Liver and gall patient condition evolution prediction method based on deep learning

Abstract

This invention discloses a deep learning-based method for predicting the disease evolution of hepatobiliary patients, belonging to the fields of medical imaging and deep learning technology. The invention first acquires multi-temporal abdominal CT images, time-series clinical laboratory indicators, and medical history follow-up data of hepatobiliary patients; it then uses a variant U-Net network with Lagrange pseudo-ordering structure constraints to accurately segment the CT images and extract radiomics features; based on the segmentation results, a three-dimensional model is constructed, and numerical simulations of hepatobiliary blood perfusion and bile fluid dynamics are performed using local entropy production theory to obtain dynamic biomechanical features; cross-dimensional feature alignment and fusion are achieved by fusing attention mechanisms and a multimodal network for complex community discovery, and then chaotic theory is introduced for phase space reconstruction to enhance dynamic features; this invention achieves an upgrade from statistical correlation to physiological causality in prediction, improving model generalization and clinical interpretability, and is suitable for prognostic assessment and intelligent early warning of disease progression in hepatobiliary diseases.

Description

Technical Field

[0001] This invention discloses a method for predicting the evolution of hepatobiliary disease in patients based on deep learning, belonging to the fields of medical imaging and deep learning technology. Background Technology

[0002] Existing methods for predicting the progression of hepatobiliary disease typically employ a single or superficially combined data-driven approach. In practice, the system acquires computed tomography (CT) images and time-series clinical laboratory indicators from patients. It then uses conventional image segmentation networks to extract the anatomical morphology of the hepatobiliary region and calculates radiomics features based on these extracted anatomical regions. Subsequently, the system directly concatenates the radiomics features with the time-series clinical laboratory indicators along the feature dimension and inputs this data into a time-series prediction deep learning network. The network model, by fitting statistical correlations from a large amount of historical sample data, outputs the probability of the patient developing endpoint events such as liver failure or malignant transformation at subsequent time points. Such approaches rely on the statistical co-occurrence frequency between features for pattern recognition.

[0003] The aforementioned existing technical solutions suffer from a core technical problem: the models cannot distinguish between the true causal characteristics of disease progression and confounding variables, resulting in predictions lacking physiological support and limited generalization ability. Because existing solutions rely solely on statistical correlation fitting between static anatomical radiomics features and clinical indicators, without incorporating dynamic physiological features such as hepatobiliary blood perfusion and bile hydrodynamics, the feature associations captured by the models often originate from accompanying confounding factors rather than the intrinsic physical mechanisms driving disease progression. When faced with data distribution shifts across different medical centers or disease stages, the prediction network built on statistical correlation cannot maintain a stable causal mapping, leading to overfitting. Summary of the Invention

[0004] The purpose of this invention is to provide a solution that can effectively address the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based method for predicting the disease progression of hepatobiliary patients includes the following steps: To obtain multi-phase abdominal CT images, time-series clinical laboratory indicators, and medical history follow-up data of hepatobiliary patients; Based on the U-Net variant network, segmentation of multi-temporal abdominal CT images was performed to extract radiomics features of liver parenchyma, biliary system and vascular branches; Based on the segmentation results, a three-dimensional model is constructed for biomechanical numerical simulation, and the dynamic biomechanical characteristics of the organ are output. By using a multimodal deep learning network that integrates attention mechanisms, cross-dimensional feature alignment and fusion are performed on radiomics features, dynamic biomechanical features, and time-series clinical laboratory indicators to generate a full-dimensional feature representation. A semi-supervised causal inference method is introduced, and a directed acyclic graph is constructed based on full-dimensional feature representation to screen the core feature subset that has a causal relationship with the disease evolution; Input a subset of core features into a time-series prediction Transformer network to output the probability of the endpoint event at different time points and the corresponding key causal features.

[0006] Preferably, segmentation of multi-temporal abdominal CT images based on a U-Net variant network includes: Pixel-level temporal variation curves of contrast agent concentration were constructed between different scanning stages of multi-phase enhanced CT. The Lagrange pseudo-order structure identification method in fluid mechanics is introduced into the loss function construction process of the U-Net variant network. The local topological curl of the above time-series change curve is used as a spatial regularization constraint term to make the segmentation boundary adaptively converge to the physiological boundary of blood perfusion. High-dimensional image omics features are extracted based on the constrained segmentation results.

[0007] Preferably, a three-dimensional model is constructed based on the segmentation results for biomechanical numerical simulation, including: In the numerical simulation of hepatobiliary blood perfusion and bile fluid dynamics, the local entropy production theory in thermodynamics is introduced to adaptively adjust the finite element mesh. Calculate the local entropy yield distribution of velocity and pressure gradients within the flow field. When the local entropy yield exceeds a preset dynamic threshold, trigger local mesh refinement iteration to reduce numerical discretization errors in high gradient regions. Dynamic biomechanical features at the node level are extracted based on the encrypted flow field mesh.

[0008] Preferably, a multi-modal deep learning network incorporating attention mechanisms generates full-dimensional feature representations, including: Radiomics features, dynamic biomechanical features, and time-series clinical laboratory indicators are mapped to heterogeneous nodes in a complex network, and the statistical correlation between features is transformed into edge weights. We use streaming community detection methods in complex networks to cluster heterogeneous nodes and use the community structure as a priori mask matrix for cross-modal attention allocation. Guided by the prior mask matrix, cross-dimensional feature alignment is achieved through a multi-head cross-attention mechanism, outputting a full-dimensional feature representation.

[0009] Preferably, a semi-supervised causal inference method is introduced to screen a subset of core features, including: The clinical intervention events and endpoint events in the patient history follow-up data are modeled as queuing service processes in operations research, and Markov arrival processes are constructed based on patient state transitions. On unlabeled samples, the steady-state distribution probability of the queuing service process is used as the prior distribution constraint on the causal connection strength in the directed acyclic graph. By maximizing the sum of the causal likelihood function on labeled samples and the prior distribution constraints on unlabeled samples, the directed acyclic graph structure is iteratively updated, and a subset of core features is output.

[0010] Preferably, after outputting the full-dimensional feature representation, it also includes: The phase space reconstruction method based on chaos theory in signal processing is introduced to handle full-dimensional feature representation; Calculate the time series embedding dimension and delay time of the full-dimensional feature representation, and map the one-dimensional feature sequence to a high-dimensional phase space using the Teikens theorem; Lyapunov exponential trajectories of feature evolution are extracted in high-dimensional phase space as supplementary dynamic features. The supplementary dynamic features are then concatenated with the full-dimensional feature representation and input into a semi-supervised causal inference method.

[0011] Preferably, iterative updates to the directed acyclic graph structure include: In the process of backpropagation of gradients for prior distribution constraints using unlabeled samples, a feature privacy protection mechanism is constructed by introducing the concept of zero-knowledge proof. The original feature values ​​of unlabeled samples are locally converted into homomorphic encrypted noise tensors, and the gradient direction and sign of the directed acyclic graph structure update are only passed to the central node, thus blocking the reverse inference path of the original feature values. Based on the desensitized gradient update of the directed acyclic graph structure, confusing variables are removed.

[0012] Preferably, the core feature subset is input into the temporal prediction Transformer network, including: The core feature subsets at different time points are regarded as orthogonal subcarriers in the communication channel, and the feature importance is regarded as the subcarrier signal-to-noise ratio; An adaptive bit and power allocation method from orthogonal frequency division multiplexing is introduced to dynamically allocate the computational bit width and computing resources of the self-attention matrix in the Transformer network for time-series prediction. High-precision floating-point computing resources are allocated to the core feature subset corresponding to high signal-to-noise ratio, and sparse attention truncation is performed on the features corresponding to low signal-to-noise ratio.

[0013] Preferably, the training process of the temporal prediction Transformer network includes: Introducing robust model predictive control from cybernetics to construct a closed-loop training optimization objective; In each training batch, a set of virtual perturbations following a Gaussian distribution is injected into the input of the time-series prediction Transformer network to simulate the distribution shift of clinical data; By constraining the minimization of the closed-loop cost function between the predicted states in the future multi-step process, including virtual perturbations, and the actual endpoint event, the network weight parameters are continuously optimized until the closed-loop cost function converges.

[0014] Preferably, the corresponding key causal features are output, including: A counterfactual inference mechanism is constructed by introducing the difference-in-differences method from econometrics at the output of the time series prediction Transformer network. A virtual state erasure operation is applied to the selected key causal features to generate counterfactual feature inputs, and forward propagation is re-executed to obtain the counterfactual prediction probability. Calculate the difference between the true prediction probability and the counterfactual prediction probability over time, and use this difference as a quantitative indicator of the marginal effect of the key causal feature driving the occurrence of the endpoint event, and output it synchronously.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the technical problem of existing technologies' inability to extract true causal features by introducing biomechanical numerical simulation and semi-supervised causal inference. The system constructs a 3D model based on segmentation results and incorporates local entropy production theory for adaptive mesh adjustment in the finite element method, extracting node-level dynamic biomechanical features reflecting the dynamic states of blood flow and bile. After aligning radiomics features, dynamic biomechanical features, and clinical indicators using a network with an attention-integrated mechanism, the system introduces a semi-supervised causal inference method to construct a directed acyclic graph. It then combines Markov arrival processes from queuing theory as prior distribution constraints to screen a subset of core features causally related to disease evolution without the participation of unlabeled samples. This mechanism eliminates confounding variables, transforming the model input from statistically relevant features to causal features, establishing a direct mapping from physiological and physical mechanisms to disease evolution.

[0016] 2. This invention optimizes feature extraction accuracy and model computation logic through the cross-integration of multidisciplinary methods. In the image segmentation stage, the Lagrange pseudo-order structure constraint is incorporated to adaptively converge the segmentation boundary to the physiological boundary of blood perfusion. In the feature fusion stage, a complex network community detection method is used to generate a priori mask matrix to guide attention allocation, reducing the computational cost of cross-dimensional alignment. Before causal inference, Lyapunov exponential trajectories are extracted using chaotic theory phase space reconstruction as supplementary dynamic features. During transformer network inference, the orthogonal frequency division multiplexing concept is introduced to dynamically allocate the computational bit width and computational resources of the self-attention matrix based on the feature signal-to-noise ratio, achieving on-demand scheduling of computational resources. In the output stage, a counterfactual inference mechanism is constructed using the double difference method. By calculating the difference between the true prediction probability and the counterfactual prediction probability, a quantitative index of the marginal effect of key causal features driving the occurrence of the endpoint event is output, transforming the prediction result into a causal explanation logic with quantifiable numerical values. Attached Figure Description

[0017] Figure 1 This is the overall main flowchart of the present invention; Figure 2 This is a flowchart of the CT image segmentation process of the present invention; Figure 3 This is a flowchart of the biomechanical numerical simulation of the present invention; Figure 4 This is a flowchart of the multimodal feature fusion process of the present invention; Figure 5 This is a flowchart of the semi-supervised causal inference process of the present invention; Figure 6 This is a flowchart of the timing prediction and output process of the present invention. Detailed Implementation

[0018] This specific embodiment discloses a deep learning-based method for predicting the disease evolution of hepatobiliary patients. This method is applied to the clinical prognostic assessment and disease progression early warning of hepatobiliary disease patients. The executing entity is an electronic device with medical image processing, numerical computation, and deep learning inference capabilities, including but not limited to medical imaging workstations, server clusters, or embedded medical auxiliary diagnostic devices. The method can be implemented through computer program instructions, which are stored in a computer-readable storage medium and can be loaded and executed by a processor.

[0019] Example 1: Please refer to the appendix Figure 1First, multi-phase abdominal CT images, time-series clinical laboratory indicators, and medical history follow-up data of hepatobiliary patients were acquired. The multi-phase abdominal CT images were dynamic contrast-enhanced CT sequences of the hepatobiliary system, including four scanning phases: plain scan, arterial phase, portal venous phase, and delayed phase. The image slice thickness was 0.625mm-1.25mm, the pixel matrix was 512×512, and the format was DICOM 3.0. Preprocessing was performed on the acquired multi-phase CT images, including liver window width and level adjustment (window width 150-200HU, window level 30-50HU), grayscale normalization, metal artifact suppression, and three-dimensional rigid respiratory motion registration based on mutual information. The objective function of the registration was to maximize the mutual information value between images of different phases, as shown in the following formula: in, , These are the image to be registered and the reference image, respectively. , This represents the probability distribution of grayscale edges. This represents the joint grayscale probability distribution.

[0020] The time-series clinical laboratory indicators include liver function indicators, kidney function indicators, tumor markers, and inflammatory markers, specifically ALT, AST, TBIL, ALB, PT, INR, AFP, CA19-9, WBC, CRP, and PCT. Data collection time points are at admission, 1 week post-surgery, 1 month, 3 months, 6 months, and 12 months post-surgery. The indicators at each time point constitute a multivariate time-series sequence. Preprocessing of the time-series clinical laboratory indicators includes multiple imputation for missing values ​​based on chain equations and Z-score standardization. The standardization formula is as follows: in, The original value of the indicator. This represents the mean of the metric in the training set. This represents the corresponding standard deviation.

[0021] The medical history follow-up data includes patient baseline information, clinical intervention events, and endpoint event data. The baseline information includes age, gender, underlying diseases, and etiological diagnosis. The clinical intervention events include surgery, interventional therapy, drug therapy, and endoscopic therapy. The endpoint events include liver failure, decompensated cirrhosis, biliary obstruction, malignant transformation, and all-cause mortality. The follow-up period is no less than 12 months.

[0022] The U-Net variant network was used to segment preprocessed multi-temporal abdominal CT images, extracting radiomics features of the liver parenchyma, biliary system, and vascular branches. The U-Net variant network uses a 3D U-Net++ architecture, with a ResNet3D residual block encoder and a dense skip connection decoder. The input is registered multi-temporal CT images from four phases, and the output is a six-class segmentation mask for the liver parenchyma, intrahepatic bile ducts, extrahepatic bile ducts, hepatic artery, portal vein, and hepatic vein. Based on the output segmentation mask, radiomics features were extracted for each anatomical structure. These features include first-order statistical features, second-order texture features, and higher-order shape features. First-order statistical features include mean, variance, skewness, kurtosis, and percentiles. Second-order texture features include contrast, correlation, energy, and entropy of the gray-level co-occurrence matrix, as well as short-run emphasis, long-run emphasis, and gray-level non-uniformity of the gray-level run length matrix. Higher-order shape features include volume, surface area, sphericity, compactness, and three-dimensional Euler number. Each anatomical structure yielded 107 initial radiomics features, which were then filtered using variance thresholding and mutual information features to obtain high-dimensional radiomics feature vectors. , The selected feature dimension has a value of 256.

[0023] Based on the segmentation results, a three-dimensional model was constructed for biomechanical numerical simulation, outputting the dynamic biomechanical characteristics of the organs. Using a six-class segmentation mask, the Marching Cubes algorithm was employed for surface reconstruction, generating a triangular mesh model. After mesh repair and smoothing, a tetrahedral volumetric mesh was generated as the base mesh for finite element calculations. Numerical simulations of hepatobiliary blood perfusion and bile hydrodynamics were performed based on this volumetric mesh. The governing equations included the continuity equation and the Navier-Stokes momentum equation, as shown below: in, For fluid velocity vector, For fluid density, blood density is taken as 1060 kg / m³. 3 Bile density is taken as 1010 kg / m³ 3 , For fluid pressure, For dynamic viscosity, blood viscosity is taken as 3.5 × 10⁻⁶. -3 Pa·s, bile viscosity is taken as 1.2 × 10⁻⁶. -3 Pa·s, For volumetric forces, based on the finite element method (FEM) results, the flow field features corresponding to each anatomical structure are extracted, including velocity vector, pressure value, wall shear stress, total pressure loss, and flow velocity distribution uniformity index. After statistical analysis and feature encoding, the dynamic biomechanical feature vector is obtained. , The value is 128.

[0024] By employing a multimodal deep learning network with an attention mechanism, cross-dimensional feature alignment and fusion are performed on radiomics features, dynamic biomechanical features, and time-series clinical laboratory indicators to generate a comprehensive feature representation. For multivariate time-series sequences of time-series clinical laboratory indicators, a two-layer LSTM encoder is used to extract time-dependent features, resulting in a clinical time-series feature vector. , The value is 128. , , The feature vectors from three modalities are input into a multimodal deep learning network with a fusion attention mechanism. This network includes a feature mapping layer, a multi-head cross-attention layer, and a feedforward neural network layer. The feature mapping layer maps the features from the three modalities to a unified dimensional space. The multi-head cross-attention layer aligns the cross-dimensional features, eliminating dimensional bias and semantic shift between modalities. The feedforward neural network layer then performs feature transformation, ultimately outputting a full-dimensional feature representation. , The value is 512.

[0025] A semi-supervised causal inference method is introduced, constructing a directed acyclic graph (DAG) based on full-dimensional feature representation to screen a subset of core features causally related to disease progression. Each feature dimension of the full-dimensional feature representation is used as a node in the DAG, and patient disease state transitions, clinical intervention events, and endpoint events are used as state variables for each node, constructing a Bayesian network structure. The training set of the semi-supervised causal inference method includes labeled and unlabeled samples. Labeled samples are patient data with complete follow-up endpoint events, while unlabeled samples are patient data with incomplete follow-up and no clear endpoint events. The adjacency matrix of the DAG is iteratively optimized using the objective function of maximizing the sum of the causal likelihood function of labeled samples and the prior constraint terms of unlabeled samples, while simultaneously satisfying the acyclic constraint. After convergence, features with direct or indirect causal connections to endpoint events are screened, and confounding variables without causal association are removed to obtain the core feature subset. , The core feature dimension has a value of 32-64.

[0026] A subset of core features is input into a temporal prediction Transformer network, which outputs the probability of the endpoint event occurring at different time points and the corresponding key causal features. The temporal prediction Transformer network adopts a Temporal FusionTransformer architecture. The input is a temporal sequence of core feature subsets at different time points, and the output is the probability of the endpoint event occurring at four time points: 1 month, 3 months, 6 months, and 12 months. Simultaneously, based on attention weights and causal path weights, it outputs key causal features that drive the prediction of the endpoint event.

[0027] As a preferred embodiment, refer to Figure 2 In the process of segmenting multi-temporal abdominal CT images based on a U-Net variant network, pixel-level contrast agent concentration temporal variation curves are first constructed across different scanning phases of multi-temporal enhanced CT. Based on the linear mapping relationship between contrast agent concentration and CT value, the contrast agent concentration of each pixel in different scanning phases is calculated using the following formula: in, for Real-time pixel-level contrast agent concentration The CT values ​​for the corresponding scanning phases. This is the CT value of that pixel during the plain scan. The CT value conversion factor for the contrast agent is 1 HU·mL / mg. A temporal curve of contrast agent concentration variation is constructed for each pixel, with the scan time of the four scan phases as the x-axis and the corresponding contrast agent concentration as the y-axis. , These correspond to the scanning times for the plain scan phase, arterial phase, portal venous phase, and delayed phase, respectively.

[0028] Secondly, the Lagrange pseudo-order structure recognition method from fluid mechanics is introduced into the loss function construction process of the U-Net variant network. The local topological curl of the temporal variation curve is used as a spatial regularization constraint term, making the segmentation boundary adaptively converge to the physiological boundary of blood perfusion. The local topological curl of the temporal variation curve is calculated for pixels in three-dimensional space. its neighborhood The time gradient field of contrast agent concentration within is The formula for calculating curl is as follows: in, Contrast agent concentrations at Spatial partial derivatives of the direction. Calculate the modulus of the local topological curl. The normalized curl feature is obtained by performing 0-1 normalization. , The mutation region corresponds to the physiological boundary of blood perfusion.

[0029] The total loss function for constructing the U-Net variant network is as follows: in, For Dice's loss, For cross-entropy loss, This is the regularization weight coefficient, with a value ranging from 0.1 to 0.5. The Lagrange pseudo-order structure regularization term is calculated using the following formula: in, Total number of pixels The segmentation boundary probability map output by the network is shown, with a value of 1 for boundary regions and 0 for non-boundary regions. A variant U-Net network is iteratively trained based on the aforementioned total loss function to adaptively converge the segmentation boundary to the physiological boundary of blood perfusion, thereby improving segmentation accuracy.

[0030] Finally, based on the constrained segmentation results, high-dimensional image omics features are extracted according to the described feature extraction process.

[0031] As a preferred embodiment, refer to Figure 3 In the process of constructing a three-dimensional model based on the segmentation results for biomechanical numerical simulation, the local entropy production theory from thermodynamics is introduced to adaptively adjust the finite element mesh during the numerical simulation of hepatobiliary blood perfusion and bile fluid dynamics. First, the local entropy production rate distribution of the velocity and pressure gradients within the flow field is calculated. For isothermal incompressible viscous fluid flow, the formula for calculating the local entropy production rate per unit volume is as follows: in, The thermodynamic temperature of the human body is taken as 310K. This is a tensor double dot product operation, representing the inner product of the velocity gradient tensor. The remaining parameters are consistent with the parameter definitions of the Navier-Stokes equation in Example 1.

[0032] Secondly, a preset dynamic threshold for the local entropy productivity is set, as shown in the following formula: in, Let be the average entropy yield of the entire flow field region. Let the standard deviation of the entropy production rate be the total entropy production rate of the entire flow field region. This is a threshold coefficient, with a value ranging from 1.5 to 3.0. It is used when the local entropy productivity of the grid cell... When the local encryption iteration of the cell is triggered, the encryption level is 1-3. Each encryption reduces the cell side length to half of the original. The number of iterations does not exceed 5 until the entropy production rate change rate of the high gradient region is less than 5%, thereby reducing the numerical discretization error of the high gradient region.

[0033] Finally, based on the encrypted flow field mesh, the dynamic biomechanical features at the node level are extracted, including the velocity vector, pressure value, wall shear stress, total pressure loss, local entropy yield, and flow velocity distribution uniformity index of each mesh node. After statistical encoding, the dynamic biomechanical feature vector is obtained.

[0034] As a preferred embodiment, refer to Figure 4 In the process of generating full-dimensional feature representations through a multimodal deep learning network that integrates attention mechanisms, radiomics features, dynamic biomechanical features, and time-series clinical test indicators are first mapped to heterogeneous nodes in a complex network, and the statistical correlation between features is transformed into edge weights. This process constructs a heterogeneous feature complex network. , where the node set Each node corresponds to a feature dimension; edge set For connections between nodes, an edge connection exists when the absolute value of the Pearson correlation coefficient between the corresponding features of two nodes is greater than a preset threshold of 0.3; edge weights... The absolute value of the Pearson correlation coefficient between the two features is 0-1.

[0035] Secondly, a streaming community detection method for complex networks is used to cluster heterogeneous nodes, with the community structure serving as a priori mask matrix for cross-modal attention allocation. This streaming community detection method aims to maximize network modularity. To optimize the objective, the modularity calculation formula is as follows: in, The sum of the weights of all edges in the network. For nodes The weighting degree, For nodes The community to which it belongs Let Kronecker function be used when The value is 1 if the condition is met, and 0 otherwise. After clustering, the following results are obtained: Each community has a characteristic feature, and the nodes within each community are highly correlated features, while the nodes across communities are less correlated.

[0036] Generate a priori mask matrix based on the community structure obtained from clustering. , The total dimension of the features, matrix elements The rule for determining the value is: when the node With nodes When they belong to the same community, ,otherwise .

[0037] Guided by a priori mask matrix, cross-dimensional feature alignment is achieved through a multi-head cross-attention mechanism. This maps radiomics features to a query matrix. Dynamic biomechanical characteristics and clinical time-series characteristics are concatenated and mapped into a bond matrix. AND-value matrix The mapping formula is as follows: in, This is a learnable weight matrix. The formula for calculating the attention weights is as follows: in, For Hadama accumulation, Scaling factor Key matrix The prior mask matrix constrains attention weights to be allocated only among features within the same community, reducing the computational cost of cross-dimensional alignment and improving the physiological rationality of feature fusion. After processing by multi-head cross-attention and a feedforward neural network, a full-dimensional feature representation is output.

[0038] As a preferred embodiment, refer to Figure 5 After outputting the full-dimensional feature representation, a phase space reconstruction method based on chaotic theory in signal processing is introduced to process the full-dimensional feature representation. First, the full-dimensional feature representation is organized into a multivariate time series according to the follow-up time nodes. For each feature dimension, a one-dimensional time series is then processed. Calculate the embedding dimension of the time series With delay time The delay time The autocorrelation function method is used for calculation. When the autocorrelation function... Decrease to the initial value At that time, the corresponding For the optimal delay time, the autocorrelation function formula is as follows: in, The mean of the time series. The standard deviation of the time series. This is a mathematical expectation operation. The embedded dimension... The spurious nearest neighbor method is used for calculation. When the proportion of spurious nearest neighbors is less than 1%, the corresponding... The optimal embedding dimension.

[0039] According to the Teikens theorem, the one-dimensional feature sequence is mapped to a high-dimensional phase space, and the reconstructed phase space vector is formulated as follows: in, .

[0040] Lyapunov exponential trajectories of feature evolution are extracted in high-dimensional phase space as supplementary dynamic features. The Lyapunov exponent characterizes the divergence or convergence rate of adjacent trajectories in phase space. The maximum Lyapunov exponent is calculated using a small data method, as shown in the following formula: in, for The nearest neighbor in phase space For time step, The number of trajectory points is given. The Lyapunov index is extracted for each time point to construct the Lyapunov index trajectory as a supplementary dynamic feature. The supplementary dynamic features are concatenated with the full-dimensional feature representation to obtain the enhanced full-dimensional feature representation. ,Will Input the subsequent semi-supervised causal inference method.

[0041] As a preferred embodiment, refer to Figure 6 In the process of introducing a semi-supervised causal inference method to screen the core feature subset, the clinical intervention events and endpoint events in the patient history follow-up data are first modeled as queuing service processes in operations research, and a Markov arrival process is constructed based on patient state transitions. The patient's condition state space is defined. ,in For a healthy state, For different stages of disease progression, The endpoint event state is represented by the service counter in a queuing service process, the endpoint event by the customer's departure process, and the patient's condition state transition by the state evolution of the queuing system, thus constructing a Markov arrival process. This Markov arrival process consists of two matrices. and Description, in which The state transition rate matrix is ​​the state transition rate matrix when no events arrive. This is a state transition rate matrix with events arriving, including clinical intervention events and endpoint events, and a state transition generation matrix. .

[0042] Calculate the steady-state distribution probability and steady-state distribution vector of the queuing service process. The following constraints must be satisfied: in, It is a vector of all 1s. For the patient to be in a state The steady-state probability. On unlabeled samples, the steady-state distribution probability is used as a prior distribution constraint for the causal connection strength in the directed acyclic graph. Obey The Dirichlet distribution with parameters, i.e. ,in For hyperparameters, Let be the adjacency matrix of a directed acyclic graph.

[0043] The objective function for semi-supervised causal inference is constructed by iteratively updating the directed acyclic graph structure by maximizing the sum of the causal likelihood function on labeled samples and the prior distribution constraints on unlabeled samples. The objective function formula is as follows: in, To define the log-likelihood function of the Bayesian network on the labeled samples, For the labeled sample set, For the prior constraints on unlabeled samples, The logarithmic probability, For the unlabeled sample set, The balancing coefficient is set to 0.5-2.0. The adjacency matrix is ​​iteratively optimized using the gradient descent method. It also satisfies the acyclic constraint of a directed acyclic graph. , The feature dimension is defined as follows. After convergence, based on the optimized directed acyclic graph structure, a subset of core features that are causally related to the evolution of the disease are selected.

[0044] As a preferred embodiment, during the iterative update of the directed acyclic graph structure, a feature privacy protection mechanism is constructed by introducing the concept of zero-knowledge proof during the backpropagation process of calculating the gradient of prior distribution constraints using unlabeled samples. First, the original feature values ​​of the unlabeled samples are transformed into homomorphically encrypted noise tensors on the local terminal. Then, the Paillier homomorphic encryption algorithm is used to process the feature vectors of the unlabeled samples. Encryption is performed using the following formula: in, The noise tensor follows a Gaussian distribution and can only be decrypted by the local terminal holding the private key. The central node cannot reverse the original value of the feature through the encrypted tensor.

[0045] When calculating the gradient of the prior distribution constraint term, the direction and sign of the gradient are calculated only based on the encrypted tensor, i.e. It only transmits the direction and sign of the gradient to the central node, but not the magnitude of the gradient or the original feature values, thus blocking the reverse inference path of the original feature values. The central node is based on the desensitized gradient. Update the adjacency matrix of the directed acyclic graph During the iteration process, confounding variables that have no causal relationship with the endpoint event are removed, and the core feature subset is finally output.

[0046] As a preferred embodiment, in the process of inputting the core feature subset into the time-series prediction Transformer network, the core feature subsets at different time points are regarded as orthogonal subcarriers in the communication channel, and the feature importance is regarded as the subcarrier signal-to-noise ratio. An adaptive bit and power allocation method in orthogonal frequency division multiplexing is introduced to dynamically allocate the computational bit width and computing power resources of the self-attention matrix in the time-series prediction Transformer network.

[0047] First, calculate the signal-to-noise ratio (SNR) for each core feature. The SNR is the sum of the weights of all causal paths from that feature to the endpoint event in the directed acyclic graph, as shown in the following formula: in, Features The set of all causal paths leading to the endpoint event. For path Causal weights.

[0048] Set the signal-to-noise ratio threshold The core features are divided into high signal-to-noise ratio (SNR) features and low SNR features, based on the mean SNR of all features. The high SNR features satisfy the following condition: The low signal-to-noise ratio feature satisfies For the core feature subset corresponding to high signal-to-noise ratio, high-precision floating-point computing resources are allocated, and FP32 computing bit width is adopted to fully preserve the computational precision of the self-attention matrix; for features corresponding to low signal-to-noise ratio, sparse attention truncation is performed, retaining only the first few features. The highest weighted attention connection is selected, and the remaining weights are reset to 0. At the same time, FP16 or INT8 low-precision calculation bit width is used to reduce computing power consumption.

[0049] The adaptive allocation formula for calculating the bit width is as follows: in, Features The corresponding calculated bit width, The maximum calculation bit width is set to 32 bits. It represents the maximum signal-to-noise ratio among all features.

[0050] As a preferred embodiment, robust model predictive control from cybernetics is introduced during the training of the temporal prediction Transformer network to construct a closed-loop training optimization objective. In each training batch, a virtual perturbation set following a Gaussian distribution is injected into the input of the temporal prediction Transformer network to simulate the distribution shift of clinical data. The perturbed input feature formula is as follows: in, For virtual perturbation set, The disturbance strength is represented by a value between 0.05 and 0.2. It is an identity matrix.

[0051] A closed-loop cost function is constructed, constrained to minimize the closed-loop cost function between the predicted states in multiple future steps, including virtual perturbations, and the actual endpoint event. The network weight parameters are then continuously optimized. The formula for the closed-loop cost function is as follows: in, To predict the step size, For the first The probability of predicting the endpoint event in each step. Labels for actual endpoint events. This refers to the update amount of network weights. Here is the regularization coefficient. Gradient descent is used, with the network weight parameters optimized once per training batch until the closed-loop cost function is achieved. The rate of change is less than It reaches a convergent state.

[0052] As a preferred embodiment, in the process of outputting the corresponding key causal features, a counterfactual inference mechanism is constructed by introducing the difference-in-differences method from econometrics at the output of the time-series prediction Transformer network. This mechanism is then applied to the selected key causal features. A virtual state erasure operation is applied, setting the value of the feature at all time points to the mean value of the feature in the healthy population. Generate counterfactual feature input The values ​​of the other features remain unchanged.

[0053] Input counterfactual features Input the trained temporal prediction Transformer network, re-execute the forward propagation, and obtain the counterfactual prediction probabilities. , The prediction time point is defined. The difference between the true prediction probability and the counterfactual prediction probability along the time dimension is calculated to obtain the difference-in-differences estimator, as shown in the following formula: in, As the baseline time point, To predict the true probability, Features The average treatment effect, i.e., the quantitative indicator of the marginal effect of this feature in driving the occurrence of the endpoint event. Sort by size from largest to smallest, and output the first... The system simultaneously outputs the corresponding marginal effect quantification index and the probability of the endpoint event occurring at different time points, based on key causal characteristics.

[0054] All the above embodiments can be combined arbitrarily according to clinical application scenarios, and the combined technical solutions all fall within the protection scope of this invention.

Claims

1. A deep learning-based method for predicting the disease progression of hepatobiliary patients, characterized in that, Includes the following steps: To obtain multi-phase abdominal CT images, time-series clinical laboratory indicators, and medical history follow-up data of hepatobiliary patients; Based on the U-Net variant network, segmentation of multi-temporal abdominal CT images was performed to extract radiomics features of liver parenchyma, biliary system and vascular branches; Based on the segmentation results, a three-dimensional model is constructed for biomechanical numerical simulation, and the dynamic biomechanical characteristics of the organ are output. By using a multimodal deep learning network that integrates attention mechanisms, cross-dimensional feature alignment and fusion are performed on radiomics features, dynamic biomechanical features, and time-series clinical laboratory indicators to generate a full-dimensional feature representation. A semi-supervised causal inference method is introduced, and a directed acyclic graph is constructed based on full-dimensional feature representation to screen the core feature subset that has a causal relationship with the disease evolution; Input a subset of core features into a time-series prediction Transformer network to output the probability of the endpoint event at different time points and the corresponding key causal features.

2. The method according to claim 1, characterized in that, Segmentation of multi-temporal abdominal CT images based on a variant of the U-Net network, including: Pixel-level temporal variation curves of contrast agent concentration were constructed between different scanning stages of multi-phase enhanced CT. The Lagrange pseudo-order structure identification method in fluid mechanics is introduced into the loss function construction process of the U-Net variant network. The local topological curl of the above time-series change curve is used as a spatial regularization constraint term to make the segmentation boundary adaptively converge to the physiological boundary of blood perfusion. High-dimensional image omics features are extracted based on the constrained segmentation results.

3. The method according to claim 1, characterized in that, Based on the segmentation results, a three-dimensional model is constructed for biomechanical numerical simulation, including: In the numerical simulation of hepatobiliary blood perfusion and bile fluid dynamics, the local entropy production theory in thermodynamics is introduced to adaptively adjust the finite element mesh. Calculate the local entropy yield distribution of velocity and pressure gradients within the flow field. When the local entropy yield exceeds a preset dynamic threshold, trigger local mesh refinement iteration to reduce numerical discretization errors in high gradient regions. Dynamic biomechanical features at the node level are extracted based on the encrypted flow field mesh.

4. The method according to claim 1, characterized in that, A multi-modal deep learning network incorporating attention mechanisms generates full-dimensional feature representations, including: Radiomics features, dynamic biomechanical features, and time-series clinical laboratory indicators are mapped to heterogeneous nodes in a complex network, and the statistical correlation between features is transformed into edge weights. We use streaming community detection methods in complex networks to cluster heterogeneous nodes and use the community structure as a priori mask matrix for cross-modal attention allocation. Guided by the prior mask matrix, cross-dimensional feature alignment is achieved through a multi-head cross-attention mechanism, outputting a full-dimensional feature representation.

5. The method according to claim 1, characterized in that, A semi-supervised causal inference method is introduced to screen a subset of core features, including: The clinical intervention events and endpoint events in the patient history follow-up data are modeled as queuing service processes in operations research, and Markov arrival processes are constructed based on patient state transitions. On unlabeled samples, the steady-state distribution probability of the queuing service process is used as the prior distribution constraint on the causal connection strength in the directed acyclic graph. By maximizing the sum of the causal likelihood function on labeled samples and the prior distribution constraints on unlabeled samples, the directed acyclic graph structure is iteratively updated, and a subset of core features is output.

6. The method according to claim 4, characterized in that, After outputting the full-dimensional feature representation, it also includes: The phase space reconstruction method based on chaos theory in signal processing is introduced to handle full-dimensional feature representation; Calculate the time series embedding dimension and delay time of the full-dimensional feature representation, and map the one-dimensional feature sequence to a high-dimensional phase space using the Teikens theorem; Lyapunov exponential trajectories of feature evolution are extracted in high-dimensional phase space as supplementary dynamic features. The supplementary dynamic features are then concatenated with the full-dimensional feature representation and input into a semi-supervised causal inference method.

7. The method according to claim 5, characterized in that, Iterative updates to the structure of a directed acyclic graph include: In the process of backpropagation of gradients for prior distribution constraints using unlabeled samples, a feature privacy protection mechanism is constructed by introducing the concept of zero-knowledge proof. The original feature values ​​of unlabeled samples are locally converted into homomorphic encrypted noise tensors, and the gradient direction and sign of the directed acyclic graph structure update are only passed to the central node, thus blocking the reverse inference path of the original feature values. Based on the desensitized gradient update of the directed acyclic graph structure, confusing variables are removed.

8. The method according to claim 6, characterized in that, Input a subset of core features into the temporal prediction Transformer network, including: The core feature subsets at different time points are regarded as orthogonal subcarriers in the communication channel, and the feature importance is regarded as the subcarrier signal-to-noise ratio; An adaptive bit and power allocation method from orthogonal frequency division multiplexing is introduced to dynamically allocate the computational bit width and computing resources of the self-attention matrix in the Transformer network for time-series prediction. High-precision floating-point computing resources are allocated to the core feature subset corresponding to high signal-to-noise ratio, and sparse attention truncation is performed on the features corresponding to low signal-to-noise ratio.

9. The method according to claim 8, characterized in that, The training process of a Transformer network for time-series prediction includes: Introducing robust model predictive control from cybernetics to construct closed-loop training and optimize the objective; In each training batch, a set of virtual perturbations following a Gaussian distribution is injected into the input of the time-series prediction Transformer network to simulate the distribution shift of clinical data; By constraining the minimization of the closed-loop cost function between the predicted states in the future multi-step process, including virtual perturbations, and the actual endpoint event, the network weight parameters are continuously optimized until the closed-loop cost function converges.

10. The method according to claim 9, characterized in that, Output the corresponding key causal features, including: A counterfactual inference mechanism is constructed by introducing the difference-in-differences method from econometrics at the output of the time series prediction Transformer network. A virtual state erasure operation is applied to the selected key causal features to generate counterfactual feature inputs, and forward propagation is re-executed to obtain the counterfactual prediction probability. Calculate the difference between the true prediction probability and the counterfactual prediction probability over time, and use this difference as a quantitative indicator of the marginal effect of the key causal feature driving the occurrence of the endpoint event, and output it synchronously.

Images