Device for quantifying esophageal and gastric varices based on CT images

Through non-invasive detection technology based on CT images, the problems of frequent gastroscopy and inaccurate replacement examinations are solved, and the accurate quantification of esophageal fundus varicocele and bleeding risk assessment are achieved, which reduces the frequency and risk of examinations and improves patient safety and medical efficiency.

CN115564741BActive Publication Date: 2024-11-26WEST CHINA HOSPITAL SICHUAN UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211269642.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-10-18
Publication Date
2024-11-26
Estimated Expiration
2042-10-18

AI Technical Summary

Technical Problem

In the prior art, gastroscopy frequently performs esophageal fundus varicose detection in patients with cirrhosis with frequent gastrointestinal varicose detection, which has an invasive, high cost and complication risk, and the results of alternative examination methods are not reliable enough to effectively reduce the risk of varicose rupture and bleeding.

Method used

Using a non-invasive device based on CT images, quantifying the degree of varicose veins and bleeding risks through data acquisition, data augmentation, image segmentation, imagingomics analysis and machine learning models, providing physician references to determine whether to perform gastroscopy and reduce the risk of bleeding. Necessary inspection frequency and risks.

Benefits of technology

Non-invasive and accurate esophageal fundus varicose veins detection has been achieved, which has reduced the frequency of examinations and medical expenses of patients, improved the accuracy of bleeding risk assessment, and reduced the occurrence of varicose rupture and bleeding events.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115564741B_ABST
    Figure CN115564741B_ABST
Patent Text Reader

Abstract

The present invention discloses a device for quantifying esophageal varices based on CT images, and relates to the technical field of medical image processing devices. The present invention includes a data acquisition module, a data enhancement module, an image segmentation and extraction module, a vascular region image extraction module, an imaging genomics analysis module, a varicose vein quantification module, and a bleeding risk quantification module. The device of the present invention processes CT images and finally obtains the quantitative value of esophageal varices for reference by doctors. For example, doctors can give patients suggestions for further endoscopic examinations based on the quantitative value of the device and the patient's historical medical records, or give patients suggestions for regular review based on the quantitative value of the device and the patient's historical medical records. Excessive endoscopic examinations can be avoided, reducing the patient's medical expenses, and the additional risks brought by endoscopic examinations can also be avoided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing devices, and more specifically to a device for quantifying esophageal and gastric fundus varices based on CT images. Background Art

[0002] Cirrhosis is the terminal stage of various chronic liver diseases and one of the main causes of death. Varicose veins are one of the most common complications of cirrhosis. Esophageal and gastric varices account for 30-40% of compensated cirrhosis, and further increase to more than 85% in the decompensated stage. When liver function is at Child-Pugh A level, the incidence of esophageal and gastric varices is 40%, and when it is Child-Pugh B / C level, the incidence further increases to more than 70%. In fact, the degree of varicose veins is closely related to the progression of cirrhosis. The risk of bleeding is extremely low in mild and moderate cases, but when it progresses to severe or even combined with red signs, the risk of rupture and bleeding increases sharply. Therefore, the purpose of repeated gastroscopy monitoring is only to detect high-risk bleeding events such as severe varicose veins and to provide targeted pretreatment in time. As the disease progresses, the pressure of the portal vein system gradually increases, the degree of esophageal and gastric fundus varices gradually worsens, and the risk of rupture and bleeding continues to rise. Once rupture and bleeding occur, it can lead to accelerated deterioration of the disease, and severe cases can endanger the patient's life. According to incomplete statistics, about 1 / 3 of patients with esophageal varices will progress to esophageal varicose bleeding, with a mortality rate of up to 15%-25% 6 weeks after acute bleeding and up to 40% 1 year after bleeding. Therefore, dynamic monitoring of the extent of esophageal and gastric varices is to quantify the risk of rupture and bleeding, timely select appropriate intervention measures and treat high-risk varicose veins in advance, which can effectively avoid the occurrence of varicose bleeding events, delay the progression of cirrhosis, and reduce mortality. At present, the most effective measure to quantify the extent of varicose veins and predict the risk of bleeding is still gastroscopy. The guidelines and consensus on the diagnosis and treatment of cirrhosis recommend that once a patient is diagnosed with cirrhosis, it is necessary to undergo a gastroscopy and review it every 1-3 years so that varicose veins with a high risk of bleeding can be identified and treated early.

[0003] However, gastroscopy is an invasive procedure, and the purpose of repeated gastroscopy monitoring is only to detect high-risk bleeding events such as severe varicose veins, and to provide targeted pretreatment in a timely manner. Not only does the patient tolerance is poor, there is a risk of complications such as cardia tear, esophageal and gastric perforation, and bleeding. The invasive procedure itself increases the patient's pain index and induces the risk of varicose vein rupture and bleeding. However, among patients who undergo repeated gastroscopy monitoring, in most cases the patients do not progress to the high-risk stage of bleeding such as severe varicose veins. They do not benefit from it, but have to endure the pain of repeated gastroscopy monitoring, and bear additional examination risks and medical expenses.

[0004] There are many alternatives to gastroscopy in the existing technology, and the results of these tests are used by doctors to determine whether patients need further gastroscopy. For example, the risk of esophageal varicose vein bleeding can be predicted by the hepatic venous pressure gradient (HVPG) test. However, HVPG testing is an invasive interventional procedure that is expensive and requires professional personnel and equipment to carry out. The stability of the results is easily affected by multiple factors such as position, respiration, anesthetic drugs, and intrahepatic shunts. In addition, many scholars have proposed non-invasive methods to construct prediction models for esophageal variceal bleeding based on serum markers, such as FIB-4 (Fibrosis Index Based on Four Factors), Lok score, APRI (Aspartate aminotransferase to platelet count ratio index), AAR (Aspartate aminotransferase toalanineaminotransferaseratio), VITRO score (VonWillebrand factor antigen / platelet ratio, von Willebrand factor antigen level to platelet count ratio), etc.

[0005] However, there are some alternatives to gastroscopy in the above-mentioned prior art, some of which are invasive and some are non-invasive, and the cost of the examination is not lower than that of gastroscopy. Moreover, the examination results of the existing alternatives to gastroscopy cannot provide effective reference indicators for doctors and cannot meet clinical needs. Summary of the invention

[0006] In order to overcome the defects and deficiencies in the above-mentioned prior art, the present invention provides a device for quantifying esophageal varices based on CT images. The purpose of the present invention is to provide a means for non-invasively quantifying esophageal varices, and to quantify the esophageal varices of patients using CT images. The detection results can provide a reference for doctors. Doctors can determine whether the patient needs to undergo a gastroscopy based on the detection results of the device of the present invention to truly determine the condition of esophageal varices. Thereby avoiding invasive examinations for some patients, and avoiding the additional risks and high medical expenses brought about by frequent endoscopic examinations. At the same time, combined with clinical data and bleeding events, the risk of varicose vein bleeding is predicted. Doctors can recommend patients with a higher risk of bleeding to choose a gastroscopy for further confirmation, timely select appropriate intervention measures and treat high-risk varicose veins in advance, which can effectively avoid the occurrence of varicose vein rupture and bleeding events, and reduce the frequency of gastroscopy examinations for patients.

[0007] The device of the present invention is based on the volume quantification of varicose veins in the intramural blood vessel area of ​​the esophageal fundus region, and can also be combined with indicators such as vein volume, vein volume ratio, and maximum vein diameter to comprehensively evaluate and quantify esophageal fundus varices in patients with portal hypertension; and then a comprehensive analysis and judgment is made based on the vein volume, vein volume ratio, maximum vein diameter, and clinical special diagnosis in the patient's historical medical records to quantify the risk of rupture and bleeding of esophageal fundus varices in the patient. The test result value is for the reference of physicians. For example, the physician can give the patient a recommendation for further endoscopic examination based on the quantified value of the device and the patient's historical medical records, or give the patient a recommendation for regular review based on the quantified value of the device and the patient's historical medical records. Excessive endoscopic examinations can be avoided, reducing the patient's medical expenses, and avoiding the additional risks brought by endoscopic examinations.

[0008] In order to solve the above problems existing in the prior art, the present invention is implemented through the following technical solutions.

[0009] The present invention provides a device for quantifying esophageal and gastric fundus varices based on CT images, the device comprising:

[0010] The data acquisition module is used to obtain a complete CT image of the subject from the lower esophagus to the upper abdomen, and transmit the acquired complete CT image to the data enhancement module; obtain the clinical characteristics in the subject's medical record information, and transmit the acquired clinical characteristics to the bleeding risk quantification module;

[0011] The data acquisition module can be a data interface connected to the hospital HIS system, used to receive the CT image of the subject in the HIS system and the patient's medical record information in the HIS system; or the data acquisition module can be directly connected to the CT tomography device to directly obtain the complete CT image of the subject's lower esophagus to the upper abdomen from the CT tomography device;

[0012] The data enhancement module is used to perform data enhancement processing on the complete CT image acquired by the data acquisition module, and transmit the CT image after data enhancement processing to the image segmentation and extraction module;

[0013] The image segmentation and extraction module uses a pre-trained image segmentation model to segment and extract the region of interest in the CT image after data enhancement processing, and segment and extract the liver region image, spleen region image, gastric fundus region image and lower esophageal region image in the CT image of the subject; the image segmentation and extraction module transmits the extracted gastric fundus region image and lower esophageal region image to the vascular region image extraction module, and transmits the extracted liver region image to the imaging omics analysis module;

[0014] The vascular region image extraction module extracts the vascular region in the gastric wall, the gastric wall and other region images from the gastric fundus region image according to the size of the CT value; extracts the vascular region in the esophageal wall, the esophageal wall and other region images from the lower esophageal region image according to the size of the CT value; and transmits the extracted vascular region in the gastric wall, the gastric wall region, the vascular region in the esophageal wall and the esophageal wall region images to the imaging omics analysis module;

[0015] The imaging omics analysis module measures the inner diameter of the outer segment of the portal vein at 1 to 2 cm below the first hepatic portal from the extracted liver region image;

[0016] Calculate the volume of the esophageal wall from the extracted esophageal wall area image ; Calculate the volume of blood vessels in the esophageal wall from the extracted blood vessel area image ; Based on the volume of the esophageal wall and the volume of blood vessels in the esophageal wall Calculate the volume ratio of blood vessels in the esophageal wall ; Take the esophageal region within the set length range, and calculate the volume of the varicose veins in the esophageal wall within this range and its volume proportion;

[0017] Calculate the volume of the stomach wall in the gastric fundus region from the extracted stomach wall region image ; Calculate the volume of blood vessels in the stomach wall from the extracted blood vessel area image in the stomach wall ; According to the volume of the stomach wall and the volume of blood vessels in the stomach wall Calculate the volume ratio of blood vessels in the gastric wall in the gastric fundus area ; Take the gastric fundus region within a set length range, and calculate the volume of varices in the gastric fundus wall cavity within this range and its volume proportion;

[0018] The imaging genomics analysis module calculates the inner diameter of the portal vein and the volume of the esophageal wall. , the volume of blood vessels in the esophageal wall , the volume ratio of blood vessels in the esophageal wall , the volume of the stomach wall , the volume of blood vessels in the stomach wall , the volume ratio of blood vessels in the stomach wall , the volume of varicose veins in the esophageal wall and their volume proportion, and the volume of varicose veins in the gastric fundus wall and their volume proportion constitute an imaging omics feature group, and the imaging omics feature group is transmitted to the varicose vein quantification module and the bleeding risk quantification module;

[0019] The varicose vein quantification module uses the radiomics feature group transmitted by the radiomics analysis module to input into the pre-trained lower esophageal varicose vein quantification classification model and gastric fundus varicose vein quantification classification model for quantification, and outputs the classification quantification result value;

[0020] The bleeding risk quantification module uses the clinical characteristics and imaging feature groups in the subject's medical record information, and inputs them into the pre-trained esophageal varicose vein bleeding risk quantification classification model and gastric fundus varicose vein bleeding risk quantification classification model for quantification, and outputs the bleeding risk quantification result value.

[0021] It should be specially noted that the classification quantification result value and bleeding risk quantification result value output by the device for quantifying esophageal and gastric varices based on CT images of the present invention cannot be used as the result value for doctors to judge the degree of varicose veins in patients. The result value is only used as a reference value to guide doctors to confirm whether it is necessary to perform endoscopic examination on patients. When the quantification result value is large, doctors may recommend patients to undergo endoscopic examination to actually confirm the actual situation of varicose veins in patients.

[0022] The quantification classification model of lower esophageal varicose veins and the quantification classification model of gastric fundus varicose veins in the varicose vein quantification module are both obtained by training based on the SVM classification model.

[0023] Specifically, multi-classification SVM was used to classify the multi-dimensional radiomic features of the lower esophagus and gastric fundus regions respectively; in the lower esophageal varicose vein classification model, the training set data had 4 classes, namely, class 1, class 2, class 3, and class 4;

[0024] A pairwise classification method is used to construct a binary classification SVM between every two categories in the above four categories, that is, 6 binary classification SVMs are constructed; after the image to be detected is input, a voting strategy is used for classification; that is, each binary classification SVM votes for the imaging features in the image to be detected according to its decision function. After all 6 binary classification SVMs have completed the voting, the category with the most votes is the category of the input image to be detected.

[0025] In the above classification process, after the six binary SVMs have completed voting, if there is a tie, the category with the smaller index among the tie categories is selected as the category of the input image to be detected.

[0026] In the data collection module, the medical records of the subjects who have experienced bleeding are collected, where the time of the medical records is limited to one month before the bleeding event to one month after the bleeding event, and the time after receiving varicose vein treatment is excluded from the time range of the medical records.

[0027] The clinical characteristics include, but are not limited to, any one or more combinations of platelets, Child-Pugh liver function classification, coagulation function, liver function, renal function, hemoglobin, cause of cirrhosis, complications, use of anticoagulant / antiplatelet drugs, liver stiffness, portal vein width and spleen stiffness.

[0028] The medical record information also includes historical varicose vein quantification result values. The current quantification result value is compared with the historical quantification result value to determine the change in the quantification result value.

[0029] The device also includes a varicose vein rupture high risk area quantification module, which quantifies the varicose vein rupture high risk area in the CT image with a higher quantification result value according to the quantification result value of the varicose vein quantification module. Specifically, the quantification is performed by the ratio of the blood vessel volume within each varicose vein area in the image to the total volume.

[0030] Specifically, according to the radiomics feature group calculated in the radiomics analysis module, all varicose vein areas in the lower esophageal region images are distinguished, and the ratio of the vascular volume within each varicose vein area to the total volume in all lower esophageal region images is calculated respectively. The varicose vein area with the largest ratio is the high-risk area for varicose vein rupture and bleeding within the lower esophageal region.

[0031] All varicose vein areas in the gastric fundus region image are distinguished, and the ratio of the vascular volume within each varicose vein area to the total volume in all gastric fundus region images is calculated respectively. The varicose vein area with the largest ratio is the high-risk area for varicose vein rupture and bleeding within the gastric fundus region.

[0032] Further preferably, the data acquisition module obtains a complete CT image of the lower esophagus to the upper abdomen of the subject during the portal venous phase.

[0033] More preferably, when the data acquisition module is unable to obtain a complete CT image of the lower esophagus to the upper abdomen of the subject in the portal phase, a complete CT image of the lower esophagus to the upper abdomen of the subject in the venous phase is obtained.

[0034] The complete CT image from the lower esophagus to the upper abdomen at least shows the area of ​​the lower esophagus and the gastric fundus.

[0035] In the vascular region image extraction module, other regions in the gastric fundus region image refer to water, air and gastric fundus content regions; other regions in the lower esophageal region image refer to water, air and esophageal content regions.

[0036] The volume of the esophageal wall It is calculated by multiplying the spacing between esophageal wall voxels and the number of esophageal wall voxels within the set length range;

[0037] The volume of blood vessels in the esophagus It is calculated by multiplying the distance between esophageal wall lumen vascular voxels and the number of esophageal wall vascular voxels within the set length range;

[0038] Volume of the stomach wall in the fundic area It is obtained by multiplying the distance between the voxels of the stomach wall in the stomach fundus region and the number of voxels of the stomach wall in the stomach fundus region within the set length range;

[0039] The volume of blood vessels in the stomach wall in the gastric fundus area It is obtained by multiplying the distance between the vascular voxels in the gastric wall in the gastric fundus region and the number of vascular voxels in the gastric wall in the gastric fundus region within a set length range.

[0040] Compared with the prior art, the beneficial technical effects brought by the present invention are as follows:

[0041] 1. Compared with the prior art, the present invention uses the volume of the esophageal wall, the volume of the blood vessels in the esophageal wall, and the volume ratio of the blood vessels in the esophageal wall as the features of the classification model for esophageal varicose veins to quantify and classify esophageal varicose veins. At the same time, the volume of the gastric wall in the gastric fundus region, the volume of the blood vessels in the gastric wall in the gastric fundus region, and the volume ratio of the blood vessels in the gastric wall in the gastric fundus region are used as the features of the classification model for gastric varicose veins to quantify and classify gastric varicose veins. The obtained classification and quantification results are more accurate, which is conducive to confirming whether further endoscopic examination is needed and reducing the frequency of endoscopic examination. At the same time, the present invention also comprehensively evaluates the bleeding risk in combination with the clinical data characteristics of the subject in the past month, so that the bleeding risk quantification results are more clinically referenced. Patients with higher bleeding risk quantification result values ​​can undergo further endoscopic examinations; patients with relatively low bleeding risk quantification result values ​​can avoid frequent endoscopic examinations and perform non-invasive monitoring.

[0042] 2. The present invention adopts the SVM classification model to train the classification models of varicose veins degree of the lower esophagus and gastric fundus respectively, adopts the pairwise classification method to construct a binary SVM between every two categories, and adopts the voting strategy for classification, thereby ensuring the accuracy of the classification and quantification results, and at the same time improving the data processing efficiency, so as to obtain the classification and quantification results more accurately and quickly, and classify and quantify the varicose veins of the gastric fundus and the lower esophagus.

[0043] 3. The clinical characteristics collected in the present invention include but are not limited to platelets, Child-Pugh liver function classification, coagulation function, liver function, kidney function, hemoglobin, cause of cirrhosis, complications, use of anticoagulant / antiplatelet drugs, liver stiffness, portal vein width and spleen stiffness. The introduction of clinical characteristics enhances the accuracy and reference of bleeding risk assessment classification.

[0044] 4. The present invention also screens key areas of concern by the ratio of the blood vessel volume to the total volume of the varicose vein area. The areas with a larger ratio have a relatively higher risk of rupture. Therefore, the high-risk areas of varicose vein rupture and bleeding can be obtained by the ratio of the blood vessel volume to the total volume in the varicose vein area, and key attention and monitoring can be carried out in the later stage.

[0045] 5. In the present invention, a complete CT image of the lower esophagus to the upper abdomen of the subject is collected during the portal phase or the venous phase, preferably during the portal phase. When there is no portal phase, the venous phase is collected. The CT image of the portal phase is more conducive to the quantification of varicose vein classification and bleeding risk. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0047] Figure 1 It is a schematic diagram of the structure of the device for quantifying esophageal and gastric fundus varices based on CT images of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings of the present invention specification to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Embodiment 1 As a preferred embodiment of the present invention, refer to the attached specification Figure 1 As shown, this embodiment discloses a device for quantifying esophageal and gastric fundus varices based on CT images, the device comprising:

[0050] The data acquisition module is used to obtain a complete CT image of the subject from the lower esophagus to the upper abdomen, and transmit the acquired complete CT image to the data enhancement module; obtain the clinical characteristics in the subject's medical record information, and transmit the acquired clinical characteristics to the bleeding risk quantification module.

[0051] In this embodiment, the data acquisition module may be a data interface connected to the hospital HIS system, and is used to receive the CT image of the subject in the HIS system and the medical record information of the patient in the HIS system.

[0052] As an example, the data acquisition module can be directly connected to the data of the CT tomography device to directly obtain a complete CT image of the lower esophagus to the upper abdomen of the subject from the CT tomography device.

[0053] The data enhancement module is used to perform data enhancement processing on the complete CT image acquired by the data acquisition module, and transmit the CT image after data enhancement processing to the image segmentation and extraction module.

[0054] The image segmentation and extraction module uses a pre-trained image segmentation model to segment and extract the region of interest in the CT image after data enhancement processing, and segments and extracts the liver area image, spleen area image, gastric fundus area image and lower esophageal area image in the CT image of the subject; the image segmentation and extraction module transmits the extracted gastric fundus area image and lower esophageal area image to the vascular area image extraction module, and transmits the extracted liver area image to the imaging omics analysis module; the neural network used to train the above-mentioned image segmentation model can be selected from VGG19FCN, U-net, etc.

[0055] The vascular area image extraction module extracts the vascular area in the gastric wall of the gastric fundus area, the gastric wall and other area images from the gastric fundus area image according to the size of the CT value; for the lower esophageal area image, the vascular area in the esophageal wall, the esophageal wall and other area images are extracted from the lower esophageal area image according to the size of the CT value; and transmits the extracted vascular area in the gastric wall of the gastric fundus area, the gastric wall area, the vascular area in the esophageal wall and the esophageal wall area images to the imaging genomics analysis module.

[0056] The imaging omics analysis module measures the inner diameter of the outer segment of the portal vein at 1 to 2 cm below the first hepatic portal from the extracted liver region image;

[0057] Calculate the volume of the esophageal wall from the extracted esophageal wall area image ; Calculate the volume of blood vessels in the esophageal wall from the extracted blood vessel area image ; Based on the volume of the esophageal wall and the volume of blood vessels in the esophageal wall Calculate the volume ratio of blood vessels in the esophageal wall ; Take the esophageal region within the set length range, and calculate the volume of the varicose veins in the esophageal wall within this range and its volume proportion;

[0058] Calculate the volume of the stomach wall in the gastric fundus region from the extracted stomach wall region image ; Calculate the volume of blood vessels in the stomach wall from the extracted blood vessel area image in the stomach wall ; According to the volume of the stomach wall and the volume of blood vessels in the stomach wall Calculate the volume ratio of blood vessels in the gastric wall in the gastric fundus area ; Take the gastric fundus region within a set length range, and calculate the volume of varices in the gastric fundus wall cavity within this range and its volume proportion;

[0059] The imaging genomics analysis module calculates the inner diameter of the portal vein and the volume of the esophageal wall. , the volume of blood vessels in the esophageal wall , the volume ratio of blood vessels in the esophageal wall , the volume of the stomach wall , the volume of blood vessels in the stomach wall , the volume ratio of blood vessels in the stomach wall The volume of varicose veins in the esophageal wall and their volume proportion, and the volume of varicose veins in the gastric fundus wall and their volume proportion constitute an imaging feature group, and the imaging feature group is transmitted to the varicose vein quantification module and the bleeding risk quantification module.

[0060] The varicose vein quantification module uses the radiomics feature group transmitted by the radiomics analysis module to input the pre-trained lower esophageal varicose vein quantification classification model and gastric fundus varicose vein quantification classification model for quantification, and outputs the classification quantification result value; the classification quantification result value includes one category, two categories, three categories and four categories. The larger the number of categories, the higher the varicose vein quantification value, and the higher the possibility of varicose veins. The specific setting of the quantification result value can be defined according to the usage habits, such as category I, category II, category III and category IV.

[0061] The bleeding risk quantification module uses the clinical characteristics and imaging feature groups in the patient's medical records to input the pre-trained classification models for quantification of bleeding risk of varicose vein rupture in the lower esophagus and varicose vein rupture in the gastric fundus for quantification, and outputs the bleeding risk quantification result value. The bleeding risk quantification result value can be classified into Class I, Class II and Class III. The higher the category, the higher the risk.

[0062] After using this device to quantify the CT image of the subject, and combining it with the historical medical information of the subject, if the medical record information contains the quantification result value of the last CT scan, the two quantification result values ​​are compared. If the quantification result value changes from Class II to Class III, that is, the quantification result value is significantly increased, the subject can be recommended to further confirm the esophageal varices through endoscopic examination; if there is no change or it decreases, it can be continuously monitored and endoscopic examination is not recommended.

[0063] Example 2

[0064] As another preferred embodiment of the present invention, refer to the attached specification Figure 1As shown, this embodiment is a supplement to the specific implementation of the bleeding risk quantification module in the above embodiment 1. In this embodiment, the specific processing method of the bleeding risk quantification module is:

[0065] For cases with varicose vein bleeding, the CT images before bleeding occurred were extracted as positive data, and the bleeding risk was set to 1; the normal CT images were taken as negative data, and the bleeding risk was set to 0.

[0066] The region of interest is segmented using a pre-trained image segmentation model, and then the original CT images of the region of interest and the surrounding cube area are correspondingly transmitted to the lower esophageal varicose vein bleeding risk quantification classification model and the gastric fundus varicose vein bleeding risk quantification classification model for quantification, and the bleeding risk quantification result value is output.

[0067] Use deep learning classification to classify and quantify the risk of varicose vein bleeding in the lower esophagus and gastric fundus (the training method of the quantification classification model for the risk of varicose vein bleeding in the lower esophagus is the same as that of the risk of varicose vein bleeding in the gastric fundus). The specific steps are as follows:

[0068] Preprocessing operations (image data enhancement processing) were performed on the CT images, i.e., the pixel spacing (spacing) of the CT images was resampled (resampling) to be consistent, and the images were cropped to patches of a fixed size. The window width and window position were transformed, normalized, and zero-valued centered. Data enhancement was performed using rotation (90 degrees), flipping, and transposition. The preprocessed CT images were divided into a lower esophageal image set and a gastric fundus image set according to the esophagus and gastric fundus regions. The lower esophageal image set was used to train a quantitative classification model for the risk of varicose vein rupture and bleeding in the lower esophagus, and the gastric fundus image set was used to train a quantitative classification model for the risk of varicose vein rupture and bleeding in the gastric fundus. The data in the lower esophageal image set and the gastric fundus image set were divided into a training set and a test set in a ratio of 8:2.

[0069] Use 3D convolution to build a classification model (a quantitative classification model for the risk of varicose vein rupture and bleeding in the lower esophagus and a quantitative classification model for the risk of varicose vein rupture and bleeding in the gastric fundus are built in the same way). After deep learning of the training set data preprocessed by the classification model, the risk probability of varicose vein bleeding in the lower esophagus and gastric fundus can be output.

[0070] Set the loss function and optimization method to optimize the model. The loss function is set to the cross entropy loss function, and the optimization method uses the Stochastic Gradient Descent (SGD) optimizer.

[0071] The model effect was evaluated by using the test set data to evaluate the performance of the varicose vein bleeding risk quantitative classification model, that is, the predicted results were compared with the true category labels, and the calculated indexes included accuracy, precision, and recall. The higher the index, the better the model training effect.

[0072] Example 3

[0073] As another preferred embodiment of the present invention, refer to the attached specification Figure 1 As shown, this embodiment is a detailed supplement to another specific implementation of the bleeding risk quantification module in the above embodiment 1. In this embodiment, the processing process of the bleeding risk quantification module is specifically as follows:

[0074] 1. CT images: For cases with varicose vein bleeding, the CT images before bleeding were extracted as positive data, and the bleeding risk was set to 1; the normal CT images were taken as negative data, and the bleeding risk was set to 0. 2. Clinical characteristics: smoking, drinking, blood pressure, anticoagulation / antiplatelet, ALT, AST, TBIL, DBIL, ALB, PT, INR, ascites, platelets, hemoglobin, blood ammonia, liver function grade, cause of cirrhosis, liver stiffness (kPa), HVPG, complications, etc.

[0075] Use deep learning classification to classify and quantify the risk of varicose vein bleeding in the lower esophagus and gastric fundus (the training method of the quantification classification model for the risk of varicose vein bleeding in the lower esophagus is the same as that of the risk of varicose vein bleeding in the gastric fundus). The specific steps are as follows:

[0076] The CT images are preprocessed (image data enhancement), that is, the pixel spacing (spacing) of the CT images is resampled (resampling) to be consistent, cropped to a patch of fixed size, and the window width and window position are transformed, normalized, and zero-centered. The data is enhanced by rotation (90 degrees), flipping, and transposition. The preprocessed CT images are divided into the lower esophageal image set and the gastric fundus image set according to the esophagus and gastric fundus regions. The lower esophageal image set is used to train the quantitative classification model for the risk of varicose vein rupture and bleeding in the lower esophagus, and the gastric fundus image set is used to train the quantitative classification model for the risk of varicose vein rupture and bleeding in the gastric fundus. The data in the lower esophageal image set and the gastric fundus image set are divided into the training set and the test set in a ratio of 8:2. Each clinical feature is quantitatively represented and normalized to form a 20-dimensional medical indicator feature vector;

[0077] Use 3D convolution to build a classification model (a quantitative classification model for the risk of varicose vein rupture and bleeding in the lower esophagus and a quantitative classification model for the risk of varicose vein rupture and bleeding in the gastric fundus are built respectively, and the two are built in the same way). After the classification model has deep learning preprocessed the training set data, if some clinical features (such as liver stiffness, etc.) are found to have a significant impact on the risk of bleeding in the study of bleeding risk prediction, then on the basis of the above classification model, the gradient boosting decision tree (GBDT) is further used to combine the image with the clinical features. The features (64 dimensions) and clinical feature vectors (20 dimensions) finally obtained by the classification model are spliced ​​into an 84-dimensional feature vector, and the labels are the corresponding categories. Input the decision tree model for training, and finally the risk probability of varicose vein rupture and bleeding in the lower esophagus and gastric fundus can be output;

[0078] Set the loss function and optimization method to optimize the model. The loss function is set to the cross entropy loss function, and the optimization method uses the Stochastic Gradient Descent (SGD) optimizer.

[0079] The model effect was evaluated and the performance of the bleeding risk classification model was evaluated using the test set data. The predicted results were compared with the true category labels and the calculated indexes included accuracy, precision and recall. The higher the index, the better the model training effect.

[0080] Example 4

[0081] As another preferred embodiment of the present invention, refer to the attached specification Figure 1 As shown, this embodiment is a detailed supplement to the specific implementation method of the varicose vein quantification module in the above-mentioned embodiment 1. In this embodiment, the neural network model used to train the quantification classification model of the lower esophageal varicose veins and the quantification classification model of the gastric fundus varicose veins can be Inception V3, ResNet50, LeNet-5, SVM, 3D convolution, etc.

[0082] As a preferred implementation in this embodiment, the SVM model is preferably used to train the neural network model of the quantitative classification model of esophageal varicose veins and the quantitative classification model of gastric fundus varicose veins. The specific training process is:

[0083] Collect a large number of (more than 500) enhanced CT images of the lower esophagus to the upper abdomen of patients with liver cirrhosis, and use the image segmentation model to segment and extract the region of interest, namely the lower esophagus area and the gastric fundus area;

[0084] The extracted images of the lower esophagus region and the gastric fundus region are stored and analyzed respectively;

[0085] Then, for the lower esophageal region image, the volume of the esophageal wall is extracted , the volume of blood vessels in the esophageal wall , the volume ratio of blood vessels in the esophageal wall Features: As the original data set for training the quantitative classification model of lower esophageal varicose veins, the original data set is divided into a training set and a test set in a ratio of 8:2. The training set is used to train the quantitative classification model of lower esophageal varicose veins based on the SVM model, and then the test set is used to test the trained quantitative classification model of lower esophageal varicose veins, and the model parameters are optimized.

[0086] Similarly, for the extracted gastric fundus area image, the volume of the gastric wall in the gastric fundus area is extracted. , the volume of blood vessels in the stomach wall in the gastric fundus area , the volume ratio of blood vessels in the gastric wall in the gastric fundus area Features; As the original data set for training the quantitative classification model of gastric varices, the original data set is divided into a training set and a test set in a ratio of 8:2. The training set is used to train the quantitative classification model of gastric varices based on the SVM model, and then the test set is used to test the trained quantitative classification model of gastric varices, and the model parameters are optimized.

[0087] As another preferred implementation mode of the present embodiment, the present implementation mode specifically uses a multi-classification SVM to classify the multi-dimensional imaging genomics features of the lower esophagus and the gastric fundus area respectively; in the classification model of varicose veins of the lower esophagus, the training set data has a total of 4 categories, namely, category one, category two, category three and category four; a pairwise classification method is used to construct a binary classification SVM between every two categories in the above four categories, that is, 6 binary classification SVMs are constructed; after the image to be detected is input, a voting strategy is used for classification; that is, each binary classification SVM performs category voting on the imaging genomics features in the image to be detected according to its decision function, and after the 6 binary classification SVMs have completed the voting, the category with the most votes is the category of the input image to be detected.

[0088] In the above classification process, after the six binary SVMs have completed voting, if there is a tie, the category with the smaller index among the tie categories is selected as the category of the input image to be detected.

[0089] Example 5

[0090] As another preferred embodiment of the present invention, refer to the attached specification Figure 1 As shown, this embodiment is a detailed supplement to another specific implementation of the varicose vein quantification module in the above embodiment 1. In this embodiment, the neural network model used to train the quantification classification model of the lower esophageal varicose veins and the quantification classification model of the gastric fundus varicose veins is 3D convolution. The specific training process is as follows:

[0091] Preprocessing operations (image data enhancement processing) were performed on the CT images, i.e., the pixel spacing (spacing) of the CT images was resampled (resampling) to be consistent, and the images were cropped to patches of a fixed size. The window width and window position were transformed, normalized, and zero-valued centered. Data enhancement was performed using rotation (90 degrees), flipping, and transposition. The preprocessed CT images were divided into a lower esophageal image set and a gastric fundus image set according to the esophagus and gastric fundus regions. The lower esophageal image set was used to train a quantitative classification model for lower esophageal varicose veins, and the gastric fundus image set was used to train a quantitative classification model for gastric fundus varicose veins. The data in the lower esophageal image set and the gastric fundus image set were divided into a training set and a test set in a ratio of 8:2.

[0092] A classification model is constructed using 3D convolution. The classification model uses deep learning to preprocess the training set data, and can ultimately output the probability of four types of quantitative results for varicose veins in the lower esophagus and gastric fundus.

[0093] Set the loss function and optimization method to optimize the model. The loss function is set to the cross entropy loss function, and the optimization method is to use the Stochastic Gradient Descent (SGD) optimizer.

[0094] The model effect was evaluated by using the test set data to evaluate the performance of the varicose vein quantification classification model, that is, the predicted results were compared with the true category labels, and the calculation indexes included accuracy, precision, and recall. The higher the index, the better the model training effect.

[0095] Example 6

[0096] As another preferred embodiment of the present invention, refer to the attached specification Figure 1 As shown, this embodiment is a further supplement to the specific implementation method of the bleeding risk quantification module based on the above-mentioned embodiment 1. In this embodiment, the medical record information of the subject is collected, and the clinical characteristics and imaging features in the medical record information of the subject are combined to construct a quantification classification model for the risk of varicose vein rupture and bleeding in the lower esophagus and a quantification classification model for the risk of varicose vein rupture and bleeding in the gastric fundus, respectively. The clinical characteristics of the subject and the gastric fundus area image obtained by segmentation and extraction from the current CT image are input into the quantification classification model for the risk of varicose vein rupture and bleeding in the gastric fundus, and the clinical characteristics of the subject and the lower esophageal area image obtained by segmentation and extraction from the current CT image are input into the quantification classification model for the risk of varicose vein rupture and bleeding in the lower esophagus, and the risk of rupture and bleeding of varicose veins in the gastric fundus and esophageal of the subject is classified and quantified.

[0097] When the subject has a varicose vein bleeding event, the medical record information of the subject with bleeding is collected. The time range of the medical record information is limited to one month before the bleeding event to one month after the bleeding event, and the time after receiving varicose vein treatment is excluded from the time range of the medical record information.

[0098] The clinical characteristics in the medical record information include, but are not limited to, any one or more combinations of platelets, Child-Pugh liver function class, coagulation function, liver function, renal function, hemoglobin, cause of cirrhosis, comorbidities, use of anticoagulants / antiplatelets, liver stiffness, portal vein width, and spleen stiffness.

[0099] Further, such as Figure 1 As shown, the method further includes a varicose vein rupture high risk area quantification module, and according to the quantification result value of the varicose vein quantification module, the varicose vein rupture high risk area in the CT image with a higher quantification result value is quantified. Specifically, the quantification is performed by the ratio of the blood vessel volume within each varicose vein area in the image to the total volume.

[0100] According to the radiomics feature group calculated in the radiomics analysis module, all varicose vein areas in the lower esophageal region images are distinguished, and the ratio of the vascular volume within each varicose vein area to the total volume in all lower esophageal region images is calculated. The varicose vein area with the largest ratio is the high-risk area for varicose vein rupture and bleeding within the lower esophageal region.

[0101] All varicose vein areas in the gastric fundus region image are distinguished, and the ratio of the vascular volume within each varicose vein area to the total volume in all gastric fundus region images is calculated respectively. The varicose vein area with the largest ratio is the high-risk area for varicose vein rupture and bleeding within the gastric fundus region.

[0102] Example 7

[0103] As another preferred embodiment of the present invention, this embodiment discloses a device for quantifying esophageal and gastric fundus varices based on CT images, the device comprising:

[0104] The data acquisition module is used to obtain a complete CT image of the subject from the lower esophagus to the upper abdomen, and transmit the acquired complete CT image to the data enhancement module; obtain the clinical characteristics in the subject's medical record information, and transmit the acquired clinical characteristics to the bleeding risk quantification module;

[0105] As an example, the data acquisition module may be a data interface connected to a hospital HIS system, for receiving CT images of the subject in the HIS system and medical records of the patient in the HIS system;

[0106] As another example, the data acquisition module is directly connected to the CT tomography device to obtain a complete CT image of the lower esophagus to the upper abdomen of the subject directly from the CT tomography device;

[0107] It is preferred to obtain a complete CT image of the lower esophagus to the upper abdomen of the subject in the portal phase. When the data acquisition module cannot obtain a complete CT image of the lower esophagus to the upper abdomen of the subject in the portal phase, a complete CT image of the lower esophagus to the upper abdomen of the subject in the venous phase is obtained. In the complete CT image of the lower esophagus to the upper abdomen, at least the lower esophagus and the gastric fundus are displayed.

[0108] The data enhancement module is used to perform data enhancement processing on the complete CT image acquired by the data acquisition module, and transmit the CT image after data enhancement processing to the image segmentation and extraction module;

[0109] As an example, the specific operation of the data enhancement module for CT images is to resample (resampling) the pixel spacing (spacing) of the CT images to be consistent, crop them to patches of fixed size, transform the window width and window position, normalize and zero-center them, and use rotation (90 degrees), flipping and transposition to perform data enhancement.

[0110] The image segmentation and extraction module uses a pre-trained image segmentation model to segment and extract the region of interest in the CT image after data enhancement processing, and segment and extract the liver region image, spleen region image, gastric fundus region image and lower esophageal region image in the CT image of the subject; the image segmentation and extraction module transmits the extracted gastric fundus region image and lower esophageal region image to the vascular region image extraction module, and transmits the extracted liver region image to the imaging omics analysis module;

[0111] The neural network used to train the above-mentioned image segmentation model can be VGG19FCN, U-net, etc.

[0112] The vascular region image extraction module extracts the vascular region in the gastric wall, the gastric wall and other region images from the gastric fundus region image according to the size of the CT value; extracts the vascular region in the esophageal wall, the esophageal wall and other region images from the lower esophageal region image according to the size of the CT value; and transmits the extracted vascular region in the gastric wall, the gastric wall region, the vascular region in the esophageal wall and the esophageal wall region images to the imaging omics analysis module;

[0113] As an example, the vascular region image extraction module determines which areas belong to the vascular region, which areas belong to the gastric fundus inner wall region, which areas belong to the esophageal inner wall region, and which areas belong to water, air, esophageal containers or gastric contents based on the CT values.

[0114] The imaging omics analysis module measures the inner diameter of the outer segment of the portal vein at 1 to 2 cm below the first hepatic portal from the extracted liver region image;

[0115] Calculate the volume of the esophageal wall from the extracted esophageal wall area image ; Calculate the volume of blood vessels in the esophageal wall from the extracted blood vessel area image ; Based on the volume of the esophageal wall and the volume of blood vessels in the esophageal wall Calculate the volume ratio of blood vessels in the esophageal wall ; Take the esophageal region within the set length range, and calculate the volume of the varicose veins in the esophageal wall within this range and its volume proportion;

[0116] Calculate the volume of the stomach wall in the gastric fundus region from the extracted stomach wall region image ; Calculate the volume of blood vessels in the stomach wall from the extracted blood vessel area image in the stomach wall ; According to the volume of the stomach wall and the volume of blood vessels in the stomach wall Calculate the volume ratio of blood vessels in the gastric wall in the gastric fundus area ; Take the gastric fundus area within a set length range and calculate the volume of varices in the gastric fundus wall cavity within this range and its volume proportion.

[0117] The imaging genomics analysis module calculates the inner diameter of the portal vein and the volume of the esophageal wall. , the volume of blood vessels in the esophageal wall , the volume ratio of blood vessels in the esophageal wall , the volume of the stomach wall , the volume of blood vessels in the stomach wall , the volume ratio of blood vessels in the stomach wall The volume of varicose veins in the esophageal wall and their volume proportion, and the volume of varicose veins in the gastric fundus wall and their volume proportion constitute an imaging feature group, and the imaging feature group is transmitted to the varicose vein quantification module and the bleeding risk quantification module.

[0118] The volume of the esophageal wall It is calculated by multiplying the spacing between esophageal wall voxels and the number of esophageal wall voxels within the set length range;

[0119] The volume of blood vessels in the esophagus It is calculated by multiplying the distance between esophageal wall lumen vascular voxels and the number of esophageal wall vascular voxels within the set length range;

[0120] Volume of the stomach wall in the fundic area It is obtained by multiplying the distance between the voxels of the stomach wall in the stomach fundus region and the number of voxels of the stomach wall in the stomach fundus region within the set length range;

[0121] The volume of blood vessels in the stomach wall in the gastric fundus area It is obtained by multiplying the distance between the vascular voxels in the gastric wall in the gastric fundus region and the number of vascular voxels in the gastric wall in the gastric fundus region within a set length range.

[0122] The varicose vein quantification module uses the radiomics feature group transmitted by the radiomics analysis module to input into the pre-trained lower esophageal varicose vein quantification classification model and gastric fundus varicose vein quantification classification model for quantification, and outputs the classification quantification result value;

[0123] As an example, a multi-classification SVM is used to classify the multi-dimensional radiomic features of the lower esophagus and gastric fundus. In the classification model of lower esophageal varicose veins, the training set data has 4 classes, namely, class 1, class 2, class 3, and class 4.

[0124] A pairwise classification method is used to construct a binary SVM between every two categories in the above four categories, that is, 6 binary SVMs are constructed; after the image to be detected is input, a voting strategy is used for classification; that is, each binary SVM votes for the imaging features in the image to be detected according to its decision function. After all 6 binary SVMs have completed the voting, the category with the most votes is the category of the input image to be detected;

[0125] In the above classification process, after the six binary SVMs have completed voting, if there is a tie, the category with the smaller index among the tie categories is selected as the category of the input image to be detected.

[0126] The bleeding risk quantification module uses the clinical characteristics and imaging feature groups in the patient's medical records to input into the pre-trained esophageal varicose vein bleeding risk quantification classification model and gastric varicose vein bleeding risk quantification classification model for quantification, and outputs the bleeding risk quantification result value;

[0127] As an example, in the data collection module, medical records of subjects who have experienced bleeding are collected, wherein the time of the medical records is limited to one month before the bleeding event and one month after the bleeding event, and the time after receiving varicose vein treatment is excluded from the time range of the medical records;

[0128] The clinical characteristics include, but are not limited to, any one or more combinations of platelets, Child-Pugh liver function classification, coagulation function, liver function, renal function, hemoglobin, cause of cirrhosis, complications, use of anticoagulant / antiplatelet drugs, liver stiffness, portal vein width and spleen stiffness.

[0129] Furthermore, the medical record information also includes historical varicose vein quantification result values, and the current quantification result value is compared with the historical quantification result value to determine the change in the quantification result value.

[0130] The device also includes a varicose vein rupture high risk area quantification module, which quantifies the varicose vein rupture high risk area in the CT image with a higher quantification result value according to the quantification result value of the varicose vein quantification module. Specifically, the quantification is performed by the ratio of the blood vessel volume within each varicose vein area in the image to the total volume.

[0131] Specifically, according to the radiomics feature group calculated in the radiomics analysis module, all varicose vein areas in the lower esophageal region images are distinguished, and the ratio of the vascular volume within each varicose vein area to the total volume in all lower esophageal region images is calculated respectively. The varicose vein area with the largest ratio is the high-risk area for varicose vein rupture and bleeding within the lower esophageal region.

[0132] All varicose vein areas in the gastric fundus region image are distinguished, and the ratio of the vascular volume within each varicose vein area to the total volume in all gastric fundus region images is calculated respectively. The varicose vein area with the largest ratio is the high-risk area for varicose vein rupture and bleeding within the gastric fundus region.

Claims

1. A device for quantifying esophageal and gastric varices based on CT images, characterized in that: The device includes: The data acquisition module is used to obtain a complete CT image of the subject from the lower esophagus to the upper abdomen, and transmit the acquired complete CT image to the data enhancement module; obtain clinical characteristics from the subject's medical record information, and transmit the acquired clinical characteristics to the bleeding risk quantification module; The data enhancement module is used to perform data enhancement processing on the complete CT image acquired by the data acquisition module, and transmit the CT image after data enhancement processing to the image segmentation and extraction module; The image segmentation and extraction module uses a pre-trained image segmentation model to segment and extract regions of interest from the data-enhanced CT image, segmenting and extracting the liver region image, spleen region image, gastric fundus region image, and lower esophageal region image from the subject's CT image. The image segmentation and extraction module transmits the extracted gastric fundus region image and lower esophageal region image to the vascular region image extraction module, and transmits the extracted liver region image to the imaging omics analysis module. The vascular region image extraction module extracts the vascular region in the gastric wall, the gastric wall, and other regions from the gastric fundus region image based on the size of the CT value; and extracts the vascular region in the esophageal wall, the esophageal wall, and other regions from the lower esophageal region image based on the size of the CT value; and transmits the extracted images of the vascular region in the gastric wall, the gastric wall region, the vascular region in the esophageal wall, and the esophageal wall region to the imaging omics analysis module; The imaging omics analysis module measures the inner diameter of the external segment of the portal vein at a point 1 to 2 cm below the first hepatic portal in the extracted liver region image. Calculate the volume of the esophageal wall from the extracted esophageal wall area image ; Calculate the volume of blood vessels in the esophageal wall from the extracted blood vessel area image ; According to the volume of the esophageal wall and the volume of blood vessels within the esophageal wall Calculate the volume ratio of blood vessels in the esophageal wall Take the esophageal region within a set length range and calculate the volume of varicose veins in the esophageal wall within this range and their volume ratio; Calculate the volume of the stomach fundus from the extracted stomach wall area image ; Calculate the volume of blood vessels in the stomach wall from the extracted blood vessel area image ; According to the volume of the stomach wall and the volume of blood vessels in the stomach wall Calculate the volume ratio of blood vessels in the gastric wall in the gastric fundus area ; Take the gastric fundus region within a set length range and calculate the volume of varices in the gastric fundus wall cavity within this range and its volume ratio; The imaging omics analysis module calculates the inner diameter of the portal vein and the volume of the esophageal wall , the volume of blood vessels in the esophageal wall , the volume ratio of blood vessels in the esophageal wall , the volume of the stomach wall , the volume of blood vessels in the stomach wall , the volume ratio of blood vessels in the gastric wall , the volume of varicose veins in the esophageal wall and their volume ratio, and the volume of varicose veins in the gastric fundus wall and their volume ratio constitute an imaging omics feature group, and the imaging omics feature group is transmitted to the varicose vein quantification module and the bleeding risk quantification module; The varicose vein quantification module uses the radiomics feature set transmitted by the radiomics analysis module and inputs it into the pre-trained lower esophageal varicose vein quantification classification model and gastric fundus varicose vein quantification classification model for quantification, and outputs the classification and quantification result value; The bleeding risk quantification module uses the clinical characteristics and imaging feature groups in the subject's medical records to input them into the pre-trained esophageal varicose vein bleeding risk quantification classification model and gastric varicose vein bleeding risk quantification classification model for quantification, and outputs the bleeding risk quantification result value.

2. The device for quantifying esophageal and gastric varices based on CT images according to claim 1, characterized in that: In the bleeding risk quantification module, the pre-trained classification models for esophageal varicose vein bleeding risk and gastric varicose vein bleeding risk were trained in the following way: For cases with varicose vein bleeding, CT images before bleeding were extracted as positive data, with the bleeding risk set to 1, and normal CT images were used as negative data, with the bleeding risk set to 0. The image segmentation and extraction module was used to segment the regions of interest from the positive and negative CT images, dividing them into a lower esophageal image set and a gastric fundus image set. A quantitative classification model for the risk of lower esophageal varicose vein bleeding was trained using the lower esophageal image set, and a quantitative classification model for the risk of gastric fundus varicose vein bleeding was trained using the gastric fundus image set. The specific training process is: The CT images were preprocessed by resampling the pixel spacing to a uniform value, cropping them into fixed-size patches, transforming the window width and window position, normalizing them, and performing zero-centering operations. Data augmentation was performed using rotation, flipping, and transposition. The data from the lower esophageal image set and the gastric fundus image set were divided into training and test sets in an 8:2 ratio. 3D convolution is used to build a quantitative classification model for varicose vein bleeding risk. The quantitative classification model uses deep learning pre-processed lower esophageal training set data and gastric fundus training set data to output the risk probability of varicose vein bleeding in the lower esophagus and gastric fundus, respectively. Set the loss function and optimization method to optimize the model. The loss function is set to the cross entropy loss function, and the optimization method uses the stochastic gradient descent optimizer. The model effect was evaluated by using the test set data to evaluate the performance of the varicose vein bleeding risk quantitative classification model. The predicted results were compared with the true category labels, and the calculated indices included accuracy, precision, and recall. The higher the index, the better the model training effect.

3. The device for quantifying esophageal and gastric varices based on CT images according to claim 1, characterized in that: In the bleeding risk quantification module, the pre-trained classification models for esophageal varicose vein bleeding risk and gastric varicose vein bleeding risk were trained in the following way: For cases with varicose vein bleeding, the CT images before the bleeding occurred were extracted as positive data, and the bleeding risk was set to 1; Normal CT images were used as negative data, and their bleeding risk was set to 0. Clinical characteristics in the medical records included platelets, Child-Pugh liver function class, coagulation function, liver function, renal function, hemoglobin, cause of cirrhosis, comorbidities, use of anticoagulants / antiplatelets, liver stiffness, portal vein width, and spleen stiffness; The image segmentation and extraction module is used to segment the regions of interest from the positive and negative CT images, dividing them into a lower esophageal image set and a gastric fundus image set. The lower esophageal image set is used to train a quantitative classification model for the risk of lower esophageal varicose bleeding, and the gastric fundus image set is used to train a quantitative classification model for the risk of gastric varicose bleeding. The specific training process is: CT images were preprocessed by resampling the pixel spacing to a consistent level and cropping them into fixed-size patches. Window width and window position were transformed, normalized, and zero-centered. Data augmentation was performed using rotation, flipping, and transposition. The data from the lower esophageal and gastric fundus image sets were divided into training and test sets in an 8:2 ratio. Clinical features were quantitatively represented and normalized to form a 20-dimensional medical indicator feature vector. A 3D convolution algorithm was used to construct a quantitative classification model for varicose vein bleeding risk. Based on this model, a gradient boosting decision tree was used to combine CT images with clinical features in the clinical feature vector. The varicose vein bleeding risk quantitative classification model ultimately generated an 84-dimensional feature vector composed of 64-dimensional image features and 20-dimensional clinical feature vectors. The labels were assigned to the corresponding categories and fed into a decision tree model for training. The final output was the risk probability of varicose vein bleeding in the lower esophagus and gastric fundus. Set the loss function and optimization method to optimize the model. The loss function is set to the cross entropy loss function, and the optimization method uses the stochastic gradient descent optimizer. The model effect was evaluated by using the test set data to evaluate the performance of the varicose vein bleeding risk quantitative classification model. The predicted results were compared with the true category labels, and the calculated indices included accuracy, precision, and recall. The higher the index, the better the model training effect.

4. The device for quantifying esophageal and gastric varices based on CT images according to any one of claims 1 to 3, characterized in that: The varices quantification classification model for the lower esophageal segment and the varices quantification classification model for the gastric fundus in the varices quantification module are both obtained through training based on the SVM classification model.

5. The device for quantifying esophageal and gastric varices based on CT images according to claim 4, characterized in that: Multi-classification SVM was used to classify the multidimensional radiomic features of the lower esophagus and gastric fundus. In the lower esophageal varicose vein classification model, the training set data had four categories: class 1, class 2, class 3, and class 4. A pairwise classification method is used to construct a binary classification SVM between every two categories in the above four categories, that is, 6 binary classification SVMs are constructed; after the image to be detected is input, a voting strategy is adopted for classification; that is, each binary classification SVM votes on the imaging omics features in the image to be detected according to its decision function. After all 6 binary classification SVMs complete the voting, the category with the most votes is the category of the input image to be detected.

6. The device for quantifying esophageal and gastric varices based on CT images according to claim 5, characterized in that: After all six binary SVMs have completed voting, if there is a tie, the category with the smaller index among the tied categories is selected as the category of the input image to be detected.

7. The device for quantifying esophageal and gastric varices based on CT images according to any one of claims 1 to 3, characterized in that: In the data collection module, the medical records of the subjects who have experienced bleeding are collected, where the time range of the medical records is limited to one month before the bleeding event to one month after the bleeding event, and the time after receiving varicose vein treatment is excluded from the time range of the medical records.

8. The device for quantifying esophageal and gastric varices based on CT images according to claim 7, characterized in that: The clinical characteristics include any one or more combinations of platelets, Child-Pugh liver function test, coagulation function, liver function, renal function, hemoglobin, cause of cirrhosis, complications, use of anticoagulant / antiplatelet drugs, liver stiffness, portal vein width and spleen stiffness.

9. The device for quantifying esophageal and gastric varices based on CT images according to claim 1, characterized in that: The device also includes a module for quantifying high-risk areas for varicose vein rupture and bleeding, which quantifies high-risk areas for varicose vein rupture in CT images with higher quantification result values ​​based on the quantification result values ​​of the varicose vein quantification module; specifically, quantification is performed by the ratio of the blood vessel volume within each varicose vein area in the image to the total volume.

10. The device for quantifying esophageal and gastric varices based on CT images according to claim 9, characterized in that: Based on the radiomics feature set calculated in the radiomics analysis module, all varicose vein areas in the lower esophageal region are identified. The ratio of the vascular volume to the total volume within each varicose vein area in all lower esophageal region images is calculated. The varicose vein area with the largest ratio is the high-risk area for variceal bleeding within the lower esophageal region. All varicose vein areas in the gastric fundus region image were distinguished, and the ratio of the vascular volume within each varicose vein area to the total volume in all gastric fundus region images was calculated. The varicose vein area with the largest ratio was the high-risk area for varicose vein rupture and bleeding within the gastric fundus region.

11. The device for quantifying esophageal and gastric varices based on CT images according to claim 1, characterized in that: The data acquisition module obtains a complete CT image of the lower esophagus to the upper abdomen of the subject during the portal venous phase.

12. The device for quantifying esophageal and gastric varices based on CT images according to claim 11, characterized in that: When the data acquisition module cannot obtain a complete CT image of the lower esophagus to the upper abdomen of the subject during the portal phase, a complete CT image of the lower esophagus to the upper abdomen of the subject during the venous phase is obtained.

13. The device for quantifying esophageal and gastric varices based on CT images according to claim 1, characterized in that: The complete CT image from the lower esophagus to the upper abdomen at least shows the area of ​​the lower esophagus and the gastric fundus.

14. The device for quantifying esophageal and gastric varices based on CT images according to claim 1, characterized in that: In the vascular region image extraction module, the other regions in the gastric fundus region image refer to the water, air, and gastric fundus content regions; the other regions in the lower esophageal region image refer to the water, air, and esophageal content regions.

15. The device for quantifying esophageal and gastric varices based on CT images according to claim 1, characterized in that: Esophageal wall volume It is calculated by multiplying the spacing between esophageal wall voxels and the number of esophageal wall voxels within the set length range; Volume of blood vessels in the esophageal wall It is calculated by multiplying the distance between esophageal wall lumen vascular voxels and the number of esophageal wall vascular voxels within the set length range; Volume of the stomach wall in the fundus area It is obtained by multiplying the distance between gastric wall voxels in the gastric fundus region and the number of gastric wall voxels in the gastric fundus region within the set length range; Volume of blood vessels in the gastric wall in the gastric fundus area It is obtained by multiplying the distance between the vascular voxels in the gastric wall of the gastric fundus region and the number of vascular voxels in the gastric wall of the gastric fundus region within the set length range.