CN114708200A - Method for extracting and screening CT (computed tomography) image characteristics for constructing chronic hepatitis B cirrhosis prediction model - Google Patents

Abstract

An image omics feature extraction and screening method for a CT image used for constructing a chronic hepatitis B cirrhosis prediction model. The existing liver fibrosis and cirrhosis prediction methods cannot be widely applied due to the cost problem, and the selection method of the characteristics used for prediction is not sound enough. The invention sets exclusion criteria for patients infected by HBV and having liver fibrosis pathological results of flat-scan CT examination, and counts the total number of patients capable of being included in CT image acquisition, and the acquired patient adopts an upper supine position to receive CT examination; dividing the obtained flat-scan CT image into two parts according to the total number of statistics, wherein one part is used as a training set and used for constructing a chronic hepatitis B cirrhosis prediction model, and the other part is used as a verification set and used for verifying the effect of the chronic hepatitis B cirrhosis prediction model; and carrying out image segmentation and image omics feature extraction. The method can accurately extract and screen the characteristics for constructing and predicting the liver cirrhosis model.

Description

Method for extracting and screening iconomics characteristics of CT (computed tomography) image for constructing chronic hepatitis B cirrhosis prediction model

Technical Field

The invention relates to an image omics feature extraction and screening method for a CT image for constructing a chronic hepatitis B cirrhosis prediction model.

Background

Current noninvasive methods for predicting the extent of liver fibrosis include serum indices and elastography. TE and MRE are known to have excellent diagnostic properties for liver fibrosis staging. However, these well-behaved equipment methods have not been widely used due to their high cost. In china, enhanced CT examination is usually recommended for HBV carriers to clarify their presence or absence of tumorigenesis, but many patients receive only flat scan CT examination due to limited cost effectiveness. While enhanced CT or MRI may provide more information than flat-scan CT, it is desirable to study noninvasive models to predict cirrhosis of the liver at a relatively low cost based on readily available data. In addition, the selection method of the characteristics used for constructing the prediction model is not sound enough.

Disclosure of Invention

The invention aims to solve the problems that the existing liver fibrosis and cirrhosis prediction method cannot be widely applied due to the cost problem and the selection method of the characteristics used for prediction is not sound enough, and provides an image omics characteristic extraction and screening method for constructing a CT image of a chronic hepatitis B cirrhosis prediction model.

A method for extracting and screening the characteristic of the imaging group of CT images used for constructing a chronic hepatitis B cirrhosis prediction model comprises the following steps:

step one, setting exclusion criteria aiming at patients infected by HBV and having liver fibrosis pathological results of flat-scan CT examination, carrying out next CT image acquisition on patients not in the exclusion criteria, and counting the total number of patients capable of being included in CT image acquisition, wherein the exclusion criteria are as follows:

(1) lack of detailed pathological records of liver fibrosis;

(2) lack of abdominal plain scan CT images of 1.5mm layer thickness;

(3) the interval between the flat scan CT examination and the biopsy is more than 3 months;

(4) poor image quality; the poor image quality refers to an image with a low score estimated by a PSNR (peak signal to noise ratio), structure similarity SSIM (structural similarity) or multi-scale structure similarity method;

(5) co-infecting with other viruses including one or more of hepatitis C virus, hepatitis D virus or human immunodeficiency virus;

(6) the liver cancer has a focal liver disease focus, and comprises one or more of HCC and hepatic tuberculosis;

(7) ingestion of large amounts of alcohol; said substantial amounts are greater than 20 grams per day;

(8) incomplete clinical data (n 39);

step two, the person to be collected adopts the supine position to receive CT examination; the parameters of the CT scan are: tube voltage is 120kVp, tube current is 250-350 mA, collimation slice thickness is 5mm, reconstruction slice thickness is 1.25mm, slice interval is 5mm, rotation time is 0.6s, spiral thread pitch is 1.375, visual field is 35-40 cm, and matrix is 512 multiplied by 512;

dividing the obtained flat-scan CT image into two parts according to the total number of statistics, wherein one part is used as a training set and used for constructing a chronic hepatitis B cirrhosis prediction model, and the other part is used as a verification set and used for verifying the effect of the chronic hepatitis B cirrhosis prediction model;

and step four, image segmentation and image omics feature extraction are carried out.

Preferably, the image omics feature extraction step comprises:

performing image preprocessing and feature extraction by adopting an open source Pyradiomics software package;

second, the voxel spacing is normalized to a size of 1 × 1 × 1mm, and the voxel intensity values are discretized with a bin width of 25 HU;

thirdly, extracting 828 imagery omics features from each ROI, including 18 first order statistics, 74 texture features and 736 wavelet-based transformation features;

fourthly, normalizing the characteristic values by using z scores in a training queue; verifying the standard score applied in the queue using the mean and standard deviation determined in the training queue;

fifthly, the step of selecting the image omics characteristics by repeatability among observers and lasso regression specifically comprises the following steps:

(1) randomly selecting a right portal horizontal CT scout image of the liver, and quantifying the repeatability of each radiological feature using intra-and inter-group correlation coefficients; one quantitative method is to repeat the ROI segmentation twice a week, another quantitative method is to perform the ROI segmentation independently and then calculate the inter-observer and inter-observer repeatability separately, wherein the minimum acceptable threshold of ICC is 0.8;

wherein, the intra-group and inter-group correlation coefficients are ICC for short and are called intraclass correlation coefficient in English;

(2) selecting an image omics feature;

first, features with high reproducibility are retained for subsequent analysis; the characteristic of high reproducibility means that the value of ICC inside an observer and between observers is more than 0.8;

and secondly, using a minimum absolute shrinkage and selection operator logistic regression algorithm, adjusting the penalty parameters through 10-time cross validation, and taking the characteristics of the non-zero coefficients as the characteristics related to the cirrhosis.

The invention has the beneficial effects that:

the invention is used for constructing a chronic hepatitis B cirrhosis prediction model, particularly provides a data base for constructing the chronic hepatitis B cirrhosis prediction model, makes up for the technical defect of feature selection used for constructing the model, and realizes the construction of the chronic hepatitis B cirrhosis prediction model through the specific image omics feature extraction and screening of CT images.

The method of the invention selects the characteristics of the image group by repeatability between observers and lasso regression, establishes repeatability tests, comprises ICC calculation between an observer and an observer, and has a minimum threshold value of 0.80. By selecting a smaller subset of the imaging features with higher repeatability, the functional clustering reproducibility of the imaging features can be obviously improved, important features related to liver cirrhosis prediction can be screened out, and the method has the advantage of accurate feature selection, thereby providing a basis for constructing a chronic hepatitis B cirrhosis prediction model.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a schematic representation of the principle of image omics feature extraction and screening according to the present invention;

FIG. 3a is a schematic representation of the LASSO model of the present invention determining the optimal λ value by minimum criteria using 10-fold cross-validation;

fig. 3b is a LASSO coefficient plot listing 85 selected features to which the present invention relates.

Detailed Description

The first embodiment is as follows:

in the embodiment, as shown in the flowchart of fig. 1, the method for extracting and screening the omics features of the CT image used for constructing the model for predicting chronic hepatitis b cirrhosis is as follows:

step one, setting exclusion criteria aiming at patients infected by HBV and having liver fibrosis pathological results of flat-scan CT examination, carrying out next CT image acquisition on patients not in the exclusion criteria, and counting the total number of patients capable of being included in CT image acquisition, wherein the exclusion criteria are as follows:

(1) lack of detailed pathological records of liver fibrosis (n-27);

(2) abdominal plain scan CT images lacking a 1.5mm layer thickness (n-128);

(3) the interval between the flat scan CT examination and the biopsy exceeds 3 months (n-16);

(4) poor image quality (n-42); the poor image quality refers to an image with a low score estimated by a PSNR (peak signal to noise ratio), structure similarity SSIM (structural similarity) or multi-scale structure similarity method;

(5) co-infecting with other viruses, including one or more of hepatitis C virus, hepatitis D virus or human immunodeficiency virus (n ═ 17);

(6) the liver cancer has a focal liver disease focus, and comprises one or more of HCC and hepatic tuberculosis (n is 45);

(7) ingestion of large amounts of alcohol; said plurality means greater than 20 grams per day (n-24);

(8) incomplete clinical data (n 39);

finally, 294 patients were included in the study. 144 patients between 2018 and 12 months were assigned to the training cohort and 150 patients between 2019 and 1 month to 2019 and 12 months were assigned to the validation cohort according to the biopsy date.

Step two, regarding the rest of the collected persons after elimination as the collected persons, the collected persons adopt the supine position to receive CT examination; the parameters of the CT scan are: tube voltage is 120kVp, tube current is 250-350 mA, collimation section thickness is 5mm, reconstruction section thickness is 1.25mm, section interval is 5mm, rotation time is 0.6s, spiral thread pitch is 1.375, visual field is between 35 and 40cm, and matrix is 512 multiplied by 512;

dividing the obtained flat-scan CT image into two parts according to the total number of statistics, wherein one part is used as a training set and used for constructing a chronic hepatitis B cirrhosis prediction model, and the other part is used as a verification set and used for verifying the effect of the chronic hepatitis B cirrhosis prediction model;

and step four, image segmentation and image omics feature extraction are carried out.

The second embodiment is as follows:

different from the specific embodiment, the method for extracting and screening the characteristic of the imaging group of the CT image for constructing the chronic hepatitis B cirrhosis prediction model of the embodiment,

the image omics feature extraction step comprises the following steps:

performing image preprocessing and feature extraction by adopting an open source Pyradiomics software package (http:// www.radiomics.io/Pyradiomics. html);

secondly, the voxel spacing is normalized in a size of 1 × 1 × 1mm, and the voxel intensity values are discretized by a bin width of 25HU to reduce the interference of image noise and normalize the intensity;

thirdly, extracting 828 image group characteristics from each ROI, including 18 first order statistics, 74 texture characteristics and 736 wavelet-based transformation characteristics;

fourthly, normalizing the characteristic values by using z scores in a training queue; the standard score (z-score) applied in the cohort was validated using the mean and standard deviation determined in the training cohort.

As shown in fig. 2, of the 828 extracted features, 85 features (8 first order statistics, 21 texture features and 56 wavelet-based transforms) with high repeatability were selected for subsequent analysis. 25 cirrhosis-related features with non-zero coefficients in the lasso regression model were selected according to the training cohort.

Fifthly, the step of selecting the image omics characteristics by repeatability among observers and lasso regression specifically comprises the following steps:

(1) randomly selecting right portal horizontal CT panned images of the liver, quantifying the reproducibility of each radiological feature using intra-and inter-group correlation coefficients based on 50 randomly selected patients; one quantitative method is to repeat the ROI segmentation twice a week, another quantitative method is to perform the ROI segmentation independently and then calculate the inter-observer and inter-observer repeatability separately, wherein the minimum acceptable threshold of ICC is 0.8;

wherein, the intra-group and inter-group correlation coefficients are ICC for short and are called intraclass correlation coefficient in English;

(2) selecting important image omics characteristics;

first, features with high reproducibility are retained for subsequent analysis; the characteristic of high reproducibility means that the value of ICC inside the observer and between the observers is more than 0.8;

next, a Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression algorithm is used, and the penalty parameters are adjusted by 10-fold cross validation to take the features of non-zero coefficients as the liver cirrhosis related features, as shown in fig. 3a and 3 b.

Figures 3a and 3b are diagrams showing selection of imagery omics features using LASSO regression; FIG. 3a shows the LASSO model determining the optimal λ value using 10-fold cross validation with the minimum norm, plotting a binomial deviation curve against log (λ), and plotting the dashed vertical line at the optimal value using the minimum norm and 1 standard error of the minimum norm (1-standard error condition), where the optimal λ value is 0.0383. Fig. 3b shows a LASSO coefficient curve listing 85 selected features.

Claims (2)

1. A method for extracting and screening the characteristic of the imagery omics of the CT image used for constructing the model for predicting the chronic hepatitis B cirrhosis is characterized in that: the method comprises the following steps:

step one, setting exclusion criteria aiming at patients infected by HBV and having liver fibrosis pathological results of flat-scan CT examination, carrying out next CT image acquisition on patients not in the exclusion criteria, and counting the total number of patients capable of being included in CT image acquisition, wherein the exclusion criteria are as follows:

(1) lack of detailed pathological records of liver fibrosis;

(2) abdominal scout CT images lacking a 1.5mm layer thickness;

(3) the interval between the flat scan CT examination and the biopsy is more than 3 months;

(4) poor image quality; the poor image quality refers to an image with a low score estimated by a PSNR (peak signal to noise ratio), structure similarity SSIM (structural similarity) or multi-scale structure similarity method;

(5) co-infecting with other viruses including one or more of hepatitis C virus, hepatitis D virus or human immunodeficiency virus;

(6) the liver cancer has a focal liver disease focus, and comprises one or more of HCC and hepatic tuberculosis;

(7) ingestion of large amounts of alcohol; said substantial amounts are greater than 20 grams per day;

(8) incomplete clinical data (n = 39);

step two, the person to be collected adopts the supine position to receive CT examination; the parameters of the CT scan are: tube voltage is 120kVp, tube current is 250-350 mA, collimation slice thickness is 5mm, reconstruction slice thickness is 1.25mm, slice interval is 5mm, rotation time is 0.6s, spiral thread pitch is 1.375, visual field is 35-40 cm, and matrix is 512 multiplied by 512;

dividing the obtained flat-scan CT image into two parts according to the total number of statistics, wherein one part is used as a training set and used for constructing a chronic hepatitis B cirrhosis prediction model, and the other part is used as a verification set and used for verifying the effect of the chronic hepatitis B cirrhosis prediction model;

and step four, image segmentation and image omics feature extraction are carried out.

2. The method for extracting and screening the iconomics features of the CT images for constructing the model for predicting chronic hepatitis b cirrhosis according to claim 1, wherein:

the image omics feature extraction step comprises the following steps:

performing image preprocessing and feature extraction by adopting an open source Pyradiomics software package;

second, the voxel spacing is normalized in size 1 × 1 × 1mm, and the voxel intensity values are discretized with a bin width of 25 HU;

thirdly, extracting 828 image group characteristics from each ROI, including 18 first order statistics, 74 texture characteristics and 736 wavelet-based transformation characteristics;

fourthly, normalizing the characteristic values by using z scores in a training queue; verifying the standard score applied in the queue using the mean and standard deviation determined in the training queue;

fifthly, the step of selecting the image omics characteristics by repeatability among observers and lasso regression specifically comprises the following steps:

(1) randomly selecting a right portal horizontal CT scout image of the liver, and quantifying the repeatability of each radiological feature using intra-and inter-group correlation coefficients; one quantitative method is to repeat the ROI segmentation twice a week, another quantitative method is to perform the ROI segmentation independently and then calculate the inter-observer and inter-observer repeatability separately, wherein the minimum acceptable threshold of ICC is 0.8;

wherein, the intra-group and inter-group correlation coefficients are ICC for short and are called intraclass correlation coefficient in English;

(2) selecting an image omics feature;

first, features with high reproducibility are retained for subsequent analysis; the characteristic of high reproducibility means that the value of ICC inside the observer and between the observers is more than 0.8;

and secondly, using a minimum absolute shrinkage and selection operator logistic regression algorithm, adjusting the penalty parameters through 10-time cross validation, and taking the characteristics of the non-zero coefficients as the characteristics related to the cirrhosis.

Images