Method, device and computer equipment for predicting the risk of recurrence of intracranial atherosclerotic plaque stroke

Abstract

The application relates to an intracranial atherosclerotic plaque stroke recurrence risk prediction method, device and computer equipment. The method acquires a magnetic resonance T1 weighted imaging sequence and a contrast-enhanced T1 weighted imaging sequence of an intracranial atherosclerotic plaque region of a subject; each sequence is divided into habitat subregions, micro features of each habitat subregion and macro features of the whole plaque region are extracted; the micro features and the macro features are fused through a bidirectional cross-region attention mechanism in the same sequence to obtain in-sequence fusion features; the in-sequence fusion features are interactively fused through a bidirectional cross-attention mechanism between two sequences to obtain cross-sequence interaction features; and the cross-sequence interaction features are input into a Transformer encoder for global modeling to output a stroke recurrence risk prediction result. The application can finely depict plaque heterogeneity and fully mine multi-sequence complementary information, and significantly improve the stroke recurrence risk prediction accuracy.

Description

Technical Field

[0001] This application relates to the fields of image processing and artificial intelligence technology, and in particular to a method, apparatus and computer device for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques. Background Technology

[0002] Intracranial arteriosclerosis is one of the main causes of ischemic stroke and transient ischemic attack. Even with standard antiplatelet therapy, statins, and risk factor control, a high percentage of patients still experience stroke recurrence. High-resolution magnetic resonance imaging of the arterial wall can non-invasively assess plaque composition, morphology, and inflammatory activity, providing important information for risk prediction.

[0003] However, existing technologies have the following shortcomings: 1) Traditional radiomics methods treat plaques as a homogeneous whole, ignoring the spatial heterogeneity of components such as the lipid core, hemorrhage, calcification, and fibrous cap within the plaque, resulting in information dilution; 2) Conventional single-sequence analysis fails to fully utilize the functional complementarity between T1-weighted images (showing intraplaque hemorrhage) and contrast-enhanced T1-weighted images (showing inflammatory activity); 3) Existing methods mostly employ simple feature splicing or post-fusion strategies, which cannot effectively model the dynamic interaction between micro-habitat and macro-plaque features, as well as cross-sequence high-order dependencies.

[0004] Therefore, there is an urgent need for a prediction method that can accurately characterize patch heterogeneity, actively mine complementary information from multiple sequences, and achieve end-to-end hierarchical fusion. Summary of the Invention

[0005] Based on this, it is necessary to provide a method, device, computer equipment, computer-readable storage medium, and computer program product for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques that can finely characterize plaque heterogeneity, actively mine multi-sequence complementary information, and achieve end-to-end hierarchical fusion, in order to address the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques, the method comprising: The first imaging sequence and the second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject were acquired. The first imaging sequence refers to the magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to the contrast-enhanced T1-weighted imaging sequence. Based on the local descriptions of each voxel within the patch region of the first imaging sequence and the second imaging sequence, the patch region of each imaging sequence is clustered to obtain multiple habitat subregions. Microscopic features of each habitat subregion and macroscopic features of the entire patch region were extracted separately. Within the same imaging sequence, the microscopic features and the macroscopic features are fused through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence; Between the first imaging sequence and the second imaging sequence, the intra-sequence fusion features of each sequence are interactively fused through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features; The cross-sequence interaction features are input into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction results.

[0007] In one embodiment, the step of clustering the patch regions of each imaging sequence based on the local descriptions of each voxel within the patch regions of the first imaging sequence and the second imaging sequence, respectively, includes: A sliding window is used to compute a local descriptor for each voxel within the patch region; wherein the local descriptor includes at least first-order intensity statistical features and texture features based on the gray-level co-occurrence matrix, gray-level size region matrix and neighborhood gray-level difference matrix; Based on the local descriptor, voxels within the patch region are clustered to obtain multiple habitat subregions.

[0008] In one embodiment, K-means clustering is used to cluster voxels within the patch region, and the Calinski-Harabasz index is used to determine the number of clusters.

[0009] In one embodiment, the optimal number of clusters for the first imaging sequence is 3, and the optimal number of clusters for the second imaging sequence is 2.

[0010] In one embodiment, the fusion of the microscopic features and the macroscopic features within the same imaging sequence using a bidirectional cross-regional attention mechanism to obtain the intra-sequence fused features of the imaging sequence includes: The micro-features of each habitat subregion are mapped to micro-representations, and the macro-features of the entire patch region are mapped to macro-representations. Based on the micro-representation and the macro-representation, calculate the first attention weight of the micro-representation to the macro-representation, and the second attention weight of the macro-representation to the micro-representation; The macroscopic representation is weighted using the first attention weight and then residually connected with the macroscopic representation to obtain the first fused representation; The micro-representation is weighted using the second attention weight and then residually connected with the micro-representation to obtain the second fused representation; The first fusion representation and the second fusion representation are spliced ​​together to obtain the intra-sequence fusion feature.

[0011] In one embodiment, the step of interactively fusing the intra-sequence fusion features of the first imaging sequence and the second imaging sequence through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features includes: Using the intra-sequence fusion features of the first imaging sequence as the first query vector, and the intra-sequence fusion features of the second imaging sequence as the first key vector and the first value vector, the first directional cross-attention output is obtained; Using the intra-sequence fusion features of the second imaging sequence as the second query vector, and the intra-sequence fusion features of the first imaging sequence as the second key vector and the second value vector, the second directional cross-attention output is obtained. The cross-sequence interaction features are obtained by fusing the first-direction cross-attention output and the second-direction cross-attention output.

[0012] In one embodiment, the Transformer encoder includes: a multi-head self-attention layer and a feedforward neural network layer, each of the self-attention layer and the feedforward neural network layer being followed by a residual connection and layer normalization; Specifically, all tokens output by the Transformer encoder are globally averaged and pooled to obtain a fixed-dimensional patient-level representation, which is then passed through a fully connected layer and a sigmoid activation function to output the stroke recurrence risk probability.

[0013] In one embodiment, the method further includes: The stroke recurrence risk prediction results are fused with at least one clinically independent predictor through logistic regression to obtain the final stroke recurrence risk probability; wherein, the clinically independent predictor is selected from one or more of the following: responsible vessel, triglycerides, total cholesterol, glycated hemoglobin, body mass index, and degree of stenosis.

[0014] Secondly, this application also provides a device for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques, the device comprising: The image acquisition module is used to acquire a first imaging sequence and a second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject. The first imaging sequence refers to a magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to a contrast-enhanced T1-weighted imaging sequence. The habitat segmentation module is used to cluster and segment the patch regions of each imaging sequence based on the local description of each voxel within the patch region of the first imaging sequence and the second imaging sequence, respectively, to obtain multiple habitat sub-regions; The feature extraction module is used to extract the microscopic features of each habitat subregion and the macroscopic features of the entire patch region, respectively. The intra-sequence fusion module is used to fuse the microscopic features and the macroscopic features within the same imaging sequence through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence. The inter-sequence interaction module is used to perform interactive fusion of the intra-sequence fusion features of the first imaging sequence and the second imaging sequence through a bidirectional cross-attention mechanism to obtain cross-sequence interaction features. The global prediction module is used to input the cross-sequence interaction features into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction result.

[0015] In one embodiment, it further includes: The clinical feature fusion module is used to perform logistic regression fusion of the stroke recurrence risk prediction results with clinical independent predictors to output the final risk probability; the clinical independent predictors are selected from one or more of the following: responsible vessel, triglycerides, total cholesterol, glycated hemoglobin, body mass index, and degree of stenosis.

[0016] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: The first imaging sequence and the second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject were acquired. The first imaging sequence refers to the magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to the contrast-enhanced T1-weighted imaging sequence. Based on the local descriptions of each voxel within the patch region of the first imaging sequence and the second imaging sequence, the patch region of each imaging sequence is clustered to obtain multiple habitat subregions. Microscopic features of each habitat subregion and macroscopic features of the entire patch region were extracted separately. Within the same imaging sequence, the microscopic features and the macroscopic features are fused through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence; Between the first imaging sequence and the second imaging sequence, the intra-sequence fusion features of each sequence are interactively fused through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features; The cross-sequence interaction features are input into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction results.

[0017] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps: The first imaging sequence and the second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject were acquired. The first imaging sequence refers to the magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to the contrast-enhanced T1-weighted imaging sequence. Based on the local descriptions of each voxel within the patch region of the first imaging sequence and the second imaging sequence, the patch region of each imaging sequence is clustered to obtain multiple habitat subregions. Microscopic features of each habitat subregion and macroscopic features of the entire patch region were extracted separately. Within the same imaging sequence, the microscopic features and the macroscopic features are fused through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence; Between the first imaging sequence and the second imaging sequence, the intra-sequence fusion features of each sequence are interactively fused through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features; The cross-sequence interaction features are input into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction results.

[0018] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps: The first imaging sequence and the second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject were acquired. The first imaging sequence refers to the magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to the contrast-enhanced T1-weighted imaging sequence. Based on the local descriptions of each voxel within the patch region of the first imaging sequence and the second imaging sequence, the patch region of each imaging sequence is clustered to obtain multiple habitat subregions. Microscopic features of each habitat subregion and macroscopic features of the entire patch region were extracted separately. Within the same imaging sequence, the microscopic features and the macroscopic features are fused through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence; Between the first imaging sequence and the second imaging sequence, the intra-sequence fusion features of each sequence are interactively fused through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features; The cross-sequence interaction features are input into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction results.

[0019] The aforementioned method, device, computer equipment, computer-readable storage medium, and computer program product for predicting stroke recurrence risk in intracranial atherosclerotic plaques acquire first and second imaging sequences corresponding to the intracranial atherosclerotic plaque region of the subject. The first imaging sequence refers to a magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to a contrast-enhanced T1-weighted imaging sequence, thereby comprehensively acquiring image data from two dimensions: plaque composition and inflammatory activity. Subsequently, habitat subregions are divided for each sequence of plaque regions, overcoming the drawback of traditional holistic feature extraction leading to dilution of heterogeneous information within the plaque. Based on this, microscopic features of each habitat subregion and macroscopic features of the entire plaque region are extracted to form... This paper describes multi-scale features that complement local details and global structure. Then, within the same sequence, a bidirectional cross-regional attention mechanism enables repeated information exchange and mutual calibration between microscopic and macroscopic representations, achieving dynamic adaptive fusion of local and global features and avoiding information fragmentation caused by simple splicing. Next, a bidirectional cross-attention mechanism between different sequences allows for bidirectional interaction between T1WI and contrast-enhanced T1WI features, actively exploring the functional complementarity between the two sequences to achieve cross-sequence feature alignment and enhancement. Finally, the multi-head self-attention mechanism of the Transformer encoder captures high-order nonlinear dependencies within cross-sequence features and performs global modeling, outputting stroke recurrence risk prediction results. This application achieves accurate prediction of stroke recurrence risk in intracranial atherosclerotic plaques, significantly outperforming traditional holistic radiomics methods and single-sequence analysis methods. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques in one embodiment. Figure 2 This is a schematic diagram illustrating the principle of patch habitat subregion division in one embodiment; Figure 3 Here is a structural diagram of a bidirectional cross-regional attention fusion module in one embodiment; Figure 4 Here is a structural diagram of the Transformer global fusion module in one embodiment; Figure 5 This is a structural block diagram of a device for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques, as shown in one embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0024] Intracranial arteriosclerosis is one of the main causes of ischemic stroke and transient ischemic attack. High-resolution magnetic resonance imaging (MRI) of the arterial wall can non-invasively assess plaque composition, morphology, and inflammatory activity, providing important imaging evidence for predicting stroke recurrence risk. With the development of radiomics and deep learning technologies, quantitative plaque analysis based on MRI images has become a hot research area in risk prediction.

[0025] Currently, the prediction of stroke recurrence risk based on magnetic resonance imaging of intracranial atherosclerotic plaques mainly faces the following technical challenges: First, existing methods typically treat plaques as a homogeneous whole for feature extraction, neglecting the spatial heterogeneity within the plaque. Pathologically, intracranial atherosclerotic plaques comprise multiple components, including a lipid core, intraplaque hemorrhage, calcification, and a fibrous cap. These components are irregularly distributed spatially, forming multiple functional subregions with distinct microenvironmental characteristics. Traditional radiomics methods extract global features from the entire plaque region, averaging or macroscopically statistically analyzing the signal intensity and texture information of various tissue components. This dilutes the unique microscopic information of different subregions, failing to accurately characterize the heterogeneous features of the plaque and thus limiting the ability of risk prediction models to identify high-risk plaques.

[0026] Second, conventional single-sequence magnetic resonance imaging (MRI) analysis fails to fully utilize the functional complementarity information between multiple imaging sequences. T1-weighted imaging sequences can effectively visualize intraplaque hemorrhage (manifested as high signal), while contrast-enhanced T1-weighted imaging sequences after contrast agent injection can assess the degree of neovascularization and inflammatory activity within the plaque (manifested as enhancement). Both reflect the biological stability of the plaque from different dimensions. However, existing methods mostly use single sequences for analysis or simply splice multiple sequence features before inputting them into a classifier, failing to actively explore the inherent functional complementarity between the two sequences, resulting in the ineffective utilization of cross-sequence synergistic information.

[0027] Third, existing methods employ relatively simple feature fusion strategies, failing to effectively model the dynamic interactions between microscopic and macroscopic features, as well as higher-order dependencies across sequences. Current research typically uses serial splicing or post-fusion strategies to combine microscopic habitat features with macroscopic whole-patch features, lacking information exchange and bidirectional calibration mechanisms between the two. Furthermore, complex nonlinear higher-order dependencies exist between features extracted from different magnetic resonance sequences. Traditional machine learning methods (such as support vector machines and random forests) or shallow neural networks struggle to fully capture these higher-order semantic associations, limiting the expressive power and generalization performance of prediction models.

[0028] To address the aforementioned issues, this application aims to propose a method for predicting stroke recurrence risk in intracranial atherosclerotic plaques. Unsupervised clustering is used to finely divide plaques into multiple habitat subregions with different microenvironmental characteristics, extracting microscopic features from each subregion and macroscopic features from the entire plaque. Within the same sequence, a bidirectional cross-regional attention mechanism is used to achieve interactive fusion of microscopic and macroscopic features. Between different sequences, a bidirectional cross-attention mechanism is used to mine complementary information from T1-weighted imaging and contrast-enhanced T1-weighted imaging. Finally, a Transformer encoder is used for global high-order dependency modeling to achieve end-to-end stroke recurrence risk prediction. This invention can effectively characterize plaque heterogeneity, actively mine complementary information from multiple sequences, and significantly improve the accuracy and generalization ability of risk prediction.

[0029] To facilitate understanding of the technical solutions in the various embodiments of this application, a brief explanation of the technical terms that may appear in the various embodiments of this application is provided first: 1) T1-weighted imaging (T1WI) is one of the most fundamental pulse sequences in magnetic resonance imaging (MRI). Its image contrast primarily reflects the differences in longitudinal relaxation time (i.e., T1 relaxation time) between different tissues. In T1WI images, tissues with short T1 relaxation times (such as fat, lipids, and methemoglobin from subacute hemorrhage) appear as high signal (bright), while tissues with long T1 relaxation times (such as water, cerebrospinal fluid, and edema) appear as low signal (dark). Hemorrhagic components within intracranial atherosclerotic plaques exhibit characteristic high signal on T1WI. Therefore, T1WI sequences are widely used for the detection and assessment of intraplaque hemorrhage and are one of the important imaging techniques for identifying high-risk plaques.

[0030] 2) Contrast-enhanced T1-weighted imaging (T1WI+C) refers to T1-weighted imaging sequences acquired after intravenous injection of a paramagnetic contrast agent (usually a gadolinium-containing contrast agent, such as Gd-DTPA). After the contrast agent reaches the neovascularization region of the vessel wall via blood circulation, it shortens the T1 relaxation time of surrounding tissues, resulting in significant enhancement of the richly blood-supply area on T1WI images. Neovascularization of intracranial atherosclerotic plaques is closely related to inflammatory activity and is one of the important markers of plaque instability. By assessing the degree of plaque enhancement on T1WI+C sequences, the level of inflammation and neovascularization density within the plaque can be indirectly reflected, providing crucial functional information for predicting plaque rupture and stroke recurrence risk.

[0031] 3) Proton density-weighted imaging (PDWI) is another fundamental weighting method in magnetic resonance imaging. Its image contrast primarily reflects the differences in hydrogen proton content between different tissues, while minimizing the influence of T1 and T2 relaxation times on image contrast. In PDWI images, tissues with high hydrogen proton density (such as brain parenchyma) appear as high signal, while tissues with low hydrogen proton density (such as cortical bone and air) appear as low signal. PDWI is commonly used to assist in displaying anatomical structures, distinguishing tissue types, and serving as a localization reference for other sequences. In some optional embodiments of this application, PDWI sequences can be used as a supplement to T1WI and T1WI+C sequences for auxiliary localization and boundary identification of plaque regions.

[0032] In this application, the first imaging sequence is a magnetic resonance T1-weighted imaging sequence (T1WI), primarily used to visualize intraplaque hemorrhage; the second imaging sequence is a contrast-enhanced T1-weighted imaging sequence (T1WI+C), primarily used to assess plaque inflammatory activity. The two sequences provide complementary information from the dimensions of plaque composition and plaque biological activity, respectively, and are the core data foundation for multi-sequence fusion prediction in this application's technical solution. Proton density-weighted imaging (PDWI) is only used as an optional auxiliary sequence in some embodiments for auxiliary localization of the plaque region and is not a necessary core sequence in the method of this application.

[0033] 4) The Transformer Encoder is the core component of the Transformer neural network architecture, used for feature encoding and context modeling of the input sequence. Its basic structure consists of stacked multi-head self-attention layers and feed-forward neural network layers, each followed by a residual connection and layer normalization. The multi-head self-attention layer captures global dependencies within the sequence by calculating the attention weights between each element and all other elements in the input sequence; the feed-forward neural network layer performs independent nonlinear transformations on the output at each position to further extract higher-order semantic features. Residual connections alleviate the vanishing gradient problem in deep networks, and layer normalization accelerates training convergence and stabilizes the network optimization process.

[0034] For example, Figure 1 This is a flowchart of a method for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques according to an embodiment of this application, as shown below. Figure 1 As shown, the method in this embodiment may include: Step S101: Obtain the first imaging sequence and the second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject. The first imaging sequence refers to the magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to the contrast-enhanced T1-weighted imaging sequence.

[0035] In this step, a high-field magnetic resonance imaging (MRI) device (e.g., a 3.0T MRI scanner) is used to perform intracranial arterial wall imaging scans on the subject, acquiring MRI image data of the intracranial atherosclerotic plaque region. The acquired image data includes at least two sequences: the first imaging sequence is a magnetic resonance T1-weighted imaging sequence (T1WI), and the second imaging sequence is a contrast-enhanced T1-weighted imaging sequence (T1WI+C). The contrast-enhanced scan is performed after intravenous injection of a paramagnetic contrast agent (e.g., gadolinium-containing contrast agent Gd-DTPA).

[0036] For example, the scanning parameters for high-resolution tube wall imaging can be set to an isotropic resolution of 0.5 mm. The specific parameters can be adjusted according to the actual equipment and clinical needs, and this application does not limit them.

[0037] In one possible implementation, the acquired image data is in DICOM format, and the acquired raw image can be further preprocessed, including: resampling the raw image to an isotropic resolution (e.g., 1×1×1mm). 3The image intensity values ​​were normalized using a standardization method to eliminate grayscale differences between different scanning conditions and different subjects, thereby improving the generalization ability of subsequent cross-subject analysis.

[0038] Step S102: Based on the local description of each voxel within the patch region of the first imaging sequence and the second imaging sequence, the patch region of each imaging sequence is clustered to obtain multiple habitat subregions.

[0039] In this step, the habitat subregions of the patch regions are divided into the T1WI and T1WI+C sequences obtained in step S101. First, the extent of the patch regions in each sequence needs to be determined, which can be achieved through manual delineation, semi-automatic segmentation, or a deep learning-based segmentation model. This application does not limit the segmentation method for the patch regions.

[0040] For example, Figure 2 This is a schematic diagram illustrating the principle of patch habitat subregion division in one embodiment; as shown below. Figure 2 As shown, taking a single patch as an example, the process of generating habitat subregions through voxel-level clustering is demonstrated. After the patch region in the original MRI image undergoes the intermediate voxel feature extraction process, the clustering results are output, and the patch is divided into three regions with different texture fillings (labeled as A, B, and C, respectively). Different regions can be distinguished by different profile lines or dot-like fillings.

[0041] For example, a local descriptor for each voxel is computed within the patch volume using a sliding window. The local descriptor includes: first-order intensity statistics, and higher-order texture features based on the Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Size Zone Matrix (GLSZM), and Neighborhood Gray-Tone Difference Matrix (NGTDM). In one specific implementation, a 3D sliding window of 5×5×5 voxels can be used to compute a 39-dimensional local descriptor per voxel. The first-order intensity statistics include basic statistical indicators such as mean and variance, used to describe the signal intensity distribution characteristics of the voxel's local area; the higher-order texture features are used to describe the texture patterns and spatial distribution patterns within the voxel's neighborhood.

[0042] Then, K-means clustering is performed on the local descriptors of all voxels to divide the patches into multiple habitat subregions. K-means clustering is a classic unsupervised clustering algorithm that iteratively optimizes the allocation of voxels into K clusters, minimizing the distance between voxels within the same cluster in the feature space and maximizing the distance between voxels in different clusters. The Calinski-Harabasz index is used to determine the optimal number of clusters. This index evaluates the clustering effect by calculating the ratio of inter-cluster dispersion to intra-cluster dispersion; a larger value indicates higher clustering quality. In one specific implementation, the number of clusters can be iterated from 2 to 10, and the K value with the largest Calinski-Harabasz index can be selected as the optimal number of clusters.

[0043] Optionally, the optimal number of clusters for the first imaging sequence is 3, and the optimal number of clusters for the second imaging sequence is 2. It should be understood that these values ​​are preferred values ​​obtained based on experimental data from this application. In practical applications, the optimal number of clusters may vary due to factors such as imaging parameters and subject population characteristics, and this application does not limit this number.

[0044] Through the above habitat subregion division, the plaque area that originally appeared as a whole on imaging was finely divided into multiple functional subregions with different signal intensities and texture features. The voxels within each subregion have high consistency in microenvironmental features, while there are significant differences between different subregions, corresponding to different pathological tissue components such as plaque hemorrhage, lipid core, fibrous tissue, calcification, or areas of active inflammation.

[0045] Step S103: Extract the microscopic features of each habitat subregion and the macroscopic features of the entire patch region.

[0046] In this step, for each habitat subregion obtained in step S102, the microscopic features of that subregion are extracted; simultaneously, for the entire patch region, macroscopic features are extracted. Microscopic features reflect the local organizational characteristics within each subregion, while macroscopic features reflect the global attributes of the entire patch. Together, they constitute a multi-scale patch feature description system.

[0047] For example, all features can be computed using the PyRadiomics tool. PyRadiomics is an open-source radiomics feature extraction toolkit that can compute large amounts of quantitative image features according to standardized definitions.

[0048] The micro-features include, but are not limited to: texture features within each habitat subregion (such as energy, contrast, entropy, and correlation based on the gray-level co-occurrence matrix; small-area emphasis, large-area emphasis, and gray-level non-uniformity based on the gray-level region matrix; and roughness, contrast, and complexity based on the neighborhood gray-level difference matrix), intensity distribution features (such as mean, standard deviation, skewness, kurtosis, and percentiles), and local morphological features. The micro-features of each subregion can be concatenated into a high-dimensional feature vector.

[0049] The macroscopic features include, but are not limited to: morphological descriptors of the entire patch region (such as volume, surface area, sphericity, elongation, flatness, etc.), first-order statistics (such as the mean, median, standard deviation, skewness, kurtosis, uniformity, entropy, etc. of the overall intensity), and global texture features. The macroscopic features of the entire patch are concatenated into a global feature vector.

[0050] In one specific implementation, a total of 1561 dimensions of features can be extracted from micro and macro features. The specific dimensions depend on the selected feature type and parameter settings, which are not limited in this application.

[0051] Step S104: Within the same imaging sequence, microscopic and macroscopic features are fused through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence.

[0052] In this step, for each of the T1WI and T1WI+C sequences, the microscopic and macroscopic features of the sequence extracted in step S103 are fused using a bidirectional cross-regional attention mechanism. The core idea of ​​this mechanism is to enable bidirectional information exchange and mutual calibration between microscopic and macroscopic representations in the feature space, achieving dynamic adaptive fusion of local details and global structure.

[0053] For example, Figure 3 Here is a structural diagram of a bidirectional cross-regional attention fusion module in one embodiment; as shown below. Figure 3 As shown in the figure, the microscopic and macroscopic representations input on the left side of the figure undergo operations such as attention weight calculation, element-level weighting, residual connection, and concatenation; the right side of the figure shows the calculation path of cross-attention for two feature vectors from T1WI and T1WIC as Q, K, and V, respectively.

[0054] For example, the microscopic features of each habitat subregion are mapped to microscopic representations using a Multi-Layer Perceptron (MLP), and the macroscopic features of the entire patch region are mapped to macroscopic representations using the same MLP. In one specific implementation, the MLP has a three-layer structure with hidden layer dimensions of 1024, 512, and 256 respectively, and the activation function is ReLU (Rectified Linear Unit). The microscopic and macroscopic representations are mapped to feature spaces of the same dimension to facilitate subsequent attention computation.

[0055] Then, the first attention weight of the micro-representation on the macro-representation and the second attention weight of the macro-representation on the micro-representation are determined. The attention weights are calculated as follows: the representation vector is transformed linearly to obtain the query and key, the dot product of the query and key is calculated, and then normalized by Softmax to obtain the attention weight distribution.

[0056] Furthermore, the macroscopic representation is element-wise weighted using the first attention weight and then connected to the macroscopic representation residual. Similarly, the microscopic representation is element-wise weighted using the second attention weight and then connected to the microscopic representation residual (ResidualConnection). This allows the attention-weighted features to be added to the original features, effectively mitigating the vanishing gradient problem while preserving original feature information, making the network easier to optimize during training. Finally, the weighted macroscopic and microscopic representations are concatenated to obtain the intra-sequence fused features.

[0057] Step S105: Between the first imaging sequence and the second imaging sequence, the intra-sequence fusion features of each sequence are interactively fused through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features.

[0058] In this step, the intra-sequence fusion features of the T1WI sequence and the intra-sequence fusion features of the T1WI+C sequence obtained in step S104 are input into the bidirectional cross-attention module to perform interactive fusion between the two sequences.

[0059] For example, the intra-sequence fusion features of the first imaging sequence are used as the first query, first key, and first value, and the intra-sequence fusion features of the second imaging sequence are used as the second query, second key, and second value. The first-direction cross-attention output and the second-direction cross-attention output are determined respectively. Specifically, the first-direction cross-attention output is: calculating the attention weight using the first query and the second key, and then applying the attention weight to the second value, resulting in an output that uses the features of the first imaging sequence as the query and focuses on the features of the second imaging sequence; the second-direction cross-attention output is: calculating the attention weight using the second query and the first key, and then applying the attention weight to the first value, resulting in an output that uses the features of the second imaging sequence as the query and focuses on the features of the first imaging sequence. The calculation of the two-direction cross-attention outputs can be performed in parallel. The first-direction cross-attention output and the second-direction cross-attention output are added together to obtain the cross-sequence interaction feature. The addition operation converges and fuses the information from the two directions, so that the final cross-sequence interaction feature simultaneously contains the interaction information from both directions: "T1WI focuses on T1WI+C" and "T1WI+C focuses on T1WI".

[0060] Through a bidirectional cross-attention mechanism, the features of T1WI and T1WI+C sequences exchange information and enhance each other, actively exploring the functional complementarity between the two sequences: T1WI sequences contribute plaque composition information, while T1WI+C sequences contribute inflammatory activity information. The two complement and align with each other, forming a cross-sequence joint representation that is more discriminative than any single sequence feature or simple splicing feature.

[0061] After processing in step S105, a fixed number of feature tokens can be obtained. In one specific implementation, the two sequences each correspond to an intra-sequence fused feature. After bidirectional cross-attention interaction, two feature tokens are obtained, each corresponding to an enhanced representation of the two sequences. These tokens are then added together to obtain cross-sequence interactive features.

[0062] Step S106: Input the cross-sequence interaction features into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction results.

[0063] In this step, the cross-sequence interaction features obtained in step S105 are input into the Transformer encoder, which performs global modeling of the high-order nonlinear dependencies and long-range semantic associations within the features.

[0064] For example, the Transformer encoder includes a multi-head self-attention layer and a feed-forward neural network layer. The multi-head self-attention layer uses multiple parallel attention heads to calculate attention weights between features in multiple semantic subspaces, then concatenates the outputs of each head and performs a linear transformation to capture the dependencies between features from different perspectives. The feed-forward neural network layer typically consists of two fully connected layers with ReLU activation function in between, performing independent nonlinear transformations on the output at each position to further extract higher-order semantic features. Residual connections are used to alleviate the vanishing gradient problem in deep networks, and layer normalization is used to accelerate training convergence and stabilize the network optimization process. In one specific implementation, the Transformer encoder can stack multiple identical encoder layers (e.g., 4 layers) to enhance the model's expressive power.

[0065] In this process, all tokens output by the Transformer encoder are globally averaged and pooled to obtain a fixed-dimensional patient-level representation. This representation is then passed through a fully connected layer and a sigmoid activation function to output the stroke recurrence risk probability. The sigmoid activation function maps the output of the fully connected layer to a probability value between 0 and 1, representing the predicted probability that a subject will experience a stroke recurrence within a predetermined time window (e.g., within one year).

[0066] For example, Figure 4 This is a structural diagram of the Transformer global fusion module in one embodiment; such as... Figure 4 As shown, it includes: an input embedding layer (receiving the interaction features of two sequences and adding positional encoding), a multi-head self-attention layer (multiple attention heads side by side), residual connections and layer normalization, a feedforward neural network layer (two fully connected layers with ReLU in between), another set of residual connections and layer normalization, a global average pooling layer, and a fully connected output layer (containing Sigmoid).

[0067] Step S106 achieves end-to-end mapping from image features to risk prediction results, avoiding the limitations of traditional methods that require manual design of classifiers or reliance on doctors' subjective interpretation.

[0068] As an optional embodiment, after step S106 above, the stroke recurrence risk prediction results can be fused with at least one clinically independent predictor through logistic regression to obtain the final stroke recurrence risk probability; wherein, the clinically independent predictor is selected from one or more of the following: responsible vessel, triglyceride (TG), total cholesterol (TC), glycated hemoglobin (HbA1c), body mass index, and degree of stenosis.

[0069] In one specific implementation, the inputs to the logistic regression model include: the image risk probability value output in step S106 and the measured or assigned values ​​of the aforementioned clinical independent predictors (e.g., the responsible vessel can be coded using multi-class classification, and the degree of stenosis can be expressed as a percentage). The logistic regression model learns the optimal fusion weights for each input feature through training and outputs the final comprehensive risk probability. Experimental verification shows that the joint model after fusing image features and clinical features outperforms the single-modality model in predictive performance.

[0070] The method flow of this application is described in detail below with reference to a specific implementation method. It should be understood that the following description is only exemplary and does not constitute a limitation on the scope of protection of this application.

[0071] For example, in a validation experiment, high-resolution magnetic resonance imaging (MRI) of the vessel wall in patients with intracranial atherosclerosis was first acquired, including T1WI and T1WI+C sequences. Scans were performed on a 3.0T MRI scanner with an isotropic resolution of 0.5 mm. Contrast-enhanced scans were performed after intravenous injection of gadolinium contrast agent. The acquired raw images were preprocessed, including resampling to 1×1×1 mm. 3Isotropic resolution and image intensity normalization were implemented. Habitat subregions were defined for patch regions in T1WI and T1WI+C sequences, respectively. Specifically, a 39-dimensional local descriptor for each voxel was calculated within the patch volume using a 5×5×5 voxel sliding window, including first-order statistics and higher-order texture features based on GLCM, GLSZM, and NGTDM. K-means clustering was performed on the descriptors of all voxels. The Calinski-Harabasz index was used to determine the optimal number of clusters: 3 clusters for the T1WI sequence and 2 clusters for the T1WI+C sequence. Microscopic features of each habitat subregion and macroscopic features of the entire patch region were extracted using PyRadiomics, totaling 1561 dimensions. The HFusionNet network was constructed for feature fusion and prediction. Microscopic and macroscopic features are mapped to microscopic and macroscopic representations respectively through a three-layer MLP (1024→512→256). Within the same sequence, the attention weights of microscopic representations on macroscopic representations and macroscopic representations on microscopic representations are calculated through a bidirectional cross-regional attention mechanism. After element-wise weighting and residual connections, they are concatenated to obtain intra-sequence fused features. Between two sequences, interactive fusion is performed through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features. The cross-sequence interactive features are input into a four-layer stacked Transformer encoder (each layer contains a multi-head self-attention layer and a feedforward neural network layer, with residual connections and layer normalization following each layer). After global average pooling and fully connected layers, the image risk probability is output. Optionally, the image risk probability is fused with clinical factors such as the responsible vessel, triglycerides, total cholesterol, glycated hemoglobin, body mass index, and degree of stenosis through logistic regression to obtain the final risk probability.

[0072] The model training employed the AdamW optimizer (β1=0.9, β2=0.999), cosine annealing learning rate scheduling (initial learning rate 0.001), batch size 32, and dropout rate 0.3. Five-fold cross-validation was used on the training set, and the optimal hyperparameters were selected based on the average AUC.

[0073] Experimental results demonstrate that the method described in this application achieves excellent discriminative performance and calibration in predicting the risk of stroke recurrence in patients with intracranial atherosclerotic plaques after standard drug treatment. The imaging model achieved an AUC of 0.834 and a sensitivity of 1.000 on the external test set; the combined model incorporating clinical variables achieved an AUC of 0.890, exhibiting good calibration and a decision curve showing a clear clinical net benefit across the entire threshold probability range. This is significantly superior to traditional radiomics models and single-sequence habitat models, validating the effectiveness and clinical applicability of the proposed technical solution.

[0074] It should be understood that the numbering of the above steps is for descriptive convenience only and does not constitute a limitation on the execution order of the steps. In specific implementations, the steps can be executed in parallel or in a manner different from the described order. For example, the intra-sequence fusion of the T1WI sequence and the T1WI+C sequence in step S104 can be performed in parallel. Those skilled in the art can merge, split, or adjust the execution order of the steps according to actual needs, and these variations are all within the protection scope of this application.

[0075] It should also be understood that the numerical values ​​in the embodiments of this application (such as sliding window size 5×5×5, MLP hidden layer dimensions 1024 / 512 / 256, total number of features 1561, number of clusters 3 and 2, number of Transformer encoder layers 4, etc.) are all illustrative examples and can be adjusted according to specific tasks and data characteristics in actual applications. This application does not limit them.

[0076] For example, Figure 5 This application provides an embodiment of a device for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques. The device in this embodiment may include: The image acquisition module 501 is used to acquire a first imaging sequence and a second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject. The first imaging sequence refers to the magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to the contrast-enhanced T1-weighted imaging sequence. The habitat segmentation module 502 is used to cluster and segment the patch regions of each imaging sequence based on the local description of each voxel within the patch region of the first imaging sequence and the second imaging sequence, respectively, to obtain multiple habitat sub-regions. The feature extraction module 503 is used to extract the micro-features of each habitat sub-region and the macro-features of the entire patch region, respectively. The intra-sequence fusion module 504 is used to fuse microscopic and macroscopic features within the same imaging sequence through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence. The inter-sequence interaction module 505 is used to perform interactive fusion of the intra-sequence fusion features of the first imaging sequence and the second imaging sequence through a bidirectional cross-attention mechanism to obtain cross-sequence interaction features. The global prediction module 506 is used to input cross-sequence interaction features into the Transformer encoder for global modeling to obtain stroke recurrence risk prediction results.

[0077] In some embodiments, the habitat segmentation module 502 is specifically used to: calculate a local descriptor for each voxel within a patch region using a sliding window; wherein the local descriptor includes at least first-order intensity statistical features and texture features based on a gray-level co-occurrence matrix, a gray-level size region matrix, and a neighborhood gray-level difference matrix; and cluster voxels within the patch region based on the local descriptor to obtain multiple habitat subregions.

[0078] In some embodiments, K-means clustering is used to cluster voxels within the patch region, and the Calinski-Harabasz index is used to determine the number of clusters.

[0079] In some embodiments, the optimal number of clusters for the first imaging sequence is 3, and the optimal number of clusters for the second imaging sequence is 2.

[0080] In some embodiments, the intra-sequence fusion module 504 is specifically configured to: map the micro-features of each habitat subregion to a micro-representation, and map the macro-features of the entire patch region to a macro-representation; based on the micro-representation and the macro-representation, calculate a first attention weight of the micro-representation on the macro-representation, and a second attention weight of the macro-representation on the micro-representation; weight the macro-representation using the first attention weight and perform a residual connection with the macro-representation to obtain a first fused representation; weight the micro-representation using the second attention weight and perform a residual connection with the micro-representation to obtain a second fused representation; and concatenate the first fused representation and the second fused representation to obtain an intra-sequence fused feature.

[0081] In some embodiments, the inter-sequence interaction module 505 is specifically configured to: use the intra-sequence fusion features of the first imaging sequence as a first query vector, and use the intra-sequence fusion features of the second imaging sequence as a first key vector and a first value vector to obtain a first directional cross-attention output; use the intra-sequence fusion features of the second imaging sequence as a second query vector, and use the intra-sequence fusion features of the first imaging sequence as a second key vector and a second value vector to obtain a second directional cross-attention output; and fuse the first directional cross-attention output and the second directional cross-attention output to obtain cross-sequence interaction features.

[0082] In some embodiments, the Transformer encoder includes: a multi-head self-attention layer and a feedforward neural network layer, each of which is followed by a residual connection and layer normalization; wherein, all tokens output by the Transformer encoder are subjected to global average pooling to obtain a fixed-dimensional patient-level representation, which is then passed through a fully connected layer and a sigmoid activation function to output the stroke recurrence risk probability.

[0083] In some embodiments, the above-mentioned device may further include: a clinical feature fusion module 507, used to perform logistic regression fusion of stroke recurrence risk prediction results and clinical independent predictive factors to output the final risk probability; the clinical independent predictive factors are selected from one or more of the following: responsible vessel, triglycerides, total cholesterol, glycated hemoglobin, body mass index, and degree of stenosis.

[0084] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method steps of the various embodiments described above.

[0085] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the method steps of the various embodiments described above.

[0086] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0087] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0088] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques, characterized in that, The method includes: The first imaging sequence and the second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject were acquired. The first imaging sequence refers to the magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to the contrast-enhanced T1-weighted imaging sequence. Based on the local descriptions of each voxel within the patch region of the first imaging sequence and the second imaging sequence, the patch region of each imaging sequence is clustered to obtain multiple habitat subregions. Microscopic features of each habitat subregion and macroscopic features of the entire patch region were extracted separately. Within the same imaging sequence, the microscopic features and the macroscopic features are fused through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence; Between the first imaging sequence and the second imaging sequence, the intra-sequence fusion features of each sequence are interactively fused through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features; The cross-sequence interaction features are input into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction results.

2. The method according to claim 1, characterized in that, The step of clustering the patch regions of each imaging sequence based on the local descriptions of each voxel within the patch regions of the first imaging sequence and the second imaging sequence, respectively, includes: A sliding window is used to compute the local descriptor for each voxel within the patch region; wherein the local descriptor includes at least first-order intensity statistical features and texture features based on the gray-level co-occurrence matrix, gray-level size region matrix and neighborhood gray-level difference matrix; Based on the local descriptor, voxels within the patch region are clustered to obtain multiple habitat subregions.

3. The method according to claim 2, characterized in that, K-means clustering was used to cluster voxels within the patch region, and the Calinski-Harabasz index was used to determine the number of clusters.

4. The method according to claim 3, characterized in that, The optimal number of clusters for the first imaging sequence is 3, and the optimal number of clusters for the second imaging sequence is 2.

5. The method according to claim 1, characterized in that, Within the same imaging sequence, the microscopic features and macroscopic features are fused using a bidirectional cross-regional attention mechanism to obtain the intra-sequence fused features of the imaging sequence, including: The micro-features of each habitat subregion are mapped to micro-representations, and the macro-features of the entire patch region are mapped to macro-representations. Based on the micro-representation and the macro-representation, calculate the first attention weight of the micro-representation to the macro-representation, and the second attention weight of the macro-representation to the micro-representation; The macroscopic representation is weighted using the first attention weight and then residually connected with the macroscopic representation to obtain the first fused representation; The micro-representation is weighted using the second attention weight and then residually connected with the micro-representation to obtain the second fused representation; The first fusion representation and the second fusion representation are spliced ​​together to obtain the intra-sequence fusion feature.

6. The method according to claim 1, characterized in that, The step involves interactively fusing the intra-sequence fusion features of the first imaging sequence and the second imaging sequence through a bidirectional cross-attention mechanism to obtain cross-sequence interactive features, including: Using the intra-sequence fusion features of the first imaging sequence as the first query vector, and the intra-sequence fusion features of the second imaging sequence as the first key vector and the first value vector, the first directional cross-attention output is obtained; Using the intra-sequence fusion features of the second imaging sequence as the second query vector, and the intra-sequence fusion features of the first imaging sequence as the second key vector and the second value vector, the second directional cross-attention output is obtained. The cross-sequence interaction features are obtained by fusing the first-direction cross-attention output and the second-direction cross-attention output.

7. The method according to claim 1, characterized in that, The Transformer encoder includes: a multi-head self-attention layer and a feedforward neural network layer, each of the self-attention layer and the feedforward neural network layer being followed by a residual connection and layer normalization; Specifically, all tokens output by the Transformer encoder are globally averaged and pooled to obtain a fixed-dimensional patient-level representation, which is then passed through a fully connected layer and a sigmoid activation function to output the stroke recurrence risk probability.

8. The method according to any one of claims 1 to 7, characterized in that, The method further includes: The stroke recurrence risk prediction results are fused with at least one clinically independent predictor through logistic regression to obtain the final stroke recurrence risk probability; wherein, the clinically independent predictor is selected from one or more of the following: responsible vessel, triglycerides, total cholesterol, glycated hemoglobin, body mass index, and degree of stenosis.

9. A device for predicting the risk of stroke recurrence in intracranial atherosclerotic plaques, characterized in that, The device includes: The image acquisition module is used to acquire a first imaging sequence and a second imaging sequence corresponding to the intracranial atherosclerotic plaque region of the subject. The first imaging sequence refers to a magnetic resonance T1-weighted imaging sequence, and the second imaging sequence refers to a contrast-enhanced T1-weighted imaging sequence. The habitat segmentation module is used to cluster and segment the patch regions of each imaging sequence based on the local description of each voxel within the patch region of the first imaging sequence and the second imaging sequence, respectively, to obtain multiple habitat sub-regions; The feature extraction module is used to extract the microscopic features of each habitat subregion and the macroscopic features of the entire patch region, respectively. The intra-sequence fusion module is used to fuse the microscopic features and the macroscopic features within the same imaging sequence through a bidirectional cross-regional attention mechanism to obtain the intra-sequence fusion features of the imaging sequence. The inter-sequence interaction module is used to perform interactive fusion of the intra-sequence fusion features of the first imaging sequence and the second imaging sequence through a bidirectional cross-attention mechanism to obtain cross-sequence interaction features. The global prediction module is used to input the cross-sequence interaction features into the Transformer encoder for global modeling to obtain the stroke recurrence risk prediction result.

10. The apparatus according to claim 9, characterized in that, Also includes: The clinical feature fusion module is used to perform logistic regression fusion of the stroke recurrence risk prediction results with clinical independent predictors to output the final risk probability. The clinically independent predictive factors are selected from one or more of the following: responsible vessel, triglycerides, total cholesterol, glycated hemoglobin, body mass index, and degree of stenosis.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images