A method and application of radiomics feature processing of upper abdominal enhanced CT in patients with liver cirrhosis

Abstract

The present invention provides a method for processing the image omics features of the upper abdomen of patients with liver cirrhosis and its application in judging esophageal varices in patients with liver cirrhosis. The processing method comprises the following steps: S1. Collecting enhanced CT images of the upper abdomen of patients with liver cirrhosis; S2. Marking the liver, spleen and lower esophageal regions in the enhanced CT images layer by layer; S3. Three-dimensional reconstruction of the marked regions respectively; S4. Extracting image omics features from each three-dimensional region, and then fusing them to obtain the original image omics features of the patient; S5. Applying the principal component analysis method to perform dimensionality reduction processing on the original image omics features to obtain the image omics features after dimensionality reduction; S6. Using a classification algorithm to construct an image omics classification model for patients with esophageal varices.

Description

Technical Field

[0001] The present invention relates to the fields of medicine, nuclear medicine and radiomics, and in particular to a method for processing radiomics features of upper abdominal enhanced CT images of patients with liver cirrhosis and an application thereof in determining esophageal varices in patients with liver cirrhosis. Background Art

[0002] Cirrhosis is the terminal stage of the progression of various chronic liver diseases and is the 11th most common cause of death worldwide. Variceal bleeding is a serious and fatal complication of cirrhosis. The mortality rate caused by esophageal variceal bleeding is relatively higher than that of gastric varices. About 1 / 3 of patients with esophageal varices will experience esophageal variceal bleeding, with a mortality rate of up to 15%-25% 6 weeks after acute bleeding and up to 40% 1 year after bleeding. Therefore, early diagnosis of variceal, especially esophageal varices, predicting the risk of bleeding, and choosing the appropriate treatment are the key to reducing the mortality rate of variceal bleeding. At present, gastroscopy is the "gold standard" for diagnosing varicose veins and predicting the risk of bleeding. The guidelines and consensus on the diagnosis and treatment of cirrhosis recommend that all patients, once diagnosed with cirrhosis, must undergo gastroscopy and be reviewed every 1-3 years to detect varicose veins that need treatment as early as possible.

[0003] However, gastroscopy is an invasive procedure, which carries the risk of complications such as cardia tear, esophageal and gastric perforation, and bleeding. In addition, the incidence and severity of varicose veins are related to the severity of liver disease. Although the proportion of esophageal and gastric varices in patients with decompensated cirrhosis is as high as 85%, only 30-40% of patients with compensated cirrhosis have esophageal and gastric varices. In patients with Child-Pugh B / C liver function, the incidence of esophageal and gastric varices is 70%, while the incidence of esophageal and gastric varices in patients with Child-Pugh A grade is only 40%. Therefore, among patients who undergo gastroscopy, a considerable number of patients do not need to undergo the examination. They do not benefit from it, but have to endure the discomfort during the gastroscopy and bear the risks brought by the examination.

[0004] There have been many attempts to replace gastroscopy to determine esophageal varices in patients with cirrhosis, and corresponding methods have been established. For example, the hepatic venous pressure gradient (HVPG) is used to determine the risk of varicose veins. However, HVPG measurement is an invasive angiographic interventional operation and requires special equipment, which is relatively expensive; at the same time, it is easily affected by some factors: such as body position, respiration, anesthetic drugs and intrahepatic shunts. Some scholars have proposed several diagnostic and predictive models based on serum markers, such as FIB-4 (Fibrosis Index Based on Four Factors, liver fibrosis index), Lok score, APRI (Aspartate aminotransferase to platelet count ratio index, aspartate aminotransferase to platelet count ratio index), AAR (Aspartate aminotransferase to alanine aminotransferaseratio, aspartate aminotransferase to alanine aminotransferase ratio), VITRO score (Von Willebrand factor antigen / platelet ratio, von Willebrand factor antigen level to platelet count ratio), etc., but the accuracy of these models is not high.

[0005] In CT imaging, in addition to the size, shape, edge morphology, enhanced structure and other features of organs or lesions that can be observed by the naked eye, a large amount of image data is also generated, which cannot be completely read and interpreted by doctors. The application of radiomics analysis methods to extract and quantify data from medical images and then use appropriate algorithms for analysis can assist in the judgment of disease conditions in clinical practice. For example, Chinese invention patent application CN11028223A (publication date April 17, 2020) discloses a method for processing image omics features of microsatellite unstable colorectal cancer spectral CT iodine water images, including: collecting colorectal cancer sample images; outlining the region of interest in the colorectal cancer sample images, and extracting image features of the region of interest; performing dummy variable processing on the extracted image features to generate a dummy variable feature with a value of 0 or 1; citing the chi-square test, and using the hypothesis test method to filter out statistically significant omics features from the image features after dummy variable processing; using the correlation analysis method to filter out the redundancy between omics features, citing the Kendall correlation analysis, calculating the correlation coefficient, and obtaining the filtered omics features; based on the model level, introducing the LASSO algorithm to perform dimensionality reduction processing on the omics features; based on the optimization theory of machine learning methods, introducing a 5-fold cross-validation technique, and through the L2 regularization model construction idea, optimizing the L2 regularized Logistic regression model, and using the ROC method to evaluate the L2 regularized Logistic regression model. Using machine learning methods to judge the status of diseases based on medical images such as CT and thus reduce invasive examinations has become a research hotspot and trend. Summary of the invention

[0006] The present invention provides a method for processing the radiomics features of the upper abdomen enhanced CT images of patients with liver cirrhosis and application of the method in determining esophageal varices in patients with liver cirrhosis.

[0007] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0008] A method for processing radiomics features of upper abdominal enhanced CT images of patients with liver cirrhosis, comprising the following steps:

[0009] S1. Collect enhanced CT images of the upper abdomen of patients with liver cirrhosis and convert the portal venous phase CT images with the most obvious enhancement of the liver, spleen and esophageal veins into DICOM format;

[0010] S2. marking the liver, spleen and lower esophagus regions in the enhanced CT image in DICOM format obtained in step S1 layer by layer, and separating the regions from other tissues;

[0011] S3. Perform three-dimensional reconstruction on the regions of interest marked in step S2;

[0012] S4. extracting radiomic features from each three-dimensional region obtained in step S3, and fusing the radiomic features of the three regions to obtain the original radiomic features of the patient;

[0013] S5. Applying the principal component analysis method to perform dimensionality reduction processing on the original imaging features obtained in step S4 to obtain the imaging features after dimensionality reduction;

[0014] S6. Based on the dimensionality-reduced omics features obtained in step S5, a classification algorithm is used to construct an imaging omics classification model for esophageal varices in patients.

[0015] Preferably, in step S2, the lower esophagus refers to the area from the cardia to 5 cm above it.

[0016] Preferably, in step S2, the doctor first marks the liver, spleen and lower esophagus in the enhanced CT image in DICOM format layer by layer, and extracts all calibration data from the enhanced CT image in DICOM format; the specific operation is:

[0017] The image data is parsed to extract the marked contour of the area, and then the extracted contour is expanded and filled to obtain a binary image with all 1s in the marked area and all 0s outside the marked area. The binary image has the same shape and size as the enhanced CT image in DICOM format. Finally, the enhanced CT image data in DICOM format and the binary image are multiplied, all calibration data are extracted from the original enhanced CT image, and the area is separated from other tissues.

[0018] Preferably, in step S3, the three-dimensional reconstruction adopts Marching Cubes (MC) algorithm.

[0019] The basic principle of the MC algorithm is to construct an isosurface in a three-dimensional data field, find the voxel passing through the isosurface, find the isosurface in the voxel and calculate the relevant parameters. The algorithm has a fast calculation speed and the reconstructed image has high quality. The 3D reconstruction of the data can well restore the true shape of the target area, which plays a key role in the training of the radiomics classification model for accurately predicting esophageal varices in patients.

[0020] Preferably, in step S4, the radiomics features extracted from each region include: 42-dimensional texture features, 540-dimensional gradient direction histogram features, 48-dimensional wavelet features and 156-dimensional statistical features, for a total of 786 dimensions.

[0021] Therefore, after the radiomic features of the liver, spleen and lower esophagus were fused, the original radiomic features of the patient obtained had a total of 2358 dimensions.

[0022] Preferably, in step S5, the specific process of the dimensionality reduction processing is:

[0023] S5-1. Standardize the original radiomics features obtained in step S4 using formula I.

[0024] Xi'=(Xi-Xmin) / (Xmax-Xmin) I,

[0025] Among them, Xi' is the image feature of the i-th patient after standardization in this dimension,

[0026] Xi is the image feature of the patient that has not been standardized in this dimension.

[0027] Xmax is the maximum value of this dimension,

[0028] Xmin is the minimum value of this dimension;

[0029] Then the covariance matrix of the multi-dimensional features is calculated to obtain the eigenvalues ​​and eigenvectors of the matrix;

[0030] S5-2. Sort the original radiomics feature values ​​obtained in step S4 from large to small, select the principal component, find its corresponding feature vector, and project the original radiomics feature data onto the selected feature vector. The dimension of the original radiomics feature is reduced to obtain the radiomics feature after dimensionality reduction.

[0031] Preferably, in step S6, the classification algorithm refers to SVM (Support vector machine).

[0032] Further preferably, in step S6, the dimensionality-reduced radiomics features and clinical features are integrated, and a radiomics classification model for esophageal varices of patients is constructed using a SVM classification method; further preferably, the specific process is:

[0033] The patients included in the model were randomly divided into 5 groups, each group of patients was used as a validation set, and the other 4 groups were used as training sets. Based on the optimization theory of machine learning methods, 5-fold cross validation was performed to optimize the SVM model.

[0034] Preferably, step S6 also includes evaluating the SVM model by discrimination, calibration and clinical practical value.

[0035] Preferably, the discrimination is evaluated by the AUC (area under the curve) value of the ROC (Receiver Operating Characteristic Curves), and the normal range of the AUC value is between 0.50 and 1.00. The larger the value, the better the discrimination of the model.

[0036] Preferably, the calibration is evaluated by a decision curve, with the degree of esophageal varices determined under gastroscopy as the gold standard. The higher the degree of agreement between the predicted esophageal varices of the patient and the gold standard, the better the model calibration.

[0037] The goal of SVM is to find the best separation hyperplane in the feature space so that the interval between positive and negative samples in the training set is maximized. The objective function of the model is as follows:

[0038]

[0039] The goal is to minimize the value of L(ω,b,α) through training and optimize the model. It is a regular term, which is used to reduce the impact of overfitting; Used to construct the algorithm "hyperplane" for classification.

[0040] The present invention also provides a device for judging the degree of esophageal varices in patients with liver cirrhosis based on the above-mentioned method for processing the image genomics features of the upper abdominal enhanced CT images of patients with liver cirrhosis, the device comprising:

[0041] (1) Input module

[0042] The input module is configured to input the following enhanced CT images of the upper abdomen of a patient with liver cirrhosis to be diagnosed:

[0043] DICOM format portal venous phase CT images of the liver, spleen, and esophageal veins with most prominent enhancement;

[0044] (2) Esophageal varices degree judgment module

[0045] The esophageal varices degree judgment module is configured to include a region of interest marking module, a region of interest three-dimensional reconstruction module, an imaging genomics feature extraction module, an imaging genomics feature dimension reduction module, and a patient esophageal varices imaging genomics feature classification model;

[0046] (3) Output module

[0047] The output module is configured to output the obtained degree of esophageal varices in the patient with liver cirrhosis.

[0048] Preferably, the region of interest marking module is configured to mark the liver, spleen and lower esophagus regions layer by layer in the enhanced CT image in DICOM format, and separate the regions from other tissues.

[0049] Further preferably, the region of interest marking module is configured to perform the following processing on the enhanced CT image in DICOM format:

[0050] First, the enhanced CT images of the liver, spleen and lower esophagus in DICOM format, which are marked layer by layer by a doctor, are analyzed to extract the marked contour of the area. Then, the extracted contour is expanded and filled to obtain a binary image with all 1s in the marked area and all 0s outside the marked area. The binary image has the same shape and size as the enhanced CT image in DICOM format. Finally, the original enhanced CT image data and the binary image are multiplied, all calibration data are extracted from the enhanced CT image in DICOM format, and the area is separated from other tissues.

[0051] Further preferably, the lower esophagus refers to the area from the cardia to 5 cm above it.

[0052] Preferably, the region of interest three-dimensional reconstruction module is configured to perform three-dimensional reconstruction on the marked regions respectively to obtain a three-dimensional image.

[0053] Preferably, the radiomics feature extraction module is configured to extract texture features, gradient direction histogram features, wavelet features and statistical features from three-dimensional images to obtain original radiomics features.

[0054] Further preferably, the imaging genomics feature extraction module is configured to extract 42-dimensional texture features, 540-dimensional gradient direction histogram features, 48-dimensional wavelet features and 156-dimensional statistical features from three-dimensional images, respectively, to obtain the imaging genomics features of the target area, and to fuse the imaging genomics features of the three candidate areas to obtain the original imaging genomics features of the patient.

[0055] Preferably, the radiomics feature dimensionality reduction module is configured to process the original radiomics features as follows:

[0056] I. Standardize the original radiomics features obtained in step S4 using formula I,

[0057] Xi'=(Xi-Xmin) / (Xmax-Xmin) I,

[0058] Among them, Xi' is the image feature of the i-th patient after standardization in this dimension,

[0059] Xi is the image feature of the patient that has not been standardized in this dimension.

[0060] Xmax is the maximum value of this dimension,

[0061] Xmin is the minimum value of this dimension;

[0062] Calculate the covariance matrix of multidimensional features and obtain the eigenvalues ​​and eigenvectors of the matrix;

[0063] II. Sort the original imaging genomics feature values ​​obtained in step S4 from large to small, select the principal component, find its corresponding eigenvector, and project the original imaging genomics feature data onto the selected eigenvector. The dimension of the original imaging genomics feature is reduced to obtain the dimensionality-reduced genomics feature.

[0064] Preferably, the esophageal varices imaging feature model is established by the following process:

[0065] The patients included in the model were randomly divided into 5 groups, each group of patients was used as a validation set, and the other 4 groups were used as training sets. Based on the optimization theory of machine learning methods, 5-fold cross-validation was performed to optimize the SVM model; the SVM model was evaluated by discrimination, calibration and clinical practical value.

[0066] The discrimination is evaluated by the AUC (area under the curve) value of the ROC (Receiver Operating Characteristic Curves); the normal range of the AUC value is between 0.50 and 1.0, and the larger the value, the better the discrimination of the model.

[0067] The calibration is evaluated by a decision curve, with the degree of esophageal varices determined under gastroscopy as the gold standard. The higher the degree of agreement between the predicted esophageal varices of the patient and the gold standard, the better the model calibration.

[0068] The device for judging the degree of esophageal varices in patients with cirrhosis based on the above-mentioned method for processing the image genomics features of the upper abdomen of patients with cirrhosis provided by the present invention can be a computer device. The module for judging the degree of esophageal varices can be included in the memory of the computer device; the computer device also includes a processor, which can run the computer program of the three-dimensional reconstruction module of the region of interest, the image genomics feature extraction module, the image genomics feature dimensionality reduction module and the image genomics feature model of esophageal varices.

[0069] In addition, another object of the present invention is to provide a method for grouping the degree of esophageal varices in patients with liver cirrhosis using the above device, the method comprising:

[0070] (1) Obtain enhanced CT images of the upper abdomen of cirrhotic patients to be grouped;

[0071] (2) inputting the portal venous phase CT images of the liver, spleen and esophageal veins with the most obvious enhancement obtained in step (1) into the input module of the device in DICOM format;

[0072] (3) using the esophageal varices degree judgment module to judge the degree of esophageal varices in patients with liver cirrhosis, and outputting the result from the output module; the degree of esophageal varices is mild esophageal varices or moderate to severe esophageal varices;

[0073] (4) The cirrhotic patients with mild esophageal varices are identified as a low bleeding risk group that does not require gastroscopy, and the cirrhotic patients with moderate to severe esophageal varices are identified as a high bleeding risk group that requires gastroscopy.

[0074] The grouping method provided by the present invention is not used for diagnosing diseases, but can provide intermediate results or auxiliary information for clinical follow-up and treatment of patients with liver cirrhosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The present invention will be further described below in conjunction with the accompanying drawings.

[0076] Figure 1 1 is a flow chart of the method for processing the radiomics features of the upper abdominal enhanced CT images of patients with cirrhosis described in Example 2.

[0077] Figure 2 It is a schematic diagram of the enhanced CT images of the liver, spleen and lower esophagus that have been marked in Example 2, wherein AD are marked CT images of patients without esophageal varices, marked CT images of patients with mild esophageal varices, marked CT images of patients with moderate esophageal varices, and marked CT images of patients with severe esophageal varices, respectively.

[0078] Figure 3 It is the receiver operating characteristic (ROC) curve of the imaging omics training set and validation set of mild esophageal varices (EV) in Example 2, wherein A is the ROC curve of the training set, and B is the ROC curve of the validation set.

[0079] Figure 4 These are the ROC curves of the imaging genomics training set and validation set for moderate to severe EV in Example 2, where A is the ROC curve of the training set and B is the ROC curve of the validation set.

[0080] Figure 5 It is the decision curve for discriminating the validation set in Example 2; wherein A is the decision curve for the validation set of imaging omics for model discrimination of mild EV, and B is the decision curve for the validation set of imaging omics for model discrimination of moderate to severe EV. DETAILED DESCRIPTION

[0081] The present invention is described below with reference to specific examples. It will be appreciated by those skilled in the art that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention in any way.

[0082] The experimental methods in the following examples are conventional methods unless otherwise specified. The reagents and materials used in the following examples are commercially available products unless otherwise specified.

[0083] The abbreviations of the embodiments are:

[0084] EV: esophageal varices;

[0085] ROC (Receiver Operating Characteristic Curves): Receiver operating characteristic curve;

[0086] AUC: Area under the curve.

[0087] In the following examples,

[0088] Gastroscopy:

[0089] Fast for more than 8 hours before the examination. Use the Japanese Olympus GIF260 or GIF290 electronic gastroscope to observe and record the esophageal varices (EV). EV grading criteria: (1) No EV. (2) Mild EV, varicose veins are straight, diameter ≤0.3cm, and red sign is negative. (3) Moderate EV, mild EV red sign is positive; or varicose veins are serpentine and tortuous, 0.3cm<diameter ≤1.0cm, and red sign is negative. (4) Severe EV, moderate EV red sign is positive; or varicose veins are beaded, nodular or tumor-like; or diameter is greater than 1.0cm, regardless of whether there is a red sign.

[0090] Enhanced CT scan of the upper abdomen:

[0091] Fasting for more than 8 hours before the examination, the examination was performed using a GE Lightspeed VCT 64-slice scanner, with the following scanning parameters: collimator width 0.625 mm, tube current 380 mA, tube voltage 120 kV, slice thickness 5 mm, and slice spacing 5 mm. For enhanced scanning, 100 ml of contrast agent iopromide (370 mg I / ml) was injected through the antecubital vein with a high-pressure syringe at a flow rate of 3 ml / s. All patients underwent three-phase scanning, with Smart prep contrast agent tracking technology used in the arterial phase. The scan was triggered when the CT value of the abdominal aorta reached or exceeded the set threshold (150 HU), the venous phase was 65s to 70s, and the delayed phase was 180s to 300s. Multiplanar reconstruction and post-processing were performed on the ADW 4.3 workstation.

[0092] Data Analysis:

[0093] IBM SPSS Statistics 23 statistical software was used, and the measurement data were expressed as mean ± standard deviation. The count data were expressed as frequency and percentage. The measurement data between the two groups were compared by t-test for normal distribution data, Mann-Whitney U test for non-normal distribution data, and chi-square test for inter-group count data. The inter-group differences in basic clinical data were compared, and P<0.05 indicated that the differences were statistically significant. AUC, accuracy, sensitivity, specificity, PPV, and NPV were calculated as indicators to evaluate the classification accuracy of the imaging omics feature model. The clinical effectiveness of the imaging omics model was evaluated by quantifying the net benefits of different thresholds in the patients included in the study and drawing a decision curve.

[0094] Programs and systems involved:

[0095] Python 3.5 (Ubuntu system, version 16.04.4) was used for feature extraction and model building. The packages used included Numpy, Pandas, Scikit-learn, Scipy, and CV2.

[0096] Example 1 Basic information of patients with cirrhosis included in the study

[0097] 1.1 Research subjects

[0098] Methods: Clinical and laboratory data, as well as gastroscopy and upper abdominal enhanced CT images of patients with cirrhosis hospitalized in Beijing You'an Hospital, Capital Medical University from January 2016 to December 2017 were retrospectively collected. The diagnosis of cirrhosis is mainly based on medical history, pathological tissue biopsy, typical clinical symptoms, laboratory tests, and imaging manifestations. The diagnostic criteria for cirrhosis in this study were in accordance with the criteria of the Guidelines for the Diagnosis and Treatment of Cirrhosis (2019 edition).

[0099] Inclusion criteria included: (1) age 18-85 years; (2) gender was not limited; (3) met the diagnostic criteria for cirrhosis, regardless of etiology; (4) completed gastroscopy and upper abdominal enhanced CT examination, and the time difference between the two was no more than 3 months.

[0100] Exclusion criteria included: (1) non-cirrhotic portal hypertension; (2) patients who had undergone splenectomy, splenic embolization, or TIPS; (3) patients who had undergone endoscopic treatment; (4) patients with incomplete endoscopic or CT data; (5) patients with primary or metastatic liver cancer; patients with tumors in other parts of the body; (6) patients with uncontrolled serious cardiovascular, respiratory, urinary, neurological, psychiatric, or other systemic diseases; (7) breastfeeding women; and (8) other patients who were considered unsuitable for inclusion by the researchers.

[0101] A total of 391 patients were included in the study, including 257 males and 134 females, with an average age of (54.40±10.71) years. The detailed clinical basic data characteristics of the included patients are shown in Table 1.

[0102] Table 1 Clinical data of patients included in the study

[0103]

[0104]

[0105] a The statistical method was Mann-Whitney U test; b Statistical methods Pearson chi-square test; c The statistical method was Fisher's exact test; d The statistical method was two independent sample t test; # It indicates the intergroup comparison between the no EV group and the mild EV group; ## Indicates comparison between the mild EV group and the moderate to severe EV group; P<0.05 was used as the standard to judge whether there was a statistical difference, and * indicates that the difference was statistically significant.

[0106] The etiology of cirrhosis was classified as follows: alcoholic liver disease in 75 cases (19.18%), hepatitis B in 179 cases (45.78%), hepatitis C in 32 cases (8.18%), hepatitis B combined with hepatitis C in 7 cases (1.79%), autoimmune liver disease in 30 cases (7.67%), non-alcoholic fatty liver disease in 3 cases (0.77%), mixed factors (including hepatitis B combined with alcoholic liver disease, hepatitis C combined with alcoholic liver disease, autoimmune liver disease combined with alcoholic liver disease) in 26 cases (6.65%), and other causes (including cryptogenic, cardiac, and drug-induced cirrhosis) in 39 cases (9.97%). Child-Pugh classification was performed according to the severity of liver function damage: 160 cases (40.92%) were grade A, 152 cases (38.87%) were grade B, and 79 cases (20.20%) were grade C. The MELD score was 10.35±5.47.

[0107] According to the results of gastroscopy, the patients were divided into groups, including 109 patients (27.88%) in the no EV group, 120 patients (30.69%) in the mild EV group, and 162 patients (41.43%) in the moderate to severe EV group.

[0108] Compared with the mild EV group, the PLT, ALB and PTA of patients in the mild EV group were decreased, AST, INR, TB, PT, and MELD scores were increased, and the ascites was more severe. There were no significant differences in age, gender composition, etiology composition, ALT, and Cr between the two groups. There were no statistically significant differences in age, gender, etiology composition, ALB, Cr, INR, PT, PTA, MELD score, and Child-Pugh grade between the mild EV group and the moderate-to-severe EV group. The PLT, ALT, AST, and TB of patients in the mild EV group were higher than those in the moderate-to-severe EV group. The ascites of patients in the moderate-to-severe EV group was more severe than that of patients in the mild EV group, and the difference was statistically significant (P<0.05).

[0109] Example 2 A method for processing radiomics features of upper abdominal enhanced CT in patients with liver cirrhosis

[0110] The process of the method described in this embodiment is as follows Figure 1 As shown, based on the research object described in Example 1, the specific establishment process is:

[0111] S1. Collect enhanced CT images of the upper abdomen of patients with cirrhosis, and upload the portal venous phase CT images with the most obvious enhancement of the liver, spleen, and esophageal veins in DICOM format to the precision medicine open platform (developed by Beijing Imprint Information Technology (Beijing) Co., Ltd.).

[0112] S2. First, the doctor marks the liver, spleen and lower esophagus in the enhanced CT images in DICOM format obtained in step S1 layer by layer. Since the most serious varicose veins mostly occur within 5 cm above the cardia, the marking range of the esophagus is the lower esophagus, that is, the area from the cardia to 5 cm above it. The marked CT images of patients without esophageal varices, CT images of patients with mild esophageal varices, CT images of patients with moderate esophageal varices and CT images of patients with severe esophageal varices are as follows: Figure 2 As shown in AD.

[0113] Then, all calibration data of the marked area are extracted from the enhanced CT image in DICOM format, so that the area is separated from other tissues. The specific operation process is:

[0114] The image data is parsed to extract the marked contour of the area, and then the extracted contour is expanded and filled to obtain a binary image with all 1s in the marked area and all 0s outside the marked area. The binary image has the same shape and size as the enhanced CT image in DICOM format. Finally, the enhanced CT image data in DICOM format and the binary image are multiplied, all calibration data are extracted from the enhanced CT image in DICOM format, and the area is separated from other tissues.

[0115] S3. Using the Marching Cubes (MC) algorithm, perform three-dimensional reconstruction on the regions of interest marked in step S2.

[0116] The basic principle of the MC algorithm is to construct an isosurface in a three-dimensional data field, find the voxel passing through the isosurface, find the isosurface in the voxel and calculate the relevant parameters. The algorithm has a fast calculation speed and the reconstructed image has high quality. The 3D reconstruction of the data can well restore the true shape of the target area, which plays a key role in the training of the radiomics classification model for accurately predicting esophageal varices in patients.

[0117] S4. For each 3D image obtained in step S3, 42-dimensional image texture features, 540-dimensional gradient direction histogram features, 48-dimensional wavelet features, and 156-dimensional statistical features are extracted. Each part has 786-dimensional features, and the three-part features are fused to form a total of 2358 dimensions, which is the original imaging omics feature.

[0118] Texture feature is a global feature that describes the surface properties of the object corresponding to the image area. In pattern matching, the features of this area have great advantages and are more resistant to noise. Statistical features are histograms constructed based on the probability of pixels in the area of ​​interest, and then the histogram features are calculated. Wavelet features are to extract the spatial information of three-dimensional images with different frequency components, mainly extracting high-frequency and low-frequency information of three-dimensional images. The high-frequency information component reflects the detail information of the image, and the low-frequency information reflects the structural information of the image. Since the edge and shape of the three-dimensional image can be well described by the gradient direction density distribution, the gradient direction histogram feature can accurately reflect the edge and shape information of the three-dimensional image. At the same time, the three-dimensional gradient direction histogram decomposes the target into many unit blocks for feature extraction, so it can accurately reflect the local information and spatial gradient information of the image.

[0119] S5. Apply the principal component analysis method to perform dimensionality reduction processing on the original imaging omics features obtained in step S4 to obtain the imaging omics features after dimensionality reduction.

[0120] A large number of variables will make the classification algorithm very complicated, and there may be correlation between the variables. Therefore, in order to reduce overfitting and increase the generalization ability of the classifier, the principal component analysis method is used to reduce the feature dimension of the original imaging features obtained from the image. First, the original imaging features obtained in step S4 are standardized using formula I, and the covariance matrix of the multidimensional features is calculated to obtain the eigenvalues ​​and eigenvectors of the matrix;

[0121] Xi'=(Xi-Xmin) / (Xmax-Xmin) I,

[0122] Among them, Xi' is the image feature of the i-th patient after standardization in this dimension,

[0123] Xi is the image feature of the patient that has not been standardized in this dimension.

[0124] Xmax is the maximum value of this dimension,

[0125] Xmin is the minimum value of this dimension;

[0126] Then sort the eigenvalues ​​from large to small, select the principal component, find its corresponding eigenvector, project the original dimensional data onto the selected eigenvector, and the dimension of the original data feature is reduced.

[0127] When constructing a diagnostic model for mild EV, the original radiomics features were reduced in dimension by principal component analysis, and 163-dimensional features were included in the model. When constructing a diagnostic model for moderate to severe EV, the original radiomics features were reduced in dimension by principal component analysis, and 177-dimensional features were included in the model.

[0128] S6. Based on the omics features after dimension reduction obtained in step S5, a classification algorithm is used to construct an esophageal varices imaging omics feature model.

[0129] The classic classification algorithm SVM (Support vector machine) was used to classify the degree of esophageal varicose veins, and an imaging genomics model for distinguishing the degree of EV was established.

[0130] The basic model of SVM is to find the best separation hyperplane in the feature space so that the interval between positive and negative samples in the training set is maximized. The objective function of the model is as follows:

[0131]

[0132] The goal is to minimize the value of L(ω,b,α) through training and optimize the model. It is a regular term, which is used to reduce the impact of overfitting; Used to construct the algorithm "hyperplane" for classification.

[0133] The number of cases in each group of patients used for model development in this embodiment is unbalanced, the number of cases in the no EV group is less than that in the mild EV group, and the number of cases in the mild EV group is less than that in the moderate to severe EV group. This imbalance may cause the model to reduce the accuracy of classifying a group with a smaller number. To overcome this problem, when the model was trained with the imaging genomics features of mild EV, 109 patients were randomly selected from the mild EV group and matched 1:1 with the no EV group. When the model was trained with the imaging genomics features of moderate to severe EV, 120 patients were randomly selected from the moderate to severe EV group and matched 1:1 with the mild EV group. The patients included in the model were randomly divided into 5 groups, each group of patients was used as a validation set, and the remaining 4 groups were used as training sets. A 5-fold cross-validation optimization model was performed to obtain the optimal error penalty parameter C of 0.8. Traverse all possible kernel functions and select the Sigmoid kernel function with the best classification effect. The kernel function coefficient gamma is "auto", which means that its value is the inverse of the number of sample features.

[0134] Finally, ROC and decision curves were drawn to evaluate the prediction accuracy of the model.

[0135] The ROC curves of the training and validation sets of radiomics for mild EV are shown in Figure 3 As shown in Figure 2, A is the ROC curve of the training set and B is the ROC curve of the validation set. The ROC curves of the training set and validation set of the imaging omics of moderate and severe EV are shown in Figure 2. Figure 4 As shown, A is the ROC curve of the training set and B is the ROC curve of the validation set.

[0136] After training and optimization, for mild EV, the average AUC of each training set was 0.993, the average sensitivity was 1.000, the average specificity was 0.977, the average PPV was 0.978, the average NPV was 1.000, and the average accuracy was 0.989; the average AUC of each validation set was 0.893, the average sensitivity was 0.952, the average specificity was 0.571, the average PPV was 0.690, the average NPV was 0.923, and the average accuracy was 0.762. For moderate and severe EV, the average AUC of each training set was 0.997, the average sensitivity was 0.990, the average specificity was 0.979, the average PPV was 0.979, the average NPV was 0.989, and the average accuracy was 0.984; the average AUC of each validation set was 0.984, the average sensitivity was 0.833, the average specificity was 0.958, the average PP was 0.952, the average NPV was 0.852, and the average accuracy was 0.896.

[0137] The decision curve analysis results of the validation set of the imaging omics model established in this example for discriminating mild EV are as follows: Figure 5Figure A shows the decision curve analysis results of the validation group of radiomics for the discriminant model of severe EV. Figure 5 As shown in Figure B. The x-axis represents the threshold and the y-axis represents the net benefit. The area under the solid line minus the area under the dotted line is the clinical net benefit. If the net benefit > 0, it indicates that the model has clinical practical value within the threshold range.

[0138] The consistency between the radiomics model and gastroscopy is shown in Table 2.

[0139] Table 2 Comparison of consistency between radiomics model and gastroscopy diagnosis

[0140]

[0141] Among them: TP: True Positive, true positive, is determined by the model as a positive sample, and is actually a positive sample;

[0142] FP: False Positive, which is judged as a positive sample by the model, but is actually a negative sample;

[0143] FN: False Negative, which is judged as a negative sample by the model, but is actually a positive sample;

[0144] TN: True Negative, which is judged as a negative sample by the model and is actually a negative sample.

[0145] In the above four groups, because FP and FN are much smaller than TP and TN, that is, there are more correct predictions and fewer wrong predictions, which is consistent with the diagnostic effect of gastroscopy, it can be judged that the model judgment is accurate.

[0146] Comparative Example 1 A method for processing the imaging features of enhanced CT images of the upper abdomen of patients with liver cirrhosis

[0147] The process of the method described in this comparative example is basically as follows Figure 1 As shown, based on the research object described in Example 1, the specific establishment process is basically the same as that of Example 2, except that:

[0148] S2. Label the liver and lower esophagus regions layer by layer in the enhanced CT image in DICOM format obtained in step S1.

[0149] S4. Extract 42-dimensional image texture features, 540-dimensional gradient direction histogram features, 48-dimensional wavelet features, and 156-dimensional statistical features from the two 3D images obtained in step S3. Each part has 786-dimensional features. The three-part features are fused to form a total of 1572 dimensions, which is the original imaging omics feature.

[0150] S5. Apply the principal component analysis method to reduce the dimensionality of the original imaging features obtained in step S4 to obtain 30-dimensional features for model construction.

[0151] S6.5-fold cross validation optimization model, the optimal error penalty parameter C is 0.6. In this comparative example, for mild EV, the average AUC of each training set is 0.835, the average sensitivity is 0.910, the average specificity is 0.945, the average PPV is 0.922, the average NPV is 0.935, and the average accuracy is 0.941; the average AUC of each validation set is 0.802, the average sensitivity is 0.895, the average specificity is 0.532, the average PPV is 0.555, the average NPV is 0.843, and the average accuracy is 0.653. For moderate and severe EV, the average AUC of each training set was 0.897, the average sensitivity was 0.883, the average specificity was 0.825, the average PPV was 0.876, the average NPV was 0.855, and the average accuracy was 0.874; the average AUC of each validation set was 0.914, the average sensitivity was 0.763, the average specificity was 0.882, the average PP was 0.876, the average NPV was 0.769, and the average accuracy was 0.635.

[0152] All the indicators in the model established in this comparative example are lower than those in the model established in Example 2, indicating that the accuracy of the model established in Comparative Example 1 is poor.

Claims

1. A method for processing the radiomics features of the upper abdominal enhanced CT images of patients with liver cirrhosis, comprising the following steps: S1. According to the results of gastroscopy, patients with cirrhosis were divided into groups without esophageal varices, mild esophageal varices, and moderate to severe esophageal varices. Enhanced CT images of the upper abdomen of patients with cirrhosis were collected, and the portal venous phase CT images with the most obvious enhancement of the liver, spleen, and esophageal veins were converted into DICOM format. S2. Labeling the liver, spleen and lower esophagus region in the enhanced CT image in DICOM format obtained in step S1 layer by layer, and separating the region from other tissues; wherein, The lower esophagus refers to the area from the cardia to 5 cm above it; S3. Perform three-dimensional reconstruction on the marked areas in step S2; S4. Extracting radiomic features from each three-dimensional region obtained in step S3, and fusing the radiomic features of the three regions to obtain the original radiomic features of the patient; wherein the radiomic features extracted from each region include: 42-dimensional texture features, 540-dimensional gradient direction histogram features, 48-dimensional wavelet features, and 156-dimensional statistical features, for a total of 786 dimensions; S5. Applying the principal component analysis method to perform dimensionality reduction processing on the original imaging features obtained in step S4 to obtain the imaging features after dimensionality reduction; the specific process of the dimensionality reduction processing is: S5-1. Standardize the original radiomics features obtained in step S4 using formula I. Xi'= (Xi - Xmin) / (Xmax - Xmin) I, Wherein, Xi' is a standardized imaging feature among the original imaging features of the i-th patient, Xi is the original imaging feature of the patient that has not been standardized. Xmax is the maximum value of the original radiomics feature, Xmin is the minimum value of the original imaging omics feature; Then the covariance matrix of the multi-dimensional features is calculated to obtain the eigenvalues ​​and eigenvectors of the matrix; S5-2. Sort the original radiomics feature values ​​obtained in step S4 from large to small, select the principal component, find its corresponding feature vector, project the original radiomics feature onto the selected feature vector, and reduce the dimension of the original radiomics feature to obtain the radiomics feature after dimensionality reduction; S6. Based on the reduced-dimensional imaging features obtained in step S5, the reduced-dimensional imaging features and clinical features are integrated, and the SVM classification method is used to construct imaging classification models for patients with mild esophageal varices and imaging classification models for patients with moderate to severe esophageal varices. When the imaging features of patients with mild esophageal varices are used to train the model, patients are randomly selected from the mild esophageal varices group and matched 1:1 with the group without esophageal varices. When the imaging features of moderate to severe esophageal varices are used to train the model, patients are randomly selected from the moderate to severe esophageal varices group and matched 1:1 with the patients in the mild esophageal varices group. Then, the patients included in the model are randomly divided into 5 groups, and each group of patients is used as a validation set. , and the remaining 4 groups were used as training sets. Based on the optimization theory of machine learning methods, 5-fold cross-validation was performed to optimize the imaging genomics classification model for mild esophageal varices in patients and the imaging genomics classification model for moderate to severe esophageal varices in patients; the imaging genomics classification model for mild esophageal varices in patients and the imaging genomics classification model for moderate to severe esophageal varices in patients were evaluated by discrimination, calibration and clinical practical value; wherein, the discrimination was evaluated by the AUC value of ROC, and the larger the AUC value, the better the discrimination of the model; the calibration was evaluated by the decision curve, and the degree of esophageal varices judged under gastroscopy was used as the gold standard. The higher the degree of consistency between the predicted esophageal varices in patients and the gold standard, the better the calibration of the model.

2. The processing method according to claim 1, characterized in that: The specific operation of step S2 is as follows: first, the doctor marks the liver, spleen and lower esophagus areas in the enhanced CT image in DICOM format layer by layer, analyzes the image data, extracts the marked contour of the area, and then expands and fills the extracted contour to obtain a binary image with all 1s in the marked area and all 0s outside the marked area, and the binary image has the same shape and size as the enhanced CT image in DICOM format; finally, the enhanced CT image data in DICOM format and the binary image are multiplied, all calibration data are extracted from the original enhanced CT image, and the area is separated from other tissues.

3. The processing method according to claim 1, characterized in that In step S3, the three-dimensional reconstruction adopts Marching Cubes algorithm.

4. A device for determining the degree of esophageal varices in patients with liver cirrhosis, the device being based on the method for processing the image genomics features of the upper abdominal enhanced CT image of patients with liver cirrhosis according to any one of claims 1 to 3, the device comprising: (1) Input module The input module is configured to input the following enhanced CT images of the upper abdomen of a patient with liver cirrhosis to be diagnosed: DICOM format portal venous phase CT images of the liver, spleen, and esophageal veins with most prominent enhancement; (2) Esophageal varices degree judgment module The esophageal varices degree judgment module is configured to include a region of interest marking module, a region of interest three-dimensional reconstruction module, an imaging genomics feature extraction module, an imaging genomics feature dimension reduction module, and a patient mild esophageal varices imaging genomics classification model and a patient moderate to severe esophageal varices imaging genomics classification model; The region of interest marking module is configured to mark the liver, spleen and lower esophagus region layer by layer in the enhanced CT image in DICOM format, and separate the region from other tissues; wherein the lower esophagus refers to the region from the cardia to 5 cm above it; The region of interest 3D reconstruction module is configured to perform 3D reconstruction on the marked regions respectively to obtain a 3D image; The radiomics feature extraction module is configured to extract radiomics features from each three-dimensional image region, and fuse the radiomics features of the three regions to obtain the original radiomics features of the patient; wherein the radiomics features extracted from each region include: 42 dimensions of texture features, 540 dimensions of gradient direction histogram features, 48 ​​dimensions of wavelet features, and 156 dimensions of statistical features, for a total of 786 dimensions; The radiomics feature dimensionality reduction module is configured to process the original radiomics features as follows: i. Standardizing the original radiomics features obtained by the radiomics feature extraction module using formula I, Xi'= (Xi - Xmin) / (Xmax - Xmin) I, Wherein, Xi' is a standardized imaging feature among the original imaging features of the i-th patient, Xi is the original imaging feature of the patient that has not been standardized. Xmax is the maximum value of the original radiomics feature, Xmin is the minimum value of the original imaging omics feature; Calculate the covariance matrix of multidimensional features and obtain the eigenvalues ​​and eigenvectors of the matrix; ii. sorting the original radiomics feature values ​​obtained by the radiomics feature extraction module from large to small, selecting the principal component, finding its corresponding feature vector, projecting the original radiomics feature onto the selected feature vector, and reducing the dimension of the original radiomics feature to obtain the reduced-dimensional radiomics feature; The patient's mild esophageal varices imaging classification model and the patient's moderate to severe esophageal varices imaging classification model are established through the following process: The models were trained using the imaging features of mild esophageal varices and the imaging features of moderate to severe esophageal varices, respectively. When the model was trained using the imaging features of patients with mild esophageal varices, patients were randomly selected from the mild esophageal varices group and matched 1:1 with the group without esophageal varices. When the model was trained using the imaging features of moderate to severe esophageal varices, patients were randomly selected from the moderate to severe esophageal varices group and matched 1:1 with the patients in the mild esophageal varices group. The patients included in the model were randomly divided into 5 groups, each group of patients was used as a validation set, and the remaining 4 groups were used as training sets. Based on the optimization theory of machine learning methods, 5-fold cross-validation was performed to optimize the imaging classification model of mild esophageal varices in patients and the imaging classification model of moderate to severe esophageal varices in patients. The imaging classification model of mild esophageal varices in patients and the imaging classification model of moderate to severe esophageal varices in patients were evaluated by discrimination, calibration and clinical practical value. The discrimination is evaluated by the AUC value of ROC; the larger the AUC value, the better the discrimination of the model; The calibration is evaluated by a decision curve, with the degree of esophageal varices determined under gastroscopy as the gold standard. The higher the degree of agreement between the predicted esophageal varices of the patient and the gold standard, the better the calibration of the model. (3) Output module The output module is configured to output the degree of esophageal varices in patients with liver cirrhosis.

5. The device according to claim 4, characterized in that The region of interest marking module is configured to process the enhanced CT image in DICOM format as follows: First, the enhanced CT images of the liver, spleen and lower esophagus in DICOM format, which are marked layer by layer by a doctor, are analyzed to extract the marked contours of the areas. The extracted contours are then expanded and filled to obtain a binary image in which all values ​​are 1 in the marked areas and all values ​​are 0 outside the marked areas. The binary image has the same shape and size as the enhanced CT image in DICOM format. Finally, the original enhanced CT image data and the binary image are multiplied to extract all calibration data from the enhanced CT image in DICOM format, and the areas are separated from other tissues.

6. The device according to claim 4 or 5, characterized in that The apparatus is a computer device.

7. The device according to claim 6, characterized in that The esophageal varices degree determination module is included in the memory of the computer device.

8. The device according to claim 6, characterized in that The computer device also includes a processor capable of running the computer programs of the region of interest three-dimensional reconstruction module, the imaging genomics feature extraction module, the imaging genomics feature dimensionality reduction module, and the patient's mild esophageal varices imaging genomics classification model and the patient's moderate to severe esophageal varices imaging genomics classification model.

9. A method for grouping the degree of esophageal varices in patients with liver cirrhosis using the device according to any one of claims 4 to 8, the method comprising: (1) Obtain enhanced CT images of the upper abdomen of cirrhotic patients to be grouped; (2) inputting the portal venous phase CT images of the liver, spleen and esophageal veins with the most obvious enhancement obtained in step (1) into the input module of the device in DICOM format; (3) using the esophageal varices degree judgment module to judge the degree of esophageal varices in patients with liver cirrhosis, and outputting the result from the output module; the degree of esophageal varices is mild esophageal varices or moderate to severe esophageal varices; (4) Patients with cirrhosis whose output results are mild esophageal varices are identified as a low bleeding risk group who do not need gastroscopy, and patients with cirrhosis whose output results are moderate to severe esophageal varices are identified as a high bleeding risk group who need gastroscopy.

Images