A thrombus image data intelligent processing method and system

Abstract

The application discloses a kind of based on thrombus image data intelligent processing method and system, comprising: the longitudinal thrombus image data of patient is collected and thrombus stage is judged, and three-dimensional image sequence is established;Three-dimensional reconstruction is carried out to each time point image, and the three-dimensional model of blood vessel and thrombus is generated and hemodynamics simulation is carried out, to calculate key biomechanics parameter;Image features and biomechanics parameters are fused into multidimensional space-time feature tensor, construct space-time graph data structure, and design space-time graph neural network learns thrombus evolution mode under the action of biomechanics;Based on the learned mode, the thrombus morphological state of patient future time point and the quantitative embolism risk probability are predicted.The scheme realizes the automatic learning of thrombus evolution law from time sequence image, and completes risk prediction and treatment simulation.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and in particular to an intelligent processing method and system based on thrombosis imaging data. Background Technology

[0002] In clinical practice, the precise management of thrombotic diseases heavily relies on medical imaging assessment. Currently, clinical diagnosis is mainly based on computed tomography (CT) angiography or magnetic resonance angiography at single or a few time points, with physicians visually assessing the size, location, and patency of the thrombus. This method only provides static morphological information of the thrombus at a specific moment, failing to quantify its dynamic evolution and revealing the biological impact of the hemodynamic environment on thrombus development. Due to the lack of objective analysis of the temporal patterns of the entire thrombus growth-stabilization-dissolution process, treatment decisions often lag behind changes in the patient's condition, resulting in insufficient predictive ability for acute embolic events and hindering individualized risk stratification and prospective intervention.

[0003] With the development of medical imaging technology and computational fluid dynamics, researchers have begun to attempt to combine 3D image reconstruction and blood flow simulation to obtain local biomechanical parameters of thrombi. However, existing methods are mostly limited to mechanical analysis at isolated time points, failing to effectively integrate temporal image features with the dynamic biomechanical environment to construct predictable evolution models. Meanwhile, traditional machine learning methods face challenges in processing such data with complex spatial topology and temporal dependencies, struggling to simultaneously learn the local spatial interactions of thrombus morphology and its global evolution over time. Summary of the Invention

[0004] This application provides an intelligent processing method based on thrombosis imaging data, which integrates multi-time series and multi-modal data and can automatically learn the spatiotemporal evolution pattern of thrombosis under biomechanical action, ultimately realizing intelligent analysis for quantitative prediction of future states and risks.

[0005] This application provides a method for intelligent processing of thrombosis imaging data, including:

[0006] S1. Based on the longitudinal thrombosis image data collected from the patient, determine the cycle of the thrombosis and establish a three-dimensional image of the thrombosis development cycle and its surrounding tissues.

[0007] S2 performs three-dimensional geometric reconstruction on the images at each time point to generate a three-dimensional model of the thrombus and blood vessel wall, and performs hemodynamic simulation and calculates biomechanical parameters based on the model;

[0008] S3 integrates the image features and biomechanical parameters at each time point into a multidimensional feature tensor to construct a cross-time dimension data representation; and designs a dedicated spatiotemporal convolutional neural network to learn the temporal pattern of biomechanical effects on thrombus evolution.

[0009] S4, based on the evolutionary time pattern, predicts the thrombotic status and embolic event risk probability of patients at future time points based on hemodynamic simulation and 3D model.

[0010] The learning of the temporal pattern of thrombus evolution by biomechanical effects specifically includes: modeling the entire spatiotemporal dataset as a spatiotemporal graph, where the node set contains the vertices of all three-dimensional grids at all time points, the spatial edge set is constructed based on the topological adjacency relationship of the three-dimensional grids to connect directly connected vertex pairs at the same time point, and the temporal edge set connects vertices at corresponding anatomical locations between aligned grids at adjacent time points using a non-rigid registration algorithm, with the connection weights of the temporal edges dynamically adjusted according to the thrombus characteristics of the region where the vertex is located; inputting the spatiotemporal graph into a neural network, aggregating vertex neighbor features in the spatial dimension through a graph attention mechanism, and capturing cross-time point dependencies in the temporal dimension through a Transformer encoder, outputting a vertex spatiotemporal feature sequence; predicting thrombus volume changes and staging probabilities through graph pooling and temporal models; and predicting the probability of embolic events through attention pooling and multilayer perceptrons.

[0011] Preferably, the embolic event risk probability includes: extracting the vertex feature set of the last time point from the spatiotemporal graph neural network, aggregating it into a risk representation vector through a risk-aware attention pooling layer; the attention pooling layer calculates a weight for each vertex, the weight value being jointly determined by the vertex feature and the most salient feature in its first-order spatial neighborhood, and normalizing the sum of all weights to one; inputting the weighted and aggregated risk representation vector into a multilayer perceptron for mapping; wherein the multilayer perceptron contains at least one hidden layer with a nonlinear activation function, and its output layer maps the final calculation result to the embolic event probability through a sigmoid function; wherein the formula for calculating the embolic event probability is: ,in, For the predicted probability of embolism, This is the risk representation vector. and Here are the weight matrix and bias vector of the first hidden layer of the multilayer perceptron. To modify the activation function of the linear unit, , The weight vector and bias scalar of the output layer. This is the Sigmoid activation function.

[0012] Preferably, the prediction of the patient's thrombotic status at future points in time specifically includes:

[0013] S41, using longitudinal multi-parameter MRI sequence data of patients and super-resolution reconstruction, analyzes the changing patterns of thrombus component proportions at each time point;

[0014] S42, segment and extract sub-regions inside the thrombus from the super-resolution image at each time point, and combine the magnetic resonance physical model to map the pixel-level signal intensity to the component ratio to generate a component quantitative spectrum;

[0015] S43 constructs a spatiotemporal graph structure with thrombus voxels as nodes, integrates component ratios and image features, learns the non-uniform evolution pattern of components through a spatiotemporal graph neural network, and predicts the component distribution of thrombus status at future time points.

[0016] S44, Analyze the degree of impact on treatment measures based on the predicted component distribution results.

[0017] Preferably, the super-resolution reconstruction specifically includes: periodically acquiring longitudinal multi-parameter magnetic resonance images of the patient to obtain multiple sequence data with different sensitivities to thrombus components, including T1-weighted, T2-weighted, and proton density-weighted data; and performing super-resolution reconstruction processing on the low-resolution images acquired for each sequence using a deep learning model based on generative adversarial networks.

[0018] Preferably, the analysis of the variation law of thrombus component proportion at each time point specifically includes: performing temporal registration on the acquired longitudinal multi-parameter magnetic resonance images; using a non-rigid registration algorithm to precisely align all images of different sequences and different time points in three-dimensional space at the sub-voxel level to establish a unified voxel-level spatial correspondence; within the registered space, extracting the signal intensity of each voxel in the multi-sequence images for each voxel in the thrombus region to form a multi-dimensional feature vector; based on the magnetic resonance physical model, decomposing the multi-dimensional feature vector of each voxel into the volume proportion of several basic components by solving a constrained linear inversion problem; generating quantitative thrombus component maps at each time point, and quantifying the dynamic change of each component proportion over time by comparing maps at adjacent time points.

[0019] Preferably, in step S41, analyzing the changing patterns of thrombus component proportions at each time point further includes:

[0020] S411 records the time of change in the proportion of thrombus components and the corresponding state of thrombus components, and marks the critical point of change in the historical record based on event backtracking;

[0021] S412, calculate the dynamic change rate of components, spatial distribution entropy and inter-component interaction coefficient at each time point, and integrate them with biomechanical parameters to construct a thrombus dynamic state vector;

[0022] S413 employs a general critical point prediction model for thrombus component evolution using neural differential equations, and introduces eigenvalue analysis and anomaly detection algorithms for critical point detection.

[0023] S414 generates graded early warning signals for changes in thrombosis status based on the critical point prediction results.

[0024] Preferably, the event retrospective includes: using the current recorded time point as a baseline, retrospectively reviewing all quantitative spectra of thrombus components and their derived kinetic curves within the time window prior to the event.

[0025] Preferably, the critical point detection specifically includes:

[0026] Acquire initial thrombotic images and corresponding basic clinical information of new target patients, and use anatomical segmentation models to accurately segment the thrombotic region and the parametric features of the vascular geometry in which it is located from the images;

[0027] Using parametric features as conditions for individualized anatomical features, a diffusion model is used to generate synthetic thrombotic images that conform to the specific anatomical background of the patient.

[0028] For each synthetic thrombus image, simulate its dynamic evolution over time to generate corresponding dynamic thrombus time series data;

[0029] The generated individualized synthetic data is matched and calibrated with existing general critical point prediction models to form a predictive model for adjusting the critical point among different patients.

[0030] This application also provides an intelligent processing system based on thrombosis image data, the system comprising: an acquisition and processing module, a reconstruction calculation module, a fusion modeling module, a risk prediction module, and an image prediction module;

[0031] The acquisition and processing module is used to determine the period of thrombus based on the longitudinal thrombus image data of the patient and to establish a three-dimensional image of the thrombus development cycle and its surrounding tissues.

[0032] The reconstruction calculation module is used to perform three-dimensional geometric reconstruction of images at each time point, generate three-dimensional models of thrombi and blood vessel walls, and perform hemodynamic simulation and calculate biomechanical parameters based on the models.

[0033] The fusion modeling module is used to fuse image features and biomechanical parameters at each time point into a multidimensional feature tensor to construct a cross-time dimension data representation; a dedicated spatiotemporal convolutional neural network is designed to learn the temporal pattern of biomechanical effects on thrombus evolution.

[0034] The risk prediction module is used to predict the patient's thrombotic status and the probability of embolic events at future time points based on evolving time patterns, hemodynamic simulations, and 3D models.

[0035] Preferably, the system further includes an image prediction module, which quantifies the treatment measures into input parameters of a spatiotemporal convolutional neural network model, simulates the changes in thrombus status after intervention through a conditional generative adversarial network, and generates image feature predictions of thrombi at future time points.

[0036] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0037] By integrating spatiotemporal biomechanical modeling and dynamic prediction, the efficiency of intelligent processing of thrombotic imaging data is significantly improved. Its core advantage lies in its ability to predict thrombus evolution trajectories based on multi-timepoint imaging data and quantify embolism risk, thereby providing a basis for prospective clinical intervention. This method achieves high-precision simulation of the dynamic evolution of thrombi through intelligent integration of biomechanical parameters and imaging features, effectively overcoming the limitations of traditional static staging, supporting personalized treatment strategy optimization, and ultimately enhancing the accuracy and practicality of clinical decision-making.

[0038] By quantitatively analyzing the spatiotemporal distribution and evolution of thrombus components, this approach achieves a deeper understanding, moving from macroscopic morphological prediction to microscopic pathological mechanism analysis. Its core advantage lies in its ability to accurately predict the component distribution of thrombi at future time points, revealing their intrinsic evolutionary trends such as organization and dissolution in advance. This provides crucial evidence for assessing treatment response and predicting the risk of recanalization or re-occlusion.

[0039] By tracking the dynamic evolution of thrombus components in real time and accurately predicting critical points, early warning of thrombus stability mutations is achieved. This method can effectively identify critical moments of component changes and generate graded warning signals, thereby assisting physicians in taking early intervention measures and significantly reducing the risk of embolism. Compared with traditional static analysis, this approach improves the initiative and accuracy of thrombus management and has high clinical applicability and decision support value.

[0040] A personalized model calibration mechanism was constructed by generating synthetic thrombotic images and evolution time-series data that match the anatomical features of target patients. This mechanism effectively addresses the problem of limited model generalization ability caused by the scarcity of longitudinal time-series data for patients in clinical practice. Using synthetic data, a general prediction model can be accurately calibrated for the vascular geometry and blood flow environment of specific patients, thereby transforming population statistical regularities into individualized risk assessments. This method significantly improves the specificity and reliability of critical point warnings and enables advanced risk quantification and intervention strategy simulation in the initial stage of treatment. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating an intelligent processing method based on thrombosis imaging data according to an embodiment of the present invention.

[0042] Figure 2This is a schematic diagram of the structure of an intelligent processing system based on thrombosis imaging data according to an embodiment of the present invention. Detailed Implementation

[0043] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0045] Example 1: Figure 1 This is a flowchart illustrating an intelligent processing method for thrombosis imaging data according to an embodiment of the present invention.

[0046] like Figure 1 As shown, an intelligent processing method based on thrombosis imaging data includes the following steps:

[0047] A method for intelligent processing of thrombosis imaging data includes the following steps:

[0048] S1. Based on the longitudinal thrombosis image data collected from the patient, determine the cycle of the thrombosis and establish a three-dimensional image of the thrombosis development cycle and its surrounding tissues.

[0049] The thrombotic cycle is defined according to the standard terminology widely used in clinical practice as the acute phase, subacute phase, and chronic phase of thrombosis.

[0050] S2 performs three-dimensional geometric reconstruction on the images at each time point to generate a three-dimensional model of the thrombus and blood vessel wall, and performs hemodynamic simulation and calculates biomechanical parameters based on the model.

[0051] The determination of time points requires the integration of multidimensional concepts including absolute physical time, relative physiological stages, and clinical event markers. Specifically, an immutable zero point is first established as the moment of image acquisition when a thrombus is first discovered and diagnosed. Each time point primarily corresponds to a precise physical time of image acquisition, and each physical time point is further labeled with its corresponding physiological stage of thrombus evolution.

[0052] Specifically, based on the standardized 3D images at each time point, precise segmentation of the thrombus and vessel wall is performed. A fully convolutional neural network with an encoder-decoder structure is used to perform voxel-level semantic segmentation of the thrombus and vessel wall, outputting a high-quality binary 3D mask of the thrombus and vessel wall. The moving cube algorithm is applied to reconstruct the surface mesh of the binary 3D mask. Each voxel cell is traversed, and a predefined configuration table is queried based on the binary state of its eight vertices. Triangular patches are generated inside the cell to approximate the target isosurface, thereby generating a smooth 3D surface mesh model composed of a set of triangular patches.

[0053] Hemodynamic simulations were conducted using a smooth 3D surface mesh model as the computational domain. The mesh model was discretized to generate either a structured volume mesh or an unstructured tetrahedral mesh. In the simulation environment construction, a time-varying velocity waveform conforming to the physiological cycle was set as the velocity boundary condition at the model inlet, while a constant pressure boundary condition was applied at the outlet. The vessel wall and thrombus surface were set as fixed wall boundary conditions with no slippage. Within the discretized fluid computational domain, the governing equations describing the motion of incompressible Newtonian fluids, namely the Navier-Stokes equations and the continuity equation, were solved. The finite volume method was used to spatially discretize the governing equations, and a semi-implicit method for the pressure coupling equations was applied for time-progression and iterative solution until the flow field converged, thus obtaining a stable 3D velocity vector field and pressure scalar field.

[0054] Based on the convergent velocity and pressure fields, key biomechanical parameters are calculated. The focus is on calculating the wall shear force acting on the surface of blood vessels and thrombi. This parameter is obtained by calculating the gradient of the velocity component parallel to the wall at the wall surface along the wall normal direction and multiplying it by the blood dynamic viscosity; the magnitude of this vector characterizes the intensity of the tangential force exerted by the blood flow on the wall. Simultaneously, pressure values ​​at each node on the model surface are directly extracted to form a pressure distribution field.

[0055] The wall shear force scalar field, pressure scalar field, and velocity vector field obtained from the simulation at each time point are associated with and assigned to each vertex of the three-dimensional surface mesh model at the corresponding time point through spatial coordinate mapping relationship, resulting in a sequence arranged in the order of acquisition time. Each element contains the three-dimensional geometric model at that moment and the complete set of biomechanical parameter fields attached to it, which together constitute a spatiotemporal dataset describing the dynamic evolution of the thrombus and its surrounding mechanical environment.

[0056] S3 integrates the image features and biomechanical parameters at each time point into a multidimensional feature tensor to construct a cross-time dimension data representation; it also designs a dedicated spatiotemporal convolutional neural network to learn the temporal pattern of biomechanical effects on thrombus evolution.

[0057] Specifically, for each time point, a multimodal feature fusion at the vertex level is performed on the generated 3D surface mesh model containing biomechanical parameter fields. For each vertex on the mesh, image features are extracted from its corresponding original 3D medical image. From the biomechanical parameter field mapped to that vertex, its wall shear force scalar, pressure value, and 3D velocity vector are directly obtained to form a biomechanical feature vector. Through vector concatenation, the image feature vector and biomechanical feature vector of the same vertex are fused to form a comprehensive multidimensional feature tensor, which serves as the fused feature representation of that vertex at that moment.

[0058] To study the temporal patterns of biomechanical effects on thrombus evolution, the entire spatiotemporal dataset is modeled as a spatiotemporal graph. The node set contains the vertices of all 3D meshes at all time points. The spatial edge set is constructed based on the adjacency relationships of the 3D meshes, connecting pairs of vertices directly connected by mesh edges at the same time point, capturing the interactions of local anatomical structures. The temporal edge set is established using a non-rigid registration algorithm. After aligning the mesh models of adjacent time points, a corresponding vertex is found on the mesh of its neighboring time points and a connection is established. The temporal edge weights are dynamically adjusted according to the thrombus characteristics of the region where the vertex is located, with a focus on slowly evolving regions.

[0059] Subsequently, a spatiotemporal graph convolutional neural network is used for forward propagation. In the spatial dimension, a graph attention mechanism is employed for feature aggregation. The first-order spatial neighbor features of each vertex are weighted and summed. The attention weights are obtained by concatenating learnable vectors with neighbor features, followed by nonlinear transformation and normalization, resulting in a spatially aggregated feature sequence. In the temporal dimension, the spatially aggregated feature sequence of each vertex is input into a Transformer encoder. First, learnable positional encodings are added to inject temporal order information. Then, a multi-head self-attention mechanism is used to capture cross-temporal dependencies. Each attention head independently calculates features, and the resulting concatenations are then processed by a feedforward network to output a vertex feature sequence containing temporal context information.

[0060] Finally, multi-task prediction output is achieved. Two parallel decoding heads are set up: In the thrombus evolution trajectory prediction task, graph pooling is first performed on the features of all vertices at each time point to obtain a global graph representation. Then, the time series is input into the time series model to predict the global state vector of the thrombus at future time points. This vector is then mapped to the thrombus volume change rate and the probability distribution of each stage through a fully connected layer. In the embolism risk assessment task, features related to high-risk biomechanical patterns are extracted from the vertex features of the last time point. These features are then aggregated into a risk representation vector through attention pooling and output as the probability of embolic events through a multilayer perceptron.

[0061] S4, based on the evolutionary time pattern, predicts the thrombotic status and embolic event risk probability of patients at future time points based on hemodynamic simulation and 3D model.

[0062] Specifically, based on the spatiotemporal graph dataset constructed in the aforementioned steps, the designed spatiotemporal graph neural network is subjected to end-to-end supervised training. The training process aims to minimize a composite loss function that integrates multiple task objectives. This function consists of a weighted sum of the losses from the thrombosis trajectory prediction task, the embolism risk assessment task, and a regularization term used to control model complexity. The trajectory prediction task loss is itself a combination of the mean squared error of continuous variable prediction and the cross-entropy loss of classification prediction, while the risk assessment task loss uses a binary cross-entropy form. Training calculates gradients using the backpropagation algorithm and dynamically adjusts the learning rate using the adaptive moment estimation algorithm to update all network weight parameters until the model converges.

[0063] In the thrombus evolution trajectory prediction branch of the model, firstly, all vertex features obtained from spatiotemporal graph neural network encoding at each time point are globally aggregated based on attention weights to generate a series of temporal vectors representing the overall state at each time point. These temporal vectors are then input into a long short-term memory network to encode long-term dependencies in historical evolution. The hidden state of the long short-term memory network at the last moment is extracted, concatenated with the embedding vector representing the future prediction time interval, and jointly input into a multilayer perceptron for decoding. The decoded vector is then mapped through two independent linear transformation layers to a specific thrombus volume prediction value and a discrete distribution representing the probability of different evolution stages, respectively.

[0064] In the embolism risk assessment branch, the processing focuses on the vertex feature set at the last observation time point. These features are first aggregated through a risk-aware attention pooling layer. The attention weights in this layer consider both the vertex's own features and the most salient features in its spatial neighborhood, thus highlighting local regions with high-risk patterns. The resulting global risk representation vector is then fed into a multilayer perceptron with two hidden layers. This network ultimately outputs a scalar value between 0 and 1 through a sigmoid activation function, representing the predicted probability of an embolic event occurring at the current or future specified time.

[0065] During the model inference and application phase, the complete spatiotemporal map data of the patient to be analyzed is input into the trained model for forward computation. The model executes the above two branches in parallel: the trajectory prediction branch outputs quantitative and probabilistic predictions of thrombus volume and evolution stage at future target time points, and the risk assessment branch outputs an estimate of the probability of embolic events.

[0066] The formula for calculating the probability of an embolic event is as follows: The formula for calculating the risk representation vector r is: attention weight and its normalized form The calculation method is as follows:

[0067]

[0068]

[0069] This represents the feature vector of vertex i after encoding by the spatiotemporal graph neural network at the last observation time point T; This represents the set of first-order spatial neighbors of vertex i in a 3D mesh; This represents the eigenvectors of all spatial neighbor vertices j of vertex i. Performing element-wise maximization results in a new vector that captures the most salient features in the local neighborhood of vertex i. This represents a vector concatenation operation; This represents a trainable parameter vector used to map the concatenated feature vector to a scalar attention score; This represents the Sigmoid activation function, used here to compress the attention scores to the (0,1) interval to obtain the initial attention weights. ; This represents the initial weights of all vertices at the same time point. The final attention weights obtained by performing Softmax normalization satisfy the following conditions: ; The risk representation vector is the feature vector of all vertices at the last time point. Based on its final attention weight The global feature vector obtained after weighted summation; , This represents the weight matrix and bias vector of the first layer (hidden layer) of a multilayer perceptron; This represents the modified linear unit activation function; , This represents the weight vector and bias scalar of the second layer (output layer) of the multilayer perceptron; The probability of an embolic event predicted by the model is represented by a scalar value between 0 and 1. The closer the value is to 1, the higher the predicted risk of embolism.

[0070] Preferably, the method further includes: S5, quantifying the treatment measures into input parameters of a spatiotemporal convolutional neural network model, simulating the thrombus state changes after intervention through a conditional generative adversarial network, and generating image feature predictions of thrombi at future time points.

[0071] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0072] By integrating spatiotemporal biomechanical modeling and dynamic prediction, the efficiency of intelligent processing of thrombotic imaging data is significantly improved. Its core advantage lies in its ability to predict thrombus evolution trajectories based on multi-timepoint imaging data and quantify embolism risk, thereby providing a basis for prospective clinical intervention. This method achieves high-precision simulation of the dynamic evolution of thrombi through intelligent integration of biomechanical parameters and imaging features, effectively overcoming the limitations of traditional static staging, supporting personalized treatment strategy optimization, and ultimately enhancing the accuracy and practicality of clinical decision-making.

[0073] Example 2: By fusing macroscopic imaging features and biomechanical parameters to construct a spatiotemporal graph neural network, the overall morphological evolution of thrombi and embolic risk were predicted. However, this method treats the thrombus as a homogeneous or quasi-homogeneous whole, and its prediction is based on spatial averaging, failing to fully consider and quantify the microscopic heterogeneity within the thrombus. In fact, thrombi are complex structures composed of fibrin, erythrocytes, platelets, and other components mixed in a non-homogeneous manner. The distribution and evolution of these different components directly affect pathological processes such as organization and dissolution, as well as treatment responses. A uniform prediction model struggles to accurately capture this non-homogeneous evolution pattern originating from differences in internal components, resulting in insufficient predictive sensitivity for key local evolutions (such as fibrin network enhancement and erythrocyte core dissolution), thus limiting the accuracy of risk warning and treatment assessment. To further refine the prediction of thrombus evolution and obtain more accurate component quantification results, it is necessary to simultaneously consider the proportional changes and spatial distribution of components within the thrombus and their impact on treatment strategies, requiring further optimization and improvement.

[0074] In some embodiments, to predict the patient's thrombotic status at future points in time, step S4 further includes:

[0075] S41, by using longitudinal multiparameter MRI sequence data of patients and performing super-resolution reconstruction, analyzes the changing patterns of thrombus component proportions at each time point.

[0076] Super-resolution reconstruction involves periodically acquiring longitudinal multi-parameter magnetic resonance imaging (MRI) images of patients to obtain multiple sequence data with varying sensitivities to thrombus components, including T1-weighted, T2-weighted, and proton density-weighted images. For the low-resolution images acquired in each sequence, super-resolution reconstruction is performed using a deep learning model based on generative adversarial networks.

[0077] The study analyzed the variation patterns of thrombus component proportions at each time point. Specifically, it established precise voxel-level correspondences between multiple image sequences and between images at different time points, and processed all images using a non-rigid registration algorithm based on a B-spline free deformation model. This algorithm maximizes the similarity metric between the image to be registered and the selected reference image by optimizing a continuous deformation field defined in three-dimensional space. Simultaneously, it constrains the smoothness of the deformation field through a regularization term, ultimately ensuring that all images from different sequences and at different time points achieve sub-voxel-level precise alignment in anatomical space.

[0078] Under a unified coordinate system with spatial registration, for each voxel within the thrombus region, its signal intensity values ​​across all registered different magnetic resonance sequences are extracted and combined to form a multidimensional vector representing the voxel's characteristics in the multi-parameter image space. Based on the principles of magnetic resonance physics, this multi-sequence signal feature vector for each voxel is substituted into a predefined linear mixture model for thrombus component quantification. This model assumes that the observed signal is the result of a weighted mixture of the characteristic signals of several basic components (such as fibrin and erythrocytes) according to their volume fractions, and estimates the volume proportion of various thrombus components within the voxel by solving a linear least squares problem with nonnegativity and a sum-of-one constraint.

[0079] The aforementioned component quantification calculations are performed on all voxels of the thrombus region at each time point, generating a component quantification map for that time point. This map records a multidimensional vector at each voxel location, representing the estimated proportion of various components at that location. By comparing and analyzing the component quantification maps generated at consecutive time points, the change in the average volume proportion of each component within the entire thrombus or a predefined specific sub-region (such as the core region) over time is calculated. Furthermore, by fitting time-series data on the evolution of each component's proportion over time, its dynamic change patterns are quantitatively described. For example, an autoregressive model or exponential model can be established to characterize the accumulation or clearance rate of specific components, thereby completing the quantitative analysis of the spatiotemporal evolution of components within the thrombus.

[0080] S42 segments and extracts sub-regions within the thrombus from the super-resolution image at each time point, and combines the magnetic resonance physical model to map the pixel-level signal intensity to the component ratio, generating a quantitative component spectrum.

[0081] The pixel-level signal intensity is mapped to component proportions. Specifically, in registered multi-sequence super-resolution images, for each voxel within the thrombus region, its signal intensity in each sequence (e.g., T1- and T2-weighted) is extracted to form a multi-parameter feature vector. This feature vector is then input into a pre-calibrated quantitative magnetic resonance imaging (MRI) model. This model is a linear mixture of equations, whose coefficient matrix defines the theoretical signal values ​​of various pure thrombus components (e.g., fibrin and erythrocytes). By solving a constrained non-negative least squares problem, an optimal set of component volume fractions is found, such that their linear combination best fits the observed feature vector.

[0082] This solution process is applied to each voxel in the thrombus region, calculating a set of values ​​representing the volume percentage of each component within each voxel. The results from all voxels are then integrated to generate a quantitative component map for that time point. This map is a multi-channel 3D dataset, with each channel corresponding to one component, and the value of each voxel representing the estimated proportion of that component at that location.

[0083] S43 constructs a spatiotemporal graph structure using thrombus voxels as nodes, integrates component proportions and image features, and learns the non-uniform evolution pattern of components through a spatiotemporal graph neural network to predict the component distribution of thrombus status at future time points.

[0084] To predict the component distribution of thrombus status at future time points, a unified spatiotemporal graph data structure is constructed based on the quantitative thrombus component atlases generated in the previous stage and the corresponding original image data at all historical time points. Each voxel within the thrombus's three-dimensional space at each time point is defined as a graph node. The feature vector of each node is composed of the voxel's multi-sequence image features and the calculated proportions of various components (such as fibrin and erythrocytes). In the spatial dimension, nodes are connected according to the three-dimensional adjacency relationships between voxels (such as six-neighbor or twenty-six-neighbor domains) to form spatial edges, encoding the interactions of local anatomical structures. In the temporal dimension, through precise image registration techniques, voxel nodes corresponding to the same anatomical location at adjacent time points are connected by temporal edges, forming cross-temporal chains. The connection weights of the temporal edges can be designed as adaptive parameters, dynamically adjusted according to the evolution rate of node features (such as the rate of change in component proportions), allowing the model to focus more on active regions where significant component changes occur.

[0085] The constructed spatiotemporal graph is then input into a specially designed spatiotemporal graph neural network for training and learning. In the spatial dimension, graph convolution operations are used to aggregate the spatial neighbor information of each node, learning the local spatial patterns and heterogeneity of the distribution of components within the thrombus. In the temporal dimension, a sequence model integrating an attention mechanism processes the feature sequences of each node along the time edge, capturing the long-term dependencies and non-uniform changes in component proportions over time. The network is trained end-to-end, learning the intrinsic mapping function from historical spatiotemporal graph features to features at a future target time point.

[0086] The spatiotemporal graph data, containing all information up to the current moment, is input into the trained model. The model, through its learned spatial aggregation and temporal evolution patterns, extrapolates one step forward, outputting a set of predicted values ​​for each node (i.e., each voxel) in the graph. These values ​​represent the predicted proportions of various thrombus components at that spatial location at the future target time point. By re-integrating the prediction results of all nodes according to the original three-dimensional spatial coordinates, a complete, voxel-level accurate prediction map of the distribution of thrombus components at future time points is generated, presented as a heatmap of the proportions of each component.

[0087] S44, Analyze the degree of impact on treatment measures based on the predicted component distribution results.

[0088] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0089] By quantitatively analyzing the spatiotemporal distribution and evolution of thrombus components, this approach achieves a deeper understanding, moving from macroscopic morphological prediction to microscopic pathological mechanism analysis. Its core advantage lies in its ability to accurately predict the component distribution of thrombi at future time points, revealing their intrinsic evolutionary trends such as organization and dissolution in advance. This provides crucial evidence for assessing treatment response and predicting the risk of recanalization or re-occlusion.

[0090] Example 3: In Example 2, when real-time quantification and visualization of the non-uniform evolution of thrombus components based on multi-timepoint thrombus imaging data are performed, dynamic monitoring of component changes can be achieved by constructing a spatiotemporal map structure and predicting component distribution trends. However, the degree of nonlinearity in the evolution of thrombus components varies greatly among different patients or at different times, and there may be critical points in component changes. The unified prediction model focuses on continuous trend analysis and cannot sensitively identify such mutation patterns, resulting in insufficient early warning capability for embolism risk. Since the occurrence of critical points in thrombus components is closely related to the biomechanical environment and individual differences, the indicators required for risk warning must be dynamically changing. In order to further prospectively optimize the prediction of component evolution to capture mutation signals, it is necessary to consider both the dynamic characteristics of component changes and the characteristics of critical points, and make further improvements.

[0091] In some embodiments, step S41, analyzing the changing patterns of thrombus component proportions at each time point, further includes:

[0092] S411 records the time of change in the proportion of thrombus components and the corresponding state of thrombus components, and marks the critical point of change in the historical record based on event backtracking.

[0093] Among them, event retrospective is based on the current recorded time point and retrospectively reviews all quantitative spectra of thrombus components and their derived kinetic curves within a period of time before the event occurred.

[0094] The critical point is defined as follows: after this point in time, the rate of change (first derivative) of the proportion of at least one key component (such as fibrin) or the value of a spatial heterogeneity index (such as entropy) undergoes a statistically significant mutation (e.g., exceeding two standard deviations from the baseline mean) compared to its previous stable baseline; secondly, this mutation pattern persists over time and ultimately leads to a documented clinical event. Experts determine the earliest time point that meets these criteria as the critical point of the event.

[0095] S412 calculates the dynamic change rate of components, spatial distribution entropy, and inter-component interaction coefficients at each time point, and integrates them with biomechanical parameters to construct a thrombus dynamic state vector.

[0096] Specifically, from the quantitative thrombus component profile at each time point, the average volume percentage of various key components within the entire thrombus region is first extracted. Based on the value of this average percentage over time, the difference between the percentage at the previous and next time points is calculated, and then divided by twice the time interval to obtain the instantaneous rate of change of the component at the current moment, thereby quantifying its evolution speed.

[0097] The uniformity of distribution of each component in three-dimensional space is assessed. The three-dimensional space of the thrombus is divided into numerous small units. The volume proportion of each component in each unit is statistically analyzed, and the proportions of all units are normalized to a sum of one, forming a probability distribution describing the spatial distribution of the component. Based on this probability distribution, the degree of disorder in the distribution is quantified by calculating its information entropy value. The entropy value directly corresponds to the uniformity or clustering of the spatial distribution.

[0098] Analyze the interactions between different components. Summarize the average proportion of all key components at this moment to form a multidimensional proportion vector. By calculating the covariance between the dimensions of this vector, analyze the synergistic or antagonistic correlation patterns between the changes in the proportions of different components, and organize this complex correlation information into a feature set.

[0099] Key mechanical parameters were extracted from hemodynamic simulation data at the same time point. The maximum value and fluctuation of the shear force on the thrombus surface were selected as one feature, and the average intensity of pressure change inside the thrombus was calculated as another feature, together forming a feature set describing the current mechanical environment.

[0100] All features obtained from the above steps, including the rate of change of each component, spatial distribution entropy, inter-component interaction feature set, and biomechanical feature set, are sequentially concatenated into a comprehensive high-dimensional feature vector. This vector is then standardized by subtracting the average value of each feature over all time points and dividing by its standard deviation, thereby eliminating the influence of dimensions. The final vector obtained after this processing is the state vector characterizing the dynamic state of thrombosis at that moment.

[0101] S413 employs a general critical point prediction model for thrombus component evolution using neural differential equations, and introduces eigenvalue analysis and anomaly detection algorithms for critical point detection.

[0102] For example, suppose a patient is receiving anticoagulation therapy for left atrial thrombosis caused by atrial fibrillation. Multiparametric MRI data were collected before treatment (day 0) and on days 3, 7, and 14 after treatment, generating corresponding quantitative thrombus component maps and biomechanical simulation results. The thrombus dynamic state vectors for these four time points have been calculated using S412. For instance, the state vector on day 7 shows characteristics such as: fibrin ratio change rate (+0.8% / day), spatial distribution entropy (1.2), covariance with erythrocyte ratio (-0.15), and maximum wall shear force (5.2 Pa).

[0103] A neural differential equation model was trained using the state vector sequences from the patient's first three time points (days 0, 3, and 7). The model learned the following dynamic law: under the patient's current treatment conditions, the thrombotic state typically evolves in a relatively gradual manner, with fibrin accumulating slowly, its distribution becoming more uniform, and the mechanical environment stable. When the scan data from day 14 was input, the trained neural differential equation was first used to predict the state on day 14 based on the state on day 7. An anomaly was found: the actual state vector on day 14 deviated significantly from the predicted value. Specifically, the fibrin proportion change rate surged to +3.5% / day (far exceeding the predicted +1.0% / day), and the fibrin spatial distribution entropy plummeted to 0.5 (predicted 1.3), indicating a sudden high accumulation of fibrin in a localized area of ​​the thrombus. The maximum local wall shear force jumped to 12.5 Pa (predicted 5.5 Pa). The Jacobian matrix eigenvalues ​​corresponding to the current state vector were immediately calculated. Calculations revealed that the real part of the largest eigenvalue changed from a negative value (indicating stability) to a near-zero value of -0.02 (a critical instability signal).

[0104] At this point, the isolated forest anomaly detection algorithm was triggered. This algorithm compared the patient's component change rate pattern on day 14 with a database of historical normal evolution patterns, determining that the rapid fibrin aggregation pattern was a statistically rare anomaly (anomaly score of 0.92, threshold > 0.8, is considered abnormal). Both the eigenvalue approaching zero and the anomaly detection alarm were simultaneously met. Based on this, it was determined that the patient's thrombotic dynamics had deviated from their previous stable trajectory between day 7 and day 14, entering a pre-instability state. Day 10.5 (estimated by interpolation at the onset of instability) was marked as the critical point for embolism risk.

[0105] It should be noted that the above cases are only simulation examples to help understand the concepts and are not intended as specific numerical limitations. In actual applications, the predicted values ​​that may deviate from the actual values ​​need to be determined based on clinical treatment.

[0106] S414 generates graded early warning signals for changes in thrombosis status based on the critical point prediction results.

[0107] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0108] By tracking the dynamic evolution of thrombus components in real time and accurately predicting critical points, early warning of thrombus stability mutations is achieved. This method can effectively identify critical moments of component changes and generate graded warning signals, thereby assisting physicians in taking early intervention measures and significantly reducing the risk of embolism. Compared with traditional static analysis, this approach improves the initiative and accuracy of thrombus management and has high clinical applicability and decision support value.

[0109] Example 4: In Example 3, a predictive warning of the risk of thrombotic system instability was achieved through neural differential equation modeling and critical point detection mechanisms. However, the training and predictive efficacy of this model highly depend on the evolutionary patterns contained in historical real patient data; its essence is a general model based on common group patterns. When faced with new patients with unique anatomical features (such as rare vascular variations, special initial thrombus morphology and location), the general model may exhibit systematic bias in its prediction results because it has not fully learned such specific patterns during training. This bias stems from the model's failure to accurately adapt the dynamic laws derived from the group to the individual's unique blood flow environment and structural constraints. To effectively transform the group predictive capability of the general model into personalized and accurate early warning applicable to any new patient, in some embodiments, the critical point detection in step S413 further includes:

[0110] 3A: Acquire initial thrombotic images and corresponding basic clinical information of new target patients, and use anatomical segmentation models to accurately segment the thrombotic region and the parametric features of the vascular geometry in which it is located from the images.

[0111] Among them, the parameter features are the parameterization of the segmented vascular geometry (such as diameter and curvature) and the initial thrombus morphology (such as volume and approximate location) to form a set of vectors describing the individual anatomical features.

[0112] 3B uses parametric features as conditions for individualized anatomical features and utilizes a diffusion model to generate synthetic thrombotic images that conform to the specific anatomical background of the patient.

[0113] The diffusion model uses individualized anatomical feature vectors as conditional inputs to guide the generation of thrombus images that match the patient's real vascular environment in terms of morphology and location. By adjusting noise sampling and conditional intensity, the diversity of generated thrombi is controlled, such as generating thrombus variants of different sizes and shapes (e.g., attached and floating) to simulate various possible situations.

[0114] Specifically, the individualized anatomical feature vector obtained in step 3A is used as a condition and input into a pre-trained 3D image generator based on a denoising diffusion probability model architecture. The generation process begins with a randomly sampled 3D Gaussian noise field. At each step of the reverse denoising process, the model guides noise prediction based on the anatomical feature conditions, ensuring that the final generated 3D thrombus image strictly conforms to the volume and distribution length defined in the feature vector in terms of spatial morphology (e.g., volume, aspect ratio), and that the thrombus centroid is located at a specified relative position, precisely fitting the geometry of the synthesized vessel wall generated based on vascular parameters. By adjusting the random seed of the initial noise and the intensity weights of the condition vector, the diversity of the generated results can be controlled. This allows for the batch synthesis of a series of hypothetical thrombus variants with varying morphologies, such as fully attached, partially floating, or thrombi with different surface roughnesses, while adhering to the patient's actual anatomical background, thus constructing a synthetic dataset covering multiple possible initial states.

[0115] 3C simulates the dynamic process of each synthetic thrombus image over time, generating corresponding dynamic thrombus time-series data.

[0116] The dynamic thrombus time-series data integrates a simplified computational fluid dynamics (CFD) model and a coagulation response model. Using the initial morphology of the synthetic thrombus and individual blood flow parameters (estimated based on anatomical features) as input, it simulates the dynamic changes of the thrombus's internal components (fibrin, erythrocytes) over time. The simulator generates a series of synthetic thrombus states at multiple future time points, including changes in component proportions and morphological evolution, thus forming a complete synthetic evolution trajectory, labeled with virtual timestamps and component tags.

[0117] 3D involves matching and calibrating the generated individualized synthetic data with existing general critical point prediction models to form a predictive model that adjusts the critical point among different patients.

[0118] Specifically, a calibration dataset is constructed based on a large number of synthetic thrombus dynamic evolution trajectories generated in step 3C. Each trajectory includes a corresponding individualized anatomical feature vector, a thrombus state sequence at multiple time points generated by a physical simulator, and the simulated true critical point time for stability loss. The thrombus state sequence of the above synthetic trajectories is input into a pre-trained general critical point prediction model (step S413) to obtain the model's predicted critical point time and risk probability value. Subsequently, the absolute deviation between the predicted critical point time and the true critical point time is calculated, and the probability error between the predicted risk probability and the binary label of whether the simulated event occurred is evaluated.

[0119] Using the above calculation results, a mapping relationship from individualized anatomical feature vectors to model prediction bias is established as a calibration function. The anatomical feature vector of each trajectory is paired with its corresponding time deviation and probability error as training samples. Regression models such as gradient boosting trees or shallow neural networks are used for training. By optimizing the loss function, the model can accurately estimate the prediction bias based on anatomical features. After training, a trained calibration function is obtained, which can output the prediction correction amount for specific anatomical features.

[0120] For a new target patient, an individualized anatomical feature vector is first extracted from their initial images and input into a general critical point prediction model to obtain the original critical point time and risk probability predictions. Simultaneously, this feature vector is input into a trained calibration function to calculate the prediction time correction and probability adjustment factor specific to that patient. Combining the original prediction values ​​with the corrections, the critical point time is adjusted by addition or subtraction, and the risk probability is scaled and adjusted in logarithmic probability space, thus obtaining the final personalized critical point time and risk probability predictions.

[0121] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0122] A personalized model calibration mechanism was constructed by generating synthetic thrombotic images and evolution time-series data that match the anatomical features of target patients. This mechanism effectively addresses the problem of limited model generalization ability caused by the scarcity of longitudinal time-series data for patients in clinical practice. Using synthetic data, a general prediction model can be accurately calibrated for the vascular geometry and blood flow environment of specific patients, thereby transforming population statistical regularities into individualized risk assessments. This method significantly improves the specificity and reliability of critical point warnings and enables advanced risk quantification and intervention strategy simulation in the initial stage of treatment.

[0123] Figure 2 This is a schematic diagram of the structure of an intelligent processing system based on thrombosis imaging data according to an embodiment of the present invention.

[0124] like Figure 2As shown, an intelligent processing system based on thrombosis imaging data includes: an acquisition and processing module, a reconstruction calculation module, a fusion modeling module, and a risk prediction module.

[0125] The acquisition and processing module is used to determine the period of thrombus based on the longitudinal thrombus image data of the patient and to establish a three-dimensional image of the thrombus development cycle and its surrounding tissues.

[0126] The reconstruction calculation module is used to perform three-dimensional geometric reconstruction of images at each time point, generate three-dimensional models of thrombi and blood vessel walls, and perform hemodynamic simulation and calculate biomechanical parameters based on the models.

[0127] The fusion modeling module is used to fuse image features and biomechanical parameters at each time point into a multidimensional feature tensor to construct a cross-time dimension data representation; a dedicated spatiotemporal convolutional neural network is designed to learn the temporal pattern of biomechanical effects on thrombus evolution.

[0128] The risk prediction module is used to predict the patient's thrombotic status and the probability of embolic events at future time points based on evolving time patterns, hemodynamic simulations, and 3D models. Each module of this intelligent thrombotic imaging data processing system is responsible for executing steps S1-S4 of the above method.

[0129] Preferably, the system may further include an image prediction module, which quantifies the treatment measures into input parameters of a spatiotemporal convolutional neural network model, simulates the changes in thrombus status after intervention through a conditional generative adversarial network, and generates image feature predictions of thrombi at future time points.

[0130] It should be noted that other specific implementations of the intelligent processing system based on thrombosis image data in this invention can refer to the above-described intelligent processing method based on thrombosis image data.

[0131] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent processing of thrombosis imaging data, characterized in that, include: S1. Based on the longitudinal thrombosis image data collected from the patient, determine the cycle of the thrombosis and establish a three-dimensional image of the thrombosis development cycle and its surrounding tissues. S2 performs three-dimensional geometric reconstruction on the images at each time point to generate a three-dimensional model of the thrombus and blood vessel wall, and performs hemodynamic simulation and calculates biomechanical parameters based on the model; S3 integrates the image features and biomechanical parameters at each time point into a multidimensional feature tensor to construct a cross-time dimension data representation; and designs a dedicated spatiotemporal convolutional neural network to learn the temporal pattern of biomechanical effects on thrombus evolution. S4, based on the evolutionary time pattern, predicts the thrombotic status and embolic event risk probability of patients at future time points based on hemodynamic simulation and 3D model.

2. The intelligent processing method based on thrombosis imaging data as described in claim 1, characterized in that, The learning of the temporal pattern of thrombus evolution by biomechanical effects specifically includes: modeling the entire spatiotemporal dataset as a spatiotemporal graph, where the node set contains the vertices of all three-dimensional grids at all time points, the spatial edge set is constructed based on the topological adjacency relationship of the three-dimensional grids to connect directly connected vertex pairs at the same time point, and the temporal edge set connects vertices at corresponding anatomical locations between aligned grids at adjacent time points using a non-rigid registration algorithm, with the connection weights of the temporal edges dynamically adjusted according to the thrombus characteristics of the region where the vertex is located; inputting the spatiotemporal graph into a neural network, aggregating vertex neighbor features in the spatial dimension through a graph attention mechanism, and capturing cross-time point dependencies in the temporal dimension through a Transformer encoder, outputting a vertex spatiotemporal feature sequence; predicting thrombus volume changes and staging probabilities through graph pooling and temporal models; and predicting the probability of embolic events through attention pooling and multilayer perceptrons.

3. The intelligent processing method based on thrombosis imaging data as described in claim 2, characterized in that, The embolic event risk probability includes: extracting the vertex feature set of the last time point from the spatiotemporal graph neural network, aggregating it into a risk representation vector through a risk-aware attention pooling layer; the attention pooling layer calculates a weight for each vertex, the weight value being jointly determined by the vertex feature and the most salient feature in its first-order spatial neighborhood, and normalizing the sum of all weights to one; inputting the weighted and aggregated risk representation vector into a multilayer perceptron for mapping; wherein the multilayer perceptron contains at least one hidden layer with a nonlinear activation function, and its output layer maps the final calculation result to the embolic event probability through a sigmoid function; wherein the formula for calculating the embolic event probability is: ,in, For the predicted probability of embolism, This is the risk representation vector. and Here are the weight matrix and bias vector of the first hidden layer of the multilayer perceptron. To modify the activation function of the linear unit, , The weight vector and bias scalar of the output layer. This is the Sigmoid activation function.

4. The intelligent processing method based on thrombosis imaging data as described in claim 1, characterized in that, The prediction of the patient's thrombotic status at future points in time specifically includes: S41, using longitudinal multi-parameter MRI sequence data of patients and super-resolution reconstruction, analyzes the changing patterns of thrombus component proportions at each time point; S42, segment and extract sub-regions inside the thrombus from the super-resolution image at each time point, and combine the magnetic resonance physical model to map the pixel-level signal intensity to the component ratio to generate a component quantitative spectrum; S43 constructs a spatiotemporal graph structure with thrombus voxels as nodes, integrates component ratios and image features, learns the non-uniform evolution pattern of components through a spatiotemporal graph neural network, and predicts the component distribution of thrombus status at future time points. S44, Analyze the degree of impact on treatment measures based on the predicted component distribution results.

5. The intelligent processing method based on thrombosis imaging data as described in claim 4, characterized in that, The super-resolution reconstruction specifically includes: periodically acquiring longitudinal multi-parameter magnetic resonance images of patients to obtain multiple sequence data with different sensitivities to thrombus components, including T1-weighted, T2-weighted, and proton density-weighted data; and performing super-resolution reconstruction processing on the low-resolution images acquired for each sequence using a deep learning model based on generative adversarial networks.

6. The intelligent processing method based on thrombosis imaging data as described in claim 4, characterized in that, The analysis of the variation patterns of thrombus component proportions at each time point specifically includes: performing temporal registration on the acquired longitudinal multi-parameter magnetic resonance images; using a non-rigid registration algorithm to precisely align all images from different sequences and time points in three-dimensional space at the sub-voxel level, establishing a unified voxel-level spatial correspondence; within the registered space, extracting the signal intensity of each voxel in the multi-sequence images for each voxel in the thrombus region to form a multidimensional feature vector; based on the magnetic resonance physical model, decomposing the multidimensional feature vector of each voxel into the volume proportions of several basic components by solving a constrained linear inversion problem; generating quantitative thrombus component maps at each time point, and quantifying the dynamic changes of each component proportion over time by comparing maps from adjacent time points.

7. The intelligent processing method based on thrombosis imaging data as described in claim 4, characterized in that, In S41, the analysis of the changing patterns of thrombus component proportions at each time point also includes: S411 records the time of change in the proportion of thrombus components and the corresponding state of thrombus components, and marks the critical point of change in the historical record based on event backtracking; S412, calculate the dynamic change rate of components, spatial distribution entropy and inter-component interaction coefficient at each time point, and integrate them with biomechanical parameters to construct a thrombus dynamic state vector; S413 employs a general critical point prediction model for thrombus component evolution using neural differential equations, and introduces eigenvalue analysis and anomaly detection algorithms for critical point detection. S414 generates graded early warning signals for changes in thrombosis status based on the critical point prediction results.

8. The intelligent processing method based on thrombosis imaging data as described in claim 7, characterized in that, The event retrospective includes: using the current recorded time point as a baseline, reviewing all quantitative spectra of thrombus components and their derived kinetic curves within the time window prior to the event.

9. The intelligent processing method based on thrombosis imaging data as described in claim 7, characterized in that, The critical point detection specifically includes: Acquire initial thrombotic images and corresponding basic clinical information of new target patients, and use anatomical segmentation models to accurately segment the thrombotic region and the parametric features of the vascular geometry in which it is located from the images; Using parametric features as conditions for individualized anatomical features, a diffusion model is used to generate synthetic thrombotic images that conform to the specific anatomical background of the patient. For each synthetic thrombus image, simulate its dynamic evolution over time to generate corresponding dynamic thrombus time series data; The generated individualized synthetic data is matched and calibrated with existing general critical point prediction models to form a predictive model for adjusting the critical point among different patients.

10. A system for intelligent processing of thrombosis imaging data, applied to a method for intelligent processing of thrombosis imaging data as described in any one of claims 1 to 9, characterized in that, The system includes: an acquisition and processing module, a reconstruction calculation module, a fusion modeling module, and a risk prediction module; The acquisition and processing module is used to determine the period of thrombus based on the longitudinal thrombus image data of the patient and to establish a three-dimensional image of the thrombus development cycle and its surrounding tissues. The reconstruction calculation module is used to perform three-dimensional geometric reconstruction of images at each time point, generate three-dimensional models of thrombi and blood vessel walls, and perform hemodynamic simulation and calculate biomechanical parameters based on the models. The fusion modeling module is used to fuse image features and biomechanical parameters at each time point into a multidimensional feature tensor to construct a cross-time dimension data representation; a dedicated spatiotemporal convolutional neural network is designed to learn the temporal pattern of biomechanical effects on thrombus evolution. The risk prediction module is used to predict the patient's thrombotic status and the probability of embolic events at future time points based on evolving time patterns, hemodynamic simulations, and 3D models.

Images