CN115546087A

CN115546087A - Imaging omics model for breast cancer molecular typing, construction method, medium and device

Info

Publication number: CN115546087A
Application number: CN202110648622.5A
Authority: CN
Inventors: 邵志敏; 江一舟; 蒋麟; 尤超; 顾雅佳; 肖毅; 苏冠华
Original assignee: Fudan University Shanghai Cancer Center
Current assignee: Fudan University Shanghai Cancer Center
Priority date: 2021-06-10
Filing date: 2021-06-10
Publication date: 2022-12-30

Abstract

The invention discloses an imaging omics model for breast cancer molecular typing, which comprises a TNBC model and a HER2 model; the TNBC model includes: 11 proteomics features predicting TNBC, and formula I: q = 1/(1 +exp (beta) ₀ +β ₁ x ₁ +β ₂ x ₂ +…+β _k x _k ) ); HER2 models include: 11 proteomics features predictive of HER2+/HER 2-breast cancer, and formula III: q "= 1/(1 + exp (beta))" ₀ +β” ₁ x” ₁ +β” ₂ x” ₂ +…+β” _k x” _k )). The invention also provides a model construction method, a medium and a device. The invention is based on the current maximumThe breast cancer nuclear magnetic resonance imaging omics data set constructs and verifies the imaging omics model, and provides a basis for realizing noninvasive prediction of breast cancer molecular typing so as to guide the application of clinical diagnosis and treatment decisions.

Description

Imaging omics model for breast cancer molecular typing, construction method, medium and device

Technical Field

The invention relates to the field of cancer typing and artificial intelligence imaging omics, and application of artificial intelligence in tumor diagnosis and treatment decision, in particular to an imaging omics model for breast cancer molecular typing, a construction method, a medium and a device.

Background

Breast cancer is the most common malignancy worldwide and the most common mortality among women. According to the expression states of Hormone Receptors (HR) [ including Estrogen Receptors (ER) and Progestogen Receptors (PR) ] and the expression states and whether amplification of human epidermal growth factor receptor 2 (her2) molecules occurs, breast cancers can be classified into three subtypes, HR + HER2-, HER2+ and Triple negative breast cancers (TNBC, i.e., breast cancers in which immunohistochemistry ER, PR and HER2 are negative) in which clinical characteristics and treatment strategies are significantly different. Wherein, TNBC is easier to recur and transfer and worse in prognosis than other two subtypes, the 5-year survival rate of TNBC at the stage I is 85 percent, and the 5-year survival rates of HR + HER 2-and HER2+ breast cancers at the stage I are 94 percent and 99 percent respectively; median survival for metastatic TNBC, HR + HER 2-and HER2+ breast cancers was 10-13 months, 4-5 years and 5 years, respectively. In the aspect of systemic treatment strategies, endocrine therapy plus or minus chemotherapy is mainly used for HR + HER 2-breast cancer, HER2+ breast cancer is mainly used for HER2 targeted inhibitor plus or minus chemotherapy, TNBC lacks corresponding treatment targets, and systemic treatment is mainly used for chemotherapy. It follows that different subtypes of breast cancer have distinct therapeutic strategies and clinical prognosis. Therefore, the differentiation of breast cancer molecular typing is an important basis for breast cancer clinical diagnosis and treatment.

Currently, obtaining tumor tissue by hollow needle puncture and measuring the expression of ER, PR and HER2 and the amplification status of HER2 in tumor tissue according to Immunohistochemistry (IHC) and Fluorescence In Situ Hybridization (FISH) methods are the main methods for preoperative determination of breast cancer molecular typing. In practical clinical practice, the waiting time of IHC staining results is one week after puncture, and the waiting time of FISH detection results is two weeks after puncture. Therefore, the process has the defects of large trauma to a patient, long IHC/FISH time consumption, large requirements on manpower and material resources and the like, and simultaneously has higher requirements on the quality and quantity of tumor tissues obtained by puncture of the hollow needle. Therefore, there is a need to find an alternative method for predicting molecular typing of breast cancer non-invasively, conveniently and at low cost before surgery.

In recent years, artificial intelligence technology has been developed vigorously, and imaging omics is one of the important applications of artificial intelligence technology in the field of clinical medicine. The imaging omics technology extracts quantitative characteristic information from medical image images in a high-flux manner, and the value of assisting clinical decision is realized by carrying out prediction and prognosis layering through modeling analysis. Due to the characteristics of noninvasive, rapid, convenient and repeatable medical imaging, the imaging omics technology has application value in predicting breast cancer molecular typing.

At present, no preoperative noninvasive breast cancer molecular typing scheme exists clinically. The commonly adopted strategy is to obtain tumor tissues through hollow needle puncture, and measure the expression of ER, PR and HER2 and the amplification state of HER2 in the tumor tissues according to immunohistochemistry and fluorescence in situ hybridization methods to determine the molecular typing of the breast cancer. However, the puncture with hollow needle, together with immunohistochemical staining and fluorescence in situ hybridization have the problems of large wound, long time consumption, high requirements for tissues obtained by puncture biopsy, and the like. Therefore, there is a need for a clinically useful molecular typing method that can be used to perform molecular typing noninvasively, rapidly, and inexpensively for any clinical breast cancer patient to guide the next treatment decision.

Disclosure of Invention

In order to non-invasively and conveniently classify breast cancer molecules, the invention provides an imaging omics model for breast cancer molecule classification on the one hand; on the other hand, a method for constructing a breast cancer molecular typing imaging omics model is provided; the third aspect also provides a medium; the fourth aspect provides a device for constructing an imaging omics model for breast cancer molecular typing.

In a first aspect, the invention provides a breast cancer molecular typing imaging omics model, which comprises a TNBC model and a HER2 model; the TNBC model includes:

the proteomics features of 11 predicted TNBC are shown in table 1;

and formula I: q = 1/(1 +exp (beta) ₀ +β ₁ x ₁ +β ₂ x ₂ +…+β _k x _k ) Wherein x is _k Features of the image group, beta, representing predictive TNBC _k Denotes the corresponding coefficient, β ₀ Representing a TNBC model intercept of 2.6087610 _k And beta _k Specifically, as shown in table 1:

TABLE 1 TNBC model parameters

HER2 models include:

11 proteomics features predicting HER2+/HER 2-breast cancer, as shown in table 3;

and formula III: q "= 1/(1 + exp (beta))" ₀ +β” ₁ x” ₁ +β” ₂ x” ₂ +…+β” _k x” _k ) Wherein, x " _k Shows the shadow characteristics, beta, of a predictor of HER2+/HER 2-breast cancer " _k Denotes the corresponding coefficient, β " ₀ Representing a HER2 model intercept of 1.09702336,x " _k And beta' _k Specifically, as shown in table 3:

TABLE 3 HER2 model parameters

Wherein k is an integer of 1 to 11, and exp represents an exponential function with a natural constant e as a base.

Further, the ROI sources for the image omics features of the 11 predicted TNBC are shown in table 4:

TABLE 4 ROI sources for predicting proteomics characteristics of TNBC

The 11 ROI sources that predict the imageomic signature of HER2+/HER 2-breast cancer are shown in table 6:

TABLE 6 ROI sources for predicting the imageomic characteristics of HER2+/HER 2-breast cancer

Further, formula I calculates the result: q is less than a first threshold value, TNBC is predicted, q is more than or equal to the first threshold value, and non-TNBC is predicted; the result is calculated by formula III: q '< the third threshold, is predicted to be HER2+, q' gtoreqthe first threshold, is predicted to be HER2-; the first threshold value is 0.640; the third threshold is 0.577.

Further, the system also comprises an operation module; the operation module is configured to: the patient whose prediction result is TNBC for the TNBC model outputs a result "TNBC"; in patients with a non-TNBC prediction result of the TNBC model, a HER2 model is adopted for prediction, if the HER2 model prediction result is HER2+, the result 'HER 2+ breast cancer' is output, and if the HER2 model prediction result is HER2-, the result 'HR + HER 2-breast cancer' is output.

Further, formula IV and formula VI are also included;

formula IV is p =1-q; the result is calculated by formula IV: p is greater than 1-first threshold, TNBC is predicted, p is less than or equal to 1-first threshold, and non-TNBC is predicted;

formula VI is p "=1-q"; the result is calculated by formula VI: p "> 1-the third threshold, is predicted to be HER2+, p" ≦ 1-the first threshold, is predicted to be HER2-.

Further, an HR model is also included; the HR model includes: 11 predicted imageomics characteristics of HR +/HR-breast cancer, as shown in table 2;

and formula II: q '= 1/(1 + exp (beta)' ₀ +β’ ₁ x’ ₁ +β’ ₂ x’ ₂ +…+β’ _k x’ _k ) X 'therein' _k Denotes the predicted imageomics signature, β ', of HR +/HR-breast cancer' _k Denotes a corresponding coefficient, β' ₀ Represents an HR model intercept of-0.2212531,x' _k And beta' _k Specifically, as shown in table 2:

TABLE 2 HR model parameters

The 11 ROI sources that predict the imageomic signature of HR +/HR-breast cancer are shown in table 5:

TABLE 5 ROI sources of proteomic features predicting HR +/HR-breast cancer

Further, formula II calculates the result: q 'is less than a second threshold value, the HR + is predicted, q' is more than or equal to the first threshold value, and the HR-is predicted; the second threshold value is 0.627;

also included is formula V; formula V is p '=1-q'; the result is calculated by formula IV: p '> 1-the second threshold, predicted to be HR +, p' ≦ 1-the first threshold, predicted to be HR-.

In a second aspect, the invention further provides a method for constructing a breast cancer molecular typing imaging omics model, which comprises the following steps:

acquiring a breast contrast enhancement nuclear magnetic resonance image before starting new auxiliary treatment or before a patient without new auxiliary treatment; the acquisition sequence of the contrast enhanced nuclear magnetic resonance image comprises a dynamic contrast enhanced sequence;

step two, sketching a target area of a breast tumor body;

respectively expanding and contracting the tumor target area by 5 +/-2 millimeters to form an outer boundary and an inner boundary, wherein the area between the outer boundary and the tumor boundary is a tumor peripheral target area, the area between the inner boundary and the tumor boundary is a tumor inner target area, and the tumor Zhou Baou is combined with the tumor target area to form a whole tumor target area so as to obtain four target areas of the tumor body, the tumor periphery, the whole tumor and the tumor interior;

calculating two main image omics characteristics of spatial domain characteristics and time domain characteristics by using Python software based on the four target areas; extracting the spatial domain features in Python based on a PyRadiomics toolkit; the time domain features include: the space domain characteristic change rate of the last phase enhanced sequence compared with the flat scan image adopts a formula 1-1; the algorithm of the change rate of the spatial domain characteristics of the two adjacent phases of images is a formula 1-2; the average value of each spatial domain feature among the enhancement sequences in each period has an algorithm of a formula 1-3; the variance of each spatial domain feature among the enhancement sequences in each period is calculated by a formula 1-4; the skewness of each spatial domain feature among the enhancement sequences at each stage is calculated by a formula 1-5; the kurtosis value of each spatial domain feature among the enhancement sequences at each stage is calculated by a formula 1-6; wherein F represents the characteristic numerical value of the image group, N represents the enhancement period number, and F _N Or x _N The characteristic value of the image set representing the Nth enhanced image,

is the standard deviation of the distribution;

fifthly, taking the spatial domain characteristics and the time domain characteristics of the four types of target areas as input variables; taking the breast cancer molecular typing determined by the protein expression conditions of ER and PR and the expression and amplification state of HER2 obtained by immunohistochemical staining of a patient based on a paraffin pathological section as a gold standard;

when the image omics characteristics of the TNBC model are screened, classifying the patients in the training set into a TNBC group and a non-TNBC group through a gold standard, wherein corresponding label in the TNBC model is TNBC and non-TNBC respectively; analyzing and comparing image omics characteristics corresponding to TNBC and non-TNBC of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through glmnet and cv-glmnet functions of an R software glmnet package, and screening out image omics characteristics of predicted TNBC; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected TNBC model under the lambda value;

when the image omics characteristics of the HR model are screened, patients in the training set are classified into an HR + group and an HR-group through a gold standard, and the corresponding label in the HR model is HR + and HR-; analyzing and comparing the corresponding proteomics characteristics of HR + and HR-of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through a glmnet function and a cv-glmnet function of an R software glmnet, and screening the proteomics characteristics for predicting HR +/-; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within one standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HR model under the lambda value;

when the imaging group characteristics of the HER2 model are screened, the patients in the training set are classified into a HER2+ group and a HER 2-group through a gold standard, and the corresponding labels in the HER2 model are HER2+ and HER2-; analyzing and comparing HER2+ of label and corresponding image omics characteristics of HER 2-by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through glmnet and cv-glmnet functions of an R software glmnet package, and screening out image omics characteristics for predicting HER2 +; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HER2 model under the lambda value;

and step six, the image omics characteristics obtained by screening in the step five are used as model input information, a logistic regression algorithm is used for constructing an image omics model for identifying the TNBC in a training set, and the area under the curve of the working characteristic curve of the subject is used as a main index for evaluating the efficiency of the image omics model for identifying the TNBC. In equations 1-1 to 1-6, the omics features are expressed as F or x due to the usage convention.

Preferably, the third step is: and respectively expanding and contracting the tumor target area by 5 millimeters to form an outer boundary and an inner boundary, wherein the region between the outer boundary and the tumor boundary is a tumor peripheral target area, the region between the inner boundary and the tumor boundary is a tumor internal target area, and the tumor Zhou Baou is combined with the tumor target area to form a whole tumor target area, so that four types of target areas of the tumor body, the tumor periphery, the whole tumor and the tumor are obtained.

Further, the spatial domain features include shape features, first-order histogram features, texture features, and wavelet features; shape features are a description of tumor morphological features including tumor length, surface area, volume, degree of edge smoothness; the first-order histogram features are based on the mathematical statistics of the image gray level, including the mean value, variance, kurtosis and the like of the image gray level; the texture characteristics comprise gray level co-occurrence matrix characteristics, gray level size area matrix characteristics, gray level run length matrix characteristics, adjacent gray level difference matrix characteristics and gray level dependency matrix characteristics; the wavelet feature is that the original image is subjected to wavelet decomposition into 8 frequency domains, namely LLL, LLH, LHL, HLL, HLH, HHL, LHH and HHH, wherein H represents a high-pass domain, and L represents a low-pass domain;

in the fifth step, the spatial domain features and the time domain features of the four types of target areas are spatial domain features and all time domain features of the flat scan period image source, and the feature number of each type of target area is equal;

and determining the area under the curve (AUC) of a working characteristic curve of the subject by taking the protein expression conditions of ER and PR and the molecular typing of the breast cancer determined by the expression and amplification state of HER2, which are obtained by carrying out immunohistochemical staining on the basis of paraffin pathological sections as a gold standard and using an R software pROC packet ROC function. A larger AUC value indicates a better prediction performance of the model. The present invention scores (p or q) each patient via the TNBC model; the model is then partitioned into TNBC and non-TNBC predicted patients based on a threshold (truncation point for p); the AUC will then be used to evaluate the predicted potency of the TNBC model. Therefore, AUC is an evaluation index of the prediction performance of the TNBC model, p or q is a prediction score, and a prediction result can be output by combining an interception point (aiming at p).

Similarly, the present invention scores each patient via HER2 model (p "or q"); then based on a threshold (truncation for p "), patients with HER2+ and HER 2-predictions from the compartmentalization model are assigned; the predicted potency of the HER2 model will then be evaluated by AUC. Therefore, AUC is an evaluation index of the prediction efficiency of the HER2 model, p ' or q ' is a prediction score, and a prediction result can be output by combining an interception point (aiming at p ').

Similarly, the present invention scores each patient (p 'or q') via the HR model; then based on a threshold (truncation point for p'), stratifying the patients for which the model predicts HR + and HR-; the predicted performance of the HR model will then be evaluated by AUC. Therefore, AUC is an evaluation index of the prediction efficiency of the HR model, p ' or q ' is a prediction score, and a prediction result can be output by combining the truncation point (aiming at p ').

Further, the four target regions add up to 10044 imaging omics features. Preferably, 2511 omics features are obtained for each type of target.

Further, the imaging omics model for breast cancer molecular typing is the imaging omics model for breast cancer molecular typing according to the first aspect.

In a third aspect, the present invention also provides a medium on which a method for constructing a proteomics model for molecular typing of breast cancer as described above is recorded.

Further, a method for constructing an imaging omics model of breast cancer molecular typing according to the first aspect is described; the following operational procedures are also described: when the prediction result of the TNBC model is TNBC, outputting a result 'TNBC'; and when the prediction result of the TNBC model is non-TNBC, adopting a HER2 model for prediction, if the prediction result of the HER2 model is HER2+, outputting a result of 'HER 2+ breast cancer', and if the prediction result of the HER2 model is HER2-, outputting a result of 'HR + HER 2-breast cancer'. Thereby forming a one-stop non-invasive whole flow for predicting breast cancer molecular typing.

In a fourth aspect, the present invention further provides an apparatus for constructing an imaging omics model for breast cancer molecular typing, including:

the first processing module is used for calculating spatial domain characteristics and time domain characteristics by using Python software;

the second processing module is used for taking the spatial domain characteristics and the time domain characteristics obtained by the first processing module as input variables; taking the breast cancer molecular typing determined by the protein expression conditions of ER and PR and the expression and amplification state of HER2 obtained by immunohistochemical staining of a patient based on a paraffin pathological section as a gold standard;

when the imaging group characteristics of the HER2 model are screened, the patients in the training set are classified into a HER2+ group and a HER 2-group through a gold standard, and the corresponding labels in the HER2 model are HER2+ and HER2-; analyzing and comparing HER2+ of label and corresponding image omics characteristics of HER 2-by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through glmnet and cv-glmnet functions of an R software glmnet package, and screening out image omics characteristics for predicting HER2 +; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within one standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HER2 model under the lambda value;

and the third processing module is used for taking the image omics characteristics of the selected TNBC model and the image omics characteristics of the selected HER2 model obtained by the second processing module as model input information and constructing the breast cancer molecular typing image omics model in a training set by using a logistic regression algorithm.

Further, the second processing module further comprises: when the image omics characteristics of the HR model are screened, patients in the training set are classified into an HR + group and an HR-group through a gold standard, and the corresponding label in the HR model is HR + and HR-; analyzing and comparing HR + and HR-corresponding videomics characteristics of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through a glmnet function and a cv-glmnet function of an R software glmnet package, and screening the videomics characteristics for predicting HR +/-; and taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HR model under the lambda value.

Further, the model obtained by the constructing apparatus is the model according to the first aspect of the present invention.

Further, the method further comprises a fourth processing module, configured to output a result "TNBC" when a prediction result of the TNBC model is TNBC; when the TNBC model prediction result is non-TNBC, HER2 model prediction is adopted, if the HER2 model prediction result is HER2+, a result of 'HER 2+ breast cancer' is output, and if the HER2 model prediction result is HER2-, a result of 'HR + HER 2-breast cancer' is output.

English in tables 1-6 corresponds to Chinese:

the English corresponding Chinese is:

1) Statistical indexes are as follows:

mean average value; variance of variance; skewness of skewness; minimum value of minium; maximum Correlation Coefficient of Maximal Correlation Coefficient (MCC); robust Mean Absolute development: a robust mean absolute deviation; median of Median

2) And (4) feature classification:

original (original character); wavelet (characteristic)

3) Characteristic subclass:

firstorder first order features

4) The feature name is as follows:

ZoneEntrol: regional entropy

Contrast: contrast ratio

Small Area High Gray Level Emphasis: small area high gray level enhancement

Large dependency Low Level Emphasis: large dependency low gray level enhancement

glszm: gray scale area matrix

ngtdm: adjacent gray level difference matrix

gldm: grey scale level dependent matrix

glcm: gray level co-occurrence matrix

Correction: correlation

Low Gray Level Zone Emphasis: low gray level region enhancement

Small dependency High Level Emphasis: small dependency high gray level enhancement

90percent Percentile:90 percentile

GLCM: a gray level co-occurrence matrix.

The invention applies the artificial intelligence image omics technology to predict the molecular classification of the breast cancer, tries to overcome the defects of the conventional process in the clinical diagnosis and treatment, and finally realizes the following aims:

1. help clinicians to accurately and efficiently judge the molecular typing of a specific breast cancer patient;

2. provides important basis for realizing accurate treatment of breast cancer;

3. the cloud system is expected to be deployed. By uploading the patient mammary gland MRI image, the target area identification, the feature extraction and the modeling analysis are completed on line, the prediction work from the image to the breast cancer molecule typing is completed in a one-stop way, and the development of breast cancer clinical diagnosis and treatment is greatly promoted.

At present, a convenient and fast noninvasive breast cancer molecular typing prediction method is not available, the invention is based on the largest breast cancer nuclear magnetic resonance imaging data set at present, an imaging omics model for distinguishing TNBC from non-TNBC, HR + and HR-breast cancer, HER2+ and HER 2-breast cancer is respectively constructed and verified through feature screening and model construction, and a one-station noninvasive breast cancer molecular typing prediction full flow is established. The result of molecular typing of breast cancer predicted by the imaging omics is higher in consistency with the gold standard, namely the immunohistochemical staining result. The invention provides a foundation for realizing noninvasive prediction of breast cancer molecular typing and guiding the application of clinical diagnosis and treatment decision.

The conception, the specific structure and the technical effects of the present invention will be further described with reference to the accompanying drawings to fully understand the objects, the features and the effects of the present invention.

Drawings

Figure 1 is a process for imaging omics data generation. Wherein FUSCC represents: subsidiary tumor hospital of the university of Compound Dan; BC represents: breast cancer.

FIG. 2 is a specific image feature of molecular typing of breast cancer screened using LASSO algorithm. (A) screening process of specific image characteristics of TNBC; (B) HR + breast cancer specificity image characteristic screening process; (C) a process for screening for specificity imaging characteristics of HER2+ breast cancer.

Where λ (lambda) represents the tuning parameter in the penalty function, the lower abscissa represents the log λ value, and the upper abscissa represents the number of variables. The left vertical dashed line represents the optimal lambda value (lambda.min), and the right vertical dashed line represents the maximum lambda value (lambda.1 se) that yields the most refined model within one standard deviation of the optimal lambda value. According to the method, 11 variables corresponding to lambda.1se in a graph A are selected as the characteristics of the selected image group of the TNBC model; selecting 11 variables corresponding to lambda.1se in the graph B as the characteristics of the selected HR model; and selecting 11 variables corresponding to lambda.1se in the graph C as the characteristics of the selected HER2 model in the image group.

Figure 3 is an ROC curve for TNBC prediction using training set imagery data. (A) The training set image omics model distinguishes the efficiency of TNBC and non-TNBC; (B) The effectiveness of distinguishing TNBC from non-TNBC by a verification set image omics model is verified.

FIG. 4 is a ROC curve for prediction of HR +/HR-breast cancer using training set and validation set proteomics data. (A) The effectiveness of distinguishing HR + breast cancer from HR-breast cancer by a training set imaging omics model; (B) The effectiveness of distinguishing HR + and HR-breast cancer by an image omics model is verified.

Figure 5 is a ROC curve for prediction of HER2+/HER 2-breast cancer using training set and validation set imaging data. (A) The efficiency of distinguishing HER2+ and HER 2-breast cancers by a training set imaging omics model; (B) The efficacy of the imaging omics model to distinguish HER2+ from HER 2-breast cancer was verified.

Fig. 6 is a TNBC model and HER2 model execution block diagram.

Figure 7 is a graph of the results of an implementation of the imaging omics model for molecular typing of breast cancer.

Detailed Description

In order to make the technical means, the characteristics, the purposes and the functions of the invention easy to understand, the invention is further described with reference to the specific drawings. However, the present invention is not limited to the following embodiments.

It should be understood that the structures, ratios, sizes, and the like shown in the drawings and described in the specification are only used for matching with the disclosure of the specification, so as to be understood and read by those skilled in the art, and are not used to limit the conditions under which the present invention can be implemented, so that the present invention has no technical significance, and any structural modification, ratio relationship change, or size adjustment should still fall within the scope of the present invention without affecting the efficacy and the achievable purpose of the present invention.

Example 1 imaging omics data Generation

The specific process is shown in figure 1.

The inventors of the present invention retrospectively collected 860 breast contrast-enhanced nuclear magnetic resonance (CE-MRI) images of primary breast Cancer patients diagnosed at the Fudan University affiliated tumor hospital (Fudan University Shanghai Cancer Center, FUSCC) between 8 months to 2019 before the start of neoadjuvant therapy or before patients without neoadjuvant therapy.

In a subsidiary tumor hospital of the university of Compound Dan, a patient who carries out mammary gland nuclear magnetic resonance examination takes a prone position, a main nuclear magnetic resonance imaging acquisition sequence is a dynamic contrast enhanced (dynamic contrast enhanced) sequence, three nuclear magnetic resonance models are adopted, and the imaging process and parameters are listed as follows:

siemens 3.0T magnetic resonance imaging system (hereinafter Siemens 3.0T): the contrast agent is gadopentetate dimeglumine (Gd-DTPA), the injection dosage is 0.1mmol/kg body weight, the Gd-DTPA is injected through the elbow vein by a pressure syringe at the speed of 3ml/s, and 15ml of normal saline is injected into a washing tube at the same speed after the Gd-DTPA is injected. Machine scan parameters: axial position, total scan 6 periods, including 1 period of flat scan and 5 periods of dynamic contrast enhancement, line enhancement scan 30 seconds after flat scan, continuously acquiring 6 time phases without intervals, wherein each time phase has a scan time of 44 seconds, a repetition Time (TR) of 5.1 milliseconds, an echo Time (TE) of 1.7 milliseconds, a field of view (FOV) of 260 millimeters multiplied by 260 millimeters, a layer thickness of 3 millimeters, a layer interval of 0 millimeter, and a flip angle of 15 degrees.

The nuclear magnetic resonance imaging system special for the ox Luo Rui 1.5.5T mammary gland (hereinafter referred to as Aurora): gd-DTPA is selected as a contrast agent, the injection dose is 0.2mmol/kg body weight, the Gd-DTPA is injected through the elbow vein by a pressure syringe at the speed of 2ml/s, and the Gd-DTPA is injected into a 15ml physiological saline flushing tube at the same speed after being injected. Machine scan parameters: axis position, total scanning 4 periods, including 1 period of flat scanning and 3 periods of dynamic contrast enhancement, horizontal line enhancement scanning for 90 seconds after flat scanning, continuously collecting 4 time phases without intervals, wherein the scanning time of each time phase is 180 seconds, the number of single-period scanning layers is 160, TR 5 milliseconds, TE 29 milliseconds, FOV 360 millimeters multiplied by 360 millimeters, the thickness of each layer is 1.1 millimeter, the interval of each layer is 0 millimeter, and the turning angle is 15 degrees.

General electric 1.5T magnetic resonance imaging system (hereinafter referred to as GE): gd-DTPA is selected as a contrast agent, the injection dose is 0.1mmol/kg body weight, the Gd-DTPA is injected through the elbow vein by a pressure syringe at the speed of 3ml/s, and the Gd-DTPA is injected into a 15ml physiological saline flushing tube at the same speed after being injected. Machine scan parameters: axis position, total scanning 4 periods, including 1 period of flat scanning and 3 periods of dynamic contrast enhancement, 20 seconds of line enhancement scanning after flat scanning, continuously acquiring 4 time phases without intervals, wherein the scanning time of each time phase is 35 seconds, TR 6.5 milliseconds, TE 3.5 milliseconds, FOV300 millimeters multiplied by 300 millimeters, layer thickness is 3 millimeters, layer spacing is 0 millimeter, and a flip angle is 10 degrees.

The inventors of the present invention selected the following CE-MRI image sequences for subsequent studies:

siemens 3.0T: a flat-scan and dynamic contrast enhancement sequence of T1 weighted imaging (T1 WI) is selected for 6 periods.

Aurora: selecting a flat scanning and dynamic contrast enhancement sequence for inhibiting T1WI by fat inhibition and water addition, and taking 4 stages in total.

GE: selecting flat scanning and dynamic contrast enhancement sequences of T1WI for 4 periods.

A radiodiagnosis doctor with 10-year mammary gland image diagnosis experience selects a first-stage image in a CE-MRI dynamic contrast enhancement sequence by using 3D Slicer software, and manually delineates a mammary gland tumor body target region.

On the basis of a tumor target area outlined in the early stage, MATLAB software is used for carrying out pixel value normalization on a flat scanning sequence image, the tumor target area is respectively expanded and contracted by 5 mm to form an outer boundary and an inner boundary, the area between the outer boundary and the tumor boundary is a tumor-surrounding target area, the area between the inner boundary and the tumor boundary is a tumor-internal target area, and the tumor Zhou Baou and the tumor target area are combined to form a whole tumor target area, so that the tumor, tumor-surrounding, whole tumor and four types of target areas in the tumor of breast tumor images of all patients are obtained in total

Based on the four target areas, two major image omics characteristics of spatial domain characteristics and time domain characteristics are calculated by respectively using Python software. In the aspect of spatial domain features, in each type of target area of each phase of image sequence, 14 shape features, 18 first-order histogram features, 75 texture features and 744 wavelet features are respectively extracted, and 851 spatial domain features are total. Shape features are a description of morphological features of the tumor, including but not limited to tumor length, surface area, volume, degree of edge smoothness, etc.; the first-order histogram features are based on mathematical statistics of the image gray levels, including but not limited to mean, variance, kurtosis, etc. of the image gray levels; the texture features include 24 Gray Level Co-actual Matrix (GLCM) features, 16 Gray Level Size area Matrix (GLSZM) features, 16 Gray Level Run Length Matrix (GLRLM) features, 5 adjacent Gray Level Difference Matrix (NGTDM) features, and 14 Gray Level Dependency Matrix (GLDM) features. The wavelet feature is to perform wavelet decomposition on the original image into 8 frequency domains (LLL, LLH, LHL, HLL, HLH, HHL, LHH, HHH), where H represents a high-pass domain and L represents a low-pass domain, and first-order histogram features and texture features are respectively calculated in the 8 decomposition domains, and each frequency domain includes 18 first-order histogram features and 75 texture features. The spatial domain features of the above-mentioned cinematomics are all extracted based on a Pyradiomics kit (https:// radiophotomics. In the aspect of time domain characteristics, the spatial domain characteristic change rate (formula 1-1) of the last-phase enhanced sequence compared with the flat-scan image and the spatial domain characteristic change rate (formula 1-2) of the adjacent two-phase image are respectively calculated, and the mean value (formula 1-3), the variance (formula 1-4), the skewness (formula 1-5) and the kurtosis value (formula 1-6) of each spatial domain characteristic among the enhanced sequences in each phase are calculated.

After the above processing, in each type of ROI, 10058 features including 3362 spatial domain features and 6696 temporal domain features are extracted from the medical image from the Aurora model; 7544 features are extracted from the medical image of the GE model source, wherein the 755 spatial domain features and 5019 temporal domain features are included; 13406 features are extracted from medical images from Siemens 3.0T model sources in total, wherein the 13406 features comprise 5036 spatial domain features and 8370 temporal domain features. Because of the significant difference of each dynamic contrast enhancement sequence of different nuclear magnetic resonance imaging systems, no known method can normalize and standardize enhancement period images of different imaging systems, so that the research only takes the spatial domain characteristics and all the time domain characteristics of a flat scan period image source when the images of three nuclear magnetic resonance machine types are brought into the research. The number of the image omics features meeting the condition in each type of ROI is 2511, and the image omics features which are comparable among 10044 three types of ROIs can be obtained by the four types of ROIs.

Example 2 construction and verification of imaging omics model for molecular typing of breast cancer

1. And (3) taking the breast cancer molecular typing determined by the protein expression conditions of ER and PR and the expression and amplification state of HER2 of the patient obtained by immunohistochemical staining based on paraffin pathological section as a gold standard. 860 breast cancer patients 1 who were diagnosed in the subsidiary tumor hospital of the university of fondand: 1 are randomly divided into a training set and a validation set.

2. In order to be able to distinguish breast cancer molecular typing in one-stop manner, the inventors of the present invention trained three imagemics models for distinguishing TNBC from non-TNBC, HR + and HR-breast cancer, HER2+ and HER 2-breast cancer, respectively.

3. (1) Screening the image omics characteristics of the predicted TNBC in R software and modeling: in a training set of 430 patients, 10044 proteomic features were used as input variables for the prediction model. Patients in the training set were typed by gold criteria into TNBC and non-TNBC groups, with the corresponding label (label) in the TNBC model being TNBC and non-TNBC, respectively. Analyzing and comparing image omics characteristics corresponding to TNBC and non-TNBC of label by using a 10-fold internal cross validation minimum absolute shrinkage and selection (LASSO) algorithm through a glmnet and cv-glmnet function of a glmnet package of R software, and screening out the image omics characteristics of the predicted TNBC. The maximum lambda value of the area under the curve (AUC) of the receiver operating characteristic curve (ROC) cross-validated in the training set and the optimal AUC within a standard deviation is used as the penalty coefficient of the feature selection algorithm, and 11 image omics features are correspondingly obtained under the lambda value to distinguish TNBC from non-TNBC (table 1, fig. 2).

(2) The proteomics features of the predicted HR +/-were screened in the R software and modeled: in a training set of 430 patients, 10044 proteomic features were used as input variables for the prediction model. Patients in the training set were typed by gold criteria into HR + and HR-groups, with the corresponding labels being HR + and HR-in the HR model, respectively. Analyzing and comparing the HR + and HR-corresponding image omics characteristics of label by using a 10-fold internal cross validation minimum absolute shrinkage and selection (LASSO) algorithm through a glmnet and cv-glmnet function of a glmnet package of R software, and screening out the image omics characteristics for predicting the HR +/-. The maximum lambda value of the area under the curve (AUC) of the receiver operating characteristic curve (ROC) cross-validated in the training set and the best AUC within one standard deviation is used as the penalty coefficient of the feature selection algorithm, and 11 image omics features are obtained under the lambda value to distinguish between HR + and HR-breast cancer (table 2, fig. 2).

(3) The proteomics features that predict HER2 +/-were screened in the R software and modeled: in a training set of 430 patients, 10044 proteomic features were used as input variables for the prediction model. Patients in the training set were typed by gold standard into HER2+ and HER 2-groups, with the corresponding labels being HER2+ and HER 2-respectively in the TNBC model. Analyzing and comparing HER2+ and HER 2-corresponding image omics characteristics of label by using a 10-fold internal cross validation minimum absolute shrinkage and selection (LASSO) algorithm through a glmnet and cv-glmnet function of a glmnet package of R software, and screening out image characteristics for predicting HER2 +/-from the image characteristics. The area under the curve (AUC) of the receiver operating characteristic curve (ROC) cross-validated in the training set and the maximum lambda value within a standard deviation from the optimal AUC are used as the penalty coefficients of the feature selection algorithm, and 11 image omics features corresponding to the lambda value are used for distinguishing HER2+ and HER 2-breast cancers (table 3, fig. 2).

11 proteomic features were used to distinguish TNBC from non-TNBC (table 1, fig. 2).

11 proteomic features were used to distinguish between HR + and HR-breast cancers (table 2, figure 2);

11 proteomic features were used to distinguish HER2+ and HER 2-breast cancers (table 3, figure 2).

4. 3 groups of screened image omics characteristics are respectively used as input information of three different models (hereinafter referred to as TNBC model, HR model and HER2 model) for distinguishing TNBC and non-TNBC, HR + and HR-breast cancer and HER2+ and HER 2-breast cancer, a predictive model of breast cancer molecular typing is constructed in a training set by using Logistic Regression (LR) algorithm, and the AUC value of a ROC curve is used as a main index for evaluating the efficacy of the predictive model; the sensitivity, specificity and Jaccard coefficient of the model are used as secondary indexes for evaluating the efficiency of the prediction model.

5. In terms of the main evaluation index of the prediction model,

(1) The optimum cut-off point of the ROC curve was determined to be (cut-off) 0.360 by the R software pROC package ROC function, and the AUC for the TNBC model to distinguish TNBC from non-TNBC under the condition was 0.969 (95% confidence interval, CI: 0.951-0.987). The prediction model is proved to have very good true positive rate and true negative rate for identifying TNBC and high model feasibility. The results show that with immunohistochemical staining as the gold standard for molecular typing assessment, in a validation set consisting of 430 patients, under the optimal cut-off 0.388 condition, the TNBC model recognized TNBC with an AUC of 0.92 (95% ci.

Intercept (. Beta.) included in the TNBC model ₀ ) And predicting the proteomics characteristics (x) of TNBC _k ) And its corresponding coefficient (beta) _k ) As in table 1. The score p for each patient was calculated as follows: p = 1-1/(1 +exp (beta) ₀ +β ₁ x ₁ +β ₂ x ₂ +…+β _k x _k ) And predicting breast cancer molecular typing according to the p value: p is a radical of>TNBC is predicted at 0.360, and non-TNBC is predicted at p ≦ 0.360.

(2) The optimum cutoff point for the ROC curve was determined by the R software pROC package ROC function to be (cut-off) 0.373, under which the AUC for the HR model to discriminate HR + and HR-breast cancers was 0.816 (95% CI. The results show that with immunohistochemical staining as the gold standard for molecular typing assessment, the HR model has an AUC of 0.799 (95% ci.

Intercept (. Beta. ') contained in the HR model' ₀ ) And predicting the imagination characteristics (x ') of HR +/HR-breast cancer' _k ) And its corresponding coefficient (. Beta. ')' _k ) As in table 2. Push buttonThe score p for each patient was calculated as follows: p '= 1-1/(1 + exp (beta)' ₀ +β’ ₁ x’ ₁ +β’ ₂ x’ ₂ +…+β’ _k x’ _k ) And predicting the molecular typing of the breast cancer according to the p' value: p'>HR + is predicted by 0.373 and HR-is predicted by p' ≦ 0.360.

(3) The optimum cutoff point for the ROC curve was determined by the R software pROC inclusion ROC function to be (cut-off) 0.423, under which the AUC for the HER2 model to distinguish between HER2+ and HER 2-breast cancers was 0.803 (95% ci. The results show that with immunohistochemical staining as the gold standard for molecular typing assessment, the HER2 model has an AUC of 0.717 (95% ci.

The HER2 model contains an intercept (. Beta.) " ₀ ) Prediction of the imageomics profile (x) of HER2+/HER 2-breast cancer " _k ) And the corresponding coefficient (beta) " _k ) As in table 3. The score p for each patient was calculated as follows: p "= 1-1/(1 + exp (beta)" ₀ +β” ₁ x” ₁ +β” ₂ x” ₂ +…+β” _k x” _k ) And predicting the molecular typing of the breast cancer according to the value of p': p is a radical of>0.423 is predicted to be HER2+, p.ltoreq.0.360 is predicted to be HER2-.

6. In terms of the secondary evaluation index of the predictive model,

(1) For the TNBC model, the inventors of the present invention defined TNBC as negative and non-TNBC as positive. Among 430 samples in the training set, 287 True Positives (TP), 17 False Negatives (FN), 10 False Positives (FP) and 116 True Negatives (TN) were selected. The sensitivity and specificity of the TNBC model for distinguishing TNBC and non-TNBC are respectively 0.944 and 0.921, and the Jaccard coefficient is 0.881; the verification set includes TP 285 cases, FN 25 cases, FP 21 cases and TN 99 cases. The TNBC model distinguishes TNBC from non-TNBC sensitivity and specificity of 0.9 and 0.825 respectively, and the Jaccard coefficient is 0.801.

(2) For the HR model, the applicant of the present invention defined HR + as negative and HR-as positive. Among 430 samples in the training set, 113 samples in the training set, 87 samples in the FN, 15 samples in the FP and 215 samples in the TN. The sensitivity and specificity of HR + and HR-breast cancer differentiation of the HR model are respectively 0.565 and 0.935, and the Jaccard coefficient is 0.678; in the verification set, there are 124 cases of TP, 68 cases of FN, 32 cases of FP and 206 cases of TN. The sensitivity and specificity of the HR model for distinguishing HR + and HR-breast cancers are 0.552 and 0.929 respectively, and the Jaccard coefficient is 0.68.

(3) For the HER2 model, the inventors of the present invention defined HER2+ as negative and HER 2-as positive. Among 430 samples in the training set, there are 266 samples in TP, 35 samples in FN, 51 samples in FP and 78 samples in TN. HER2 model the sensitivity and specificity for distinguishing HER2+ and HER 2-breast cancers were 0.884 and 0.605, respectively, with a jaccard coefficient of 0.667; in the verification set, there are 208 cases of TP, 83 cases of FN, 51 cases of FP and 88 cases of TN. The HER2 model distinguished HER2+ and HER 2-breast cancers by sensitivity and specificity of 0.794 and 0.54, respectively, with a jaccard coefficient of 0.547.

7. In conclusion, the inventors of the present invention found that breast cancer patients with TNBC versus non-TNBC, HR + and HR-breast cancer and HER2+ and HER 2-breast cancer can be accurately distinguished by CE-MRI based screening and modeling of imaging features. The method lays an important foundation for realizing one-stop noninvasive breast cancer molecular typing in the imaging omics.

8. In order to be able to non-invasively predict breast cancer molecular typing in one-stop manner, the inventors of the present invention designed the following procedure: in the validation set, the prediction result of a patient whose model (specifically, TNBC model) is predicted to be TNBC is output as "TNBC". In the non-TNBC patient predicted by the model (specifically, TNBC model), if the model (specifically, HER2 model) is predicted to be HER2+, the final output result is "HER2+ breast cancer", and if the model (specifically, HER2 model) is predicted to be HER2 ", the final output result is" HR + HER 2-breast cancer ", as shown in fig. 6.

Finally, the prediction results are shown in fig. 7. The abscissa represents the current clinical gold standard for molecular typing judgment, i.e., immunohistochemical typing, and the ordinate represents the molecular typing predicted by the proteomic model for molecular typing of breast cancer according to this example. Therefore, the correct cases of HR + HER2-, HER2+ and TNBC predicted by the model are 113 cases, 61 cases and 100 cases respectively. The total accuracy of the model for predicting the breast cancer molecular typing reaches 63.7 percent, the accuracy of the model for predicting TNBC reaches 87.6 percent, and the accuracy of the model for predicting HR + HER 2-breast cancer and HER2+ breast cancer is about 70 percent.

9. The invention relates to a model construction required software:

1)3D Slicer Software

2)MATLAB Software

3)Python Software

4)R Software

EXAMPLE 3 media

The medium is recorded with a construction method of an imaging omics model of breast cancer molecular typing. The construction method comprises the following steps:

step two, sketching a target area of a breast tumor body;

respectively expanding and contracting the tumor target area by 5 +/-2 millimeters to form an outer boundary and an inner boundary, wherein the area between the outer boundary and the tumor boundary is a tumor peripheral target area, the area between the inner boundary and the tumor boundary is a tumor inner target area, and the tumor Zhou Baou is combined with the tumor target area to form a whole tumor target area so as to obtain four target areas of the tumor body, the tumor periphery, the whole tumor and the tumor interior; and respectively expanding and contracting the tumor target area by 5 millimeters to form an outer boundary and an inner boundary, wherein the region between the outer boundary and the tumor boundary is a tumor peripheral target area, the region between the inner boundary and the tumor boundary is a tumor internal target area, and the tumor Zhou Baou is combined with the tumor target area to form a whole tumor target area, so that four types of target areas of the tumor body, the tumor periphery, the whole tumor and the tumor are obtained.

Based on the four target areas, calculating two major image omics characteristics of spatial domain characteristics and time domain characteristics by using Python software respectively; extracting the spatial domain features in Python based on a PyRadiomics toolkit; the time domain features include: the spatial domain characteristic change rate of the last-stage enhancement sequence compared with the flat-scan image adopts an algorithm of a formula 1-1; the algorithm of the change rate of the spatial domain characteristics of the two adjacent phases of images is a formula 1-2; the average value of each spatial domain feature among the enhancement sequences in each period has an algorithm of a formula 1-3; the variance of each spatial domain feature among the enhancement sequences in each period is calculated by a formula 1-4; each skyThe skewness of the inter-domain features among sequences is enhanced in each period, and the algorithm is a formula 1-5; the kurtosis value of each spatial domain feature among the enhancement sequences at each stage is calculated by a formula 1-6; wherein F represents the characteristic value of the image group, N represents the enhancement period number, F _N Or x _N The characteristic value of the image set representing the Nth enhanced image,

is the standard deviation of the distribution;

the spatial domain features comprise shape features, first-order histogram features, texture features and wavelet features; shape features are a description of tumor morphological features including tumor length, surface area, volume, degree of edge smoothness; the first-order histogram features are based on mathematical statistics of image gray levels, including mean, variance, kurtosis and the like of the image gray levels; the texture characteristics comprise gray level co-occurrence matrix characteristics, gray level size area matrix characteristics, gray level run length matrix characteristics, adjacent gray level difference matrix characteristics and gray level dependency matrix characteristics; the wavelet feature is to perform wavelet decomposition on the original image into 8 frequency domains, namely LLL, LLH, LHL, HLL, HLH, HHL, LHH and HHH, wherein H represents a high-pass domain, and L represents a low-pass domain.

when the image omics characteristics of the TNBC model are screened, classifying the patients in the training set into a TNBC group and a non-TNBC group through a gold standard, wherein corresponding label in the TNBC model is TNBC and non-TNBC respectively; analyzing and comparing image omics characteristics corresponding to the TNBC and non-TNBC of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through a glmnet function and a cv-glmnet function of an R software glmnet packet, and screening out the image omics characteristics of the predicted TNBC; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected TNBC model under the lambda value;

when the image omics characteristics of the HR model are screened, patients in the training set are classified into an HR + group and an HR-group through a gold standard, and the corresponding label in the HR model is HR + and HR-; analyzing and comparing HR + and HR-corresponding videomics characteristics of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through a glmnet function and a cv-glmnet function of an R software glmnet package, and screening the videomics characteristics for predicting HR +/-; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HR model under the lambda value;

when the imaging group characteristics of the HER2 model are screened, the patients in the training set are classified into a HER2+ group and a HER 2-group through a gold standard, and the corresponding labels in the HER2 model are HER2+ and HER2-; analyzing and comparing HER2+ of label and corresponding image omics characteristics of HER 2-by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through glmnet and cv-glmnet functions of an R software glmnet package, and screening out image omics characteristics for predicting HER2 +; and taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HER2 model under the lambda value.

The spatial domain features and the time domain features of the four types of target areas are spatial domain features and all time domain features of a flat scanning period image source, and the feature number of each type of target area is equal; AUC of the receiver working characteristic curve was determined by the R software pROC inclusion ROC function.

And step six, the image omics characteristics obtained by screening in the step five are used as model input information, a logistic regression algorithm is used for constructing an image omics model for identifying the TNBC in a training set, and the area under the curve of the working characteristic curve of the subject is used as a main index for evaluating the efficiency of the image omics model for identifying the TNBC.

The medium describes a method for constructing an imaging omics model for breast cancer molecular typing according to example 2; the following operational procedures are also described: when the prediction result of the TNBC model is TNBC, outputting a result of TNBC; when the TNBC model prediction result is non-TNBC, HER2 model prediction is adopted, if the HER2 model prediction result is HER2+, a result of 'HER 2+ breast cancer' is output, and if the HER2 model prediction result is HER2-, a result of 'HR + HER 2-breast cancer' is output. Thereby forming a one-stop noninvasive breast cancer molecular typing whole process.

Example 4 apparatus for constructing molecular breast cancer molecular typing proteomics model

The device for constructing the imaging omics model for breast cancer molecular typing comprises:

and the image recording module is used for recording the CE-MRI image.

And the image selecting module is used for selecting a flat scanning and dynamic contrast enhancement sequence of the CE-MRI image.

And the tumor body target area determining module is used for delineating the breast tumor body target area from the first-stage image in the CE-MRI dynamic contrast enhancement sequence.

The four types of target area determination modules are used for performing pixel value normalization on the flat scanning sequence image by using MATLAB software on the basis of a tumor target area delineated by the tumor target area determination module, and respectively expanding and contracting the tumor target area by 5 +/-2 millimeters (in the embodiment, 5 millimeters) to form an outer boundary and an inner boundary, wherein an area between the outer boundary and the tumor boundary is a tumor peripheral target area, an area between the inner boundary and the tumor boundary is an intra-tumor target area, and the tumor Zhou Baou and the tumor target area are combined to form a whole tumor target area, so that the tumor, peritumor, whole tumor and intra-tumor four types of target areas of the breast tumor image are obtained.

And the first processing module is used for calculating the spatial domain characteristics and the time domain characteristics by using Python software.

The second processing module is used for taking the spatial domain characteristics and the time domain characteristics obtained by the first processing module as input variables; and (3) taking the breast cancer molecular typing determined by the protein expression conditions of ER and PR and the expression and amplification state of HER2 of the patient obtained by immunohistochemical staining based on paraffin pathological section as a gold standard.

When the image omics characteristics of the TNBC model are screened, classifying the patients in the training set into a TNBC group and a non-TNBC group through a gold standard, wherein corresponding label in the TNBC model is TNBC and non-TNBC respectively; analyzing and comparing image omics characteristics corresponding to TNBC and non-TNBC of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through glmnet and cv-glmnet functions of an R software glmnet package, and screening out image omics characteristics of predicted TNBC; and taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected TNBC model under the lambda value.

When the imaging group characteristics of the HER2 model are screened, the patients in the training set are classified into a HER2+ group and a HER 2-group through a gold standard, and the corresponding labels in the HER2 model are HER2+ and HER2-; analyzing and comparing the corresponding imagery omics characteristics of HER2+ and HER 2-of label by using a minimum absolute contraction and selection algorithm of 10-fold internal cross validation through the glmnet and cv.glmnet functions of the R software glmnet, and screening the imagery omics characteristics for predicting HER2 +; and taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HER2 model under the lambda value.

When the image omics characteristics of the HR model are screened, patients in the training set are classified into an HR + group and an HR-group through a gold standard, and the corresponding label in the HR model is HR + and HR-; analyzing and comparing HR + and HR-corresponding videomics characteristics of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through a glmnet function and a cv-glmnet function of an R software glmnet package, and screening the videomics characteristics for predicting HR +/-; and taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the visual omics characteristic of the selected HR model under the lambda value.

The fourth processing module is used for outputting a result 'TNBC' when the prediction result of the TNBC model is TNBC; when the TNBC model prediction result is non-TNBC, HER2 model prediction is adopted, if the HER2 model prediction result is HER2+, a result of 'HER 2+ breast cancer' is output, and if the HER2 model prediction result is HER2-, a result of 'HR + HER 2-breast cancer' is output.

The imaging omics model of breast cancer molecular typing according to example 2 can be constructed by the construction device.

The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.

Claims

1. An imaging omics model for breast cancer molecular typing is characterized by comprising a TNBC model and a HER2 model; the TNBC model comprises:

the proteomics features of 11 predicted TNBC are shown in table 1;

and formula I: q = 1/(1 + exp (beta) ₀ +β ₁ x ₁ +β ₂ x ₂ +…+β _k x _k ) Wherein x) is _k Features of the image omics, beta, representing the predicted TNBC _k Denotes the corresponding coefficient, β ₀ Representing a TNBC model intercept of 2.6087610 _k And beta _k Specifically, as shown in table 1:

TABLE 1 TNBC model parameters

The HER2 model comprises:

and formula III: q "= 1/(1 + exp (beta))" ₀ +β” ₁ x” ₁ +β” ₂ x” ₂ +…+β” _k x” _k ) Wherein, x " _k Represents the imageomic feature, beta, of the predicted HER2+/HER 2-breast cancer " _k Denotes the corresponding coefficient, β " ₀ Representing a HER2 model intercept of 1.09702336,x " _k And beta' _k Specifically, as shown in table 3:

TABLE 3 HER2 model parameters

2. The proteomics model for molecular profiling of breast cancer according to claim 1, wherein the ROI sources of the 11 proteomic signatures of said predictive TNBC are shown in table 4:

TABLE 4 ROI sources for predicting proteomics characteristics of TNBC

The 11 ROI sources of the proteomic features predictive of HER2+/HER 2-breast cancer are shown in table 6:

3. The imaging omics model of breast cancer molecular typing of claim 1, wherein the results are calculated according to formula I: q is less than a first threshold value, TNBC is predicted, q is more than or equal to the first threshold value, and non-TNBC is predicted; the result is calculated by formula III: q '< the third threshold, is predicted to be HER2+, q' gtoreqthe first threshold, is predicted to be HER2-; the first threshold value is 0.640; the third threshold value is 0.577.

4. The imaging omics model of breast cancer molecular typing of claim 3, further comprising a run module; the execution module is configured to: outputting a result "TNBC" to the patient whose prediction result is TNBC for the TNBC model; and in the patients with the TNBC model prediction result of non-TNBC, adopting the HER2 model prediction to output a result of ' HER2+ breast cancer ' if the HER2 model prediction result is HER2+, and output a result of ' HR + HER 2-breast cancer ' if the HER2 model prediction result is HER2 '.

5. The proteomics model for molecular profiling of breast cancer according to claim 1, further comprising an HR model; the HR model includes: 11 proteomic features predicting HR +/HR-breast cancer, as shown in table 2;

and formula II: q '= 1/(1 + exp (beta)' ₀ +β’ ₁ x’ ₁ +β’ ₂ x’ ₂ +…+β’ _k x’ _k ) X 'in the formula' _k Beta represents the imago characteristics of the predicted HR +/HR-breast cancer' _k Denotes a corresponding coefficient, β' ₀ Represents an HR model intercept of-0.2212531,x' _k And beta' _k Specifically, as shown in table 2:

TABLE 2 HR model parameters

The 11 ROI sources of the predicted imageomic signature of HR +/HR-breast cancer are shown in table 5:

TABLE 5 ROI sources of proteomic features predicting HR +/HR-breast cancer

6. The imaging omics model of breast cancer molecular typing of claim 5, wherein the results are calculated by formula II: q 'is less than a second threshold value, the HR + is predicted, q' is more than or equal to the first threshold value, and the HR-is predicted; the second threshold is 0.627;

7. A method for constructing a breast cancer molecular typing imaging omics model is characterized by comprising the following steps:

step two, sketching a target area of a tumor body of the breast tumor;

respectively expanding and contracting the tumor target area by 5 +/-2 millimeters to form an outer boundary and an inner boundary, wherein the region between the outer boundary and the tumor boundary is a tumor-surrounding target area, the region between the inner boundary and the tumor boundary is a tumor-internal target area, and the tumor Zhou Baou is combined with the tumor target area to form a whole tumor target area so as to obtain four target areas of tumor, tumor-surrounding, whole tumor and tumor-inside;

calculating two major image omics characteristics of spatial domain characteristics and time domain characteristics by using Python software based on the four target areas; extracting the spatial domain features in Python based on a PyRadiomics toolkit; the time domain features include: the space domain characteristic change rate of the last phase enhanced sequence compared with the flat scan image adopts a formula 1-1; the algorithm of the change rate of the spatial domain characteristics of the two adjacent phases of images is a formula 1-2; the average value of each spatial domain feature among the enhancement sequences in each period has an algorithm of a formula 1-3; the variance of each spatial domain feature among the enhancement sequences in each period is calculated by a formula 1-4; the skewness of each spatial domain feature among the enhancement sequences at each stage is calculated by a formula 1-5; the kurtosis value of each spatial domain feature among the enhancement sequences at each stage is calculated byFormulas 1 to 6; wherein F represents the characteristic numerical value of the image group, N represents the enhancement period number, and F _N Or x _N The characteristic value of the image set representing the Nth enhanced image,

is the standard deviation of the distribution;

fifthly, taking the space domain characteristics and the time domain characteristics of the four types of target areas as input variables; taking the breast cancer molecular typing determined by the protein expression conditions of ER and PR and the expression and amplification state of HER2 obtained by immunohistochemical staining of a patient based on a paraffin pathological section as a gold standard;

when the image omics characteristics of the HR model are screened, patients in the training set are classified into an HR + group and an HR-group through a gold standard, and the corresponding label in the HR model is HR + and HR-; analyzing and comparing HR + and HR-corresponding videomics characteristics of label by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through a glmnet function and a cv-glmnet function of an R software glmnet package, and screening the videomics characteristics for predicting HR +/-; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within one standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HR model under the lambda value;

when the imaging group characteristics of the HER2 model are screened, the patients in the training set are classified into a HER2+ group and a HER 2-group through a gold standard, and the corresponding labels in the HER2 model are HER2+ and HER2-; analyzing and comparing the corresponding imagery omics characteristics of HER2+ and HER 2-of label by using a minimum absolute contraction and selection algorithm of 10-fold internal cross validation through the glmnet and cv.glmnet functions of the R software glmnet, and screening the imagery omics characteristics for predicting HER2 +; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HER2 model under the lambda value;

and step six, taking the image omics characteristics obtained by screening in the step five as model input information, constructing the TNBC-recognized image omics model in a training set by using a logistic regression algorithm, and taking the area under the working characteristic curve of a subject as a main index for evaluating the efficiency of the TNBC-recognized image omics model.

8. The method for constructing a proteomics model for molecular typing of breast cancer according to claim 7, wherein the proteomics model for molecular typing of breast cancer is the model according to claim 1.

9. A medium having the construction method according to claim 7 or 8 described therein.

10. The medium according to claim 9, wherein the construction method according to claim 8 is described thereon; the following operational procedures are also described: when the prediction result of the TNBC model is TNBC, outputting a result 'TNBC'; and when the TNBC model prediction result is non-TNBC, adopting the HER2 model prediction, if the HER2 model prediction result is HER2+, outputting a result of 'HER 2+ breast cancer', and if the HER2 model prediction result is HER2-, outputting a result of 'HR + HER 2-breast cancer'.

11. A device for constructing an imaging omics model for breast cancer molecular typing is characterized by comprising:

the second processing module is used for taking the spatial domain features and the time domain features obtained by the first processing module as input variables; taking the breast cancer molecular typing determined by the protein expression conditions of ER and PR and the expression and amplification state of HER2 obtained by immunohistochemical staining of a patient based on a paraffin pathological section as a gold standard;

when screening the imagemics characteristics of a HER2 model, patients in a training set are classified into a HER2+ group and a HER 2-group through gold standard, and corresponding labels in the HER2 model are respectively HER2+ and HER2-; analyzing and comparing HER2+ of label and corresponding image omics characteristics of HER 2-by using a minimum absolute shrinkage and selection algorithm of 10-fold internal cross validation through glmnet and cv-glmnet functions of an R software glmnet package, and screening out image omics characteristics for predicting HER2 +; taking the maximum lambda value of the difference between the area under the curve of the working characteristic curve of the subject subjected to cross validation in the training set and the optimal AUC within a standard deviation as a penalty coefficient of the characteristic selection algorithm, and correspondingly obtaining the image omics characteristic of the selected HER2 model under the lambda value;