in the radiology department of the Second Xiangya Hospital of Central South University within 1 month before surgery; (4) having complete clinical pathological data.

[0037] Table 1

[0038]

[0039] The case exclusion criteria are as follows: (1) being diagnosed as bilateral breast cancer; (2) receiving neoadjuvant chemotherapy before surgery; (3) having incomplete CT scan sequences, poor imaging quality, only completing plain scan, or CT examination performed in an external hospital; (4) having SLNB (sentinel lymph node biopsy) or SLNB+ALND (sentinel lymph node biopsy+axillary lymph node dissection) as the axillary surgery method; (5) having distant metastasis lesions or being combined with other malignant tumors.

[0040] The present application retrospectively collects 156 cases of unilateral invasive breast cancer patients who have completed preoperative lung high-resolution thin-slice enhanced CT examination and ALND surgery in the Department of Breast Surgery of the Second Xiangya Hospital of Central South University, wherein 125 cases are indicated by postoperative ALND pathology that axillary lymph nodes are not metastatic, 31 cases are indicated by postoperative ALND pathology that axillary lymph nodes are all metastatic, all patients have complete clinical pathological data, do not have distant metastasis, and have not received NAT (neoadjuvant therapy). The baseline characteristics of the enrolled patients are shown in Table 1, and Table 1 is the baseline characteristics of the patients grouped according to the metastatic status of breast cancer axillary lymph nodes. In Table 1, the data of serial number 1 is the average value of age in the brackets, and the standard deviation in the brackets; the data of serial numbers 2-10 are all the number of people in the brackets, and the percentage of the number of people in the brackets, so the sum of the percentages of each data of serial numbers 2-10 is 100%.

[0041] The lung high-resolution thin-slice enhanced CT of the enrolled patients is reconstructed in three dimensions by using three-dimensional visualization technology, and a three-dimensional map of axillary lymph nodes is constructed.

[0042] The position of axillary lymph nodes is located by using the map, the metastatic lymph node region of interest (ROI) is outlined on the CT image, and the imaging features are extracted from the ROI. Single factor variance analysis, correlation analysis and Lasso regression are used for imaging feature screening.

[0043] Based on the optimal features, a variety of machine learning methods are used to construct a prediction model for evaluating whether axillary lymph nodes are metastatic or not.

[0044] The performance of the prediction model is evaluated by the method of nested cross-validation, and the indicators representing the prediction performance include area under curve (AUC), specificity, sensitivity, false positive rate, and false negative rate. The AUC is defined as the area surrounded by the ROC curve and the coordinate axis, and the value of the area is obviously not greater than 1. Since the ROC curve is generally above the straight line y=x, the value range of AUC is between 0.5 and 1. The closer the AUC is to 1.0, the higher the authenticity of the detection method; when the AUC is equal to 0.5, the authenticity is the lowest and has no application value.

[0045] The development process of the prediction model is shown in FIG. 1, which sequentially includes a CT three-dimensional reconstruction diagram, a ROI region delineation diagram, a feature extraction and screening diagram, and a model construction diagram. Figure 1

[0046] The research results are as follows: a total of 156 patients met the inclusion criteria and completed the three-dimensional reconstruction, including 31 patients in the metastasis group and 125 patients in the non-metastasis group. An axillary lymph node atlas was established by using the three-dimensional reconstruction technology.

[0047] A total of 936 ROI regions of non-metastatic lymph nodes and 336 ROI regions of metastatic lymph nodes were delineated, and a total of 107 features were extracted from the ROI. After screening, a total of 22 features were included in the model. Table 2 below shows the 22 image feature categories and specific names included in the model.

[0048] Table 2

[0049]

[0050]

[0051] The single-factor and multi-factor analysis of the clinical pathological factors of the enrolled population found that there was no significant statistical difference in age, tumor size, clinical lymph node stage, and molecular subtype between the metastasis group and the non-metastasis group (P>0.05).

[0052] The 22 image features were used to establish logistic regression, linear discriminant analysis, quadratic discriminant analysis, Bayesian learning, and decision tree (classification tree) machine learning prediction models, and it was found that the logistic regression model had the best diagnostic performance in predicting axillary lymph node metastasis.

[0053] Table 3 shows the comparison of AUC values of each model, and the data in the brackets in Table 3 is the 95% confidence interval. Table 4 shows the prediction performance indicators of each model, and the data std in the brackets in Table 4 is the standard deviation.

[0054] Table 3

[0055] Machine learning method Training set Validation set Logistic regression 0.946(0.934-0.958) 0.936(0.924-0.950) Linear discriminant analysis 0.942(0.927-0.957) 0.937(0.923-0.954) Quadratic discriminant analysis 0.944(0.932-0.956) 0.924(0.912-0.942) Bayesian learning 0.925(0.910-0.940) 0.923(0.908-0.936) Decision tree 0.916(0.891-0.941) 0.860(0.834-0.878)

[0056] ​Table 4

[0057] Machine learning method Sensitivity (std) Specificity (std) False negative rate (std) False positive rate (std) Logistic regression 0.768(0.05) 0.963(0.04) 0.232(0.05) 0.037(0.04) Linear discriminant analysis 0.729(0.05) 0.973(0.05) 0.271(0.05) 0.027(0.05) Quadratic discriminant analysis 0.723(0.05) 0.957(0.03) 0.277(0.05) 0.043(0.03) Bayesian learning 0.771(0.05) 0.930(0.04) 0.229(0.05) 0.071(0.04) Decision tree 0.738(0.06) 0.943(0.05) 0.262(0.06) 0.057(0.05)

[0058] The prediction model established by the logistic regression algorithm has a high diagnostic efficiency, with an AUC of 0.946 in the training set and 0.936 in the validation set, and a specificity of 0.963. Figure 2 and Figure 3 are the ROC curve and the PR curve of the logistic regression prediction model in the application, respectively.

[0059] The application also provides a method for predicting axillary lymph node metastasis of breast cancer based on radiomics, which comprises the following specific steps:

[0060] S1, collecting the clinicopathological information and CT information of breast cancer patients; the clinicopathological factors of the patients include age, menopausal status, pathological TNM stage of the tumor, pathological report time, tumor type, pathological grade, estrogen receptor status (Estrogen Receptor, ER), progesterone receptor status (Progesterone Receptor, PR), human epidermal growth factor receptor-2 status (Human Epidermal Growth Factor Receptor 2, HER2), tumor proliferation index Ki-67, tumor molecular typing, number of axillary lymph node metastasis, and total number of axillary lymph nodes; the CT information of the patients includes hospitalization number, CT number, and CT examination time;

[0061] S2, collecting the lung high-resolution thin-slice enhanced CT images of the patients in the group, automatically identifying, segmenting and extracting the axillary lymph node information in the two-dimensional CT images by using an algorithm, and constructing an axillary lymph node three-dimensional atlas;

[0062] S3, on the axillary lymph node three-dimensional atlas of the patient, the positioning and navigation function of three-dimensional visualization technology can be used to find the target lymph node to be positioned and determine its position on the CT image;

[0063] S4, using the Segmentation module in the 3D-Slicer software, the ROI of the patient in multiple layers is outlined;

[0064] S5, using the Radiomics radiomics module in the 3D-Slicer software, the ROI mask in the CT original image is feature extracted;

[0065] S6, the patients are randomly divided into a training set and a test set in proportion, the training set is used to train the prediction model, and the test set is used to test the performance of the prediction model;

[0066] S7, the machine learning method is used to screen out the radiomics features related to the axillary lymph node metastasis of breast cancer from the radiomics features in the training set, a prediction model is trained, a radiomics prediction model related to the axillary lymph node metastasis is generated, and the model is verified in the test set, and the specific steps are as follows: using logistic regression, linear discriminant analysis, quadratic discriminant analysis, naive Bayes, and classification tree algorithm to train the prediction model respectively, generating a radiomics prediction model related to the axillary lymph node metastasis of breast cancer, calculating the sensitivity, specificity, false positive rate, false negative rate, and area under the curve of the model on the test set, drawing an AUC curve and a precision-recall curve, and measuring the performance of the model.

[0067] The present application uses three-dimensional reconstruction technology to position the axillary lymph node space, and the delineation of the ROI region (region of interest) does not require an imaging physician, thereby saving manpower. In addition, the delineated region of interest is the lesion itself, i.e., the axillary lymph node. The previous studies all delineated breast lesions, and the delineation based on the lymph node region of interest can more accurately extract meaningful radiomics features. Using a machine learning method, the inventors established a model that can efficiently predict the metastasis of the axillary lymph node, thereby providing a reference for the next step of axillary diagnosis and treatment. Notably, the model features of the present application are all based on radiomics features, without including clinical features, which means that the patient only needs to complete a CT examination to realize the prediction of the lymph node, which is economical and convenient, and suitable for promotion.

[0068] In the method for predicting the axillary lymph node metastasis of breast cancer based on radiomics according to the present application, the clinical pathological information of breast cancer patients is generally collected in step S1, but analysis shows that these clinical pathological information has no statistical significance on whether the axillary lymph node of breast cancer is metastasized. That is, when predicting the axillary lymph node metastasis of breast cancer using the model according to the present application, any patient clinical pathological features do not need to be considered, but are completely based on radiomics features. In general, the present application establishes a radiomics model based on three-dimensional reconstruction technology, which can non-invasively predict the axillary lymph node involvement of breast cancer patients. The model does not include the clinical pathological features of the patient, and the image cutting in the model is not based on the breast tumor, but on the axillary lymph node. The model according to the present application is used to determine a suitable axillary treatment plan, thereby avoiding unnecessary axillary surgery and complications, and is helpful for the development of more precise surgery and adjuvant therapy mode for breast cancer.

[0069] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, some simple deductions and substitutions can be made without departing from the concept of the present application, and all of them should be regarded as falling within the protection scope of the present application.

Claims

1. A breast cancer patient lymph node metastasis prediction device without the need to incorporate clinical pathological characteristics, the prediction device comprising a prediction model, the model is used to realize non-invasive prediction of whether breast cancer has axillary lymph node metastasis, the model comprises the following steps: after three-dimensional reconstruction of two-dimensional lung enhanced CT images, an axillary lymph node atlas is established, all axillary lymph nodes in the atlas are selected as ROI regions, and more than 9 of the following 22 image features of each axillary lymph node are combined to distinguish whether the breast cancer has axillary lymph node metastasis, and a logistic regression machine learning prediction model is used to construct the breast cancer patient axillary lymph node metastasis prediction model; the 22 image features include: elongation, flatness, minimum axis length, maximum 2D diameter, sphericity, value of the tenth percentile of the eigenvalues, quartile distance, kurtosis, mean, minimum eigenvalue, range of gray values, root mean square error, skewness, cluster significance, cluster shadow, correlation, differential entropy, correlation information measure 2, joint energy, joint distribution of large correlations with higher gray values, gray non-uniformity normalization, intensity.

2. The apparatus of claim 1, wherein, All of the 22 radiomics features are used together.

3. The apparatus of claim 1, wherein, The three-dimensional reconstruction uses Vitaworks software.

4. The apparatus of claim 1, wherein, Features of the model are all based on radiomics features, without including clinical and pathological features.

5. A method for constructing the prediction model according to any one of claims 1-4, comprising the following steps: 1) data collection: collecting and processing two-dimensional lung high-resolution thin-slice enhanced CT image data and its clinical and pathological feature data of patients; 2) construction of breast cancer axillary lymph node three-dimensional atlas: using an algorithm to automatically identify, segment and extract axillary lymph node information in two-dimensional CT images, and constructing a three-dimensional axillary lymph node atlas; 3) establishment of an imageomics prediction model: screening key radiomics features related to the status of breast cancer axillary lymph nodes in CT axillary lymph node images of breast cancer patients by single factor variance analysis, correlation analysis and Lasso regression, and establishing a corresponding machine learning prediction model according to a logistic regression algorithm.

6. The method of claim 5, wherein, Between step 2) and step 3), it further includes using the Segmentation module in 3D-Slicer software to perform ROI delineation of multiple layers of lymph nodes of the enrolled patients; and using the Radiomics imageomics module in 3D-Slicer software to extract features of the ROI mask in the CT original image.

7. The method according to claim 5 or 6, characterized in that, In step 3), after the prediction model is established, it further includes using a nested cross-validation method to evaluate the performance of the prediction model, and the indicators representing the prediction performance include area under the curve, specificity, sensitivity, false positive rate and false negative rate.

Images