A method and system for obtaining a portal vein pressure gradient by fusing multi-modal features

Abstract

The application discloses a kind of fusion multimodal feature's portal vein pressure gradient acquisition method and system, it is related to portal vein pressure acquisition technical field, the method includes: S1, the CT image of patient is obtained and is preprocessed, obtains its relevant feature data;S2, based on liver blood vessel in CT image, constructs multimodal feature model;The CT image data after pre-processing, relevant feature data and the corresponding to-be-predicted portal vein pressure gradient label value are input into model, and training is carried out;S3, the CT image data after pre-processing, relevant feature data are input into the model trained, and prediction is carried out, and the portal vein pressure gradient value corresponding to CT image is obtained.By the innovation of multimodal fusion architecture and two-stage training strategy, the geometric structure information of three-dimensional model, physical parameter information and pathological characteristics are fully utilized, and high-precision, high-robustness non-contact prediction of PPG value is realized under the condition of only CT image blood vessel model input.

Description

Technical Field

[0001] This invention relates to the field of portal vein pressure acquisition technology, and specifically to a method and system for acquiring portal vein pressure gradient by integrating multimodal features. Background Technology

[0002] Portal hypertension (PH or PHT) refers to obstruction of blood flow or abnormally increased flow within the portal venous system, resulting in pathologically high pressure levels within the system. It is one of the most common complications in patients with cirrhosis. Traditional assessment of portal hypertension mainly relies on imaging examinations, such as abdominal ultrasound, CT, and MRI portal vein imaging, indirectly assessing the condition by observing anatomical features such as portal vein diameter, collateral circulation formation, and spleen size. Accurate acquisition of portal venous pressure gradient (PPG) data requires invasive procedures, specifically the placement of a pressure catheter within the portal vein to directly measure PPG. Clinically, the hepatic venous pressure gradient (HVPG) is also commonly used as an alternative measurement method for PCG.

[0003] Current noninvasive PPG assessment studies primarily focus on the geometric characteristics and blood flow information of the portal venous system in patients with portal hypertension. HPVG, the main assessment indicator, can be influenced by multiple factors, exhibiting significant individual variability and being highly dependent on etiology; that is, in certain circumstances, it may not be significantly correlated with PPG. Furthermore, no studies have yet incorporated hepatic artery and related pathological features into a comprehensive analysis, which may be a key factor contributing to the low accuracy of current noninvasive PPG assessments. Invasive methods for obtaining accurate PPG data not only impose a continuous physiological and psychological burden on patients but also have limitations in application. Therefore, there is an urgent need for a noninvasive and accurate method to assess the degree of portal hypertension in patients with cirrhosis. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for obtaining portal vein pressure gradient by integrating multimodal features.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention discloses a method for obtaining portal vein pressure gradient by fusing multimodal features, comprising the following steps:

[0007] S1. Acquire the patient's CT images and preprocess them to obtain their relevant feature data;

[0008] S2. Based on the liver vessels in CT images, construct a multimodal feature model; input the preprocessed CT image data, relevant feature data, and the corresponding portal vein pressure gradient label values ​​to be predicted into the model for training;

[0009] S3. Input the preprocessed CT image data and related feature data into the trained model to make predictions and obtain the portal vein pressure gradient value corresponding to the CT image.

[0010] Based on the first aspect, step S1 specifically includes the following steps:

[0011] S11. Acquire the patient's CT images and preprocess them to obtain a CT dataset with uniform size and quality that meets preset standards. The preprocessing includes image enhancement and image processing.

[0012] S12. Based on a preset threshold, obtain the mask of the vascular region from the CT dataset. Dicom. The file, which describes the vascular region including the portal vein, hepatic artery, and aorta, then obtains the background for each CT dataset. Dicom. document;

[0013] S13, Background Dicom. Documents and masks Dicom. The file is converted to the first background. all. Documents and the first mask all. File, then the first background all. Documents and the first mask all. The file is cropped to a preset size to obtain a second background. all. Document and Second Mask all. document;

[0014] S14, Second Background all. Document and Second Mask all. The file is converted to an STL file, and then feature values ​​Ti are extracted from the STL file. The feature values ​​Ti include the blood vessel centerline point number ID, the blood vessel centerline point x-axis coordinate X, the blood vessel centerline point y-axis coordinate Y, the blood vessel centerline point z-axis coordinate Z, the parent node FatherNode, the left child node LeftNode, the right child node RightNode, the curvature K, the angle Angle, the area S, the maximum diameter MaxR, and the minimum diameter MinR.

[0015] S15. Extract other pathology-related radiomics features from CT images, including liver volume (LVi) and spleen volume (SVi).

[0016] Based on the first aspect, step S2 specifically includes the following steps:

[0017] S21. Data preprocessing: Based on the stl. files of portal vein, hepatic artery and aortic vessels, perform three-dimensional projection on them, that is, orthogonally project the three-dimensional models of the stl. files onto the three coordinate planes xoy, yoz and zox respectively to generate two-dimensional projection images.

[0018] S22. Image standardization: The size of the two-dimensional projection image is uniformly adjusted to the preset resolution, and they are stacked along the channel dimension to obtain the synthetic image tensor I_i ∈ R^(3×H×W) corresponding to the i-th blood vessel, where H represents the height and W represents the width.

[0019] S23. Construct a multimodal feature model, which includes an encoder, a ViT decoder, and a portal vein pressure gradient estimator;

[0020] S24. Feature encoding: The encoder includes an image encoder and a feature value encoder. The synthesized image tensor I_i obtained in step S22 is input into the multimodal feature model. The image encoder segments it into a sequence of image blocks. After masking, the feature tensor E(Ii) ∈ R^(L×D) of the image is output, where L represents the length of the image block sequence and D represents the feature dimension. At the same time, the feature value Ti obtained in step S14 is converted into a vector representation by the feature value encoder, and the feature vector E(Ti) ∈ R^D of the feature value is output.

[0021] S25. Pre-training: The model is trained using a self-supervised learning method so that it can reconstruct the masked image patch based on the features of partially visible image patches and text.

[0022] S26. Fine-tuning stage: Using the pre-trained encoder features, a multilayer perceptron is trained based on data with portal vein pressure gradient labels to predict portal vein pressure gradient values.

[0023] S27. Input the preprocessed CT image data, relevant feature data and the corresponding portal vein pressure gradient label value to be predicted into the model for training to obtain a trained multimodal feature model.

[0024] Based on the first aspect, the masking process described in step S24 includes: performing a block operation on the image, selecting a number of block regions of a preset size and the size of the block regions, randomly selecting block regions for masking, making them regions to be learned; randomly selecting a proportion of image blocks, replacing their pixel values ​​with learnable mask tokens, and recording their original pixel values ​​Ptrue.

[0025] Based on the first aspect, step S25 specifically includes: fusing the image's feature tensor E(Ii) and the feature vector E(Ti) of the eigenvalues ​​through the feature fusion module to obtain a first fused feature Fi; then inputting the first fused feature Fi and the mask token into the ViT decoder; outputting the reconstructed pixel values ​​of all image blocks through the ViT decoder, and using the formula... Calculate pixel values ​​of reconstructed image blocks Compared with the actual pixel value Error between ,in This represents the total number of masked image patches in a batch; pre-training is repeated on different vascular models.

[0026] Based on the first aspect, step S25 further includes: using the AdamW optimizer to handle weight decay and prevent overfitting during pre-training; using a cosine annealing scheduling strategy for its learning rate, and setting a warm-up period and a first learning rate; setting the batch size according to GPU memory; and setting the first training cycle of pre-training based on the amount of unlabeled data until the reconstruction loss converges.

[0027] Based on the first aspect, in step S26, all weight parameters of the pre-trained image encoder and feature encoder are frozen, and only the subsequent prediction head, including the parameters of the global average pooling layer GAP and the linear regression layer PPG estimator, are trained.

[0028] Based on the first aspect, step S27 specifically includes: obtaining a second fusion feature Fi' by using the stl file of the blood vessel and the corresponding feature values ​​through a frozen image encoder and feature value encoder; then concatenating all the second fusion features Fi' by F=Concat(Fi') to obtain a feature vector F; then performing global average pooling on F to obtain a global feature vector g; and inputting g into the newly initialized linear regression layer to obtain the portal vein pressure gradient prediction value. =W⋅g+b, where W represents the weight matrix and b represents the bias term; through the formula Calculate the predicted value of portal vein pressure gradient Compared with the true value of portal vein pressure gradient Differences between , where N represents the batch size.

[0029] Based on the first aspect, step S27 further includes: using the AdamW optimizer during training, and setting a second learning rate and a second training period, wherein the second learning rate is lower than the first learning rate and the second training period is shorter than the first training period; and monitoring the mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (r) during training.

[0030] Secondly, this invention discloses a portal vein pressure gradient acquisition system that integrates multimodal features, used in the aforementioned method for acquiring portal vein pressure gradients that integrates multimodal features, comprising:

[0031] The image processing module is used to acquire and preprocess the patient's CT images to obtain their relevant feature data.

[0032] The model building module is used to construct and train a multimodal feature model based on liver blood vessels in CT images.

[0033] The output module is used to obtain the portal vein pressure gradient value corresponding to the CT image through the trained model.

[0034] The beneficial effects of this invention are:

[0035] 1) This application fully utilizes the geometric structure information, physical parameter information and pathological features of the three-dimensional model through an innovative multimodal fusion architecture and a two-stage training strategy, and achieves high-precision and robust non-contact prediction of PPG values ​​under the condition of only CT image vascular model input. Attached Figure Description

[0036] Figure 1 This is a schematic diagram illustrating the steps of a method for obtaining portal vein pressure gradient by fusing multimodal features according to an embodiment of the present invention. Detailed Implementation

[0037] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] This invention provides a method and system for acquiring portal vein pressure gradient by fusing multimodal features. Based on liver vessels and related pathological features from CT images, the method predicts target values ​​through multimodal fusion and self-supervised pre-training. Specifically, it involves extracting visual, physical, and pathological features from multiple 3D vascular model files through a specific processing flow, effectively fusing them, and finally predicting the target value using a regressor, thus achieving accurate and non-invasive acquisition of the portal vein pressure gradient. A schematic diagram of the method steps is shown below. Figure 1 As shown, it includes the following steps:

[0039] S1. Acquire the patient's CT images and preprocess them to obtain their relevant feature data;

[0040] S2. Based on the liver vessels in CT images, construct a multimodal feature model; input the preprocessed CT image data, relevant feature data, and the corresponding portal vein pressure gradient label values ​​to be predicted into the model for training;

[0041] S3. Input the preprocessed CT image data and related feature data into the trained model to make predictions and obtain the portal vein pressure gradient value corresponding to the CT image.

[0042] Specifically, step S1 includes the following steps:

[0043] S11. Acquire the patient's CT images and preprocess them to obtain a CT dataset with uniform size and quality that meets preset standards. The preprocessing includes image enhancement and image processing, etc.

[0044] S12. Based on a preset threshold, obtain the mask of the vascular region from the CT dataset. Dicom. The file, which describes the vascular region including the portal vein, hepatic artery, and aorta, then obtains the background for each CT dataset. Dicom. document;

[0045] S13, Background Dicom. Documents and masks Dicom. The file is converted to the first background. all. Documents and the first mask all. File, then the first background all. Documents and the first mask all. The file is cropped to a preset size to obtain a second background. all. Document and Second Mask all. document;

[0046] S14, Second Background all. Document and Second Mask all. The file is converted to an STL file, and then feature values ​​Ti are extracted from the STL file. The feature values ​​Ti include the blood vessel centerline point number ID, the blood vessel centerline point x-axis coordinate X, the blood vessel centerline point y-axis coordinate Y, the blood vessel centerline point z-axis coordinate Z, the parent node FatherNode, the left child node LeftNode, the right child node RightNode, the curvature K, the angle Angle, the area S, the maximum diameter MaxR, and the minimum diameter MinR, etc.

[0047] S15. Extract other pathology-related radiomics features from CT images, including liver volume (LVi) and spleen volume (SVi).

[0048] Specifically, in step S2, visual, physical, and pathological features are extracted from multiple 3D vascular model files through a specific processing flow and effectively fused. Finally, a regressor predicts continuous target values. Model training is divided into two stages: first, self-supervised pre-training is performed on a large amount of unlabeled STL vascular model file data; then, supervised fine-tuning is performed on a small dataset with portal vein pressure gradient labels. The specific steps include:

[0049] S21. Data preprocessing: Based on the stl. files of portal vein, hepatic artery and aortic vessels, perform three-dimensional projection on them. The projection operation preferably adopts three-dimensional orthogonal projection, that is, orthogonally project the three-dimensional model of the stl. file onto the three coordinate planes xoy, yoz and zox respectively to generate two-dimensional projection images.

[0050] S22. Image standardization: The size of the two-dimensional projection image is uniformly adjusted to a preset resolution (preferably 224×224 pixels), and stacked along the channel dimension to obtain the synthetic image tensor I_i ∈ R^(3×H×W) corresponding to the i-th blood vessel, where H represents the height and W represents the width. At this time, H=224 and W=224.

[0051] S23. Construct a multimodal feature model, which includes an encoder, a ViT decoder, and a portal vein pressure gradient estimator;

[0052] For example, the ViT decoder (enabled during pre-training phase) is used to reconstruct masked image patches during the pre-training phase; its input is a fused feature tensor and a mask token, and its output is a sequence of reconstructed image patches; the portal vein pressure gradient estimator (enabled during fine-tuning phase) is a multilayer perceptron used to predict portal vein pressure gradient values ​​during the fine-tuning phase; its input is a global feature vector, and its output is the predicted portal vein pressure gradient value.

[0053] S24. Feature encoding: The encoder includes an image encoder and a feature value encoder. The image encoder uses a visual Transformer (ViT) encoder, preferably a model based on the ViT-Base architecture. The feature value encoder is a lightweight encoding network composed of multilayer perceptrons. The synthesized image tensor I_i obtained in step S22 is input into the multimodal feature model. The image encoder segments it into a sequence of image blocks. After masking, the feature tensor E(Ii)∈R^(L×D) of the output image is generated, where L represents the length of the image block sequence and D represents the feature dimension. At the same time, the feature value Ti obtained in step S14 is converted into a vector representation by the feature value encoder, and the feature vector E(Ti)∈R^D of the output feature value is generated.

[0054] For example, when the Visual Transformer (ViT) is used as an image encoder, its input is an image tensor, and its output is a feature tensor of the image. The processing is represented as: Image Feature Tensor = ViTEncoder(Image Tensor). The feature value encoder uses a Transformer-based encoder, specifically BERT, as its feature value encoder. Its input is the feature value Ti, and its output is the feature vector of the feature value. The processing is represented as: Feature Vector of Feature Value = TextEncoder(Ti). An example of the data for the feature value Ti is shown below: Ti = (ID, X, Y, Z, FatherNode, LeftNode, RightNode, K, Angle, S, MaxR, MinR, LVi, SVi) = (0 81, 0 73, 0 51, 0, -1, 1, -1, Curvature, Angle, Area, Maximum Diameter, Minimum Diameter, Liver Volume, Spleen Volume); The parent node FatherNode, left child node LeftNode, and right child node RightNode have a value of -1, indicating no corresponding node. All data records are numerical values ​​without units.

[0055] S25. Pre-training: The model is trained using a self-supervised learning method so that it can reconstruct the masked image patch based on the features of partially visible image patches and text.

[0056] S26. Fine-tuning stage: Using the pre-trained encoder features, a multilayer perceptron is trained based on data with portal vein pressure gradient labels to predict portal vein pressure gradient values.

[0057] S27. Input the preprocessed CT image data, relevant feature data and the corresponding portal vein pressure gradient label value to be predicted into the model for training to obtain a trained multimodal feature model.

[0058] Specifically, the masking process in step S24 includes: dividing the image into blocks, selecting a preset number and size of block regions, randomly selecting block regions for masking to make them regions to be learned; randomly selecting a proportion of image blocks, replacing their pixel values ​​with learnable mask tokens, and recording their original pixel values ​​Ptrue.

[0059] Specifically, step S25 includes: fusing the image's feature tensor E(Ii) and the feature vector E(Ti) of the eigenvalues ​​through the feature fusion module to obtain a first fused feature Fi; then inputting the first fused feature Fi and the mask token into the ViT decoder; outputting the reconstructed pixel values ​​of all image blocks through the ViT decoder, and using the formula... Calculate pixel values ​​of reconstructed image blocks Compared with the actual pixel value Error between ,in This represents the total number of masked image patches in a batch; pre-training is repeated on different vascular models.

[0060] For example, the feature fusion module is used to fuse image features with feature values. Since both the image encoder and the feature value encoder use the Transformer architecture, they can project both the image space and the feature value space into the same vector space, allowing direct superposition operations in the vector space. Specifically, the feature vectors of the feature values ​​are first broadcast along the sequence length dimension to obtain the feature tensor of the broadcast feature values. Then, the image feature tensor is added element-wise to it to obtain the fused feature tensor. The process can be represented as: Fusion Feature Tensor = Image Feature Tensor + Broadcast(Feature Vectors of Feature Values).

[0061] Specifically, step S25 also includes: using the AdamW optimizer to handle weight decay and prevent overfitting during pre-training; using a cosine annealing scheduling strategy for its learning rate, and setting a warm-up period and a first learning rate, preferably 1×10^−4; setting the batch size according to GPU memory, preferably 4 or 8; setting the first training cycle of pre-training based on the amount of unlabeled data, since the first training cycle is usually trained on a large amount of unlabeled data, it is set to 100-400 cycles until the reconstruction loss converges.

[0062] Specifically, in step S26, all weight parameters of the pre-trained image encoder and feature value encoder are frozen. This means that during the fine-tuning phase, the backpropagation algorithm will not update these parameters, thus preserving the general feature representations learned in pre-training. Only the parameters of the subsequent prediction head, including the global average pooling layer GAP and the linear regression layer PPG estimator, are trained. The purpose of this step is to minimize the error between the predicted and true values ​​of the portal vein pressure gradient.

[0063] Specifically, step S27 includes: obtaining a second fusion feature Fi' by passing the stl file of the blood vessel and the corresponding feature values ​​through a frozen image encoder and feature value encoder; then concatenating all the second fusion features Fi' by F=Concat(Fi') to obtain a feature vector F; then performing global average pooling on F to obtain a global feature vector g; and inputting g into the newly initialized linear regression layer to obtain the portal vein pressure gradient prediction value. =W⋅g+b, where W represents the weight matrix and b represents the bias term; through the formula Calculate the predicted value of portal vein pressure gradient Compared with the true value of portal vein pressure gradient Differences between , where N represents the batch size. L1 Loss is less sensitive to outliers than mean squared error (L2 Loss), thus resulting in more robust regression performance.

[0064] Specifically, step S27 further includes: using the AdamW optimizer during training and setting a second learning rate and a second training period, wherein the second learning rate is lower than the first learning rate, and the second learning rate is preferably set between 1×10−4 and 1×10−3 to accommodate the training of new layers; since most of the model parameters are frozen, only a small number of parameters need to be trained, so the convergence speed is fast, and the second training period is shorter than the first training period, usually set to 10-50 training periods; at the same time, during the training process, the mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (r) are monitored to monitor their performance on the validation set in order to prevent overfitting and select the best model.

[0065] Secondly, this invention discloses a portal vein pressure gradient acquisition system that integrates multimodal features, used in the aforementioned method for acquiring portal vein pressure gradients that integrates multimodal features, comprising:

[0066] The image processing module is used to acquire and preprocess the patient's CT images to obtain their relevant feature data.

[0067] The model building module is used to construct and train a multimodal feature model based on liver blood vessels in CT images.

[0068] The output module is used to obtain the portal vein pressure gradient value corresponding to the CT image through the trained model.

[0069] In summary, this application, through a multimodal fusion architecture and a two-stage training strategy, fully utilizes the geometric structure information, physical parameter information, and pathological features of the three-dimensional model to achieve high-precision and robust non-contact prediction of portal vein pressure gradient values ​​under the condition of only CT image vascular model input, thus realizing accurate and non-invasive acquisition of portal vein pressure gradient.

[0070] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for obtaining portal vein pressure gradient by fusing multimodal features, characterized in that, Includes the following steps: S1. Acquire the patient's CT images and preprocess them to obtain their relevant feature data; S2. Construct a multimodal feature model based on liver blood vessels in CT images; The preprocessed CT image data, relevant feature data, and corresponding portal vein pressure gradient label values ​​to be predicted are input into the model for training. S3. Input the preprocessed CT image data and related feature data into the trained model to make predictions and obtain the portal vein pressure gradient value corresponding to the CT image. Step S1 specifically includes the following steps: S11. Acquire the patient's CT images and preprocess them to obtain a CT dataset with uniform size and quality that meets preset standards. The preprocessing includes image enhancement and image processing. S12. Based on a preset threshold, obtain the mask of the vascular region from the CT dataset. Dicom. The file, which describes the vascular region including the portal vein, hepatic artery, and aorta, then obtains the background for each CT dataset. Dicom. document; S13, Background Dicom. Documents and masks Dicom. The file is converted to the first background. nii. Documents and the first mask nii. File, then the first background nii. Documents and the first mask nii. The file is cropped to a preset size to obtain a second background. nii. Document and Second Mask nii. document; S14, Second Background nii. Document and Second Mask nii. The file is converted to an STL file, and then feature values ​​Ti are extracted from the STL file. The feature values ​​Ti include the blood vessel centerline point number ID, the blood vessel centerline point x-axis coordinate X, the blood vessel centerline point y-axis coordinate Y, the blood vessel centerline point z-axis coordinate Z, the parent node FatherNode, the left child node LeftNode, the right child node RightNode, the curvature K, the angle Angle, the area S, the maximum diameter MaxR, and the minimum diameter MinR. S15. Extract other pathology-related radiomics features from CT images, including liver volume (LVi) and spleen volume (SVi). Step S2 specifically includes the following steps: S21. Data preprocessing: Based on the stl. files of portal vein, hepatic artery and aortic vessels, perform three-dimensional projection on them, that is, orthogonally project the three-dimensional models of the stl. files onto the three coordinate planes xoy, yoz and zox respectively to generate two-dimensional projection images. S22. Image standardization: The size of the two-dimensional projection image is uniformly adjusted to the preset resolution, and they are stacked along the channel dimension to obtain the synthetic image tensor I_i ∈ R^(3×H×W) corresponding to the i-th blood vessel, where H represents the height and W represents the width. S23. Construct a multimodal feature model, which includes an encoder, a ViT decoder, and a portal vein pressure gradient estimator; S24. Feature encoding: The encoder includes an image encoder and a feature value encoder. The synthesized image tensor I_i obtained in step S22 is input into the multimodal feature model. The image encoder segments it into a sequence of image blocks. After masking, the feature tensor E(Ii) ∈ R^(L×D) of the image is output, where L represents the length of the image block sequence and D represents the feature dimension. At the same time, the feature value Ti obtained in step S14 is converted into a vector representation by the feature value encoder, and the feature vector E(Ti) ∈ R^D of the feature value is output. S25. Pre-training: The model is trained using a self-supervised learning method so that it can reconstruct the masked image patch based on the features of partially visible image patches and text. S26. Fine-tuning stage: Using the pre-trained encoder features, a multilayer perceptron is trained based on data with portal vein pressure gradient labels to predict portal vein pressure gradient values. S27. Input the preprocessed CT image data, relevant feature data and the corresponding portal vein pressure gradient label value to be predicted into the model for training to obtain a trained multimodal feature model.

2. The method for obtaining portal vein pressure gradient by fusing multimodal features according to claim 1, characterized in that, The masking process described in step S24 includes: dividing the image into blocks, selecting a preset number and size of block regions, randomly selecting block regions for masking to make them the learning regions; randomly selecting a proportion of image blocks, replacing their pixel values ​​with learnable mask tokens, and recording their original pixel values ​​Ptrue.

3. The method for obtaining portal vein pressure gradient by fusing multimodal features according to claim 2, characterized in that, Step S25 specifically includes: fusing the image's feature tensor E(Ii) and the feature vector E(Ti) of the eigenvalues ​​through the feature fusion module to obtain a first fused feature Fi; then inputting the first fused feature Fi and the mask token into the ViT decoder; outputting the reconstructed pixel values ​​of all image blocks through the ViT decoder, and using the formula... Calculate pixel values ​​of reconstructed image blocks Compared with the actual pixel value Error between ,in This represents the total number of masked image patches in a batch; pre-training is repeated on different vascular models.

4. The method for obtaining portal vein pressure gradient by fusing multimodal features according to claim 3, characterized in that, Step S25 also includes: using the AdamW optimizer to handle weight decay and prevent overfitting during pre-training; using a cosine annealing scheduling strategy for its learning rate and setting a warm-up period and a first learning rate; setting the batch size according to GPU memory; and setting the first training cycle of pre-training based on the amount of unlabeled data until the reconstruction loss converges.

5. The method for obtaining portal vein pressure gradient by fusing multimodal features according to claim 4, characterized in that: In step S26, all weight parameters of the pre-trained image encoder and eigenvalue encoder are frozen, and only the subsequent prediction head, including the parameters of the global average pooling layer GAP and the linear regression layer PPG estimator, is trained.

6. The method for obtaining portal vein pressure gradient by fusing multimodal features according to claim 5, characterized in that: Step S27 specifically includes: obtaining a second fusion feature Fi' from the blood vessel's STL file and corresponding feature values ​​using a frozen image encoder and feature value encoder; then concatenating all second fusion features Fi' using F=Concat(Fi') to obtain a feature vector F; then performing global average pooling on F to obtain a global feature vector g; and finally inputting g into the newly initialized linear regression layer to obtain the predicted portal vein pressure gradient value. =W⋅g+b, where W represents the weight matrix and b represents the bias term; through the formula Calculate the predicted value of portal vein pressure gradient Compared with the true value of portal vein pressure gradient Differences between , where N represents the batch size.

7. The method for obtaining portal vein pressure gradient by fusing multimodal features according to claim 6, characterized in that: Step S27 further includes: using the AdamW optimizer during training, and setting a second learning rate and a second training period, wherein the second learning rate is lower than the first learning rate and the second training period is shorter than the first training period; and monitoring the mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (r) during training.

8. A portal vein pressure gradient acquisition system integrating multimodal features, used in the portal vein pressure gradient acquisition method integrating multimodal features as described in any one of claims 1-7, characterized in that, include: The image processing module is used to acquire and preprocess the patient's CT images to obtain their relevant feature data. The model building module is used to construct and train a multimodal feature model based on liver blood vessels in CT images. The output module is used to obtain the portal vein pressure gradient value corresponding to the CT image through the trained model.

Images