CN116091484A - Method and system for predicting Ki-67 expression level of lung cancer - Google Patents

Abstract

The invention belongs to the technical field of medical image processing, and particularly relates to a method and a system for predicting the Ki-67 expression level of lung cancer. The method comprises the following steps: step 1, inputting lung CT image data, and dividing to obtain lung cancer tumor areas; step 2, extracting deep learning features from the lung cancer tumor area, and constructing a deep learning score; step 3, extracting image histology characteristics from the lung cancer tumor area, and constructing an image histology score; and 4, based on the deep learning score and the image histology score, predicting the Ki-67 expression level by using a nomogram model. The invention also provides a system for realizing the method. The method has the advantages of accuracy, noninvasive detection, good result interpretation and the like, and has good application prospect.

Description

Method and system for predicting Ki-67 expression level of lung cancer

Technical Field

The invention belongs to the technical field of medical image processing, and particularly relates to a method and a system for predicting the Ki-67 expression level of lung cancer.

Background

Non-small cell lung cancer (NSCLC) is one of the leading causes of cancer-related death worldwide, with the histological types most commonly adenocarcinoma and squamous cell carcinoma, with high recurrence rates and poor patient prognosis. Ki-67 is the most widely used cell proliferation assessment marker at present, and its expression is related to the occurrence, metastasis and prognosis of lung cancer. Studies have shown that Ki-67 nuclear antigen is only present in proliferating cells, which makes it a reliable way to rapidly assess normal and abnormal cell growth scores. Thus, ki-67 Proliferation Index (PI) has been used in the diagnosis of various tumor diseases to reflect proliferation of tumor cells, evaluate prognosis, and design molecular targeted drugs. It has also been shown to be an important prognostic factor for lung cancer.

Ki-67 was reported to be expressed at different levels in lung cancer of different tissue types. Wherein the expression level of small cell lung cancer is highest, and the expression level of lung squamous cell carcinoma is higher than that of adenocarcinoma. Thus, high-precision prediction of Ki-67 PI may highlight aggressive growth patterns of tumors, allowing accurate assessment of tumor biological behavior and helping to make clinical treatment decisions for patient's personalized management. However, the clinical Ki-67 index can only be obtained by immunohistochemical staining. The collection of tissue samples is invasive, and involves some subjectivity and sampling errors. At the same time, there are also some patients who have no tendency to undergo surgery or needle biopsies, and whose expression level cannot be estimated.

Image histology analysis of large imaging datasets has been successfully applied in the oncology field for non-invasive analysis of tumor heterogeneity and there is increasing interest in designing correlations between tumor heterogeneity and imaging characteristics. This involves extracting quantitative features from digital medical images, which enables high-dimensional data that can be mined to be applied to clinical decision support to provide improved diagnostic, prognostic and predictive accuracy. Imaging histology is increasingly important in personalized cancer treatments.

In the prior art, some reports on an artificial intelligence prediction method of Ki-67 expression exist, for example, a noninvasive detection method and a noninvasive detection device for the Ki67 expression level of lung adenocarcinoma based on deep imaging of CN202111523577.7 provide a method for predicting the Ki67 expression level of lung adenocarcinoma by using a multisource feature fusion prediction model. However, artificial intelligence prediction of Ki67 expression levels in lung cancer is currently lacking in related studies, and how to select appropriate model structures and input features to obtain accurate prediction results is still a problem to be solved in the art.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a method and a system for predicting the Ki-67 expression level of non-small cell lung cancer, which can realize accurate Ki-67 expression level prediction of non-small cell lung cancer through the selection of characteristics and the selection of model types.

A method for predicting the expression level of Ki-67 for lung cancer, comprising the steps of:

step 1, inputting lung CT image data, and dividing to obtain lung cancer tumor areas;

step 2, extracting deep learning features from the lung cancer tumor area, and constructing a deep learning score;

step 3, extracting image histology characteristics from the lung cancer tumor area, and constructing an image histology score;

and 4, based on the deep learning score and the image histology score, predicting the Ki-67 expression level by using a nomogram model.

Preferably, in step 1, the model for the segmentation task is built based on a U-Net network.

Preferably, in step 2, the deep learning features are extracted using a res net network.

Preferably, in step 2, the deep learning score is obtained by the following steps:

step a, obtaining a feature set construction through a three-level feature screening algorithm;

and b, obtaining the deep learning score by adopting a logistic regression algorithm.

Preferably, in step a, the three-level feature screening algorithm sequentially includes: the Mann-Whitney U test screens features with p values less than 0.05, the Spearman test removes features with redundancy correlation coefficients |r| >0.80, and the final feature set is screened based on the random forest Boruta algorithm.

Preferably, in step 3, the image histology features are implemented based on a pyrenoiomics tool, including at least one of shape features, first order features, or texture features.

Preferably, in step 3, the image histology score is obtained by the following steps:

step A, obtaining a feature set construction through a three-level feature screening algorithm;

and step B, obtaining the image histology score by adopting a logistic regression algorithm.

Preferably, in the step a, the three-level feature screening algorithm sequentially includes: the Mann-Whitney U test screens features with p values less than 0.05, the Spearman test removes features with redundancy correlation coefficients |r| >0.80, and the final feature set is screened based on the random forest Boruta algorithm.

Preferably, in step 4, the construction of the alignment chart model is realized based on the R language by inputting the deep learning score and the image group score.

The invention also provides a system for realizing the prediction method, which comprises the following steps:

the lung CT data acquisition and storage module is used for acquiring and storing lung CT image data;

the lung cancer automatic segmentation module is used for obtaining lung cancer tumor areas by segmentation from the lung CT image data;

the deep learning scoring module is used for extracting deep learning features and calculating a deep learning score;

the image histology scoring module is used for extracting image histology characteristics and calculating image histology scores;

and the alignment chart module inputs the deep learning score and the image group score into an alignment chart model to realize the prediction of the Ki-67 expression level.

The present invention also provides a computer-readable storage medium having stored thereon a computer program for implementing the above-described prediction method or system.

The invention is based on deep learning and image histology technology, and realizes automatic segmentation of lung cancer tumor area based on lung CT image, image histology feature extraction and Ki-67 expression level intelligent diagnosis. By optimizing the construction mode of the feature set and the algorithm of the model, the invention realizes the accurate and noninvasive detection of the Ki-67 expression of the non-small cell lung cancer, assists doctors in diagnosing and prognosis evaluation of lung cancer, can meet clinical application requirements, can help doctors to carry out treatment decision reference, reduces the diagnosis cost of patients, and provides technical support for early diagnosis and early treatment of the diseases. Considering that the noninvasive detection of Ki-67 is relatively lower in cost and safer for patients, the method has good application prospect.

It should be apparent that, in light of the foregoing, various modifications, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.

The above-described aspects of the present invention will be described in further detail below with reference to specific embodiments in the form of examples. It should not be understood that the scope of the above subject matter of the present invention is limited to the following examples only. All techniques implemented based on the above description of the invention are within the scope of the invention.

Drawings

Fig. 1 is a schematic flow chart of embodiment 1 of the present invention.

Fig. 2 is a schematic flow chart of embodiment 2 of the present invention.

Detailed Description

It should be noted that, in the embodiments, algorithms of steps such as data acquisition, transmission, storage, and processing, which are not specifically described, and hardware structures, circuit connections, and the like, which are not specifically described may be implemented through the disclosure of the prior art.

Example 1 prediction method of Ki-67 expression of non-Small cell Lung cancer

As shown in fig. 1, the present embodiment provides an intelligent detection method for non-small cell lung cancer Ki-67 expression based on chest CT image data, comprising the steps of:

s1, inputting lung CT image data, and dividing to obtain lung cancer tumor areas based on a U-Net network;

specifically, the U-Net network is trained by supervised learning, and the training data is manually segmented into tumor regions of lung cancer by experienced physicians. The lung CT lung cancer database as training data was from four independent hospital centers, with 87 patient data from one center as the training set and 45 patient data from the other three centers as the test set.

S2, automatically extracting deep learning features from the segmented lung cancer tumor region, and constructing a deep learning score;

the method specifically comprises the following steps:

s21, extracting deep learning features by using a ResNet network, wherein the process is realized based on Keras;

s22, the deep learning score is constructed by a feature set obtained by a three-level feature screening algorithm, and the score is established by a logistic regression algorithm;

s221, the three-level feature screening algorithm sequentially screens features with p value smaller than 0.05 for Mann-Whitney U test, spearman test removes features with redundancy correlation coefficient |r| >0.80, and the final feature set is screened based on a random forest Boruta algorithm.

S3, automatically extracting image histology characteristics from the segmented lung cancer tumor area, and constructing an image histology score;

the method specifically comprises the following steps:

s31, realizing image histology features based on a Pyradiomics tool, wherein the image histology features comprise shape features, first-order features and texture features;

s32, establishing an image histology score by a logistic regression algorithm, wherein the image histology score is constructed by a feature set obtained by a three-level feature screening algorithm. The specific steps of the three-level feature screening algorithm are the same as those described in S221.

S4, based on the deep learning score and the image histology score, the alignment chart model is utilized to predict the Ki-67 expression level.

Specifically, the step realizes the construction of the nomogram model based on the R language by inputting the deep learning score and the image group learning score. The nomogram model is built in the training set and verified in the external test set.

Example 2 alignment System for prediction of Ki-67 expression level of non-Small cell Lung cancer

As shown in fig. 2, the present embodiment provides a non-small cell lung cancer Ki-67 expression intelligent detection system based on chest CT images, comprising:

the lung CT data acquisition and storage module is used for acquiring and storing lung CT image data;

the lung cancer automatic segmentation module is used for obtaining lung cancer tumor areas by segmentation from the lung CT image data;

the deep learning scoring module is used for extracting deep learning features and calculating a deep learning score;

the image histology scoring module is used for extracting image histology characteristics and calculating image histology scores;

and the alignment chart module inputs the deep learning score and the image group score into an alignment chart model to realize the prediction of the Ki-67 expression level.

Specific procedures for predicting the expression level of non-small cell lung cancer Ki-67 using this system are described in example 1.

The technical scheme of the invention is further described through experiments.

Experimental example 1 influence of feature selection and model species on the accuracy of diagnosis of non-small cell lung cancer Ki-67 expression

1. Experimental method

Constructing a plurality of groups of models, and comparing the prediction performance, wherein the method sequentially comprises the following steps:

experiment group 1: using only deep learning scores; the specific implementation conditions are as follows: and finally, 10 deep learning features such as F12, F32, F52 and the like are screened out through three-level feature screening, and a deep learning score is constructed through a logistic regression algorithm. Using this score, the Ki-67 expression index can be calculated by inputting the relevant features.

Experiment group 2: using only image histology scoring; the specific implementation conditions are as follows: and finally, 28 image histology characteristics such as kurtosis, correlation coefficient, skewness and the like are screened out through three-level characteristic screening, and an image histology score is constructed through a logistic regression algorithm. Using this score, the Ki-67 expression index can be calculated by inputting the relevant features.

Experiment group 3: prediction of non-small cell lung cancer Ki-67 expression was performed using the method of example 1.

2. Experimental results

The predictive performance of the above three experimental groups on non-small cell lung cancer Ki-67 expression was compared as follows:

modeling method	Training set AUC	Test set AUC
			Experiment group 1: deep learning scoring	0.95	0.80
Experiment group 2: image histology scoring	0.98	0.82
			Experiment group 3: nomogram model	0.99	0.85

The experimental result proves that the alignment chart model integrating the deep learning score and the image group score has better prediction performance than the simple use of the deep learning score and the image group score.

Experimental example 2 influence of different feature fusion methods on the accuracy of non-small cell lung cancer Ki-67 expression diagnosis

1. Experimental method

On the basis of obtaining the deep learning score and the image group score, the experimental example adopts different algorithms to construct a final fusion model, and performs prediction performance comparison, and the method sequentially comprises the following steps:

experiment group 1: prediction of non-small cell lung cancer Ki-67 expression was performed using the method of example 1.

Experiment group 2: adopting a decision tree model to fuse the deep learning score and the image group score to predict the Ki-67 expression of the non-small cell lung cancer;

experiment group 3: adopting a random forest model to fuse the deep learning score and the image group score to predict the Ki-67 expression of the non-small cell lung cancer;

experiment group 4: and adopting a support vector machine model to fuse the deep learning score and the image histology score to predict the Ki-67 expression of the non-small cell lung cancer.

Other steps not illustrated in experimental groups 2-4 were the same as in example 1.

2. Experimental results

The predictive performance of the above four experimental groups on non-small cell lung cancer Ki-67 expression was compared as follows:

model integration method	Training set AUC	Test set AUC
			Experiment group 1: nomogram model	0.99	0.85
Experiment group 2: decision tree model	0.99	0.83
			Experiment group 3: random forest model	0.98	0.80
Experiment group 4: support vector machine model	0.99	0.79

Experimental results show that in a prediction task of non-small cell lung cancer Ki-67 expression, constructing a nomogram model by integrating a deep learning score and an image histology score has the best prediction performance in a test set; on the other hand, the nomogram model is more easily applied to clinics, used and interpreted by clinicians. In the present invention, it is found that model integration using the alignment chart model is the best choice.

According to the embodiment and experimental example, the invention constructs an accurate non-small cell lung cancer Ki-67 expression level prediction method and system through improving and optimizing input characteristics and model algorithm, and has good application prospect.

Claims (11)

1. A method for predicting the expression level of Ki-67 for lung cancer, comprising the steps of:

step 1, inputting lung CT image data, and dividing to obtain lung cancer tumor areas;

step 2, extracting deep learning features from the lung cancer tumor area, and constructing a deep learning score;

step 3, extracting image histology characteristics from the lung cancer tumor area, and constructing an image histology score;

and 4, based on the deep learning score and the image histology score, predicting the Ki-67 expression level by using a nomogram model.

2. The prediction method according to claim 1, characterized in that: in step 1, a model for the segmentation task is built based on a U-Net network.

3. The prediction method according to claim 1, characterized in that: in step 2, the deep learning features are extracted by using a ResNet network.

4. The prediction method according to claim 1, characterized in that: in step 2, the deep learning score is obtained by the following steps:

step a, obtaining a feature set construction through a three-level feature screening algorithm;

and b, obtaining the deep learning score by adopting a logistic regression algorithm.

5. The prediction method according to claim 4, wherein: in the step a, the three-level feature screening algorithm sequentially comprises the following steps: the Mann-Whitney U test screens features with p values less than 0.05, the Spearman test removes features with redundancy correlation coefficients |r| >0.80, and the final feature set is screened based on the random forest Boruta algorithm.

6. The prediction method according to claim 1, characterized in that: in step 3, the image histology features are implemented based on a pyrenoiomics tool, including at least one of shape features, first order features, or texture features.

7. The prediction method according to claim 1, characterized in that: in step 3, the image histology score is obtained by the following steps:

step A, obtaining a feature set construction through a three-level feature screening algorithm;

and step B, obtaining the image histology score by adopting a logistic regression algorithm.

8. The prediction method according to claim 7, wherein: in the step a, the three-level feature screening algorithm sequentially includes: the Mann-Whitney U test screens features with p values less than 0.05, the Spearman test removes features with redundancy correlation coefficients |r| >0.80, and the final feature set is screened based on the random forest Boruta algorithm.

9. The prediction method according to claim 1, characterized in that: in step 4, the construction of the nomogram model is realized based on the R language by inputting the deep learning score and the image group score.

10. A system for implementing the prediction method of any one of claims 1-9, comprising:

the lung CT data acquisition and storage module is used for acquiring and storing lung CT image data;

the lung cancer automatic segmentation module is used for obtaining lung cancer tumor areas by segmentation from the lung CT image data;

the deep learning scoring module is used for extracting deep learning features and calculating a deep learning score;

the image histology scoring module is used for extracting image histology characteristics and calculating image histology scores;

and the alignment chart module inputs the deep learning score and the image group score into an alignment chart model to realize the prediction of the Ki-67 expression level.

11. A computer-readable storage medium, characterized by: on which a computer program for implementing the prediction method according to any of the claims 1-9 or for implementing the system according to claim 10 is stored.

Images