WO2022100497A1

WO2022100497A1 - Method for determining mutation state of epidermal growth factor receptor, and medium and electronic device

Info

Publication number: WO2022100497A1
Application number: PCT/CN2021/128446
Authority: WO
Inventors: 黄钢; 左艳; 聂生东
Original assignee: 上海健康医学院; 上海理工大学
Priority date: 2020-11-13
Filing date: 2021-11-03
Publication date: 2022-05-19
Also published as: CN112488992A; CN112488992B

Abstract

The present invention relates to a method for determining a mutation state of an epidermal growth factor receptor, and a medium and an electronic device. The method comprises the following steps: acquiring a PET image and a CT image, and respectively preprocessing same, so as to form processed images; extracting deep learning features of the processed images by using a pre-trained convolutional neural network; extracting radiomics features of the processed images; fusing the deep learning features, the radiomics features and corresponding clinical features to form fused features; and obtaining a determination result for a mutation state of an epidermal growth factor receptor by using a pre-trained determination model and on the basis of the fused features. Compared with the prior art, the present invention has the advantages of a determination result having a high accuracy, etc.

Description

Epidermal growth factor receptor mutation state determination method, medium and electronic device

technical field

The invention relates to the technical field of CT image processing, in particular to a method, medium and electronic equipment for judging the mutation state of an epidermal growth factor receptor.

Background technique

Lung cancer is the most common malignant tumor and the most important tumor-killing disease in the world. Among them, non-small cell lung cancer accounts for about 80% to 85% of the total number of lung cancers. How to reduce the mortality of lung cancer is a challenging clinical problem that needs to be solved urgently. question. Whether the prognosis of lung cancer has recurrence or metastasis may be related to a variety of genes, one of the most important genes is epidermal growth factor receptor (EGFR). Improving the accuracy of epidermal growth factor receptor mutation judgment can be used to make early lung cancer timely and effective treatment.

The current EGFR genotyping method requires biopsy and gene sequence detection, which is invasive and difficult to obtain tissue samples. In addition, biopsy testing increases the potential risk of cancer metastasis, and the relatively high cost limits the applicability of mutation sequencing. With the introduction of the concept of radiomics, a more quantitative imaging method has been provided for this problem, making it possible to non-invasively, reliably and conveniently determine the EGFR mutation status. However, the existing methods cannot fully obtain information such as tumor heterogeneity, tumor metastasis, and tumor metabolic status, and the accuracy of the judgment results needs to be improved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method, medium and electronic device for judging the mutation state of epidermal growth factor receptor with high accuracy of judgment results in order to overcome the above-mentioned defects of the prior art.

The object of the present invention can be realized through the following technical solutions:

A method for judging the mutation state of epidermal growth factor receptor, the method comprises the following steps:

Acquire PET images and CT images, and perform preprocessing respectively to form processed images;

Using a pre-trained convolutional neural network to extract the deep learning features of the processed image;

extracting radiomic features of the processed image;

Fusing the deep learning features, radiomics features and corresponding clinical features to form a fusion feature;

A pre-trained judgment model is used to obtain an epidermal growth factor receptor mutation state judgment result based on the fusion feature.

Further, the preprocessing of the PET image is specifically:

Slice sorting of PET images;

Determine the center point of each slice based on the corresponding gold standard data of tumor region segmentation;

From the center point to the surrounding area, it is evenly filled with 0 pixels until the pixel size of the slice reaches the set value.

Further, the preprocessing of the CT image is specifically:

Slice sorting of CT images;

With the center point as the center, slices with the pixel size as the set value are evenly divided into the surrounding areas.

Further, the radiomic features include intensity features, shape features, texture features, improved LBP-3D features, wavelet features and Fourier features of CT images and intensity features, shape features and texture features of PET images.

Further, the acquisition formula of the improved LBP-3D feature is:

Among them, c represents the center point, x _c represents the pixel value of the center point, x _p represents the pixel value around the center point, P represents the pixel value, R represents the distance from the adjacent pixel to the center pixel, and i represents the adjacent pixels included in the process. number, P _i represents the pixel value of i pixel number, P _c represents the pixel value of the center point, the function

Further, the formation process of the fusion feature is specifically:

Calculate the covariance matrix of deep learning features, radiomics features, and clinical features, eigenvectors of covariance matrices, and inter-class scattering matrices, respectively;

obtaining an inter-set covariance matrix and its diagonalization parameter matrix based on the inter-class scattering matrix;

The fusion feature is obtained based on the inter-class scattering matrix and the diagonalization parameter matrix.

Before the fusion feature is input into the pre-trained judgment model, the processing includes equalization processing, abnormal point detection, feature simplification and feature selection.

Further, 5-fold cross-validation is used to perform model validation on the judgment model.

The present invention also provides a computer-readable storage medium comprising one or more programs for execution by one or more processors of an electronic device, the one or more programs comprising for executing the epidermal growth factor receptor as described above Instructions for the mutation state judgment method.

The present invention also provides an electronic device, comprising:

one or more processors;

memory; and

One or more programs stored in the memory, the one or more programs including instructions for performing the method for determining the mutation status of the epidermal growth factor receptor as described above.

Compared with the prior art, the present invention has the following beneficial effects:

1. The present invention uses both the radiomics features of PET images and the deep learning features of PET/CT to perform image-based judgment of receptor mutation status, which can provide not only tumor area information, but also information on surrounding areas of the tumor, and can completely obtain tumor heterogeneity. Information such as qualitative, tumor metastasis and tumor metabolic status, etc., has good robustness.

2. The input of the judgment model of the present invention is the fusion of deep learning features, radiomics features and corresponding clinical features, which can obtain more stable and accurate judgment results.

3. The radiomics features considered in the present invention include intensity features, shape features, texture features, improved LBP-3D features, etc. of CT images. Based on the feature set that can more accurately reflect CT images, the accuracy of judgment is effectively improved. sex.

4. The present invention can obtain stable and better judgment results under the data verification of different medical centers and under the test of different feature selection and prediction algorithms.

Description of drawings

FIG. 1 is a schematic flow chart of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and provides a detailed implementation manner and a specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

Example 1

As shown in FIG. 1 , this embodiment provides a method for judging the mutation state of epidermal growth factor receptor, and the method includes:

Step 1, acquiring PET (Positron Emission Computed Tomography, positron emission computed tomography) images and CT images.

In step 2, the PET image and the CT image are respectively preprocessed to form a processed image.

The preprocessing of the CT image is specifically:

(21) Read the original DICOM file of CT, and sort each CT slice;

(22) Read the gold standard of CT tumor region segmentation and sort the corresponding slices;

(23) Determine the center point of each slice of the CT tumor region segmentation file;

(24) Determine the center point of the CT original image file according to the center point of each slice of the CT tumor region segmentation file;

(25) Taking the center point of the original DICOM file of CT as the center, slices of 224×224 pixels are evenly divided around.

The preprocessing of the PET image is specifically:

Read the original DICOM file of PET, and sort the corresponding slices for each pair of PET;

The center point of each slice was determined based on the corresponding gold standard data for segmentation of the tumor area. The original DICOM slice of PET was taken as the center, and 0 pixels were evenly filled around until the pixel size of the slice was 224×224.

Step 3, using a pre-trained convolutional neural network to extract the deep learning features of the processed image, specifically:

(31) import a pre-trained convolutional neural network, the present embodiment, the convolutional neural network is obtained based on ImageNet data training;

(32) importing the CT image preprocessing data, calculating the output features of the last fully connected layer, and deriving the deep learning features of the CT image;

(33) Importing the PET image preprocessing data, calculating the output features of the last fully connected layer, and deriving the deep learning features of the PET image;

(34) Integrate the deep learning features of CT images and the deep learning features of PET images into deep learning features.

Step 4, extracting the radiomics feature of the processed image.

(41) Import the original CT image and its corresponding gold standard image for tumor segmentation, and calculate the intensity features, shape features, texture features, improved local binary pattern L-3 dimension (LBP-3D) features, wavelet features of each CT image tumor region Features, Fourier features, and derived CT image radiomic features;

(42) Import the original PET image and its corresponding gold standard image for tumor segmentation, calculate the intensity feature, shape feature, texture feature of each PET image tumor, and derive the PET image radiomics feature;

(43) Combine CT image radiomics features and PET image radiomics features into radiomics features.

Among them, the acquisition formula of the improved LBP-3D feature is:

Step 5, fuse the deep learning features, radiomics features and corresponding clinical features to form a fusion feature. Clinical features include demographic information, smoking history, diabetes, pathological stage, treatment history, treatment effect, recurrence and survival status, CT semantic features and serum tumor marker information.

The specific process of fusion feature acquisition includes:

Assume that the deep learning feature matrix of PET/CT images is denoted as X, the radiomics feature matrix is denoted as Y, and the clinical feature matrix is denoted as Z, x _ij ∈ X, y _ij ∈ Y, z _ij ∈ Z.

(51) Taking the deep learning feature X as an example, the formula for calculating the covariance matrix is:

Among them, c represents the number of categories of features, and n represents the number of features.

(52) Calculate the eigenvector φ _b ^T φ _b of the covariance matrix, the formula is:

where P is a matrix of orthogonal eigenvectors,

A diagonal matrix of real and nonnegative eigenvalues in decreasing order.

(53) Calculate the inter-class scattering matrix I:

Suppose Q _(c×r) is the largest non-zero feature of order r of matrix P:

Q ^T (φ _bx ^T φ _bx )Q=Λ _(r×r) ,

Then the most significant feature vector S _bx of order r can be calculated from the following formula:

(φ _bx Q) ^T S _bx (φ _bx )Q=Λ _(r×r) ,

Finally, the inter-class scattering matrix I can be derived from:

Similarly, S _by , I _Y , Y′ _(r×n) , S _bz , I _Z , Z′ _(r×n) are calculated.

(54) Diagonalize the transformed inter-set covariance matrix.

First S′ _xy is defined as:

S′ _xy =X′Y′ ^T ,

S′ _xy =U∑V ^T

Then the diagonal matrix ∑ with non-zero main diagonal elements is:

U ^T S′ _xy V=∑.

(55) Calculate the features in the fusion space.

make

Available:

The features X', Y', and Z' in the fusion space are:

Step 6, using a pre-trained judgment model to obtain an epidermal growth factor receptor mutation state judgment result based on the fusion feature.

Before the fusion feature is input into the pre-trained judgment model, the processing includes equalization processing, abnormal point detection, feature simplification and feature selection. Specifically, in this embodiment, the synthetic minority oversampling technique (SMOTE) is used for equalization; the outlier detection is performed by calculating the median absolute deviation; the Pearson correlation coefficient is used for feature simplification, and the threshold is 0.86; One-way ANOVA, Recursive Feature Elimination, Relief, etc. were used for feature selection.

The judgment model can be based on support vector machine (SVM), linear discriminant analysis (LDA), autoencoder (AE), random forest (Random Forest, RF), logistic regression (Logistic Regression, LR), logistic regression algorithm combined with LASSO (LR-LASSO), Adaboost (AB), Decision Tree (DT), Gaussian Process (GP), Naive Bayes (Naive Bayes, NB) etc. .

In this embodiment, when the judgment model is constructed and trained, 5-fold cross-validation is used to perform model verification on the judgment model, so as to improve the accuracy of the model.

The functions, if implemented in the form of software functional units and sold or used as independent products, may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

This embodiment verifies the effectiveness of the above method through the following test experiments.

The experimental data used the complete clinical data of 100 lung adenocarcinoma patients and their 18F-FDG PET/CT image data as the training samples of the judgment model. The basic information of the samples is shown in Table 1.

Table 1 Basic information of samples

Table 2 lists the performance results of 30 models established based on three feature selection methods and ten machine learning methods to verify the robustness of the above methods.

Table 2 Performance results of 30 models

From the above results, it can be seen that all models can make binary prediction of EGFR mutation status (mutated or wild). ANOVA was used for feature selection, and the model predicted by LDA had the highest performance (AUC=0.8071). Other classifiers also showed better predictive power after feature space dimensionality reduction (AUC ≥ 0.6000, ACC ≥ 0.6250 in the independent trial cohort). To sum up, the present invention fuses the features extracted from the PET/CT image, processes the fused features, and then determines the state, has high accuracy, and can effectively predict the mutation state of the epidermal growth factor receptor, and Good robustness and strong anti-interference ability.

Example 2

This embodiment provides an electronic device, including one or more processors, a memory, and one or more programs stored in the memory, the one or more programs including a method for performing epidermal growth as described in Embodiment 1 Instructions for the method of determining the mutation status of factor receptors.

The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art through logical analysis, reasoning or limited experiments on the basis of the prior art according to the concept of the present invention shall fall within the protection scope determined by the claims.

Claims

A method for judging the mutation state of epidermal growth factor receptor, characterized in that the method comprises the following steps:

Acquire PET images and CT images, and perform preprocessing respectively to form processed images;

Using a pre-trained convolutional neural network to extract the deep learning features of the processed image;

extracting radiomic features of the processed image;

fusing the deep learning features, radiomics features and corresponding clinical features to form a fusion feature;

A pre-trained judgment model is used to obtain an epidermal growth factor receptor mutation state judgment result based on the fusion feature.
The method for judging the mutation state of epidermal growth factor receptor according to claim 1, wherein the preprocessing of the PET image is specifically:

Slice sorting of PET images;

Determine the center point of each slice based on the corresponding gold standard data of tumor region segmentation;

From the center point to the surrounding area, it is evenly filled with 0 pixels until the pixel size of the slice reaches the set value.
The method for judging the mutation state of epidermal growth factor receptor according to claim 1, wherein the preprocessing of the CT image is specifically:

Slice sorting of CT images;

Determine the center point of each slice based on the corresponding gold standard data of tumor region segmentation;

With the center point as the center, slices with the pixel size as the set value are evenly divided into the surrounding areas.
The method for judging the epidermal growth factor receptor mutation state according to claim 1, wherein the radiomics features include intensity features, shape features, texture features, improved LBP-3D features, wavelet features and Fourier features and intensity features, shape features, and texture features of PET images.
The method for judging the mutation state of epidermal growth factor receptor according to claim 4, wherein the acquisition formula of the improved LBP-3D feature is:

Among them, c represents the center point, x c represents the pixel value of the center point, x p represents the pixel value around the center point, P represents the pixel value, R represents the distance from the adjacent pixel to the center pixel, and i represents the adjacent pixels included in the process. number, P i represents the pixel value of i pixel number, P c represents the pixel value of the center point, the function
The method for judging the mutation state of epidermal growth factor receptor according to claim 1, wherein the formation process of the fusion feature is specifically:

Calculate the covariance matrix of deep learning features, radiomics features, and clinical features, eigenvectors of covariance matrices, and inter-class scattering matrices, respectively;

obtaining an inter-set covariance matrix and its diagonalization parameter matrix based on the inter-class scattering matrix;

The fusion feature is obtained based on the inter-class scattering matrix and the diagonalization parameter matrix.
The method for judging the mutation state of epidermal growth factor receptor according to claim 1, wherein, before the fusion feature is input into the pre-trained judgment model, the processing includes equalization processing, abnormal point detection, feature simplification and feature selection .
The method for judging the mutation state of epidermal growth factor receptor according to claim 1, wherein a 5-fold cross-validation is used to perform model validation on the judgment model.
A computer-readable storage medium, characterized in that it includes one or more programs for execution by one or more processors of an electronic device, the one or more programs including a program for executing any one of claims 1-8 Instructions for the method for determining the mutation state of the epidermal growth factor receptor.
An electronic device, comprising:

one or more processors;

memory; and

One or more programs stored in the memory, the one or more programs including instructions for executing the method for determining the mutation status of the epidermal growth factor receptor according to any one of claims 1-8.