Hierarchical prediction method and system for liver transplantation postoperative complications based on ultrasonic image, and storage medium
By using time-series datasets based on ultrasound images and a dual-channel architecture model, the problem of insufficient accuracy in predicting post-liver transplantation complications in existing technologies has been solved, enabling earlier and more accurate complication grading prediction and enhancing its clinical application value.
Patent Information
- Application Number
- CN202511096185.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2025-11-21
AI Technical Summary
Existing technologies struggle to accurately predict post-liver transplant complications using longitudinally updated ultrasound images of the transplanted liver, and traditional models fail to capture higher-order correlations and nonlinear relationships, resulting in insufficient prediction accuracy.
A complication grading prediction method based on ultrasound images was adopted. By continuously acquiring ultrasound images of transplanted livers at preset time intervals, a time-series dataset was constructed. A dual-channel architecture model that integrates time-series features and image features, including a long short-term memory network and a convolutional neural network, was used in conjunction with an expert database for feature matching and grading prediction.
It improves the completeness and dynamism of data collection, enhances the accuracy of feature recognition, and improves the timeliness and reliability of prediction. It can identify the development trend of complications at an early stage and provide a reliable basis for postoperative intervention.
Smart Images
Figure CN120997155A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, and more specifically to a method, system, and storage medium for the graded prediction of post-liver transplant complications based on ultrasound images. Background Technology
[0002] Liver transplantation is a crucial means of saving patients with end-stage liver disease, but the global shortage of donor organs is severe, and many patients die while waiting. The scarcity of donor livers makes early and accurate prediction of post-transplant complications essential. While traditional regression models can directly explain the correlation between features and outcomes, they cannot capture higher-order associations and nonlinear relationships between features after liver transplantation. Therefore, machine learning techniques are increasingly being used to predict post-transplant complications and survival.
[0003] Ultrasound, due to its non-invasive, convenient, and real-time dynamic characteristics, has become the preferred imaging method for early assessment of the transplanted liver's condition during the post-transplant monitoring phase. It is highly valuable in predicting complications, providing direct evidence of liver-related complications and indicating systemic complications through hemodynamic changes. Existing studies have confirmed that artificial intelligence models constructed based on liver parenchyma, portal vein, and arteriovenous blood flow spectrum data from transplanted liver ultrasound images can diagnose early graft dysfunction. However, other studies have shown that incorporating ultrasound images has not improved the predictive efficiency of transplanted liver fibrosis.
[0004] Existing studies largely rely on clinical characteristics at single time points, neglecting the dynamic nature of these characteristics as they change over time after transplantation. This is partly due to the incomplete preservation of bedside ultrasound data, and partly because training stable AI prediction models on longitudinal time-series data is extremely challenging. Currently, no studies have used longitudinally updated transplanted liver ultrasound images to predict postoperative complications; therefore, developing such precise grading and prediction methods has significant clinical implications. Summary of the Invention
[0005] In view of this, the present invention provides a method, system and storage medium for the graded prediction of postoperative complications of liver transplantation based on ultrasound images, which provides accurate classification and management for the selection of subsequent treatment plans for patients after liver transplantation.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A method for grading and predicting post-liver transplant complications based on ultrasound images, comprising the following steps:
[0008] Ultrasound images of the transplanted liver were acquired at preset time intervals for several consecutive days. The acquired ultrasound images of the transplanted liver were preprocessed to select regions of interest.
[0009] Based on the region of interest, the ultrasound images of the transplanted liver are reconstructed, and a time-series dataset mapping the physiological condition to the ultrasound images of the transplanted liver is constructed by combining the current physiological condition of the patient.
[0010] The time series dataset is input into the complication grading prediction model, which then identifies features of the region of interest in the time series.
[0011] The changes in feature recognition results in the time series dataset are matched with the postoperative complication features in the expert database, and the postoperative complications of the current patient are predicted based on the matching degree.
[0012] Optionally, the complication grading prediction model adopts a dual-channel architecture that integrates temporal features and imaging features. The specific structure includes an encoder module, a feature extractor module, a matching prediction module, and a grading head module connected sequentially. The encoder module is constructed using a long short-term memory network and is responsible for modeling the temporal dynamic features of the input time-series dataset. By encoding the region of interest features and corresponding physiological status data of the transplanted liver ultrasound images over multiple consecutive days, it captures the dynamic patterns of trend, periodicity, and abrupt changes in features over time, outputting a high-dimensional temporal feature vector containing temporal correlation information. The feature extractor module is composed of a convolutional neural network, performing deep feature extraction on the region of interest of the transplanted liver ultrasound images at each time point in the time series, through multi-layer convolution and pooling operations. The model extracts texture, morphological, and hemodynamic features from image regions of the liver parenchyma, portal vein blood flow spectrum, hepatic artery blood flow spectrum, and hepatic vein blood flow spectrum to generate static image feature vectors. The model then fuses the temporal feature vectors output by the encoder with the static image feature vectors output by the feature extractor to form a comprehensive feature matrix containing multi-dimensional spatiotemporal information, which is input to the matching prediction module. The matching prediction module uses an attention mechanism to weight and strengthen key features in the comprehensive feature matrix, and then calculates similarity with preset postoperative complication feature templates of different levels in the expert database. The calculated matching result is input to the grading head module. Based on preset grading rules, the grading head module processes the input matching data and finally outputs the complication grading prediction result.
[0013] Optionally, the training process for the complication grading prediction model is as follows:
[0014] Collect historical medical data that has been processed in accordance with ethical and compliant principles, including medical data of patients with different levels of complications and medical data of patients without complications;
[0015] Medical data of patients with different levels of complications are labeled according to their level type to generate different level of labeled data. Medical data of patients without complications are not labeled and generate 0-label data to construct a training dataset containing input features and corresponding labels.
[0016] The training dataset was preprocessed, including screening and reconstructing regions of interest from ultrasound images of transplanted livers;
[0017] The preprocessed labeled data and unlabeled data are divided into training, validation, and test sets according to a certain ratio, and then input into the complication grading prediction model. During the training phase, the encoder module dynamically models the temporal features, and the feature extractor module extracts deep image features. The feature vectors output by both are fused and then input into the matching prediction module. The cross-entropy loss function is used to calculate the error between the prediction result and the true label. The model parameters are iteratively optimized through the backpropagation algorithm, with a focus on adjusting the temporal weights of the encoder and the convolution kernel parameters of the feature extractor.
[0018] After each training round, the model's prediction accuracy, recall, and F1 score are evaluated using the validation set, and an early stopping mechanism is used to prevent overfitting. After training is completed, the final performance is validated using the test set, and the feature matching results of the model output are calibrated using an expert database to generate a fully trained complication grading prediction model.
[0019] Optionally, the region of interest includes a grayscale image of the liver parenchyma, the blood flow spectrum of the main portal vein, the blood flow spectrum of the hepatic artery at the first hepatic hilum, and the blood flow spectrum of the hepatic vein; wherein, the middle hepatic vein is selected for whole liver transplantation, the right hepatic vein is selected for right liver transplantation, and the left hepatic vein is selected for left liver transplantation.
[0020] Optionally, the classification head module is a linear layer, and the model is optimized by calculating the cross-entropy loss with the actual results. The loss function is L. cross Including the following formulas:
[0021]
[0022] Where N represents the total number of training samples, y i Let represent the sample label of the i-th training sample, and let represent the label of the i-th training sample.
[0023] Optionally, during the matching process, a dynamic time warping algorithm is used to align the temporal feature trajectory matrix with the feature templates in the expert database to eliminate matching bias caused by differences in examination time intervals among different patients.
[0024] A grading and prediction system for post-liver transplant complications based on ultrasound images, comprising:
[0025] Region of Interest (ROI) Filtering Module: This module is used to collect ultrasound images of the transplanted liver after liver transplantation within a preset time interval. The images are continuously collected over multiple days. The acquired ultrasound images of the transplanted liver are preprocessed to filter regions of interest.
[0026] Time-series dataset construction module: used to reconstruct transplanted liver ultrasound images based on regions of interest, and to construct a time-series dataset that maps physiological conditions to transplanted liver ultrasound images, combined with the current physiological condition of the patient.
[0027] Feature recognition module: used to input time series datasets into the complication grading prediction model, which then performs feature recognition on regions of interest in the time series.
[0028] Postoperative complication prediction module: This module matches the changes in feature recognition results in time series datasets with postoperative complication features in the expert database, and predicts the postoperative complications of the current patient based on the matching degree.
[0029] A computer storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of any one of the methods for grading and predicting post-liver transplant complications based on ultrasound images.
[0030] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method, system, and storage medium for grading and predicting postoperative complications of liver transplantation based on ultrasound images, which has the following beneficial effects:
[0031] 1. Enhanced Data Completeness and Dynamism: By acquiring ultrasound images of the transplanted liver over multiple consecutive days at preset time intervals, the limitations of traditional single-time-point data are overcome, comprehensively capturing the imaging changes at different stages after liver transplantation. A time-series dataset is constructed by combining physiological conditions, fully integrating imaging information with patient physiological dynamics, providing more comprehensive basic data support for subsequent predictions, and solving the problem of insufficient prediction accuracy caused by incomplete data preservation in clinical practice.
[0032] 2. Enhanced accuracy and depth of feature recognition: By preprocessing images to select regions of interest and reconstructing images, the model effectively focuses on feature extraction of key structures in the transplanted liver (such as liver parenchyma and vascular blood flow spectrum), reducing interference from irrelevant information. The complication grading prediction model identifies features in regions of interest within time series data, capturing subtle trends in feature changes over time. Compared to traditional models, it is more likely to discover higher-order correlations and nonlinear relationships between features, thus improving the depth and accuracy of feature recognition.
[0033] 3. Improved predictive timeliness and tiered reliability: Based on a matching mechanism between time-series feature changes and expert database features, combined with specific matching methods such as dynamic time warping, cosine similarity calculation, and multi-level threshold verification, accurate tracking of complication feature changes is achieved. By assigning higher weight to recent features and verifying the pathological logical chain of feature associations, the development trend of complications can be identified early, and tiered predictions can be completed based on the matching degree, making the prediction results more consistent with clinical reality and providing a reliable basis for timely intervention and tiered diagnosis and treatment of postoperative complications.
[0034] 4. Leveraging the clinical advantages of ultrasound technology: Fully utilizing the non-invasive, convenient, and real-time dynamic characteristics of ultrasound, continuous bedside data acquisition can be achieved during postoperative monitoring without additional invasive procedures, reducing interference with the patient. Furthermore, feature analysis based on ultrasound images directly correlates with changes in the structure and function of the transplanted liver. Compared to models relying solely on clinical data, this allows for earlier detection of complication clues, further enhancing the clinical practicality and feasibility of prediction. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0036] Figure 1 This is a schematic diagram of the method flow provided by the present invention;
[0037] Figure 2 This is a schematic diagram of the system structure provided by the present invention. Detailed Implementation
[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.
[0039] This invention discloses a method for grading and predicting post-liver transplant complications based on ultrasound images, such as... Figure 1 As shown, it includes the following steps:
[0040] Step 1: Collect ultrasound images of the transplanted liver after liver transplantation at preset time intervals. Collect images continuously for several days. Perform image preprocessing on the collected ultrasound images of the transplanted liver and select regions of interest.
[0041] Step 2: Reconstruct the ultrasound images of the transplanted liver based on the region of interest, and construct a time-series dataset that maps the physiological condition to the ultrasound images of the transplanted liver, taking into account the current physiological condition of the patient.
[0042] Step 3: Input the time series dataset into the complication grading prediction model. The complication grading prediction model will identify features of the regions of interest in the time series.
[0043] Step 4: Match the changes in the feature recognition results in the time series dataset with the postoperative complication features in the expert database, and predict the postoperative complications of the current patient based on the matching degree.
[0044] Furthermore, in step one, this embodiment uses seven days as a node to collect ultrasound images of the transplanted liver after liver transplantation; the region of interest includes grayscale images of the liver parenchyma, blood flow spectrum of the main portal vein, blood flow spectrum of the hepatic artery at the first hepatic hilum, and blood flow spectrum of the hepatic vein; wherein, the middle hepatic vein is selected for whole liver transplantation, the right hepatic vein is selected for right liver transplantation, and the left hepatic vein is selected for left liver transplantation.
[0045] Furthermore, the complication grading prediction model in this embodiment adopts a dual-channel architecture that integrates temporal features and image features. Specifically, the structure includes an encoder module, a feature extractor module, a matching prediction module, and a grading head module connected sequentially. The encoder module is constructed using a long short-term memory network and is responsible for modeling the temporal dynamic features of the input time-series dataset. By encoding the region of interest features and corresponding physiological status data of the transplanted liver ultrasound images over multiple consecutive days, it captures the dynamic patterns of trend, periodicity, and abrupt changes in features over time, outputting a high-dimensional temporal feature vector containing temporal correlation information. The feature extractor module is composed of a convolutional neural network, performing deep feature extraction on the region of interest of the transplanted liver ultrasound images at each time point in the time series. This is achieved through multi-layer convolution and... Pooling operations extract texture, morphological, and hemodynamic features from the image regions of liver parenchyma, portal vein blood flow spectrum, hepatic artery blood flow spectrum, and hepatic vein blood flow spectrum to generate static image feature vectors. The model fuses the temporal feature vector output by the encoder with the static image feature vector output by the feature extractor to form a comprehensive feature matrix containing multi-dimensional spatiotemporal information, which is then input into the matching prediction module. The matching prediction module uses an attention mechanism to weight and strengthen key features in the comprehensive feature matrix, and then calculates similarity with preset postoperative complication feature templates of different levels in the expert database. The calculated matching degree results are input into the grading head module. The grading head module processes the input matching degree data based on preset grading rules and finally outputs the complication grading prediction results.
[0046] Furthermore, the training process for the complication grading prediction model in this embodiment is as follows:
[0047] Collect historical medical data that has been processed in accordance with ethical regulations, including medical data of patients with different levels of complications and medical data of patients without complications. The patient medical data includes ultrasound images of the transplanted liver over several consecutive days, corresponding physiological indicators and clinical diagnostic records.
[0048] Medical data of patients with different levels of complications are labeled according to the preset severity level standard of complications, generating labeled data from level 1 to n (n is the highest level). Medical data of patients without complications are labeled as 0-labeled data, and a training dataset containing input features and corresponding labels is constructed.
[0049] The training dataset was preprocessed, including screening and reconstructing regions of interest from ultrasound images of transplanted livers, constructing a temporally correlated feature matrix by combining physiological condition indicators, and using data augmentation techniques to rotate, scale, and noise-perturb the image data to balance the data distribution of different level labels.
[0050] The preprocessed labeled data and unlabeled data are divided into training, validation, and test sets according to a certain ratio, and then input into the complication grading prediction model. During the training phase, the encoder module dynamically models the temporal features, and the feature extractor module extracts deep image features. The feature vectors output by both are fused and then input into the matching prediction module. The cross-entropy loss function is used to calculate the error between the prediction result and the true label. The model parameters are iteratively optimized through the backpropagation algorithm, with a focus on adjusting the temporal weights of the encoder and the convolution kernel parameters of the feature extractor.
[0051] After each training round, the model's prediction accuracy, recall, and F1 score are evaluated using the validation set, and an early stopping mechanism is used to prevent overfitting. After training is completed, the final performance is validated using the test set, and the feature matching results of the model output are calibrated using an expert database to ensure the stability and accuracy of the model in predicting different levels of complications, ultimately forming a fully trained complication grading prediction model.
[0052] Furthermore, the classification head module is a linear layer, and the cross-entropy loss is calculated with the actual results to optimize the model, with the loss function L. cross Including the following formulas:
[0053]
[0054] Where N represents the total number of training samples, y i Let represent the sample label of the i-th training sample, and let represent the label of the i-th training sample.
[0055] Furthermore, in step four, the specific matching method between the changes in feature recognition results in the time-series dataset and the postoperative complication features in the expert database is as follows: First, dynamic feature vectors are extracted from the feature recognition results processed by the encoder and feature extractor in the time-series dataset. This vector contains the trend of liver parenchyma texture changes at different time points, the evolution of dynamic parameters of portal vein / hepatic artery / hepatic vein blood flow spectra (such as resistance index, blood flow velocity fluctuation amplitude), and the temporal correlation features of physiological status indicators, forming a temporal feature trajectory matrix. The expert database pre-stores typical feature templates corresponding to different levels of complications. Each template contains the feature change threshold range, the temporal order of key feature occurrence, and the correlation weight between features (such as the correlation coefficient between abnormal hepatic artery blood flow spectrum and the occurrence of biliary complications).
[0056] During the matching process, the Dynamic Time Warping (DTW) algorithm is used to align the temporal feature trajectory matrix with the feature templates in the expert database, eliminating matching biases caused by differences in examination time intervals among different patients. Simultaneously, cosine similarity is used to calculate the similarity score between the two in the high-dimensional feature space, focusing on the matching degree of key features (such as a sudden increase in intrahepatic vascular resistance index and heterogeneous enhancement of liver parenchyma echo). To address the dynamic nature of feature changes, a time decay factor is introduced, assigning higher weight to recent feature changes and strengthening the influence of the latest ultrasound imaging features in the matching process. Furthermore, multi-level matching thresholds are set. When the similarity score exceeds the first-level threshold, feature association verification is triggered. A graph neural network (GNN) is used to calculate the matching degree of higher-order associations between features, verifying whether the combination of feature changes conforms to the pathological logical chain of complication development in the expert database (such as the association path of abnormal hepatic venous blood flow spectrum → portal hypertension → deterioration of liver function indicators). When the comprehensive matching score exceeds the second-level threshold, a secondary calibration is performed using a database of typical cases annotated by experts. Finally, the grade of postoperative complications for the current patient is determined based on the weighted total matching degree, achieving accurate graded matching prediction.
[0057] and Figure 1 Corresponding to the method shown, this invention also discloses a grading prediction system for post-liver transplant complications based on ultrasound images, used for... Figure 1 The implementation of the method, specifically its structure, is as follows: Figure 2 As shown, it includes:
[0058] Region of Interest (ROI) Filtering Module: This module is used to collect ultrasound images of the transplanted liver after liver transplantation within a preset time interval. The images are continuously collected over multiple days. The acquired ultrasound images of the transplanted liver are preprocessed to filter regions of interest.
[0059] Time-series dataset construction module: used to reconstruct transplanted liver ultrasound images based on regions of interest, and to construct a time-series dataset that maps physiological conditions to transplanted liver ultrasound images, combined with the current physiological condition of the patient.
[0060] Feature recognition module: used to input time series datasets into the complication grading prediction model, which then performs feature recognition on regions of interest in the time series.
[0061] Postoperative complication prediction module: This module matches the changes in feature recognition results in time series datasets with postoperative complication features in the expert database, and predicts the postoperative complications of the current patient based on the matching degree.
[0062] A computer storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of any one of the methods for grading and predicting post-liver transplant complications based on ultrasound images.
[0063] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0064] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for grading and predicting post-liver transplant complications based on ultrasound images, characterized in that, Includes the following steps: Ultrasound images of the transplanted liver were acquired at preset time intervals for several consecutive days. The acquired ultrasound images of the transplanted liver were preprocessed to select regions of interest. Based on the region of interest, the ultrasound images of the transplanted liver are reconstructed, and a time-series dataset mapping the physiological condition to the ultrasound images of the transplanted liver is constructed by combining the current physiological condition of the patient. The time series dataset is input into the complication grading prediction model, which then identifies features of the region of interest in the time series. The changes in feature recognition results in the time series dataset are matched with the postoperative complication features in the expert database, and the postoperative complications of the current patient are predicted based on the matching degree.
2. The method for grading and predicting post-liver transplantation complications based on ultrasound images according to claim 1, characterized in that, The complication grading prediction model adopts a dual-channel architecture that integrates temporal and imaging features. Specifically, the structure includes an encoder module, a feature extractor module, a matching prediction module, and a grading head module connected sequentially. The encoder module, constructed using a long short-term memory network, is responsible for modeling the temporal dynamic features of the input time-series dataset. By encoding the region of interest features of transplanted liver ultrasound images over multiple consecutive days and the corresponding physiological status data time-by-time, it captures the dynamic patterns of trend, periodicity, and abrupt changes in features over time, outputting a high-dimensional temporal feature vector containing temporal correlation information. The feature extractor module, composed of a convolutional neural network, targets the region of interest in the transplanted liver ultrasound images at each time point in the time series. Deep feature extraction is performed by extracting texture, morphological, and hemodynamic features from the image regions of liver parenchyma, portal vein blood flow spectrum, hepatic artery blood flow spectrum, and hepatic vein blood flow spectrum through multi-layer convolution and pooling operations, generating static image feature vectors. The model fuses the temporal feature vectors output by the encoder with the static image feature vectors output by the feature extractor to form a comprehensive feature matrix containing multi-dimensional spatiotemporal information, which is then input into the matching prediction module. The matching prediction module uses an attention mechanism to weight and strengthen the key features in the comprehensive feature matrix, and then calculates the similarity with the preset postoperative complication feature templates of different levels in the expert database. The calculated matching result is then input into the hierarchical head module. The grading head module processes the input matching data based on preset grading rules and finally outputs the complication grading prediction results.
3. A method for grading and predicting post-liver transplant complications based on ultrasound images according to claim 1 or 2, characterized in that, The training process for the complication grading prediction model is as follows: Collect historical medical data that has been processed in accordance with ethical and compliant principles, including medical data of patients with different levels of complications and medical data of patients without complications; Medical data of patients with different levels of complications are labeled according to their level type to generate different level labeled data. Medical data of patients without complications are not labeled and generate 0-label data to construct a training dataset containing input features and corresponding labels. The training dataset was preprocessed, including screening and reconstructing regions of interest from ultrasound images of transplanted livers; The preprocessed labeled data and unlabeled data were divided into training set, validation set and test set according to the proportion, and then input into the complication grading prediction model respectively; During the training phase, the encoder module dynamically models the temporal features, and the feature extractor module extracts deep features from the image. The feature vectors output by both modules are fused and then fed into the matching prediction module. The cross-entropy loss function is used to calculate the error between the prediction result and the true label. The model parameters are iteratively optimized through the backpropagation algorithm, with a focus on adjusting the temporal weights of the encoder and the convolution kernel parameters of the feature extractor. After each training round, the model's prediction accuracy, recall, and F1 score are evaluated using the validation set, and an early stopping mechanism is used to prevent overfitting. After training is completed, the final performance is validated using the test set, and the feature matching results of the model output are calibrated using an expert database to generate a fully trained complication grading prediction model.
4. The method for grading and predicting post-liver transplantation complications based on ultrasound images according to claim 1, characterized in that, The regions of interest include grayscale images of the liver parenchyma, the blood flow spectrum of the main portal vein, the blood flow spectrum of the hepatic artery at the first hepatic hilum, and the blood flow spectrum of the hepatic vein; among them, the middle hepatic vein is selected for whole liver transplantation, the right hepatic vein is selected for right liver transplantation, and the left hepatic vein is selected for left liver transplantation.
5. The method for grading and predicting post-liver transplantation complications based on ultrasound images according to claim 2, characterized in that, The classification head module is a linear layer. It calculates cross-entropy loss with the actual results to optimize the model. The loss function is L. cross Including the following formulas: Where N represents the total number of training samples, y i p represents the sample label of the i-th training sample. i This represents the label of the i-th training sample.
6. The method for grading and predicting post-liver transplantation complications based on ultrasound images according to claim 1, characterized in that, During the matching process, a dynamic time warping algorithm is used to align the temporal feature trajectory matrix with the feature templates in the expert database to eliminate matching bias caused by differences in examination time intervals among different patients.
7. A grading and prediction system for post-liver transplant complications based on ultrasound images, characterized in that, include: Region of Interest (ROI) Filtering Module: This module is used to collect ultrasound images of the transplanted liver after liver transplantation within a preset time interval. The images are continuously collected over multiple days. The acquired ultrasound images of the transplanted liver are preprocessed to filter regions of interest. Time-series dataset construction module: used to reconstruct transplanted liver ultrasound images based on regions of interest, and to construct a time-series dataset that maps physiological conditions to transplanted liver ultrasound images, combined with the current physiological condition of the patient. Feature recognition module: used to input time series datasets into the complication grading prediction model, which then performs feature recognition on regions of interest in the time series. Postoperative complication prediction module: This module matches the changes in feature recognition results in time series datasets with postoperative complication features in the expert database, and predicts the postoperative complications of the current patient based on the matching degree.
8. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for grading and predicting post-liver transplant complications based on ultrasound images as described in any one of claims 1-6.