Postoperative portal vein pressure prediction system

Abstract

The invention discloses a postoperative portal vein pressure prediction system, which relates to the technical field of portal vein pressure prediction, and is characterized in that three-dimensional structure regions of interest of the liver and the spleen are segmented based on preoperative and postoperative abdomen enhanced CT vein phase image data of a patient; image omics features including first-order statistical features, texture features and shape features are extracted from the segmented regions of interest of the three-dimensional structures of the liver and the spleen, and core feature screening is carried out on the extracted features; integrating the screened core features with clinical hemodynamic parameters and surgical parameters to construct a multi-modal prediction model of the portal vein pressure gradient; and predicting and outputting a portal vein pressure gradient predicted value by using the multi-modal prediction model of the portal vein pressure gradient, and determining a postoperative portal vein pressure gradient risk grade of the patient. The portal vein pressure gradient prediction method solves the problems of insufficient dynamic evaluation capability, multi-modal information integration and risk layering application in the prior art, and realizes non-invasive and accurate prediction of the portal vein pressure gradient.

Description

Technical Field

[0001] The present invention relates to the technical field of portal vein pressure prediction, in particular to a postoperative portal vein pressure prediction system. Background Art

[0002] Portal hypertension is a key determinant of the clinical prognosis of patients with cirrhosis. Its severity directly affects the occurrence and progression of cirrhosis-related complications (such as esophageal variceal bleeding). Transjugular intrahepatic portosystemic shunt is a minimally invasive interventional treatment technique that establishes a shunt channel by implanting a stent in the liver parenchyma between the portal vein and the hepatic vein, thereby significantly reducing the blood flow resistance of the portal vein system from a structural perspective, thereby effectively reducing portal vein pressure. Compared with drug therapy and endoscopic therapy, TIPS can fundamentally reduce portal vein pressure and is an important intervention for the treatment of esophageal variceal bleeding.

[0003] Clinically, existing technologies rely solely on simple morphological indicators such as liver volume and spleen diameter, which cannot reflect tissue heterogeneity or hemodynamic changes. Postoperative portal vein pressure is affected by multiple factors such as surgical operation and liver parenchymal regeneration capacity. Preoperative clinical parameters cannot reflect postoperative changes in real time. Traditional invasive examinations may cause complications such as bleeding and pneumothorax and cannot be monitored dynamically. At the same time, existing studies are mostly based on small samples from single centers, and the models have poor adaptability to patients with cirrhosis of different causes. They do not combine preoperative and postoperative imaging changes with surgical parameters and cannot reflect the dynamic impact of treatment on portal vein pressure.

[0004] Therefore, to address the above issues, a postoperative portal vein pressure prediction system is urgently needed. Summary of the Invention

[0005] In response to the deficiencies of the existing technology, the present invention provides a postoperative portal vein pressure prediction system, which solves the deficiencies of the existing technology in dynamic assessment capabilities, multimodal information integration and risk stratification applications.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a postoperative portal vein pressure prediction system, including: an image data acquisition module, used to obtain the patient's preoperative and postoperative abdominal enhanced CT venous phase image data; a region segmentation module, used to segment the three-dimensional structural regions of interest of the liver and spleen based on the patient's preoperative and postoperative abdominal enhanced CT venous phase image data; a feature extraction module, used to extract imaging genomics features including first-order statistical features, texture features and shape features from the segmented three-dimensional structural regions of interest of the liver and spleen, and perform core feature screening on the extracted features; a multimodal fusion module, used to integrate the screened core features with clinical hemodynamic parameters and surgical parameters to construct a multimodal prediction model of portal vein pressure gradient; a result output module, used to predict and output the portal vein pressure gradient prediction value using the multimodal prediction model of portal vein pressure gradient, and determine the patient's postoperative portal vein pressure gradient risk level.

[0007] Furthermore, the image data acquisition module is specifically analyzed as follows: by establishing a connection with the hospital's image archiving and communication system, the patient's preoperative and postoperative abdominal enhanced CT venous phase image data is acquired. The format of the patient's preoperative and postoperative abdominal enhanced CT venous phase image data is in DICOM format, which contains the patient's basic information, scanning parameters and the image data itself. The specific data acquisition analysis is as follows: based on the monitor, the patient's arterial blood pressure is measured and the average blood pressure is identified, and the patient's inferior vena cava blood pressure and portal vein blood pressure are acquired by a catheter. The blood pressure representing the hepatic sinusoidal layer is determined by the portal vein and inferior vena cava blood pressures. At the same time, Doppler ultrasound is used to determine the average flow velocity of the hepatic artery and the flow velocity in the middle of the portal vein main trunk, and the pressure of the hepatic artery system is identified. The hepatic venous pressure is equivalently replaced by the right atrial pressure and its population averaged data is taken. The flow of the hepatic venous system is obtained by summing the difference between the hepatic artery flow and the total portal vein flow and the shunt flow.

[0008] Furthermore, the regional segmentation module is specifically analyzed as follows: the patient's preoperative and postoperative abdominal enhanced CT venous phase image data is imported into 3D Slicer, and then the contrast and clarity of the liver and spleen in the image are optimized by optimizing the image window width and window position; a region of interest marker for marking the liver and a region of interest marker for marking the spleen are created, and then the drawn liver region of interest and spleen region of interest are identified layer by layer starting from the first slice of the CT image in combination with the delineation tool; after completing the delineation of the liver and spleen, three-dimensional reconstruction is performed, and then the current project data is saved to a designated storage location, and the current project data includes the imported abdominal enhanced CT venous phase image data, the delineated liver region of interest and spleen region of interest, and the generated three-dimensional structure model.

[0009] Furthermore, the feature extraction module specifically extracts feature analysis as follows: within the segmented three-dimensional structure region of interest of the liver and spleen, first-order statistical features are identified, and the first-order statistical features specifically include mean, variance, skewness, and kurtosis; texture features are extracted using the gray-level co-occurrence matrix, and the texture features include energy, entropy, contrast, and correlation; shape features are extracted based on the geometric shape of the three-dimensional structure region of interest, and the shape features include volume, surface area, sphericity, and compactness.

[0010] Furthermore, the specific analysis of the core feature screening of the extracted features is as follows: the extracted features are integrated into a feature matrix, and a portal vein pressure gradient label vector is set that corresponds one-to-one to the samples in the feature matrix, and then the relationship between the features and the portal vein pressure gradient is learned using random forest to obtain the importance score of each feature, and the features with an importance score greater than the importance threshold are marked as core features.

[0011] Furthermore, the multimodal fusion module is specifically analyzed as follows: the core features of the markers are combined and organized into a core feature matrix, and the patient's clinical hemodynamic parameters and surgical parameters are integrated into a clinical parameter matrix, and then the core feature matrix and the clinical parameter matrix are standardized respectively, and the standardized core feature matrix and the clinical parameter matrix are spliced by column to form a multimodal feature matrix, and the corresponding portal vein pressure gradient true value vector is obtained. Each element in the multimodal feature matrix corresponds to the actual portal vein pressure gradient of a patient sample; based on linear regression, a multimodal prediction model of the portal vein pressure gradient is constructed using the multimodal feature matrix and the corresponding portal vein pressure gradient true value vector.

[0012] Furthermore, the specific method of outputting the portal vein pressure gradient prediction value is: obtaining the patient's multimodal feature matrix, and then inputting the patient's multimodal feature matrix into the multimodal prediction model of the portal vein pressure gradient, and outputting the portal vein pressure gradient prediction value.

[0013] Furthermore, the specific analysis for determining the patient's postoperative portal vein pressure gradient risk level is as follows: determining the risk level thresholds corresponding to different portal vein pressure gradient intervals, wherein the risk level thresholds include a low-risk threshold, a medium-risk threshold, and a high-risk threshold; comparing the patient's portal vein pressure gradient prediction value with the risk level threshold; if the portal vein pressure gradient prediction value is less than the low-risk threshold, it is determined to be a low-risk level; if the portal vein pressure gradient prediction value is between the low-risk threshold and the medium-risk threshold, it is determined to be a medium-risk level; if the portal vein pressure gradient prediction value is greater than the medium-risk threshold, it is determined to be a high-risk level.

[0014] The present invention has the following beneficial effects:

[0015] This postoperative portal vein pressure prediction system combines imaging genomics, clinical parameters and surgical details, and integrates multi-dimensional information through machine learning, making it more accurate than predictions from a single data source. Through preoperative and postoperative CT comparisons, the actual impact of surgery on the portal vein system is quantified, enabling dynamic assessment of individualized pressure gradients. Traditional portal vein pressure gradients require pressure measurement via jugular vein cannulation, which is an invasive procedure and carries the risk of bleeding and infection. Non-invasive prediction through CT images reduces the risk of complications in patients. Postoperative patients need to have pressure changes assessed multiple times, and non-invasive methods are more easily accepted, improving follow-up compliance. The entire process from image segmentation to feature extraction is automated, reducing manual intervention and shortening analysis time. Low / medium / high risk levels are clearly divided to assist doctors in quickly formulating intervention strategies. Imaging genomics features reveal pathological mechanisms that traditional indicators cannot reflect, providing new biomarkers for the study of portal hypertension. The system supports the inclusion of new parameters or optimization algorithms to continuously improve prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a structural diagram of a postoperative portal vein pressure prediction system of the present invention.

[0017] Figure 2 This is a method flow chart of a postoperative portal vein pressure prediction system of the present invention. DETAILED DESCRIPTION

[0018] The embodiment of the present application realizes non-invasive and accurate prediction of portal vein pressure gradient through a postoperative portal vein pressure prediction system.

[0019] The overall idea of the embodiments of this application is: by integrating preoperative and postoperative imaging genomics features, clinical parameters and surgical data, using artificial intelligence technology to construct a dynamic prediction model of portal vein pressure gradient (PPG) to achieve postoperative risk stratification and individualized assessment.

[0020] See also Figure 1 , Figure 2An embodiment of the present invention provides a technical solution: a postoperative portal vein pressure prediction system, comprising: an image data acquisition module for acquiring the patient's preoperative and postoperative abdominal enhanced CT venous phase image data; a region segmentation module for segmenting the three-dimensional structural regions of interest of the liver and spleen based on the patient's preoperative and postoperative abdominal enhanced CT venous phase image data; a feature extraction module for extracting imaging genomics features including first-order statistical features, texture features, and shape features from the segmented three-dimensional structural regions of interest of the liver and spleen, and performing core feature screening on the extracted features; a multimodal fusion module for integrating the screened core features with clinical hemodynamic parameters and surgical parameters to construct a multimodal prediction model for portal vein pressure gradient; a result output module for predicting and outputting a portal vein pressure gradient prediction value using the multimodal prediction model for portal vein pressure gradient, and determining the patient's postoperative portal vein pressure gradient risk level.

[0021] Specifically, the imaging data acquisition module is specifically analyzed as follows: by establishing a connection with the hospital's imaging archiving and communication system, the patient's preoperative and postoperative abdominal enhanced CT venous phase imaging data is obtained. The format of the patient's preoperative and postoperative abdominal enhanced CT venous phase imaging data is in DICOM format, which contains the patient's basic information, scanning parameters and the image data itself. The specific data acquisition analysis is as follows: based on the monitor, the patient's arterial blood pressure is measured and the average blood pressure is identified. The patient's inferior vena cava blood pressure and portal vein blood pressure are obtained by using a catheter. The blood pressure representing the hepatic sinusoidal level is determined by the portal vein and inferior vena cava blood pressures. At the same time, Doppler ultrasound is used to determine the average flow velocity of the hepatic artery and the flow velocity in the middle of the portal vein trunk, and the pressure of the hepatic artery system is identified. The hepatic venous pressure is equivalently replaced by the right atrial pressure and its population averaged data is taken. The flow of the hepatic venous system is obtained by summing the difference between the hepatic artery flow and the total portal vein flow and the shunt flow.

[0022] In this implementation scheme, by establishing a connection with the hospital's picture archiving and communication system (PACS), the patient's preoperative and postoperative abdominal enhanced CT venous phase image data is obtained. At the same time, manual import of image data files is supported to ensure the diversity of data sources. The data format is usually DICOM format, which contains the patient's basic information (such as name, age, ID number, etc.), scanning parameters (such as layer thickness, tube voltage, tube current, etc.) and the image data itself.

[0023] Verify the integrity and accuracy of acquired imaging data to check for missing, erroneous, or corrupted data. If any issues are found, promptly communicate with relevant departments to obtain correct data. Preprocess the image data that meets the requirements, including adjusting the image window width and window position to optimize the display of venous structure; and normalize the image grayscale value range to the [0, 1] range for ease of subsequent processing.

[0024] Store verified and pre-processed image data in a local database or cloud storage system, and establish a comprehensive data indexing and management mechanism. Categorize and store data by patient ID, examination date, and other information to facilitate rapid retrieval and access by subsequent modules.

[0025] By connecting to the hospital's imaging system to obtain DICOM-formatted image data, comprehensive information about the patient's abdominal contrast-enhanced CT venous phase can be obtained, including basic information, scanning parameters, and image data, providing rich data for subsequent analysis. Furthermore, combining multiple measurement methods to obtain physiological data such as blood pressure and flow velocity reflects the patient's physiological status from different perspectives, facilitating a more accurate assessment of the disease and liver function. Using a monitor to measure arterial blood pressure and identify mean blood pressure, and using a catheter to directly obtain inferior vena cava and portal vein blood pressure, accurate pressure values in key blood vessels can be obtained. Determining the blood pressure representative of the hepatic sinusoidal level through a specific method, measuring blood flow velocity using Doppler ultrasound, and calculating hepatic arterial system pressure based on empirical formulas ensures that the obtained physiological parameters are closer to the true physiological state, providing reliable data support for in-depth research on liver hemodynamics. Substituting hepatic venous pressure for right atrial pressure and averaging the data across the population, as well as calculating the flow rate in the hepatic venous system, enables a comprehensive assessment of the liver's hemodynamic status. This multi-parameter, multi-angle analysis approach provides a more comprehensive understanding of liver function and is crucial for disease diagnosis, treatment planning, and postoperative recovery monitoring.

[0026] Specifically, the regional segmentation module is analyzed as follows: the patient's preoperative and postoperative abdominal enhanced CT venous phase image data is imported into 3DSlicer, and then the contrast and clarity of the liver and spleen in the image are optimized by optimizing the image window width and window position; a region of interest marker for marking the liver and a region of interest marker for marking the spleen are created, and then the drawn liver region of interest and spleen region of interest are identified layer by layer starting from the first slice of the CT image in combination with the outlining tool; after completing the outlining of the liver and spleen, three-dimensional reconstruction is performed, and then the current project data is saved to the specified storage location. The current project data includes the imported abdominal enhanced CT venous phase image data, the outlined liver region of interest and spleen region of interest, and the generated three-dimensional structure model.

[0027] In this embodiment, the specific steps for optimizing the image window width and window position are as follows: analyzing the normal density range of the liver and spleen in the abdominal enhanced CT venous phase image. Generally speaking, the density of the liver is about 40-60HU (Houns units) during plain scanning, which will increase after enhancement. The density of the spleen is slightly lower than that of the liver, about 30-50HU during plain scanning, and will also be enhanced to a certain extent after enhancement. According to the density range of the liver and spleen, the window width and window position are preliminarily set. For example, the window position can be set to the middle value of the average density of the liver and spleen. Assuming that the average density of the liver is 50HU and the average density of the spleen is 40HU, the window position can be set to 45HU first, and the window width can be preliminarily adjusted according to the difference between the two density ranges. Step 1: Set the value, such as 40HU, which can roughly cover the density range of the liver and spleen and make them appear in the image. In 3DSlicer software, by observing the display effect of the liver and spleen in the image, gradually adjust the window width and window level. If the boundary of the liver or spleen is not clear enough, increase the window width appropriately to increase the contrast of the image. If the overall image is too bright or too dark, adjust the window level to change the brightness of the image. At the same time, refer to the anatomical structure of the liver and spleen and the contrast of the surrounding tissues until the contrast and clarity of the liver and spleen in the image reach the optimal state, that is, the boundary, internal structure and relationship of the liver and spleen with the surrounding tissues can be clearly distinguished.

[0028] The specific steps for creating region of interest markers are as follows: in the 3DSlicer software, identify the tool or function module used to create region of interest markers; choose to create a new region of interest marker. For the region of interest marker of the liver, it can be named "Liver_ROI" to distinguish it from other markers. Similarly, for the region of interest marker of the spleen, it can be named "Spleen_ROI"; set relevant attributes for each region of interest marker, such as color, transparency, etc., so that the liver and spleen markers can be clearly identified and distinguished in subsequent operations. For example, the color of the liver region of interest marker can be set to red and the transparency can be set to 50%; the color of the spleen region of interest marker can be set to blue and the transparency can be set to 50%.

[0029] The specific steps for identifying and drawing the region of interest layer by layer with the help of the outlining tool are as follows: starting from the first slice of the CT image, activate the outlining tool in the 3DSlicer software. The tool has multiple drawing methods, such as free hand drawing and polygon drawing. The appropriate drawing method can be selected according to the actual situation; according to the image characteristics of the liver and spleen in the first slice, use the outlining tool to start drawing the region of interest along the boundary of the liver and spleen. During the drawing process, refer to the color and transparency of the region of interest marker set previously, as well as the optimized image window width and window level to ensure that the drawn area accurately contains the liver and spleen tissues, while avoiding including too much surrounding irrelevant tissue; after drawing the liver and spleen regions of interest in the first slice, switch to the next slice. In the new slice, according to the continuity and image characteristics of the liver and spleen, continue to use the outlining tool to draw the region of interest so that it is continuous with the region in the previous slice. During the drawing process, the shape and size of the drawing need to be adjusted according to the changes in organ morphology; traverse the entire CT image layer by layer until the drawing of the liver and spleen regions of interest in all slices is completed.

[0030] The specific steps for 3D reconstruction after delineating the liver and spleen are as follows: In the 3DSlicer software, confirm that the delineation of the liver and spleen regions of interest on all slices has been completed; identify the 3D reconstruction function module or tool in the software; select an appropriate 3D reconstruction algorithm. Common algorithms include the marchingcubes algorithm and the surface drawing algorithm. Based on the contour information on the slice, a 3D surface model or volume model is generated through interpolation, fitting, and other methods. For example, the marchingcubes algorithm searches for isosurfaces in the volume data and converts them into triangular patches to construct a 3D model; set relevant parameters for 3D reconstruction, such as the resolution and smoothness of the model. Higher resolution makes the reconstructed model more accurate, but it also increases the amount of calculation. and storage space, the smoothness parameter can adjust the smoothness of the model surface and reduce jagged or discontinuous phenomena. These parameters should be adjusted reasonably according to actual needs and image quality requirements; perform 3D reconstruction operations, and generate a 3D structural model using the previously outlined liver and spleen region of interest data according to the set algorithm and parameters; after the 3D reconstruction is completed, visualize and check the generated 3D model, specifically by observing various angles of the model through rotation, scaling, translation and other operations in the software to check whether the model accurately reflects the morphology and structure of the liver and spleen, and whether there are any errors or incomplete areas. If problems are found, make adjustments and modifications, which include rechecking the outlined area, adjusting reconstruction parameters, etc., until a satisfactory 3D model is obtained.

[0031] By optimizing the window width and window position, the imaging features of the liver and spleen can be highlighted, making them clearer and more identifiable in the image, providing a high-quality image foundation for subsequent regional delineation and analysis, and helping to improve the accuracy of segmentation; creating special region-of-interest markers and combining them with delineation tools to identify and draw layer by layer can accurately define the scope of the liver and spleen, reduce human errors, and improve the accuracy and reliability of segmentation; all relevant data, including original image data, delineated areas, and generated three-dimensional structural models, are saved to a designated location to ensure data integrity and traceability, facilitating subsequent review, comparison, and further analysis and research; three-dimensional reconstruction after delineation can intuitively display the three-dimensional structure of the liver and spleen, helping doctors to more comprehensively understand the morphology, position, and spatial relationship of the organs, which is helpful for preoperative planning and postoperative evaluation.

[0032] Specifically, the feature extraction module extracts features for analysis as follows: within the segmented three-dimensional structure regions of interest of the liver and spleen, first-order statistical features are identified, including mean, variance, skewness, and kurtosis; texture features are extracted using the gray-level co-occurrence matrix, including energy, entropy, contrast, and correlation; shape features are extracted based on the geometric shape of the three-dimensional structure region of interest, including volume, surface area, sphericity, and compactness.

[0033] The specific analysis of core feature screening of the extracted features is as follows: the extracted features are integrated into a feature matrix, and the portal vein pressure gradient label vector is set to correspond one-to-one with the samples in the feature matrix. Then, the relationship between the features and the portal vein pressure gradient is learned using random forest to obtain the importance score of each feature, and the features with importance scores greater than the importance threshold are marked as core features.

[0034] In this embodiment, the mean in the first-order statistical feature represents the average level of the image grayscale value, reflecting the overall brightness of the image, and is obtained by calculating the average value of the grayscale values of all pixels in the region of interest; the variance is used to measure the degree of dispersion of the image grayscale value relative to the mean. The larger the variance, the more dispersed the grayscale value distribution of the image. It is obtained by calculating the average value of the square of the difference between the grayscale value of each pixel in the region of interest and the mean; the skewness describes the asymmetry of the image grayscale distribution. A positive skewness indicates that the right tail of the grayscale distribution is longer, and a negative skewness indicates that the left tail is longer. It is obtained by calculating the third-order central moment of the pixel grayscale value in the region of interest and dividing it by the cube of the standard deviation; the kurtosis is used to characterize the steepness of the image grayscale distribution near the mean. The higher the kurtosis, the more concentrated the grayscale values are near the mean. It is obtained by calculating the fourth-order central moment of the pixel grayscale value in the region of interest and dividing it by the fourth power of the standard deviation. 3 is usually subtracted to make it meet the standard of normal distribution.

[0035] The energy in the texture feature reflects the uniformity and regularity of the image grayscale distribution. The larger the energy value, the more regular the image texture. It is obtained by calculating the sum of the squares of the elements in the grayscale co-occurrence matrix; entropy represents the complexity of the image texture. The larger the entropy value, the more complex the texture. It is obtained by performing logarithmic operation on the elements in the grayscale co-occurrence matrix and then taking the weighted sum; contrast reflects the severity of local grayscale changes in the image. The higher the contrast, the clearer the texture edge in the image. It is obtained by calculating the weighted sum of the differences between different grayscale levels in the grayscale co-occurrence matrix; correlation measures the correlation of grayscale values in the image in the horizontal and vertical directions. The higher the correlation, the stronger the consistency of the texture in the image in a certain direction. It is obtained by calculating the mean, variance and covariance of the rows and columns of the grayscale co-occurrence matrix and then calculating according to the relevant formula.

[0036] Among the shape features, volume represents the size of the space occupied by the region of interest of the three-dimensional structure. It is obtained by counting the voxels in the region of interest of the three-dimensional structure and converting them according to the actual physical size of the voxels. Surface area refers to the surface area of the region of interest of the three-dimensional structure. It is obtained by identifying and calculating the boundaries of the region of interest through a specific three-dimensional surface reconstruction algorithm. Sphericity is used to describe the degree of similarity between a three-dimensional structure and a sphere. The value range is between 0 and 1. The closer to 1, the closer it is to a sphere. It is obtained by calculating the volume and surface area of the three-dimensional structure and calculating it according to a specific formula. Compactness reflects the compactness of the three-dimensional structure. It is usually defined as the ratio of the square of the perimeter to the area (in two dimensions) or the ratio of the square of the surface area to the volume (in three dimensions). The higher the compactness, the more compact the structure.

[0037] The portal vein pressure gradient label vector is set as follows: the portal vein pressure gradient label vector is set according to the patient's actual clinical measurement data. For example, the patient's portal vein pressure and inferior vena cava pressure are obtained by invasive or non-invasive methods, and the difference between the two is calculated as the portal vein pressure gradient value. For each sample (i.e., each patient's imaging data), its corresponding portal vein pressure gradient value is used as an element in the label vector, thereby forming a portal vein pressure gradient label vector that corresponds one-to-one to the sample in the feature matrix.

[0038] The specific steps of using random forest to learn the relationship between features and portal vein pressure gradient and obtain the importance score of each feature are as follows: divide the feature matrix and the corresponding portal vein pressure gradient label vector into training set and test set according to a certain ratio; build a random forest model on the training set. The random forest consists of multiple decision trees. When building each decision tree, a part of features (usually the square root of the total number of features) and a part of samples (random sampling with replacement) are randomly selected from the feature matrix for training to increase the diversity and stability of the model; for each decision tree, when the node is split, the optimal splitting feature is selected by calculating the information gain or Gini index of different features, thereby constructing the structure of the decision tree; by constructing Build multiple decision trees to form a random forest model; use the trained random forest model to predict the test set to obtain the predicted portal vein pressure gradient value; calculate the error between the predicted value and the true value (such as mean square error, mean absolute error, etc.) to evaluate the performance of the model; randomly disrupt the order of the feature values in the test set, and then use the disrupted feature values and other undisrupted feature values to input them into the trained random forest model for prediction, calculate the prediction error at this time, and compare the error with the prediction error when the original feature value is input. The greater the increase in the error, the greater the impact of the feature on the model and the higher its importance score. By performing this operation on all features, the importance score of each feature can be obtained.

[0039] The importance threshold is set as follows: set the threshold to a relatively small value (such as 0.05), and then observe the number of core features screened out and the performance of the model. If too many core features are screened out, the model may be overfitted. In this case, the threshold can be appropriately increased; if too few core features are screened out, important information may be lost, resulting in a decrease in model performance. In this case, the threshold can be appropriately lowered. You can also refer to the distribution of feature importance scores, such as selecting a threshold that can make the number of core features screened out account for a certain proportion of the total number of features (such as 30%-50%). In addition, you can also use the cross-validation method to train and verify the model under different thresholds, and select the threshold that makes the model perform best on the validation set as the importance threshold.

[0040] Extracting multiple features from the segmented three-dimensional structures of the liver and spleen within the region of interest can comprehensively and meticulously describe the imaging characteristics of the liver and spleen, providing a rich data foundation for subsequent portal vein pressure prediction. Core feature screening through random forests can remove redundant and unimportant features, improve the efficiency and accuracy of the model, reduce the risk of overfitting, and make the model more generalizable.

[0041] Specifically, the multimodal fusion module is analyzed as follows: the core features of the markers are combined and organized into a core feature matrix, and the patient's clinical hemodynamic parameters and surgical parameters are integrated into a clinical parameter matrix. The core feature matrix and the clinical parameter matrix are then standardized respectively, and the standardized core feature matrix and the clinical parameter matrix are spliced column by column to form a multimodal feature matrix, and the corresponding portal vein pressure gradient true value vector is obtained. Each element in the multimodal feature matrix corresponds to the actual portal vein pressure gradient of a patient sample; based on linear regression, a multimodal prediction model of the portal vein pressure gradient is constructed using the multimodal feature matrix and the corresponding portal vein pressure gradient true value vector.

[0042] The specific method of outputting the predicted value of the portal vein pressure gradient is: obtaining the patient's multimodal feature matrix, and then inputting the patient's multimodal feature matrix into the multimodal prediction model of the portal vein pressure gradient, and outputting the predicted value of the portal vein pressure gradient.

[0043] The specific analysis to determine the patient's postoperative portal vein pressure gradient risk level is as follows: determine the risk level thresholds corresponding to different portal vein pressure gradient intervals, which include low-risk thresholds, medium-risk thresholds, and high-risk thresholds; compare the patient's portal vein pressure gradient prediction value with the risk level thresholds. If the portal vein pressure gradient prediction value is less than the low-risk threshold, it is judged as a low-risk level; if the portal vein pressure gradient prediction value is between the low-risk threshold and the medium-risk threshold, it is judged as a medium-risk level; if the portal vein pressure gradient prediction value is greater than the medium-risk threshold, it is judged as a high-risk level.

[0044] In this embodiment, an example of a multimodal prediction model expression for portal vein pressure gradient is: In the formula represents the predicted value of portal vein pressure gradient, x j represents the jth feature value in the multimodal feature matrix, m represents the number of features, j is the feature number, β0 represents the intercept, β j Represents the model coefficient corresponding to the j-th feature value, where the intercept and the model coefficient are obtained by fitting by minimizing the loss function.

[0045] Clinical hemodynamic parameters include but are not limited to mean arterial blood pressure, inferior vena cava blood pressure, portal vein blood pressure, mean hepatic artery flow velocity, and middle portal vein flow velocity. Mean arterial blood pressure reflects the combined effect of cardiac ejection and peripheral vascular resistance and is an important indicator for evaluating circulatory function. Systolic and diastolic blood pressures are measured by a monitor and then the mean arterial blood pressure is calculated; inferior vena cava blood pressure reflects the pressure of the venous system and is directly measured by a catheter; portal vein blood pressure directly reflects the pressure of the portal vein system and is obtained by catheter measurement; mean hepatic artery flow velocity reflects the blood supply of the hepatic artery and is obtained by Doppler ultrasound measurement; and the middle portal vein flow velocity reflects the blood flow velocity in the portal vein and is also measured by Doppler ultrasound.

[0046] Surgical parameters include but are not limited to surgical method, operation time, and intraoperative blood loss.

[0047] The true value vector of the portal vein pressure gradient is obtained by invasive measurement methods, such as using a pressure measuring device to directly measure the pressure of the portal vein and inferior vena cava during surgery or at a specific time after surgery, and then calculating the difference between the two to obtain the true value of the portal vein pressure gradient. These true values are organized into vector form and correspond one-to-one to the samples in the multimodal feature matrix.

[0048] The risk threshold is set in the following way: based on a large amount of clinical data and research results, combined with clinical experience, the risk level threshold is set, a certain number of patients are followed up for a long time, and the clinical outcomes of patients under different portal vein pressure gradients, such as the incidence of complications and survival rate, are observed to determine the boundaries of different risk levels. For example, through statistical analysis, it is found that patients with a portal vein pressure gradient within a certain range have a low incidence of postoperative complications and a better prognosis. The upper limit of this range can be set as a low-risk threshold. In another higher pressure gradient range, the probability of patients developing serious complications is significantly increased. The lower limit of this range can be set as a high-risk threshold, and the range between the two is a medium-risk threshold.

[0049] The role of determining the patient's postoperative portal vein pressure gradient risk level is as follows: patients with different risk levels can adopt different treatment strategies. For low-risk patients, routine postoperative observation and care are required. For medium-risk patients, liver function, blood routine and other indicators need to be closely monitored, and the treatment plan needs to be adjusted appropriately. For high-risk patients, more active intervention measures are needed, such as drug therapy to reduce portal vein pressure and prevent complications in advance. The risk level can help doctors and patients understand the development trend and prognosis of the disease after surgery, and help patients and their families to be mentally prepared and plan subsequent rehabilitation plans. Stratified management of patients according to risk level can enable a more reasonable allocation of medical resources, concentrate more resources on the treatment and monitoring of high-risk patients, and improve medical efficiency and quality.

[0050] It integrates imaging genomics features with clinical parameters and makes full use of data from different modalities, which can more comprehensively reflect the patient's condition and improve the accuracy of portal vein pressure gradient prediction; it standardizes and integrates different types of data to make the data comparable and consistent, which is conducive to building a more stable and reliable prediction model; by setting risk level thresholds, the predicted value of portal vein pressure gradient is converted into a specific risk level, providing clinicians with intuitive and clear assessment results, which helps to formulate personalized treatment plans and follow-up plans.

[0051] In summary, this application has at least the following effects:

[0052] By integrating radiomics features with clinical hemodynamic parameters and surgical parameters, the accuracy of postoperative portal vein pressure gradient prediction can be significantly improved, assisting clinicians in formulating individualized treatment plans; based on the predicted portal vein pressure gradient value, postoperative risk levels can be divided to guide postoperative monitoring frequency and intervention strategies; breaking through the limitations of single images or clinical parameters, through three-dimensional ROI segmentation, radiomics feature extraction and multimodal fusion models, collaborative analysis of anatomical structure, tissue heterogeneity and hemodynamic parameters can be achieved; through feature dimensionality reduction, key features can be screened to reduce redundant information interference and improve model generalization ability; the ROI segmentation and feature extraction process based on deep learning or traditional machine learning algorithms can be standardized to reduce the subjectivity of manual delineation and improve the repeatability of results; it can replace some invasive portal vein pressure measurements and reduce patient risks and medical costs.

[0053] Those skilled in the art will appreciate that embodiments of the present invention may be provided as systems. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] The present invention is described with reference to a block diagram of a system according to an embodiment of the present invention. It should be understood that the combination of each structure in the block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, such that the instructions, executed by the processor of the computer or other programmable data processing device, produce a device for implementing the functions specified in each structure in the block diagram.

[0055] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in each structure of the structural diagram.

[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in each structure in the structural diagram.

[0057] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0058] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A postoperative portal vein pressure prediction system, characterized in that: include: An imaging data acquisition module is used to obtain the patient's preoperative and postoperative abdominal enhanced CT venous phase imaging data; A regional segmentation module is used to segment the three-dimensional structural regions of interest of the liver and spleen based on the patient's preoperative and postoperative abdominal enhanced CT venous phase image data; A feature extraction module is used to extract radiomics features including first-order statistical features, texture features, and shape features from the segmented three-dimensional structural regions of interest of the liver and spleen, and to perform core feature screening on the extracted features; The multimodal fusion module is used to integrate the selected core features with clinical hemodynamic parameters and surgical parameters to construct a multimodal prediction model of portal vein pressure gradient; The result output module is used to predict and output the portal vein pressure gradient prediction value using the multimodal prediction model of the portal vein pressure gradient, and determine the patient's postoperative portal vein pressure gradient risk level.

2. A postoperative portal vein pressure prediction system according to claim 1, characterized in that: The imaging data acquisition module is specifically analyzed as follows: by establishing a connection with the hospital's imaging archiving and communication system, the patient's preoperative and postoperative abdominal enhanced CT venous phase imaging data is acquired. The format of the patient's preoperative and postoperative abdominal enhanced CT venous phase imaging data is in DICOM format, which contains the patient's basic information, scanning parameters, and the image data itself. The specific data acquisition analysis is as follows: based on the monitor, the patient's arterial blood pressure is measured and the average blood pressure is identified. The patient's inferior vena cava blood pressure and portal vein blood pressure are acquired by a catheter. The blood pressure representing the hepatic sinusoidal layer is determined by the portal vein and inferior vena cava blood pressures. At the same time, Doppler ultrasound is used to determine the average flow velocity of the hepatic artery and the flow velocity in the middle of the portal vein trunk, the pressure of the hepatic artery system is identified, the hepatic vein pressure is equivalently replaced by the right atrial pressure and the population averaged data is taken. The flow of the hepatic venous system is obtained by summing the difference between the hepatic artery flow and the total portal vein flow and the shunt flow.

3. A postoperative portal vein pressure prediction system according to claim 1, characterized in that: The regional segmentation module specifically analyzes: importing the patient's preoperative and postoperative abdominal enhanced CT venous phase image data into 3D Slicer, and then optimizing the image window width and window position to achieve the best contrast and clarity of the liver and spleen in the image; Create a region of interest marker for marking the liver and a region of interest marker for marking the spleen, and then use the delineation tool to identify the drawn liver region of interest and spleen region of interest layer by layer starting from the first slice of the CT image; After the liver and spleen are delineated, three-dimensional reconstruction is performed, and the current project data is saved to a designated storage location. The current project data includes the imported abdominal enhanced CT venous phase image data, the delineated liver region of interest and spleen region of interest, and the generated three-dimensional structure model.

4. A postoperative portal vein pressure prediction system according to claim 1, characterized in that: The feature extraction module specifically extracts feature analysis by: identifying first-order statistical features in the segmented liver and spleen three-dimensional structure regions of interest, wherein the first-order statistical features specifically include mean, variance, skewness, and kurtosis; Extracting texture features using gray-level co-occurrence matrix, where the texture features include energy, entropy, contrast, and correlation; Shape features are extracted based on the geometric shape of the region of interest of the three-dimensional structure, and the shape features include volume, surface area, sphericity, and compactness.

5. A postoperative portal vein pressure prediction system according to claim 4, characterized in that: The specific analysis of the core feature screening of the extracted features is as follows: the extracted features are integrated into a feature matrix, and a portal vein pressure gradient label vector is set that corresponds one-to-one to the samples in the feature matrix. Then, the relationship between the features and the portal vein pressure gradient is learned using a random forest to obtain the importance score of each feature, and the features whose importance score is greater than the importance threshold are marked as core features.

6. A postoperative portal vein pressure prediction system according to claim 1, characterized in that: The multimodal fusion module is specifically analyzed as follows: the core features of the markers are combined and organized into a core feature matrix, and the patient's clinical hemodynamic parameters and surgical parameters are integrated into a clinical parameter matrix. The core feature matrix and the clinical parameter matrix are then standardized respectively. The standardized core feature matrix and the clinical parameter matrix are spliced column by column to form a multimodal feature matrix, and the corresponding portal vein pressure gradient true value vector is obtained. Each element in the multimodal feature matrix corresponds to the actual portal vein pressure gradient of a patient sample. Based on linear regression, a multimodal prediction model of portal vein pressure gradient is constructed using the multimodal feature matrix and the corresponding portal vein pressure gradient true value vector.

7. A postoperative portal vein pressure prediction system according to claim 6, characterized in that: The specific method of outputting the portal vein pressure gradient prediction value is: obtaining the patient's multimodal feature matrix, and then inputting the patient's multimodal feature matrix into the multimodal prediction model of the portal vein pressure gradient, and outputting the portal vein pressure gradient prediction value.

8. The postoperative portal vein pressure prediction system according to claim 6, characterized in that: The specific analysis of determining the risk level of the patient's postoperative portal vein pressure gradient is as follows: determining risk level thresholds corresponding to different portal vein pressure gradient intervals, wherein the risk level thresholds include a low risk threshold, a medium risk threshold, and a high risk threshold; The patient's portal vein pressure gradient prediction value is compared with the risk level threshold. If the portal vein pressure gradient prediction value is less than the low-risk threshold, the patient is judged to be at a low-risk level. If the predicted value of the portal vein pressure gradient is between the low-risk threshold and the medium-risk threshold, it is judged as the medium-risk level; If the predicted value of the portal vein pressure gradient is greater than the medium-risk threshold, it is judged as a high-risk level.

Images