Fatty liver disease intelligent quantitative diagnosis model based on deep combination of multi-mode ultrasound and machine learning and application of fatty liver disease intelligent quantitative diagnosis model

Abstract

The invention discloses a fatty liver disease intelligent quantitative diagnosis model based on deep combination of multi-modal ultrasound and machine learning and application of the fatty liver disease intelligent quantitative diagnosis model based on deep combination of multi-modal ultrasound and machine learning. The four-dimensional ultrasonic sequence data acquisition module is used for acquiring four-dimensional ultrasonic sequence data, shear wave elastography data, ultrasonic fat fraction data, hemodynamic parameters and clinical data of the liver, and the spatial-temporal feature extraction module is used for performing spatial-temporal feature coding on a four-dimensional ultrasonic sequence, capturing spatial texture features and time dynamic features of liver tissues and extracting the spatial texture features and the time dynamic features of the liver tissues. Relates to the technical field of medical image diagnosis, the diagnosis model synchronously obtains multi-dimensional data of liver disease detection through a dynamic multi-modal data acquisition module, captures spatial texture and time dynamic characteristics of liver tissues by utilizing a spatial-temporal characteristic extraction mode, and comprehensively characterizes the physiological state of the liver, so that the diagnosis accuracy is improved. Dynamic capture of fatty liver disease pathological characteristics is realized, the recognition ability of the diagnosis model to early lesions is improved, and a stronger diagnosis effect is provided for disease progress monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image diagnosis, in particular to a fat liver disease intelligent quantitative diagnosis model combining multi-modal ultrasound and machine learning in depth and application. BACKGROUND

[0002] Fatty liver disease is a disease caused by abnormal accumulation of fat in the liver (more than 5% to 10% of the weight of the liver), which is divided into alcoholic fatty liver disease (related to alcohol consumption) and non-alcoholic fatty liver disease (related to metabolic abnormalities). Early symptoms may be asymptomatic, and severe cases may develop into hepatitis, cirrhosis and even liver cancer, which requires prevention and treatment through lifestyle adjustment and cause control. The fat liver disease intelligent quantitative diagnosis analysis is one of the methods for diagnosing fatty liver by quantifying the fat content in the liver using magnetic resonance special imaging technology.

[0003] Because the existing ultrasound diagnosis technology for fatty liver disease relies on static single-modal data, it lacks explicit modeling of the causal relationship between the spatial and temporal dynamic characteristics of the liver and the multi-modal characteristics, and the model parameters cannot be adaptively adjusted according to the individual characteristics of the patient. At the same time, there is a problem of not real-time data quality control, which limits the accuracy of diagnosis and the generalization ability, and it is difficult to meet the clinical demand for accurate grading and individualized diagnosis of fatty liver disease, resulting in lagging behind in early pathological process recognition, limited application in primary medical care and other technical bottlenecks. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a fat liver disease intelligent quantitative diagnosis model combining multi-modal ultrasound and machine learning in depth and application, which solves the problems mentioned in the background art.

[0005] To achieve the above purpose, the present application realizes the technical scheme as follows: a fat liver disease intelligent quantitative diagnosis model combining multi-modal ultrasound and machine learning in depth, the quantitative diagnosis model comprising:

[0006] A dynamic multi-modal data acquisition module is used to acquire four-dimensional ultrasound sequence data, shear wave elastography data, ultrasound fat fraction data, hemodynamic parameters and clinical data of the liver;

[0007] A spatio-temporal feature extraction module is used to perform spatio-temporal feature coding on the four-dimensional ultrasound sequence to capture the spatial texture features and temporal dynamic features of the liver tissue;

[0008] A cross-modal causal reasoning module is used to construct a fat liver disease pathological process causal relationship model based on domain knowledge, explicitly model the causal dependency relationship between fatty degeneration, fibrosis and inflammation, and mine the causal relationship between multi-modal features and pathological results;

[0009] The personalized adaptive learning module dynamically adjusts the model parameters in combination with the individual characteristics of the patient, thereby realizing individualized adaptation of the diagnostic model.

[0010] The explainability enhancement module presents the basis for model decision-making through visualization technology and quantifies the contribution of each modality feature to the diagnostic result.

[0011] The real-time feedback and quality control module is used for scanning quality evaluation, probe position navigation, and data quality screening.

[0012] The federated learning and edge computing module is used for collaborative training and lightweight deployment of the model under the condition of multi-center data privacy protection.

[0013] Preferably, the spatio-temporal feature extraction module adopts a combined architecture of 3DCNN and ConvLSTM network to perform spatio-temporal feature coding on the four-dimensional ultrasound sequence and output a feature vector containing spatial texture and temporal dynamic information.

[0014] The personalized adaptive learning module encodes the individual characteristics of the patient into a parameter vector through conditional normalization and dynamically adjusts the normalization parameters of the model convolution layer, thereby realizing adaptive optimization of the model for different patients.

[0015] The explainability enhancement module generates a causal heat map through Grad-CAM, visualizes the liver region of interest for decision-making, and performs causal decomposition to quantify the direct and indirect causal contribution of each modality feature to the diagnostic result.

[0016] Preferably, the spatial convolution operation algorithm of 3DCNN in the spatio-temporal feature extraction module is specifically as follows:

[0017]

[0018] where F l is the l-th feature map, W l-1 is the convolution kernel weight, σ is the activation function, a, b, and c are the spatial dimensions of the convolution kernel, and i, j, and k are the spatial coordinates of the feature map.

[0019] Preferably, the processing of the time series feature by ConvLSTM in the spatio-temporal feature extraction module is specifically as follows:

[0020]

[0021] where i t , f t , and o t are the input gate, the forget gate, and the output gate, respectively, C t is the cell state, H t is the hidden state, and X tis the input feature at time t, and is an element-wise multiplication.

[0022] Preferably, the real-time feedback and quality control module analyzes the image entropy, edge definition and other quality indicators of the ultrasound scan section in real time, generates a red-green-yellow three-color quality score and prompts the probe to adjust the direction, and automatically removes low-quality data.

[0023] Preferably, the federated learning and edge computing module deploys a lightweight model on the edge by compressing the complex model on the cloud, realizes real-time inference of the ultrasound device, and protects the privacy of multi-center data through federated parameter aggregation.

[0024] Preferably, the dynamic multi-modal data acquisition module further comprises a self-supervised data enhancement sub-module, which uses unlabeled ultrasound data to improve the generalization ability of the model through contrastive learning and puzzle game tasks.

[0025] Preferably, the cross-modal causal reasoning module constructs a causal graph G=(V,E) of pathological variables of fatty liver disease based on the causal graph theory of medical field knowledge;

[0026] Where the node set V={S,F,G,X} contains the steatosis grade S, fibrosis grade F, inflammation grade G and multi-modal feature X, and the edge set E represents the direct causal relationship between variables.

[0027] The causal effect of the intervention variable do(X) on the outcome variable Y is calculated as follows:

[0028] P(Y=y|do(X=x))=∑ Z P(Z=z|X=x)·P(Y=y|X=x,z=z)

[0029] Where Z is the set of confounding variables.

[0030] Preferably, the individualized adaptive learning module realizes dynamic adjustment of model parameters through conditional normalization algorithm, and the normalization process of the lth layer feature Fl is specifically as follows:

[0031]

[0032] Where γ(c) and β(c) are learnable scaling factors and offset factors dependent on patient feature c, μ(F l ) and σ(F l ) are the mean and standard deviation of the feature.

[0033] The application of the multi-modal ultrasound and machine learning deep combination fatty liver disease intelligent quantitative diagnosis model integrates the diagnosis model into a clinical application system composed of an ultrasound device and a cloud platform for fatty liver disease intelligent quantitative diagnosis.

[0034] The application provides an intelligent quantitative diagnosis model of fatty liver disease based on deep combination of multi-modal ultrasound and machine learning, and application thereof.

[0035] (1) The diagnosis model of the application synchronously acquires multi-dimensional data for liver disease detection through a dynamic multi-modal data acquisition module, and captures spatial texture and time dynamic characteristics of liver tissue by using a spatiotemporal feature extraction method, so as to comprehensively represent the physiological state of the liver, break through the limitations of traditional static single-modal data, realize dynamic capture of pathological characteristics of fatty liver disease, and improve the recognition ability of the diagnosis model for early subtle lesions, thereby providing more abundant information support for disease progression monitoring.

[0036] (2) The multi-modal feature and pathological result causal association is explicitly modeled, and individualized diagnosis adaptation is realized by constructing a fatty liver disease pathological process causal relationship model based on domain knowledge through a cross-modal causal reasoning module, and dynamically adjusting model parameters according to individual characteristics of patients through an individualized adaptive learning module, so as to effectively solve the problems of insufficient feature correlation analysis and weak model generalization ability in the prior art, make the diagnosis result more suitable for the actual pathological state of the patient, and improve the diagnosis accuracy of complex cases.

[0037] (3) The real-time feedback and quality control module analyzes the scanning quality in real time and provides probe adjustment suggestions, the federated learning and edge computing module realizes multi-center data privacy protection and lightweight deployment, so as to improve the data acquisition quality and the adaptability of the base-level equipment, solve the problems of data quality control lag, multi-center data sharing security risks and base-level application limitations in traditional ultrasound diagnosis, and enable the model to realize bedside real-time diagnosis on a portable ultrasound device, thereby promoting the sinking of fatty liver disease screening to base-level medical scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0038] Fig. 1 The diagnosis model framework diagram of the application of the intelligent quantitative diagnosis model of fatty liver disease based on deep combination of multi-modal ultrasound and machine learning.

[0039] Fig. 2 The diagnosis method flowchart of the application of the intelligent quantitative diagnosis model of fatty liver disease based on deep combination of multi-modal ultrasound and machine learning. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0041] Embodiment 1

[0042] Referring to Figs. 1-2 The application provides an intelligent quantitative diagnosis model of fatty liver disease based on deep combination of multi-modal ultrasound and machine learning, and application thereof. To achieve the above object, the application is implemented by the following technical scheme: the intelligent quantitative diagnosis model of fatty liver disease based on deep combination of multi-modal ultrasound and machine learning comprises:

[0043] A dynamic multi-modal data acquisition module is configured to acquire four-dimensional ultrasound sequence data, shear wave elastography data, ultrasound fat fraction data, hemodynamic parameters and clinical data of a liver;

[0044] A spatio-temporal feature extraction module is configured to perform spatio-temporal feature coding on the four-dimensional ultrasound sequence to capture spatial texture features and temporal dynamic features of liver tissue;

[0045] A cross-modal causal reasoning module is configured to construct a causal relationship model of pathological processes of fatty liver disease based on domain knowledge, explicitly model causal dependency relationships among fatty degeneration, fibrosis and inflammation, and mine causal correlations between multi-modal features and pathological results;

[0046] A personalized adaptive learning module is configured to dynamically adjust model parameters in combination with individual characteristics of a patient to realize individualized adaptation of the diagnosis model;

[0047] An explainability enhancement module is configured to present model decision basis through visualization technology and quantify contribution degrees of each modality feature to a diagnosis result;

[0048] A real-time feedback and quality control module is configured to perform scanning quality evaluation, probe position navigation and data quality screening;

[0049] A federated learning and edge computing module is configured to perform collaborative training and lightweight deployment of the model under multi-center data privacy protection.

[0050] Application of the intelligent quantitative diagnosis model of fatty liver disease based on deep combination of multi-modal ultrasound and machine learning integrates the diagnosis model into a clinical application system composed of an ultrasound device and a cloud platform to realize intelligent quantitative diagnosis of fatty liver disease and achieve the following functions:

[0051] Real-time quantitative diagnosis of fatty liver disease based on four-dimensional ultrasound sequence and hemodynamic parameters;

[0052] Pathological grading and risk prediction combined with causal reasoning and personalized adaptation;

[0053] Improvement of ultrasound data acquisition quality through real-time quality control feedback;

[0054] Multi-center data sharing and continuous optimization of the model through federated learning;

[0055] Generate a visual diagnostic report containing causal thermodynamic maps, feature contribution, and personalized intervention suggestions.

[0056] In this embodiment, the dynamic multi-modal data acquisition module is configured to synchronously acquire four-dimensional ultrasound sequence data of the liver, including B-mode grayscale image sequences, shear wave elastography elastic modulus dynamic distribution sequences, real-time ultrasound fat fraction measurement sequences, color Doppler ultrasound hemodynamic parameters, including portal vein blood flow velocity and hepatic arterial resistance index, patient clinical data, including patient BMI, blood lipids, and liver function indicators, and to perform timestamp alignment and spatial positioning calibration on the multi-modal data to ensure that each modality data corresponds to the same anatomical region and physiological state.

[0057] The real-time feedback and quality control module is installed and deployed on the ultrasound device, analyzes the image quality of the current scanning section in real time, provides probe position adjustment suggestions through the human-machine interaction interface, and automatically filters low-quality data to improve the reliability of subsequent modeling data.

[0058] The federated learning and edge computing module supports collaborative training of multi-center clinical data under privacy protection, deploys a lightweight inference model on the portable ultrasound device end through edge computing technology, realizes real-time diagnosis and analysis of ultrasound data on the patient's bedside ultrasound device, and simultaneously performs global optimization and updating of model parameters through the cloud server to ensure the accuracy of deep learning model parameters.

[0059] Through the ultrasound diagnostic instrument, high-frame-rate acquisition of four-dimensional ultrasound dynamic sequences is realized, and combined with color Doppler blood flow parameters, the elastic modulus fluctuation and blood perfusion change of the liver within the respiratory cycle can be captured. Compared with traditional static single-mode ultrasound, dynamic multi-modal data can cover multiple-dimensional pathological characteristics such as liver parenchyma echo enhancement caused by fatty degeneration, elastic modulus increase caused by fibrosis, and blood flow resistance increase accompanied by inflammation, etc., so as to improve the recognition rate of early fatty degeneration of the model, provide more comprehensive diagnostic basis for the clinic, and improve the accuracy and efficiency of intelligent quantitative diagnosis of fatty liver disease.

[0060] Through multi-dimensional technical innovation, the precision and intelligence of fatty liver disease diagnosis are realized. At the level of dynamic multi-modal data acquisition and overall architecture, through synchronous acquisition of four-dimensional ultrasound sequences, blood flow hemodynamic parameters, and clinical data, the limitations of traditional static single-mode data are broken through, so that the model can capture the elastic fluctuation and blood perfusion change of the liver within the respiratory cycle, etc. The recognition rate of early fatty degeneration is improved, and comprehensive diagnostic basis covering fatty degeneration, fibrosis, inflammation, and other multi-dimensional pathological characteristics is provided for the clinic.

[0061] Embodiment 2

[0062] Specifically, the spatio-temporal feature extraction module adopts the architecture combining 3DCNN and ConvLSTM network to perform spatio-temporal feature coding on the four-dimensional ultrasound sequence, and outputs a feature vector containing spatial texture and time dynamic information.

[0063] The personalized adaptive learning module encodes the individual characteristics of the patient into a parameter vector through conditional normalization, dynamically adjusts the normalization parameters of the model convolution layer, and realizes the adaptive optimization of the model for different patients.

[0064] The explainability enhancement module generates a causal heat map through Grad-CAM, visualizes the liver region focused by the decision, and performs causal decomposition to quantify the direct and indirect causal contributions of each modality feature to the diagnosis result.

[0065] The specific formula of the spatial convolution operation algorithm of 3DCNN in the spatio-temporal feature extraction module is as follows:

[0066]

[0067] wherein, F l is the lth feature map, W l-1 is the convolution kernel weight, σ is the activation function, a, b, and c are the spatial dimensions of the convolution kernel, and i, j, and k are the spatial coordinates of the feature map.

[0068] The processing of the time sequence feature by ConvLSTM in the spatio-temporal feature extraction module is specifically as follows:

[0069]

[0070] wherein, i t , f t , and o t are the input gate, the forget gate, and the output gate, respectively, C t is the cell state, H t is the hidden state, X t is the input feature at t time, and is the element-wise multiplication.

[0071] In this embodiment, the spatial convolution operation of 3DCNN (3D convolutional neural network) performs three-dimensional convolution on each frame of B-ultrasound image of the four-dimensional ultrasound sequence using a 3x3x3 convolution kernel, an activation function, extracts liver parenchyma texture features, and the liver parenchyma texture features include echo intensity standard deviation and vascular edge blurring degree and other image contents. Taking 5 consecutive images as a three-dimensional block, a 2048-dimensional feature vector is output, and then the feature vector output by 3DCNN is input in time sequence through ConvLSTM (convolutional long short-term memory network) time sequence processing, the forget gate threshold is set to 0.7, the liver displacement caused by respiratory motion and the periodic change of elastic modulus are captured, the hidden state dimension is set to 512, and a feature vector containing time dynamic information is output.

[0072] By the conditional normalization algorithm, the patient BMI is encoded into a parameter vector, and the normalization parameters of the convolution layer are dynamically adjusted. When the patient BMI > 28, the model automatically enhances the feature weight of the posterior segment of the right lobe of the liver to adapt to the attenuation characteristics of the ultrasound image of the obese patient.

[0073] Therefore, when processing four-dimensional ultrasound sequences with the 3DCNN+ConvLSTM architecture, the 3DCNN can extract the spatial texture consistency of different sections of the liver, and the ConvLSTM can capture the variation law of the elastic modulus with respiration. Combined with the adaptation of the conditional normalization to obese patients, the model improves the accuracy of fatty degeneration grading in patients with BMI > 30, especially in the case of image quality degradation caused by narrow intercostal space, and the self-adaptive adjustment can improve the reliability of diagnosis.

[0074] The deep combination of spatio-temporal feature extraction and personalized adaptive learning encodes the spatio-temporal features of the ultrasound sequence through the 3DCNN and ConvLSTM architecture, which can accurately capture the liver parenchyma texture consistency and dynamic variation law of the elastic modulus. Combined with the adaptive adjustment of the conditional normalization algorithm to obese, liver anatomical variation and other special cases, the accuracy of fatty degeneration grading is improved, the reliability of diagnosis in complex scenarios such as narrow intercostal space is improved, and the problem of insufficient generalization ability of traditional models is effectively solved.

[0075] Embodiment 3

[0076] Specifically, the real-time feedback and quality control module analyzes the image entropy, edge definition and other quality indicators of the ultrasound scanning section in real time, generates a red-green-yellow three-color quality score and prompts the probe adjustment direction, and automatically removes low-quality data.

[0077] The federated learning and edge computing module compresses the complex model on the cloud side, deploys a lightweight model on the edge side, realizes real-time inference of the ultrasound device, and protects the privacy of multi-center data through federated parameter aggregation.

[0078] The dynamic multi-modal data acquisition module further includes a self-supervised data enhancement sub-module that uses unlabeled ultrasound data to improve the model's generalization ability through contrastive learning and puzzle game tasks.

[0079] In this embodiment, the image entropy calculation of the real-time feedback and quality control module uses the Shannon entropy formula:

[0080]

[0081] When the entropy value is greater than 7.5, it is determined that the image is of low quality, and at this time, edge sharpness evaluation needs to be performed, the gradient amplitude of the liver capsule edge is calculated by a Canny operator, and when the amplitude is less than 15, it is prompted to adjust the probe angle, wherein the image score standard is that a quality score greater than 85 is green, i.e., qualified, a quality score of 60-85 is yellow, i.e., needing adjustment, and a quality score less than 60 is red, i.e., rejected;

[0082] Therefore, in the diagnosis process, for example, when the probe position causes the right lobe of the liver to be incompletely displayed, the system prompts to "move 1 cm to the right and up", and after adjustment, the image entropy decreases from 9.2 to 6.8, and the edge gradient amplitude increases from 12 to 22, meeting the diagnostic requirements.

[0083] Example 4

[0084] Specifically, the cross-modal causal reasoning module constructs a causal graph G=(V, E) of pathological variables of fatty liver disease based on the causal graph theory of medical field knowledge;

[0085] Wherein the node set V={S, F, G, X} contains steatosis grading S, fibrosis grading F, inflammation grading G and multi-modal feature X, and the edge set E represents the direct causal relationship between variables;

[0086] The causal effect of the intervention variable do(X) on the result variable Y is calculated, and the specific calculation formula is as follows:

[0087] P(Y=y|do(X=x))=∑ Z P(Z=z|X=x)·P(Y=y|X=x,Z=z)

[0088] Wherein, Z is a set of confounding variables.

[0089] The personalized adaptive learning module realizes dynamic adjustment of model parameters through conditional normalization algorithm, and the normalization process of the lth layer feature Fl is specifically as follows:

[0090]

[0091] Wherein, γ(c) and β(c) are learnable scaling factors and offset factors depending on the patient feature c, μ(F l ) and σ(F l ) are the mean and standard deviation of the feature;

[0092] In this embodiment, the causal graph G=(V, E) is constructed, wherein steatosis S and fibrosis F are direct causal edges, inflammation G and fibrosis F are indirect causal edges, and multi-modal features X include B-ultrasound texture, SWE elasticity value and UDFF. When calculating the intervention effect, the confounding variable Z is controlled through the backdoor adjustment formula to quantify the direct causal effect of S on F.

[0093] Personalized adaptive learning, for the first layer feature Fl, for example, when the patient has atrophy of the right lobe of the liver, the feature scaling factor γ(c) is set to 1.3, and the offset factor β(c) is set to 0.2, which enhances the attention to the left lobe feature, avoiding the diagnostic bias caused by anatomical variation;

[0094] Through the theory of causal diagram, the diagnostic model can clearly determine the causal contribution of each modality feature, for example, in fibrosis grading, the direct causal effect of SWE elasticity value accounts for 45%, and the indirect causal effect of B-ultrasound texture accounts for 28%, providing interpretable diagnostic logic for clinical practice. For patients with anatomical variation of the liver, personalized parameter adjustment improves the accuracy of fibrosis staging, especially for the differential diagnosis of F3-F4 stage, reducing the misdiagnosis caused by abnormal anatomical structure, and ensuring the accuracy of the diagnostic results, providing more accurate information data for subsequent treatment of patients.

[0095] Specifically, in the intelligent quantitative diagnosis of fatty liver disease, the ultrasound equipment collects four-dimensional ultrasound sequence of the patient's liver, synchronously acquires B-ultrasound gray scale image, shear wave elastography data, ultrasound fat fraction and color Doppler hemodynamic parameters, and collects clinical data such as patient's BMI, ALT, etc. When collecting, the patient is required to be fasting for 8 hours, take supine position, and scan along the right intercostal to obtain the maximum section dynamic sequence of the right lobe of the liver. The multi-modal data are spatio-temporally aligned through time stamp, then the dynamic multi-modal data are input into the spatio-temporal feature extraction module, 3DCNN is used to perform three-dimensional convolution on continuous 5 frames of ultrasound images, liver parenchyma texture features are extracted through 3x3x3 convolution kernel, and ConvLSTM network is used for processing to capture the time dynamic changes of liver displacement and elastic modulus in the respiratory cycle, and the feature vector containing spatial texture and time dynamic information is output;

[0096] After the feature vector is transmitted to the cross-modal causal reasoning module, a causal relationship graph containing fatty degeneration grading, fibrosis grading, inflammation grading and multi-modal features is constructed based on medical field knowledge, and then the causal effect of each modality feature on the pathological result is calculated to determine the direct influence degree of fatty degeneration on fibrosis. At the same time, the personalized adaptive learning module encodes the individual characteristics such as patient's BMI and liver anatomical structure into parameter vector through conditional normalization algorithm, dynamically adjusts the normalization parameters of model convolution layer, for example, for obese patients with BMI>28, the weight of the feature of the posterior segment of the right lobe of the liver is automatically enhanced;

[0097] During data acquisition, the real-time feedback and quality control module synchronously analyzes the image entropy and edge definition of the ultrasound section, and generates a red-green-yellow three-color quality score. When the image entropy is greater than 7.5 or the edge gradient amplitude is less than 15, the doctor is prompted to adjust the probe position in real time, and low-quality data with a score less than 60 is automatically removed. The feature data after quality control is input into the lightweight model deployed on the ultrasound device. The model completes multi-center data collaborative training in the cloud through federated learning, aggregates the parameters of each center, and is compressed to 15MB, realizing real-time reasoning at the bedside within 1.5 seconds.

[0098] Then the diagnostic model inferences output the results of steatosis grading and fibrosis grading, and the explainability enhancement module generates a causal heat map to show the liver regions that the model focuses on. The SHAP value is used for causal decomposition to quantify the contribution of each modality feature. Finally, the clinical application system integrates the diagnostic results with hemodynamic parameters and clinical data to generate a report containing pathological grading, causal analysis, and personalized intervention suggestions. For example, for a patient with S3 steatosis, a diet adjustment and exercise program are pushed based on the patient's BMI and liver function indicators, realizing intelligent diagnosis of the whole process from data acquisition to clinical decision-making.

[0099] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application.

Claims

1. A smart quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning, characterized in that: The quantitative diagnostic model includes: The dynamic multimodal data acquisition module is used to acquire four-dimensional ultrasound sequence data, shear wave elastography data, ultrasound fat fraction data, hemodynamic parameters and clinical data of the liver. The spatiotemporal feature extraction module is used to encode the spatiotemporal features of four-dimensional ultrasound sequences and capture the spatial texture features and temporal dynamic features of liver tissue. The cross-modal causal reasoning module constructs a causal relationship model of the pathological process of fatty liver disease based on domain knowledge, explicitly models the causal dependence between steatosis, fibrosis, and inflammation, and explores the causal association between multimodal features and pathological outcomes. The personalized adaptive learning module dynamically adjusts the model parameters based on individual patient characteristics to achieve individualized adaptation of the diagnostic model. The interpretability enhancement module uses visualization technology to present the basis for model decision-making and quantifies the contribution of each modality feature to the diagnostic results. The real-time feedback and quality control module is used for scan quality assessment, probe position navigation, and data quality screening. The Federated Learning and Edge Computing module is used for collaborative model training and lightweight deployment under multi-center data privacy protection.

2. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 1, characterized in that: The spatiotemporal feature extraction module adopts an architecture combining 3DCNN and ConvLSTM networks to encode the spatiotemporal features of the four-dimensional ultrasound sequence and output a feature vector containing spatial texture and temporal dynamic information. The personalized adaptive learning module encodes individual patient features into parameter vectors through conditional normalization, dynamically adjusts the normalization parameters of the model's convolutional layers, and achieves adaptive optimization of the model for different patients. The interpretability enhancement module generates a causal heatmap through Grad-CAM, visualizes the liver region of interest in decision-making, and performs causal decomposition to quantify the direct and indirect causal contributions of each modality feature to the diagnostic results.

3. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 2, characterized in that: The specific formula for the spatial convolution operation algorithm of 3DCNN in the spatiotemporal feature extraction module is as follows: Among them, F l For the feature map of layer l, W l-1 σ represents the convolution kernel weights, σ ​​is the activation function, a, b, c are the spatial dimensions of the convolution kernel, and i, j, k are the spatial coordinates of the feature map.

4. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 2, characterized in that: The specific formula for processing time series features by ConvLSTM in the spatiotemporal feature extraction module is as follows: Among them, i t f t o t These are the input gate, forget gate, and output gate, respectively. (C) t In cellular state, H t In the hidden state, X t Let t be the input feature at time t, and ⊙ denotes element-wise multiplication.

5. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 1, characterized in that: The real-time feedback and quality control module analyzes quality indicators such as image entropy and edge sharpness of the ultrasound scan section in real time, generates a red, green and yellow three-color quality score and prompts the probe to adjust its direction, and automatically removes low-quality data.

6. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 1, characterized in that: The federated learning and edge computing module compresses complex models in the cloud and deploys lightweight models at the edge to achieve real-time inference for ultrasound equipment, and protects multi-center data privacy through federated parameter aggregation.

7. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 1, characterized in that: The dynamic multimodal data acquisition module also includes a self-supervised data augmentation submodule, which improves the model's generalization ability by using unlabeled ultrasound data through comparative learning and jigsaw puzzle tasks.

8. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 1, characterized in that: The cross-modal causal reasoning module constructs a causal relationship graph G = (V, E) of pathological variables in fatty liver disease based on the causal graph theory of medical knowledge; The node set V = {S, F, G, X} contains the fatty degeneration grade S, the fibrosis grade F, the inflammation grade G, and the multimodal feature X, and the edge set E represents the direct causal relationship between variables; The specific formula for calculating the causal effect of the intervention variable do(X) on the outcome variable Y is as follows: P(Y=y∣do(X=x))=Σ z P(Z=z∣X=x)·P(Y=y∣X=x,Z=z) Where Z is the set of mixed variables.

9. The intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to claim 1, characterized in that: The personalized adaptive learning module dynamically adjusts the model parameters through a conditional normalization algorithm. The specific formula for the normalization process of the l-th layer feature Fl is as follows: Where γ(c) and β(c) are learnable scaling and offset factors dependent on patient feature c, and μ(F) l ) and σ(F l ) represents the characteristic mean and standard deviation.

10. The application of the intelligent quantitative diagnostic model for fatty liver disease that deeply integrates multimodal ultrasound and machine learning according to any one of claims 1-9, characterized in that: The diagnostic model is integrated into a clinical application system consisting of ultrasound equipment and a cloud platform for intelligent quantitative diagnosis of fatty liver disease.

Images