Cerebral hemorrhage heterogeneity quantitative evaluation system based on cyclic cross-attention clustering

Abstract

The invention discloses a cerebral hemorrhage heterogeneity quantitative evaluation system based on cyclic cross-attention clustering, and relates to the field of cerebral hemorrhage evaluation, and the system comprises an image segmentation module which is used for carrying out the three-dimensional segmentation of a brain plain-scan CT image through an nnU-NetV2 model, and outputting the binary masks of a hematoma region and an edema region; the feature extraction module is used for extracting high-dimensional pixel features from the mask region and generating an initial feature embedding matrix; the cyclic cross-attention clustering module is used for dynamically updating a clustering center through a multi-scale cyclic EM cross-attention mechanism and executing clustering analysis on the feature embedding matrix; the heterogeneity scoring module is used for calculating information entropy and generating hematoma and edema heterogeneity scores; and the prognosis prediction module is used for predicting the hematoma expansion risk and the neural function outcome through a machine learning model. According to the scheme, objective quantification of cerebral hemorrhage heterogeneity can be realized, and the prediction accuracy of the hematoma expansion risk and the neural function outcome is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cerebral hemorrhage evaluation, in particular to a cerebral hemorrhage heterogeneity quantitative evaluation system based on a recurrent cross-attention clustering. BACKGROUND

[0002] Cerebral hemorrhage is one of the acute cerebrovascular diseases with the highest mortality and disability rate in the world, and the mortality rate is more than 50% one year after the onset. The poor prognosis is mainly related to hematoma enlargement and neurological damage, and hematoma heterogeneity has been confirmed as a key imaging feature for predicting hematoma enlargement and poor prognosis. In existing clinical practice, heterogeneity evaluation relies on nine signs of plain CT, but the sensitivity of such signs is only 28%-63%, and the interpretation is highly dependent on the subjective experience of radiologists, with low consistency between different observers. In addition, traditional signs cannot quantify the subtle structural differences in the hematoma and the surrounding edema area, making it difficult to reveal the deep relationship between heterogeneity and prognosis. Therefore, developing an objective and accurate heterogeneity quantitative tool is a core challenge to improve the prognosis prediction of cerebral hemorrhage.

[0003] The existing methods mainly have four limitations: 1. Although the heterogeneity characterization method based on dual-energy CT can improve accuracy, it is difficult to be widely applied in primary medical settings due to low device penetration rate. 2. Most deep learning models achieve automatic segmentation, but only focus on the internal features of the hematoma, ignoring the spatial distribution and structural correlation of the surrounding edema area, resulting in a lack of global perspective in heterogeneity evaluation. 3. Traditional clustering methods cannot dynamically optimize feature embedding, and have not established a quantitative bridge between heterogeneity and clinical endpoints. For example, although some studies have proposed anatomically site-specific thresholds for hematoma volume, such static thresholds have limited clinical generalizability as they do not integrate dynamic evolution information of heterogeneity. 4. The self-attention mechanism of the mainstream Transformer architecture has a complexity of O(H²W²D), which is too large for processing three-dimensional medical images, making it difficult to meet the real-time decision-making needs of clinical practice. Therefore, based on the above problems, the present application proposes a cerebral hemorrhage heterogeneity quantitative evaluation system based on a recurrent cross-attention clustering. SUMMARY

[0004] OBJECTIVE To solve the above problems, the purpose of the present application is to provide a cerebral hemorrhage heterogeneity quantitative evaluation system based on a recurrent cross-attention clustering, which aims to replace the traditional nine signs of artificial interpretation with an automated algorithm, eliminate subjective bias, improve the sensitivity and consistency of heterogeneity evaluation, reduce the spatiotemporal complexity of three-dimensional medical image processing, meet the real-time analysis needs of clinical practice, and provide data-driven decision support for individualized treatment.

[0005] TECHNICAL SCHEME In order to achieve the above object, the application provides a cerebral hemorrhage heterogeneity quantitative evaluation system based on cyclic cross-attention clustering, which adopts an nnU-NetV2 model to perform three-dimensional segmentation on a plain CT image, generates a hematoma and edema region mask, extracts high-dimensional pixel features through three-dimensional convolution operation, constructs an initial feature embedding matrix, introduces a cyclic cross-attention mechanism, iteratively executes E steps and M steps, realizes dynamic optimization and multi-scale clustering of feature embedding, calculates information entropy based on the spatial distribution of the clustering results, generates a normalized hematoma and edema heterogeneity score, maps the clustering categories to the original image to generate a color space marker graph, and finally inputs the heterogeneity score and patient clinical data into a logistic regression model to output a hematoma enlargement probability and an adverse neurological outcome risk value. The scheme reduces the computational complexity through the cyclic EM mechanism, and enhances the internal structure representation of clustering by using feature scheduling weighting, and finally forms a closed-loop analysis framework.

[0006] In a first aspect, the application provides a cerebral hemorrhage heterogeneity quantitative evaluation system based on cyclic cross-attention clustering, comprising: An image segmentation module is configured to perform three-dimensional segmentation on a brain plain CT image by using a pre-trained nnU-NetV2 model, and output binary masks of hematoma and edema regions. A feature extraction module is configured to extract high-dimensional pixel features from the mask region through three-dimensional convolution operation, and generate an initial feature embedding matrix. A cyclic cross-attention clustering module is configured to dynamically update the clustering center through a multi-scale cyclic EM cross-attention mechanism, and perform clustering analysis on the feature embedding matrix. A heterogeneity scoring module is configured to calculate information entropy based on the spatial distribution of the clustering results, and generate a normalized hematoma and edema heterogeneity score. A prognosis prediction module is configured to combine the heterogeneity score with clinical data, and predict the hematoma enlargement risk and neurological outcome by using a machine learning model.

[0007] Further, the feature extraction module is further configured to synchronously fuse the hematoma internal structure feature and the edema region feature around the hematoma, and construct a global three-dimensional heterogeneity representation.

[0008] Further, the cyclic cross-attention clustering module realizes dynamic clustering optimization by iteratively executing the following operations: E step: generate a query vector based on the current clustering center, and calculate a soft clustering assignment matrix by combining the generated key vector with the image feature; M step: update the clustering center by weighting according to the soft clustering assignment matrix.

[0009] Furthermore, the cyclic cross-attention clustering module achieves computational complexity optimization by sharing query key projection weights, and its time complexity is lower than that of the traditional self-attention mechanism. The clustering results are mapped to the original CT image to generate a spatial distribution color map, marking different heterogeneous sub-regions.

[0010] Furthermore, the cluster center is initialized by selecting initial points from the image grid features through an adaptive pooling operation, and is optimized and generated using a position feedforward neural network.

[0011] Furthermore, the heterogeneity scoring module generates a heterogeneity score by the following steps: Statistical analysis of pixel distribution probabilities of each cluster category in the hematoma and edema areas; Calculating an information entropy value based on the probability distribution; The entropy values ​​are normalized to the range [0,1] to generate standardized scores.

[0012] Furthermore, the prognosis prediction module adopts a logistic regression model, whose input features include heterogeneity score, patient age, gender and medical history data, and the output is the probability of hematoma expansion or the probability of adverse neurological function outcome.

[0013] Furthermore, a cross-domain feature coupling module is included to capture the spatial conduction law of heterogeneity by calculating the density gradient field from the hematoma boundary to the edema area. The gradient field construction process of the segmented hematoma mask and edema mask is as follows:

[0014] Where, is the three-dimensional spatial gradient synergy value; is the distance attenuation weight; and They are the CT values ​​of the hematoma area and the CT values ​​of the edema area; is the density difference adjustment factor; is the natural exponential function; is the spatial attenuation coefficient; is the Euclidean distance from the current point to the hematoma boundary.

[0015] By constructing a density gradient field at the hematoma-edema interface, cross-domain structural correlation information is systematically integrated. This mechanism overcomes the limitations of traditional methods that focus on a single region, establishing a quantitative model of heterogeneous spatial conduction. This significantly improves the integrity of feature representation, enhances segmentation accuracy and the biological plausibility of prognostic prediction, while reducing the risk of boundary misjudgment and providing an anatomically interpretable basis for heterogeneity assessment.

[0016] Further, a dynamic weight distribution module is further included, and high confidence features are given stronger influence by calculating the matching confidence of the feature vector and the cluster center, wherein the confidence weight generation formula is:

[0017] In the formula, is a feature of the cluster confidence weight; is a natural exponential function; , and are adjustment parameters; is the m-th feature vector; is the m-th cluster center; is the m-th cluster center; is the m-th cluster center; is the intra-class standard deviation; is the preset total number of clusters; is the m-th cluster center; is the m-th cluster center; is the m-th cluster center; is the m-th cluster center; is the base of the natural logarithm; is a feature stability index.

[0018] Based on the feature stability dynamic distribution of cluster weight, the noise interference problem caused by the difference of feature reliability is innovatively solved. The scheduler suppresses the contribution of low-quality features through the confidence gating mechanism, synchronously accelerates the convergence of clustering, and improves the robustness of the results, optimizes the tightness and separation of clustering, and strengthens the stability of heterogeneity score and prognosis relevance.

[0019] Further, a feature scheduling sub-module is further included, which is used to calculate the similarity weight of each feature vector and the cluster center, and update the feature embedding through weighted average to enhance the internal structure representation of the cluster.

[0020] In a second aspect, the application further provides a cerebral hemorrhage heterogeneity quantification evaluation method based on cyclic cross-attention clustering, which is based on the system of the first aspect and comprises the following steps: Perform three-dimensional segmentation on the brain CT image through the nnU-NetV2 model to generate binary masks of the hematoma region and the edema region; Based on the mask, the gradient synergy field of the hematoma and edema boundary region is calculated, and the gradient field is spliced and fused with the original feature embedding extracted by three-dimensional convolution to generate enhanced features; The enhanced features are iteratively executed by E-step and M-step through the cyclic EM cross-attention mechanism; According to the spatial distribution of the clustering result, the information entropy is calculated, and the hematoma and edema heterogeneity score is normalized and output; The heterogeneity score is input into a machine learning model combined with clinical data to predict hematoma enlargement risk and neurological outcome probability.

[0021] In a third aspect, the present application further provides a computer device comprising a management platform and a memory, the management platform being connected to the memory, the memory being used to store a computer program, and the management platform being used to execute the computer program stored in the memory to enable the computer device to execute the aforementioned brain hemorrhage heterogeneity quantification evaluation method based on cyclic cross-attention clustering.

[0022] In a fourth aspect, the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a management platform to implement the aforementioned brain hemorrhage heterogeneity quantification evaluation method based on cyclic cross-attention clustering.

[0023] The nnU-NetV2 model is used to realize high-precision segmentation of the hematoma and edema regions, and a gradient synergy field is constructed to integrate the spatial conduction features of the hematoma and edema boundary regions; a cyclic cross-attention mechanism is introduced to iteratively update the clustering centers, and a confidence threshold is used to dynamically optimize the heterogeneity representation by combining the noise interference suppression; a normalized hematoma and edema heterogeneity score is constructed based on the clustering distribution information entropy to eliminate the subjective threshold dependence; and the heterogeneity score is input into a machine learning model combined with clinical data to output the hematoma enlargement risk and neurological outcome probability. This scheme replaces manual sign interpretation with automatic quantification, eliminates subjective bias, improves heterogeneity sensitivity and interpretation consistency; the cyclic EM mechanism reduces the complexity of three-dimensional image processing, meets the real-time analysis requirements in clinical practice, and realizes the dynamic correlation between heterogeneity and prognosis based on plain CT, providing interpretable decision support for individualized diagnosis and treatment; and finally, a precise analysis framework is formed from image input to prognosis output, significantly improving the scientific nature of brain hemorrhage risk stratification and intervention strategies.

[0024] Advantages By implementing the aforementioned brain hemorrhage heterogeneity quantification evaluation system based on cyclic cross-attention clustering, the following technical effects are achieved: (1) The cyclic EM mechanism is used to dynamically update the clustering centers and feature assignments, overcoming the defects of traditional static optimization clustering. Through multi-scale feature interaction, the adaptive refinement of heterogeneity representation is realized, significantly reducing the computational complexity of three-dimensional image processing, balancing the computational efficiency and feature analysis depth, and constructing a high-precision heterogeneity quantification pipeline.

[0025] (2) Based on the information entropy theory, a normalized scoring system is designed to convert the spatial clustering distribution into a scalar index. The global heterogeneity information of the hematoma and edema regions is innovatively integrated, solving the problem of relying on subjective threshold in existing methods. It can establish an objective bridge between image features and clinical endpoints, and improve the universality and operability of prognosis prediction.

[0026] (3) By constructing the density gradient field of the hematoma and edema interface, the cross-domain structural correlation information is systematically integrated. This mechanism breaks through the limitation of traditional methods focusing on a single region, establishes a quantitative model of heterogeneous spatial conduction, significantly improves the completeness of feature representation, enhances the segmentation accuracy and biological rationality of prognosis prediction, and reduces the risk of boundary misjudgment, providing anatomically interpretable basis for heterogeneity assessment.

[0027] (4) Based on the dynamic allocation of clustering weights according to feature stability, the noise interference problem caused by feature reliability difference is innovatively solved. The confidence threshold gating mechanism of this scheduler suppresses the contribution of low-quality features, simultaneously accelerates the clustering convergence and improves the result robustness, optimizes the clustering tightness and separation, and strengthens the stability of heterogeneity scoring and prognosis correlation. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to make the above-mentioned cerebral hemorrhage heterogeneity quantification evaluation system based on the cycle cross-attention clustering of the present application more obvious and easy to understand, the drawings needed in the specific embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained from these drawings without creative labor for those skilled in the art.

[0029] Figure 1 The technical roadmap of the present application is shown in the figure. Figure 2 The nnU-NetV2 model segmentation example is shown in the figure. Figure 3 The consistency between the radiologist segmented volume and the nnU-NetV2 predicted volume is shown in the figure. DETAILED DESCRIPTION

[0030] Example 1: The technical roadmap of the present application is shown in the figure. Figure 1As shown, (1) Clinical image data: 500 cases of intracerebral hemorrhage patient data were collected from A and B hospitals, ensuring anonymization and secure storage. (2) Automatic segmentation and feature extraction: The nnU-NetV2 model was applied to segment CT images, generating hematoma and edema region masks. High-dimensional features were extracted using three-dimensional convolution techniques to form feature vectors. (3) Recurrent cross-attention clustering: Each recurrent cross-attention clustering layer performs T times of cross-attention clustering and center updating. (4) Feature scheduling: Reassign feature embeddings on updated cluster centers. Perform T times of iteration within a recurrent EM cross-attention layer. (5) Global distribution analysis and heterogeneity measurement: Map the clustering results onto the original CT images to generate color-labeled maps. Calculate the information entropy and normalize it to obtain the hematoma and edema heterogeneity score. (6) Clinical relevance analysis: Use single-factor test to evaluate the association between hematoma heterogeneity score and the risk of hematoma enlargement within 24 hours. Control confounding factors through multivariate logistic regression to analyze the relationship between the score and 6-month adverse neurological outcomes, and calculate the odds ratio and its 95% confidence interval. Details are as follows.

[0031] This example intends to retrospectively collect data of 500 patients with intracerebral hemorrhage from A and B hospitals, respectively, among which 400 cases are collected from A hospital and 100 cases are provided from B hospital. All patient information has been anonymized and stored securely to ensure data security and privacy. The inclusion criteria include patients with intracerebral hemorrhage aged 18 to 80 years old, Glasgow coma score ≥5 and onset time within 6 hours. The exclusion criteria include patients with intracerebral hemorrhage secondary to aneurysm, vascular malformation, anticoagulant therapy, tumor or head trauma, as well as patients with severe renal, liver or respiratory failure and patients who cannot comply with the treatment regimen.

[0032] The collected data covers the basic demographic information, clinical characteristics, image data, laboratory test results, treatment methods and their responses, and the modified Rankin Scale score at 6 months.

[0033] Firstly, medical image automatic segmentation technology will be used to process the head CT image, and the hematoma region and perihematoma edema region will be accurately segmented. Then, for each pixel in the mask region, the image features of its surrounding neighborhood are calculated as local feature representation.

[0034] (1) Hematoma and edema automatic segmentation The pre-trained nnU-NetV2 model is applied to the head CT image to automatically output the binary mask of the hematoma region and the perihematoma edema region for subsequent feature extraction and analysis. An example of nnU-NetV2 model segmentation is shown in Figure 2 ​The left column shows the original image; the middle column shows the radiologist annotation, with orange lesions representing intraparenchymal hemorrhage, yellow lesions representing intraventricular hemorrhage, and green lesions representing perihematomal edema; and the right column shows the model prediction results, with red lesions representing intraparenchymal hemorrhage, pink lesions representing intraventricular hemorrhage, and blue lesions representing perihematomal edema.

[0035] The consistency between the radiologist segmented volume and the nnU-NetV2 predicted volume is shown in FIG. 6, where the x-axis represents the radiologist segmented volume and the y-axis represents the nnU-NetV2 predicted volume. Figure 3

[0036] (2) Feature extraction A suitable neighborhood size (3x3) is selected to capture features. A three-dimensional convolution is used to extract high-dimensional features of each pixel and its surrounding neighborhood, resulting in a feature vector.

[0037] (3) Feature embedding The extracted feature vectors are combined to form an initial feature embedding matrix, where and represent the height and width of the image, respectively, represents the dimension of the matrix, represents the dimension of the feature vector. Then, the features of the hematoma region and the edema region are extracted and combined into a whole feature dataset.

[0038] The merged feature embedding is subjected to multi-scale feature extraction and clustering using the recurrent cross-attention clustering method in the ClusterTransformer model.

[0039] (1) Initial center setting The cluster centers are initialized from the image grid through an adaptive pooling operation, selecting feature points from the feature embedding , and using a position feedforward neural network to further process these feature points to generate the final initial cluster centers , with the specific formula as follows:

[0040] In the formula, is the position feedforward neural network used to optimize the initial feature points to generate cluster centers; is the adaptive pooling operation.

[0041] (2) Recurrent EM cross-attention In each stage, the following recurrent EM cross-attention steps are performed to achieve clustering: ① E step: First, project the query vector from the cluster centers . Then, from the image features​ Projection key vector Finally, the soft clustering assignment matrix is calculated to represent the probability of each pixel belonging to each cluster, as follows:

[0042] wherein, is the soft clustering assignment matrix, representing the probability of each pixel belonging to each cluster; is normalized along the dimension of the cluster categories.

[0043] Step 2: M-step: From the image features projection value vector. According to the soft clustering assignment matrix and the value vector update the cluster centers. The above steps are repeated times until the clustering result converges.

[0044]

[0045] wherein, is the updated cluster center matrix; is the soft clustering assignment matrix generated in the E-step; is the value vector generated from the image features projection.

[0046] After completing the clustering assignment, the similarity is used to schedule the feature blocks within each cluster to optimize the feature embedding and enhance the understanding of all aspects of the CT image. The specific process is as follows: (1) Similarity calculation For each feature embedding , calculate its similarity with all cluster centers .

[0047] (2) Feature weighting According to the similarity, the cluster centers are weighted and averaged to update the feature embedding:

[0048] wherein, is the updated feature vector; is a multi-layer perceptron for nonlinear transformation; is the total number of preset clusters; is a similarity calculation function.

[0049] (3) Feature update Through the above similarity weighting process, the updated feature embedding is generated, so that the feature representation can better reflect the structural information within the cluster.

[0050] With the clustering results after feature scheduling, global distribution analysis and heterogeneity measurement are directly performed. The specific steps are as follows: (1) Clustering visualization The clustering results are mapped onto the original CT images to generate color-labeled maps, which are used to show the spatial distribution of different clustering categories in the hematoma and edema regions.

[0051] (2) Entropy calculation Use information entropy to measure the disorder degree of category distribution. The higher the information entropy, the greater the internal heterogeneity of the hematoma and edema region. First, the number of pixels of each clustering category is counted. Second, the probability of each clustering category is calculated, i.e. the number of pixels belonging to the th category divided by the total number of pixels.

[0052] (3) Maximum entropy value calculation In order to convert the entropy value into a standardized heterogeneity score, the maximum entropy value Hmax needs to be calculated first. The maximum entropy value corresponds to the case of completely uniform distribution, i.e. the probability of each category is equal.

[0053] (4) Heterogeneity score The entropy value is normalized to the range of [0, 1] to obtain the hematoma and edema heterogeneity score, which provides an intuitive indicator to reflect the heterogeneity of the hematoma and edema region. The closer the score is to 1, the higher the heterogeneity in the region; the closer the score is to 0, the lower the heterogeneity in the region. The complete heterogeneity score calculation formula is:

[0054] In the formula, is the standardized hematoma and edema heterogeneity score; is the pixel proportion of the th clustering category.

[0055] In order to explore the correlation between hematoma heterogeneity score and hematoma enlargement and neurological outcome, a retrospective data analysis will be performed using single factor test and multi-factor logistic regression to evaluate the role of hematoma heterogeneity score in the risk of hematoma enlargement within 24 hours, defined as hematoma volume increase ≥6ml or more than 33%. At the same time, the relationship between hematoma heterogeneity score and 6-month adverse neurological outcome, defined as modified Rankin Scale mRS ≥4 points, is analyzed by logistic regression model. These analyses will control other potential confounding factors such as patient age, gender and medical history to calculate values and their 95% confidence intervals, and the statistical significance is judged by value, so as to evaluate the practicability of hematoma heterogeneity score in predicting patient prognosis and further verify the clinical value of the scoring system.

[0056] Embodiment 2 On the basis of the foregoing embodiments, in view of the problem that the existing method ignores the dynamic interaction of the hematoma edema boundary area, a gradient-driven cross-domain feature coupling method is proposed. By calculating the density gradient field of the hematoma boundary to the edema area, the conduction rule of heterogeneity in space is captured, a synergistic attenuation factor is introduced to quantify the contribution of gradient direction and intensity to heterogeneity, and a mathematical model of structure correlation in three-dimensional space is established.

[0057] Perform gradient field construction on the segmented hematoma mask and edema mask:

[0058] In the formula, is the three-dimensional space gradient synergy value; is the distance attenuation weight; and are the CT values of the hematoma area and the edema area, respectively; is the density difference adjustment factor; is the natural exponential function; is the spatial attenuation coefficient; is the Euclidean distance from the current point to the hematoma boundary.

[0059] Fuse with the original feature embedding matrix:

[0060] In the formula, is the enhanced feature embedding; is the feature embedding; is the channel splicing operation; is a nonlinear activation function; is a learnable weight matrix; is a learnable bias vector.

[0061] Verification shows that in the case of obtaining similar average error as the above embodiments, the Dice coefficient is increased by 3.2%, the hematoma expansion prediction AUC is increased by 5.6%, and the misdiagnosis rate is reduced by 37%. The results show that this method models the bidirectional density gradient conduction of the hematoma edema boundary, systematically improves the anatomical rationality of heterogeneity representation, and can significantly enhance the recognition ability of key pathological regions in the feature space, effectively overcoming the feature fragmentation defect of the traditional method for the boundary area. At the same time, the introduction of the gradient synergy field significantly improves the sensitivity of the model to subtle density changes, providing a spatial basis with biological interpretability for heterogeneity scoring.

[0062] Embodiment 3 Building on the previous examples, we propose a confidence-based dynamic weight assignment strategy to address the flaw of traditional cluster center updates, which ignore differences in feature reliability. By calculating the matching confidence between feature vectors and cluster centers, we assign greater influence to high-confidence features, suppressing noise interference and accelerating cluster convergence.

[0063] For the first The confidence weights are generated by clustering centers and feature vectors:

[0064] Where, Features Clustering The confidence weight of is the natural exponential function; It is an adjustment parameter, usually 0.5; For the feature vectors; For the cluster centers; is the within-class standard deviation; is the total number of preset clusters; For the cluster centers; For the The within-cluster standard deviation of each cluster; is the base of natural logarithms; is an adjustment parameter, generally 1.2; is the characteristic stability index; It is an adjustment parameter, generally 0.8.

[0065] The weight Substitute into the M-step update:

[0066] Where, After the update cluster centers; The total number of valid feature vectors participating in the current cluster center update; is the image feature projection value vector.

[0067] For example, suppose that after feature extraction of CT images of cerebral hemorrhage, 5 feature vectors are obtained and the number of cluster centers is . Characteristic stability index Through neighborhood consistency calculation, the specific data is as follows: Table 1. Summary of eigenvector data

[0068] calculate right the confidence weight of each feature; updating the cluster centers;

[0069] Performance comparison of adaptive dynamic clustering scheduler and traditional K-means on cerebral hemorrhage dataset is shown in Table 2.

[0070] Table 2, adaptive dynamic clustering scheduler effect summary

[0071] According to the experiment table, higher weight is obtained due to high stability characteristics, effectively reducing invalid iterations; the weight of low stability characteristics is reduced, improving the anti-interference ability; and the heterogeneity score is associated with the enhanced prognosis. The experimental results show that the scheduler gives high stability characteristics stronger influence through the confidence gating mechanism, and essentially optimizes the anti-noise and convergence certainty of the clustering process; compared with the traditional method, it can effectively suppress abnormal clustering drift caused by local artifacts or imaging noise, so that the heterogeneity score result remains highly consistent between different data batches; more importantly, the design significantly speeds up the topology optimization process of the feature space, providing more stable distribution input for subsequent entropy calculation.

[0072] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems or computer program products. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable non-transitory storage media having computer-usable program code embodied in the medium.

[0073] The present application can provide computer program instructions to the management platform of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing devices to generate a machine, so that the instructions executed by the management platform of the computer or other programmable data processing devices generate means for implementing the system.

[0074] These computer program instructions can also be stored in a computer-readable storage medium that can guide the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable storage medium generate a product including instruction means, which implements the functions of the system.

[0075] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions of the system.

Claims

1. A system for quantitatively evaluating cerebral hemorrhage heterogeneity based on cyclic cross-attention clustering, characterized in that, Comprise: An image segmentation module for performing three-dimensional segmentation on brain CT images using the nnU-NetV2 model, and outputting binary masks of hematoma and edema regions; A feature extraction module for extracting high-dimensional pixel features from the mask region through three-dimensional convolution operation, and generating an initial feature embedding matrix; A recurrent cross-attention clustering module for dynamically updating the cluster center through a multi-scale recurrent EM cross-attention mechanism, and performing clustering analysis on the feature embedding matrix; A heterogeneity scoring module for calculating the information entropy based on the spatial distribution of the clustering results, and generating a hematoma and edema heterogeneity score after normalization; A prognosis prediction module for combining the heterogeneity score with clinical data, and predicting the risk of hematoma expansion and neurological outcome through a machine learning model; A cross-domain feature coupling module for capturing the spatial transmission rule of heterogeneity by calculating the density gradient field from the hematoma boundary to the edema region.

2. The system of claim 1, wherein: The feature extraction module is further configured to synchronously fuse the internal structure features of the hematoma and the edema region features around the hematoma, and construct a global three-dimensional heterogeneity representation.

3. The system of claim 1, wherein: The recurrent cross-attention clustering module implements dynamic clustering optimization by iteratively performing the following operations: Step E: Generate a query vector based on the current cluster center, and calculate a soft clustering assignment matrix by combining the generated key vector with the image features; Step M: Update the cluster center based on the soft clustering assignment matrix.

4. The system of claim 3, wherein: The recurrent cross-attention clustering module implements computational complexity optimization by projecting the shared query key weight, and maps the clustering results to the original CT image to generate a spatial distribution color label map, which labels different heterogeneity sub-regions.

5. The system of claim 3, wherein: The initialization of the cluster center is performed by selecting initial points from the image grid features through adaptive pooling operation, and optimizing the position feedforward neural network to generate.

6. The system of claim 1, wherein: The heterogeneity scoring module generates a heterogeneity score by the following steps: Statistical analysis of the pixel distribution probability of each clustering class in the hematoma and edema region; Calculate the information entropy value based on the probability distribution; Normalize the entropy value to the [0, 1] interval to generate a standardized score.

7. The system of claim 1, further comprising a dynamic weight allocation module that assigns higher influence to features with high confidence by calculating the matching confidence between feature vectors and cluster centers.

8. The system of claim 7, wherein: The confidence weight generation formula of the dynamic weight allocation module is:

9. A method for quantitatively evaluating cerebral hemorrhage heterogeneity based on recurrent cross-attention clustering, comprising: Where, Characterized by Clustering The confidence weight of is the natural exponential function; 、 and To adjust the parameters; For the feature vectors; For the cluster centers; is the within-class standard deviation; is the total number of preset clusters; For the cluster centers; For the The within-cluster standard deviation of each cluster; is the base of natural logarithms; It is a characteristic stability indicator. The implementation of the method is based on the system of any one of claims 1-8: The method comprises: Performing three-dimensional segmentation on brain CT images using the nnU-NetV2 model, and generating binary masks of hematoma and edema regions; ​ A gradient synergy field of the hematoma-perihematoma interface is calculated based on the mask, and the gradient field is spliced and fused with original features extracted by three-dimensional convolution to generate enhanced features; The enhanced features are iteratively executed in E-step and M-step by using a cyclic EM cross-attention mechanism; An information entropy is calculated according to the spatial distribution of the clustering results, and a hematoma-perihematoma heterogeneity score is normalized and output. The heterogeneity score is input into a machine learning model together with clinical data to predict the risk of hematoma enlargement and the probability of neurological outcome.

10. A computer readable storage medium having stored therein a computer program, characterized in that: The computer program is executed by a processor to perform the method of claim 9.

Images