Cerebral hemorrhage risk prediction method and system based on multi-modal learning

Abstract

The invention discloses a cerebral hemorrhage risk prediction method and system based on multi-modal learning, and relates to the technical field of medical detection. The cerebral hemorrhage risk prediction method and system based on multi-modal learning comprises the following steps: S1, collecting craniocerebral multi-modal data, and preprocessing the craniocerebral multi-modal data; s2, identifying a high-gray-scale suspicious region, quantifying bleeding features of the corresponding region, and marking a suspected hematoma region; s3, the cerebral hemorrhage risk degree of the suspected hematoma area is evaluated, the risk level is judged based on the evaluation result, and a hierarchical regulation and control strategy is generated; and S4, based on the cerebral hemorrhage risk assessment results of the current assessment period and the historical assessment period, performing quantitative assessment on the cerebral hemorrhage risk level fluctuation state, and dynamically updating the risk threshold. The problems that in early cerebral hemorrhage recognition in the prior art, tiny hemorrhage lesions are difficult to find, risk dynamic evaluation cannot be achieved, and therefore clinical judgment and intervention opportunities are affected are solved.

Description

Technical Field

[0001] This invention relates to the field of medical testing technology, and in particular to a method and system for predicting the risk of cerebral hemorrhage based on multimodal learning. Background Technology

[0002] With the continuous development of artificial intelligence technology, medical image analysis, and multimodal learning algorithms, the early identification and risk prediction of brain diseases have gradually become an important direction in intelligent medical research. Among them, cerebral hemorrhage, as one of the high-risk brain diseases, is of great significance for early detection and dynamic assessment in preventing the condition from worsening and guiding clinical intervention.

[0003] For example, invention CN114795145B discloses a method and system for predicting cerebral hemorrhage based on a PCA-BP neural network. The method includes the following steps: collecting human physiological data and arm amplitude data through a physiological information acquisition module and an arm amplitude detection module, respectively, as inputs to a cerebral hemorrhage probability prediction model and an arm amplitude detection model; establishing the cerebral hemorrhage probability prediction model and the arm amplitude detection model using a PCA-BP neural network algorithm; determining whether the tested person has experienced cerebral hemorrhage using the cerebral hemorrhage probability prediction model, and determining whether the tested person is paralyzed using the arm amplitude detection model. This invention establishes the cerebral hemorrhage probability prediction model and the arm amplitude detection model using the PCA-BP neural network algorithm, and combines these two models to predict cerebral hemorrhage, greatly improving the accuracy of cerebral hemorrhage prediction.

[0004] For example, the invention disclosed in publication number CN119326399A discloses a method, apparatus, device, medium, and product for detecting the MIT region of brain hemorrhage, relating to the field of brain image processing. The method includes dividing the brain imaging region into multiple regions with predetermined angles, using different coil arrays for different regions, acquiring magnetic field and conductivity information for different coil arrays, establishing multiple sample sets using the magnetic field and conductivity information as sample pairs, training stacked autoencoders with the multiple sample sets to obtain multiple MIT prediction models, determining the location of the lesion, and selecting an array from different coil arrays that positions the lesion within the -n° to n° region on the detection side based on the lesion location; acquiring magnetic field information and obtaining phase information using the selected array; selecting the MIT prediction model corresponding to the selected array and inputting the phase information into this MIT prediction model to predict the conductivity distribution; and completing image reconstruction based on the conductivity distribution.

[0005] However, although the above-mentioned technical solutions have made some progress in image recognition and probability prediction of cerebral hemorrhage, they still have the following shortcomings: First, there is a lack of a unified temporal alignment and scale normalization mechanism for cross-modal data; second, the assessment of hemorrhage characteristics is mostly a static result and cannot dynamically reflect the changing trend of hematoma expansion; third, the risk threshold cannot be adaptively adjusted according to individual patient differences and historical assessment results.

[0006] Therefore, in order to address the above problems, there is an urgent need for a method and system for predicting the risk of cerebral hemorrhage based on multimodal learning. Summary of the Invention

[0007] To address the technical problem that existing technologies struggle to detect minute hemorrhage foci in the early stages of cerebral hemorrhage and cannot achieve dynamic risk assessment, thus affecting clinical judgment and intervention timing, this invention provides a method and system for predicting cerebral hemorrhage risk based on multimodal learning. The technical solution is as follows:

[0008] On the one hand, a method for predicting the risk of cerebral hemorrhage based on multimodal learning is provided. This method includes: S1, collecting cranial multimodal data and performing time alignment, noise suppression, feature enhancement, continuous correction and normalization processing on the cranial multimodal data to obtain preprocessed cranial multimodal data; S2, identifying high grayscale suspicious regions based on the preprocessed cranial multimodal data, calculating the grayscale feature parameters of the high grayscale suspicious regions, quantifying the hemorrhage characteristics of the corresponding regions, marking suspected hematoma regions and generating a binary mask matrix of the suspected hematoma regions; S3, calling the binary mask matrix of the suspected hematoma regions and the temporal multimodal data to assess the degree of cerebral hemorrhage risk in the suspected hematoma regions, determining the risk level based on the assessment results and generating a graded control strategy; S4, based on the cerebral hemorrhage risk assessment results of the current assessment period and the historical assessment period, quantitatively assessing the fluctuation state of the cerebral hemorrhage risk level, dynamically updating the risk threshold, and realizing a closed loop of cerebral hemorrhage risk prediction and control.

[0009] Furthermore, multimodal data of the brain were collected, and time alignment, noise suppression, feature enhancement, continuous correction, and normalization were performed on the preprocessed multimodal data. The specific steps to obtain the preprocessed multimodal data of the brain are as follows: Using the anterior commissure of the patient as the origin, a three-dimensional spatial coordinate system of the brain was established using a DICOM-based world coordinate system-based 3D spatial reconstruction method. Multimodal data of the brain at various locations within the patient's brain were collected in real time. The multimodal data of the brain included CT pixel grayscale values, cerebral blood flow, cerebral blood volume, systolic blood pressure, diastolic blood pressure, heart rate, blood oxygen saturation, respiratory rate, body temperature, and other parameters. Prothrombin time; for the acquired cranial multimodal data, time synchronization calibration algorithm was used to unify the time alignment; adaptive anisotropic filtering algorithm was used to denoise and smooth the cranial image data to preserve structure and suppress random noise and maintain hematoma boundary information; sliding window averaging and outlier detection algorithms were used to remove artifacts and smooth trends; Kalman filtering combined with dynamic state estimation algorithm was used to continuously correct the cranial multimodal data; finally, Z-Score normalization and min-max normalization algorithms were used to unify the scale and eliminate dimensions of the cranial multimodal data.

[0010] Furthermore, the specific steps for identifying high grayscale suspicious regions based on preprocessed cranial multimodal data are as follows: extract the preprocessed cranial multimodal data, perform statistical analysis of the grayscale values ​​of CT pixels across the entire brain, calculate the grayscale frequency distribution vector, and determine the grayscale segmentation threshold using an adaptive threshold segmentation algorithm based on the grayscale frequency distribution vector, thereby filtering out high grayscale pixel regions where the grayscale values ​​of CT pixels are higher than the grayscale segmentation threshold.

[0011] Further, the specific steps for calculating the grayscale feature parameters of high grayscale suspicious regions are as follows: Perform three-dimensional connectivity analysis on the high grayscale pixel regions to identify continuous high-density regions, calculate the volume of each high-density region, and remove isolated small regions with volumes smaller than the volume threshold; calculate the mean grayscale value of CT pixels and the mean gradient value of CT edges for the retained high-density regions; and mirror the current high-density region along the midline of the brain to the contralateral brain region, extract the grayscale values ​​of CT pixels at symmetrical positions, calculate the regional standard deviation, and obtain the control grayscale standard deviation.

[0012] Further, the specific steps for quantifying the hemorrhage characteristics of the corresponding region, marking suspected hematoma areas, and generating a binary mask matrix for the suspected hematoma areas are as follows: Divide the mean grayscale value of CT pixels by the sum of the mean grayscale value of CT pixels and the standard deviation of the control grayscale value to obtain the grayscale ratio term; calculate the absolute value of the mean CT edge gradient and divide it by the sum of the absolute value of the mean CT edge gradient and a constant 1 to obtain the edge gradient adjustment term; multiply the grayscale ratio term and the edge gradient adjustment term to obtain the density structure factor; divide the cerebral blood volume by the sum of the cerebral blood flow and a constant 1, add 1 to the resulting ratio, and take the natural logarithm to obtain the cerebral circulation regulation factor; multiply the ratio of prothrombin time to reference prothrombin time and the cerebral circulation regulation factor... Multiplying the segment factors yields the coagulation status factor; dividing the difference between systolic and diastolic blood pressure by their sum and taking the hyperbolic tangent of the comparison value yields the blood pressure response factor; sequentially adding the density structure factor, coagulation status factor, and blood pressure response factor yields the bleeding characteristic assessment value; comparing the bleeding characteristic assessment value with the bleeding threshold in real time, when the bleeding characteristic assessment value is less than or equal to the bleeding threshold, the corresponding high-density area is marked as a normal area; when the bleeding characteristic assessment value is greater than the bleeding threshold, the corresponding high-density area is marked as a suspected hematoma area; performing a three-dimensional morphological closing operation on the suspected hematoma area smooths the area boundary and generates a binary mask matrix with the same spatial size as the original CT image.

[0013] Furthermore, the specific steps for assessing the risk of cerebral hemorrhage in suspected hematoma areas by calling the binary mask matrix and temporal multimodal data are as follows: The binary mask matrix is ​​used to extract the spatial locations of all suspected hematoma areas, and the corresponding intracranial multimodal data for these locations is obtained; a fixed time window is set as an assessment period, and within each assessment period, the mean and standard deviation of heart rate and systolic blood pressure are calculated, and the ratio of the corresponding standard deviation to the mean is calculated as the coefficient of variation for heart rate and systolic blood pressure; simultaneously, the volume of the suspected hematoma area is dynamically measured, the rate of change of hematoma volume is calculated, and the largest volume is extracted. The following steps are performed to obtain the dynamic change regulation term: First, add one to the coefficient of variation of heart rate, take the natural logarithm, multiply by the maximum rate of change of volume, divide the product by the maximum rate of change of volume, and add one. Second, divide cerebral blood volume by cerebral blood flow, add one to the ratio, and add the ratio of prothrombin time to reference prothrombin time to obtain the circulatory coagulation composite term. Third, add the product of blood oxygen saturation and respiratory rate to body temperature, and take the square root of the result to obtain the physiological baseline term. Fourth, divide the hemorrhage characteristic assessment value by the physiological baseline term to obtain the physiological correction term. Finally, multiply the dynamic change regulation term, the circulatory coagulation composite term, and the physiological correction term sequentially to obtain the cerebral hemorrhage risk assessment value.

[0014] Further, the specific steps for determining the risk level based on the evaluation results and generating a hierarchical control strategy are as follows: Compare the intracerebral hemorrhage risk assessment value S with the multi-level risk thresholds S1 and S2 in real time, determine the risk level and generate a hierarchical control strategy: When S ≤ S1, it is determined as a first-level risk, the evaluation cycle is shortened, and continuous detection is carried out; When S1 < S < S2, it is determined as a second-level risk, the key attention mechanism is activated, and based on the spatial position, volume, and bleeding characteristic evaluation value of the suspected hematoma area in the binary mask matrix, a structured risk prompt information is generated and synchronized to the attending doctor's management terminal; When S ≥ S2, it is determined as a third-level risk, the risk control process is started, spatial expansion simulation is performed on the suspected hematoma area, a bleeding prediction report is generated by combining the original CT image and the binary mask matrix, and pushed to the attending doctor and the nursing station, and at the same time, the alarm mechanism is triggered.

[0015] Further, the specific steps for quantitatively evaluating the fluctuation state of the intracerebral hemorrhage risk level based on the intracerebral hemorrhage risk assessment results of the current evaluation cycle and the historical evaluation cycle are as follows: After each evaluation cycle ends, extract the intracerebral hemorrhage risk assessment values of the current evaluation cycle and the previous N evaluation cycles, construct a historical intracerebral hemorrhage risk assessment value sequence, and calculate the historical risk assessment mean value; Take the natural logarithm of the current intracerebral hemorrhage risk assessment value after adding 1, subtract the natural logarithm value of the historical risk assessment mean value after adding 1, and use the obtained difference as the numerator; Add the square of the current intracerebral hemorrhage risk assessment value, the square of the historical risk assessment mean value, and the constant 1 in sequence, and take the square root of the sum value as the denominator; Divide the numerator by the denominator to obtain the risk adjustment value.

[0016] Further, the specific steps for dynamically updating the risk threshold to achieve a closed-loop control of intracerebral hemorrhage risk prediction and regulation are as follows: Adjust the risk threshold according to the positive and negative trend of the risk adjustment value. When the risk adjustment value is positive, keep the current risk threshold unchanged; When the risk adjustment value is negative, dynamically adjust the risk threshold in combination with the change trend of the historical intracerebral hemorrhage risk assessment value; Apply the updated risk threshold to the determination of the intracerebral hemorrhage risk level in the next evaluation cycle to achieve closed-loop control of the evaluation logic.

[0017] On the other hand, a multimodal learning-based intracerebral hemorrhage risk prediction system is provided. This system is applied to a multimodal learning-based intracerebral hemorrhage risk prediction method. The system includes: a data acquisition and preprocessing module, used to acquire intracranial multimodal data and perform time alignment, noise suppression, feature enhancement, continuous correction and normalization processing on the intracranial multimodal data to obtain preprocessed intracranial multimodal data; a hemorrhage feature recognition module, used to identify high grayscale suspicious regions based on the preprocessed intracranial multimodal data, calculate the grayscale feature parameters of the high grayscale suspicious regions, quantify the hemorrhage features of the corresponding regions, mark suspected hematoma regions and generate a binary mask matrix of the suspected hematoma regions; a risk assessment and determination module, used to call the binary mask matrix of the suspected hematoma regions and the temporal multimodal data to assess the intracerebral hemorrhage risk level of the suspected hematoma regions, determine the risk level based on the assessment results and generate a graded control strategy; and a dynamic threshold control module, used to quantitatively assess the fluctuation state of the intracerebral hemorrhage risk level based on the intracerebral hemorrhage risk assessment results of the current assessment period and the historical assessment period, dynamically update the risk threshold, and realize a closed loop of intracerebral hemorrhage risk prediction and control.

[0018] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0019] (1) A whole-brain gray-level statistical model is established by using gray-level frequency distribution vectors, and combined with adaptive threshold segmentation algorithm and three-dimensional connectivity analysis to achieve accurate identification of high gray-level suspicious regions. Furthermore, a mirror comparison mechanism is introduced, and an individualized control gray-level benchmark is obtained by calculating the gray-level standard deviation of symmetrical brain regions. Thus, without the need for manual intervention, the gray-level mean, edge gradient and spatial volume parameters of the hematoma region are automatically extracted to complete the quantitative identification and structured expression of the hemorrhage region.

[0020] (2) For the first time, a composite parameter system including density structure factor, circulation coagulation composite term and physiological correction term was introduced into the risk modeling of cerebral hemorrhage. Through the fusion calculation of multidimensional physiological indicators and imaging features, a nonlinear risk mapping relationship was established. This model can comprehensively reflect the dynamic interaction between local hemorrhage density, blood circulation, coagulation status and systemic physiological response, realizing the transformation from single-point measurement to global situation prediction, and improving the accuracy and clinical reference value of hemorrhage trend assessment.

[0021] (3) Through a multi-level risk threshold system, the risk of cerebral hemorrhage is divided into different levels, and the corresponding regulatory mechanism is automatically triggered based on the risk classification results. The system can generate structured risk information based on the spatial location, volume and hemorrhage characteristic parameters of the suspected hematoma area in the binary mask matrix, dynamically update the risk warning level, and synchronize data with the attending physician's management terminal to realize intelligent linkage between risk classification, information push and clinical auxiliary decision-making.

[0022] (4) A risk adjustment value calculation method based on historical risk assessment sequences is proposed. By quantifying the fluctuation trend of cerebral hemorrhage risk assessment values ​​over time, the risk threshold is adjusted in real time. When the risk assessment results fluctuate significantly, the system can automatically calibrate the threshold parameters according to the direction of the adjustment value, realizing adaptive learning and dynamic correction of the model under multi-cycle continuous monitoring. This closed-loop control mechanism enables the system to maintain the stability and sensitivity of risk judgment in long-term operation, and has the ability to continuously optimize and autonomously adjust. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a method for predicting the risk of cerebral hemorrhage based on multimodal learning, provided by an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of a brain hemorrhage risk prediction system based on multimodal learning provided in an embodiment of the present invention;

[0026] Figure 3 This is a visualization of the detection results of a suspected hematoma area provided in an embodiment of the present invention;

[0027] Figure 4 This is a flowchart for determining the risk level of cerebral hemorrhage provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0029] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0030] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0031] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0032] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0033] This invention provides a method for predicting the risk of cerebral hemorrhage based on multimodal learning, such as... Figure 1 The flowchart shown is a method for predicting the risk of intracerebral hemorrhage based on multimodal learning. The processing flow of this method can include the following steps: S1, collecting intracranial multimodal data and performing time alignment, noise suppression, feature enhancement, continuous correction and normalization on the intracranial multimodal data to obtain preprocessed intracranial multimodal data; S2, identifying high grayscale suspicious regions based on the preprocessed intracranial multimodal data, calculating the grayscale feature parameters of the high grayscale suspicious regions, quantifying the hemorrhage characteristics of the corresponding regions, marking suspected hematoma regions and generating a binary mask matrix of the suspected hematoma regions; S3, calling the binary mask matrix of the suspected hematoma regions and the temporal multimodal data to assess the intracerebral hemorrhage risk level of the suspected hematoma regions, determining the risk level based on the assessment results and generating a graded control strategy; S4, based on the intracerebral hemorrhage risk assessment results of the current assessment period and the historical assessment period, quantitatively assessing the fluctuation state of the intracerebral hemorrhage risk level, dynamically updating the risk threshold, and realizing a closed loop of intracerebral hemorrhage risk prediction and control.

[0034] Optionally, multimodal brain data is collected, and time alignment, noise suppression, feature enhancement, continuous correction, and normalization are performed on the preprocessed multimodal brain data. The specific steps for obtaining the preprocessed multimodal brain data are as follows: Using the patient's anterior commissure as the origin, a three-dimensional spatial coordinate system for the brain is established using a DICOM-based world coordinate system. During the reconstruction process, multidimensional spatial mapping relationships are established using CT image slice thickness information, pixel spacing, and slice sequence indexes to ensure the consistency and traceability of data at various brain locations in the spatial dimension. Multimodal brain data corresponding to each spatial location in the patient's brain is collected in real time to ensure that temporal and spatial resolutions simultaneously meet sampling accuracy requirements. The multimodal brain data includes CT pixel grayscale values, cerebral blood flow, cerebral blood volume, systolic blood pressure, diastolic blood pressure, heart rate, blood oxygen saturation, respiratory rate, body temperature, prothrombin time, and other multidimensional physiological and imaging parameters. All parameters are synchronously collected and timestamped using high-precision sensors for subsequent dynamic fusion and temporal calibration of multi-source data. The CT pixel grayscale values ​​were acquired by a head CT scan device and output via the DICOM protocol; cerebral blood flow and cerebral blood volume were calculated using CT perfusion imaging; systolic and diastolic blood pressure were measured in real time using a non-invasive blood pressure monitoring device; heart rate was calculated using the rate of change of pulse signal; blood oxygen saturation was acquired in real time using a fingertip pulse oximeter; respiratory rate was sampled by a thoracic pressure sensor; body temperature was monitored by an infrared body temperature detection unit; and prothrombin time was measured using a blood analyzer. For the acquired multimodal brain data, a time synchronization calibration algorithm was first used to achieve unified time alignment, ensuring the physical correspondence of each modality's data at the same sampling time point. The time synchronization calibration algorithm is based on the principles of sampling frequency matching and minimizing time offset, achieving time base unification of different modal signals through iterative solution of minimum mean square error. Subsequently, an adaptive anisotropic filtering algorithm was used to denoise and smooth the brain image data while maintaining structural fidelity. This algorithm dynamically adjusts the filtering coefficients within the pixel neighborhood, adaptively controlling the smoothing range based on local gradient direction and intensity differences. This effectively suppresses random noise while preserving the gray-level abrupt changes in hematoma boundaries, avoiding boundary blurring and loss of tissue details. Then, a sliding window averaging and outlier detection algorithm are used to perform artifact removal and trend smoothing on the multimodal time series. The sliding window averaging eliminates interference signals caused by short-term fluctuations, while outlier detection removes abnormal data points based on a statistical deviation threshold, improving signal continuity and reliability. Subsequently, Kalman filtering combined with a dynamic state estimation algorithm is used to continuously correct the cranial multimodal data. This algorithm treats the dynamic changes of each parameter as a state transition process, using prediction and update mechanisms to adaptively suppress measurement noise, ensuring the stability and trend consistency of the corrected cranial multimodal data over time. Finally, Z-Score normalization and min-max normalization algorithms are combined to perform scale unification and dimensionless elimination on the cranial multimodal data.Z-Score normalization is used to eliminate statistical bias in various cranial multimodal data and improve the comparability of features in distribution. Minimum-maximum normalization is used to map different cranial multimodal data to the same numerical range, ensuring the consistency of parameters in order of magnitude and computational stability in subsequent feature extraction and evaluation calculations.

[0035] In this implementation scheme, a DICOM-based spatial reconstruction method for the cranium was established, achieving unified spatial mapping and temporal alignment of multimodal cranial data, ensuring data consistency and comparability across spatiotemporal dimensions. By combining adaptive anisotropic filtering with a sliding window averaging method, random noise and artifact interference were effectively suppressed, maintaining the structural integrity of hematoma boundaries and the authenticity of image details. Kalman filtering and dynamic state estimation algorithms were used to continuously correct the multimodal time series, ensuring the continuity and stability of the cranium multimodal data during dynamic changes. Z-score normalization and min-max normalization algorithms were combined to achieve dimensional unification and scale normalization, providing a reliable, stable, and accurate multimodal input data foundation for subsequent high-grayscale suspicious region identification and cerebral hemorrhage risk assessment, significantly improving the robustness of data processing and the credibility of assessment results.

[0036] Optionally, the specific steps for identifying high-grayscale suspicious regions based on preprocessed cranial multimodal data are as follows: Extract cranial multimodal data after time alignment, noise suppression, feature enhancement, continuous correction, and normalization processing to ensure consistency in both time series and spatial coordinates. Perform whole-brain statistical analysis on CT pixel grayscale values, traversing the 3D image data layer by layer, accumulating and statistically analyzing the grayscale values ​​of each voxel to construct a grayscale frequency distribution vector, used to describe the overall grayscale distribution characteristics of brain tissue. After calculating the grayscale frequency distribution vector, an adaptive threshold segmentation algorithm is used to determine the grayscale segmentation threshold based on this distribution. The adaptive threshold segmentation algorithm dynamically determines the threshold based on the statistical characteristics of the whole-brain grayscale distribution, achieving the boundary between different tissue regions by optimizing the grayscale partitioning standard. Based on the determined grayscale segmentation threshold, high-grayscale pixel regions with CT pixel grayscale values ​​higher than the grayscale segmentation threshold are selected. The selection process is performed point-by-point in 3D space, recording the voxel coordinates and boundary information of the high-grayscale pixel regions, providing input data for subsequent calculation of high-density region feature parameters and suspicious region labeling.

[0037] In this implementation scheme, by utilizing the whole-brain statistical features of CT pixel grayscale values, combined with the dynamic threshold determination mechanism of the grayscale frequency distribution vector and adaptive threshold segmentation algorithm, the extraction of high grayscale regions becomes more targeted and stable. After unified time alignment and spatial consistency processing, the identified high grayscale pixel regions can provide an accurate data foundation for the subsequent quantitative calculation of hemorrhage feature parameters, ensuring the reliability and timeliness of the labeling results of suspected hematoma areas.

[0038] Optionally, the specific steps for calculating the grayscale feature parameters of high-grayscale suspicious regions are as follows: Perform three-dimensional connectivity analysis on the high-grayscale pixel regions, and determine the connectivity of the regions point by point in three-dimensional space based on the adjacency relationship between voxels to identify high-density regions with spatial continuity. Perform voxel accumulation calculation on each high-density region to obtain the volume value, and compare the volume result with the volume threshold to remove isolated small regions with volumes smaller than the volume threshold, thereby eliminating noise interference and misidentified regions. After completing the high-density region screening, calculate the CT pixel grayscale mean for each of the retained high-density regions; at the same time, use the gradient operation operator to calculate the grayscale change rate at the region edge to obtain the CT edge gradient mean, which is used to characterize the intensity change characteristics of the region edge. Mirror the current high-density region along the brain midline, determine the corresponding position in the contralateral brain region after mapping based on symmetric coordinate transformation, and extract the CT pixel grayscale values ​​at the symmetrical position. Calculate the regional standard deviation for the extracted symmetrical grayscale value set to obtain the control grayscale standard deviation. The standard deviation of grayscale values ​​reflects the grayscale differences between corresponding regions of the left and right hemispheres of the brain, providing a symmetrical reference benchmark for subsequent quantitative calculations of hemorrhage characteristics.

[0039] In this implementation scheme, three-dimensional connectivity analysis ensures the spatial integrity of high-density region identification, and the volume screening process effectively eliminates isolated artifact regions, ensuring data stability. The joint calculation of the average grayscale value of CT pixels and the average gradient value of CT edges allows for the simultaneous characterization of the internal density and edge features of the hemorrhage region. The introduction of brain midline mirror mapping and the standard deviation of control grayscale values ​​provides a symmetrical benchmark for the identification of high-density regions, thereby improving the accuracy and consistency of hemorrhage feature extraction and providing reliable feature support for the subsequent quantitative labeling of suspected hematoma areas.

[0040] Optionally, the specific steps for quantifying the bleeding characteristics of the corresponding region, marking suspected hematoma regions, and generating a binary mask matrix for the suspected hematoma regions are as follows: Divide the mean grayscale value of CT pixels by the sum of the mean grayscale value of CT pixels and the standard deviation of the control grayscale value to obtain a grayscale ratio term; in this ratio calculation, the region voxel set is used as the calculation unit; calculate the absolute value of the mean CT edge gradient by dividing it by the sum of the absolute value of the mean CT edge gradient and a constant 1 to obtain an edge gradient adjustment term; the mean CT edge gradient is calculated by the three-dimensional gradient operator at the region boundary voxels and quantile truncation is performed to suppress abnormal sharpness. Peak; Multiply the grayscale ratio term by the edge gradient adjustment term to obtain the density structure factor; the density structure factor is constrained to between zero and one under normalized scaling and bound to a timestamp to ensure temporal consistency; divide cerebral blood volume by the sum of cerebral blood flow and a constant 1, add 1 to the resulting ratio, and take the natural logarithm to obtain the cerebral circulation regulation factor; the natural logarithm operation is base e and a non-negativity check is performed before input; multiply the ratio of prothrombin time to reference prothrombin time by the cerebral circulation regulation factor to obtain the coagulation status factor; the reference prothrombin time is a clinically calibrated fixed reference value. The blood pressure remains constant within the same assessment period; the difference between systolic and diastolic blood pressure is divided by the sum of the two, and the hyperbolic tangent of the ratio is taken to obtain the blood pressure response factor; the range of this ratio is limited to -1 to 1, and the hyperbolic tangent output is located within the open interval; the density structure factor, coagulation state factor, and blood pressure response factor are added sequentially to obtain the bleeding characteristic assessment value; the calculation process is completed under a unified timestamp and the corresponding high-density region identifier is recorded; the bleeding characteristic assessment value and the bleeding threshold are compared in real time, and when the bleeding characteristic assessment value is less than or equal to the bleeding threshold, the corresponding high-density region is marked as positive. Normal regions; when the bleeding feature assessment value is greater than the bleeding threshold, the corresponding high-density region is marked as a suspected hematoma region; the bleeding threshold is determined during the training phase and calibrated through the hospital's standardized process, and is expressed as a fixed-precision decimal; a three-dimensional morphological closing operation is performed on the suspected hematoma region to smooth the region boundary and generate a binary mask matrix with the same spatial size as the original CT image; the closing operation uses a three-dimensional spherical structuring element and is performed in isotropic voxel space, and the output mask corresponds one-to-one with the original CT image in the DICOM world coordinate system, while recording the voxel index list and timestamp.

[0041] The specific formula for calculating the bleeding characteristic assessment value is as follows: ;

[0042] In the formula, Indicates the bleeding characteristic assessment value, This represents the average grayscale value of CT pixels. This represents the standard deviation of the grayscale values. This represents the mean gradient of the CT edge. Indicates cerebral blood volume. Indicates cerebral blood flow. Indicates prothrombin time. This indicates the reference prothrombin time. Indicates systolic blood pressure. It indicates diastolic blood pressure.

[0043] In this embodiment, Table 1 is a data table of hemorrhage feature assessment values, listing the key parameters and corresponding hemorrhage feature assessment values ​​for five high-density regions. The key parameters used include: mean CT pixel grayscale value, standard deviation of control grayscale value, mean CT edge gradient value, cerebral blood volume, cerebral blood flow, prothrombin time, reference prothrombin time, systolic blood pressure, and diastolic blood pressure. Specific data are as follows: In high-density region 1, the average CT pixel grayscale value is 145, the standard deviation of the control grayscale value is 30, the average CT edge gradient value is 25, the cerebral blood volume is 4.8, the cerebral blood flow is 3.2, the prothrombin time is 13.5, the reference prothrombin time is 12.0, the systolic blood pressure is 140, the diastolic blood pressure is 85, and the corresponding hemorrhage feature assessment value is 0.92; In high-density region 2, the average CT pixel grayscale value is 160, the standard deviation of the control grayscale value is 35, the average CT edge gradient value is 28, the cerebral blood volume is 5.2, the cerebral blood flow is 3.4, the prothrombin time is 15.2, the reference prothrombin time is 12.0, the systolic blood pressure is 150, the diastolic blood pressure is 90, and the corresponding hemorrhage feature assessment value is 1.03; In high-density region 3, the average CT pixel grayscale value is 130, the standard deviation of the control grayscale value is 28, the average CT edge gradient value is 20, the cerebral blood volume is 4.0, the cerebral blood flow is 3.2, the cerebral blood flow is 3.2, the prothrombin time is 15.2, the reference prothrombin time is 12.0, the systolic blood pressure is 150, the diastolic blood pressure is 90, and the corresponding hemorrhage feature assessment value is 1.03; In high-density region 3, the average CT pixel grayscale value is 130, the standard deviation of the control grayscale value is 28, the average CT edge gradient value is 20, the cerebral blood The cerebral blood flow was 3.1, the prothrombin time was 12.8, the reference prothrombin time was 12.0, the systolic blood pressure was 135, the diastolic blood pressure was 80, and the corresponding hemorrhage characteristic assessment value was 0.82. In high-density region 4, the mean CT pixel grayscale value was 170, the standard deviation of the control grayscale value was 40, the mean CT edge gradient value was 32, the cerebral blood volume was 5.5, the cerebral blood flow was 3.0, the prothrombin time was 16.0, and the reference prothrombin time was 12.0. The systolic blood pressure was 155, the diastolic blood pressure was 92, and the corresponding hemorrhage feature assessment value was 1.16. In high-density region 5, the mean CT pixel gray level was 150, the standard deviation of the control gray level was 33, the mean CT edge gradient was 26, the cerebral blood volume was 4.7, the cerebral blood flow was 3.3, the prothrombin time was 14.1, the reference prothrombin time was 12.0, the systolic blood pressure was 142, the diastolic blood pressure was 88, and the corresponding hemorrhage feature assessment value was 0.92.

[0044] Table 1. Bleeding Characteristic Assessment Values

[0045] like Figure 3As shown, the hemorrhage feature assessment values ​​and suspected hematoma area determination results for five high-density regions are displayed to identify the hemorrhage risk status at different spatial locations in cranial imaging. The figure uses a color bar chart to represent the hemorrhage feature assessment value R for each high-density region, where blue represents normal areas and red represents suspected hematoma areas, visually clearly distinguishing the differences in hemorrhage features between different regions. A black dashed line is set as the hemorrhage determination threshold to differentiate between normal areas and suspected hematoma areas. According to the results, the hemorrhage feature assessment values ​​of high-density regions 2 and 4 are both higher than the hemorrhage threshold, and are determined to be suspected hematoma areas; the hemorrhage feature assessment values ​​of the remaining high-density regions are all lower than the hemorrhage threshold, and are determined to be normal areas. Figure 3 It intuitively reflects the differential distribution of hemorrhage characteristic parameters in each high-density region, which can provide quantitative basis for the early identification of signs of cerebral hemorrhage and the localization of high-risk lesions, and support the grading and dynamic control process of the cerebral hemorrhage risk prediction module.

[0046] In this implementation scheme, the density differences and structural changes of the hemorrhage area are simultaneously expressed through the joint calculation of the gray-scale ratio term and the edge gradient adjustment term. The introduction of cerebral circulation regulatory factors and coagulation status factors allows hemodynamic status and coagulation function status to jointly participate in the assessment of hemorrhage characteristics. The embedding of the blood pressure response factor further reflects the influence of systemic circulatory pressure on local cerebrovascular stability. The hemorrhage characteristic assessment value formed by multi-factor interactive calculation can accurately determine the hemorrhage risk in high-density areas. Combined with the binary mask matrix generated by three-dimensional morphological closing operation, the spatial boundary of the suspected hematoma area can be ensured to be continuous and morphologically smooth, thereby improving the spatial consistency and structural reliability of the suspected area marking and providing a stable spatial input basis for subsequent cerebral hemorrhage risk assessment.

[0047] Optionally, the specific steps for assessing the risk of cerebral hemorrhage in suspected hematoma areas by calling the binary mask matrix and temporal multimodal data of the suspected hematoma area are as follows: Call the binary mask matrix to extract the spatial location of all suspected hematoma areas and obtain the corresponding intracranial multimodal data; during the extraction process, use the DICOM world coordinate system as a unified reference benchmark to ensure a one-to-one correspondence between the mask spatial index and the image data spatial index; set a fixed time window as an evaluation cycle, and synchronously calibrate the sampling points according to the time series within each evaluation cycle to ensure that various physiological and imaging parameters have timeliness consistency at the same moment; within each evaluation cycle, calculate the mean and standard deviation of heart rate and systolic blood pressure respectively, and calculate the ratio of the corresponding standard deviation to the mean as the coefficient of variation of heart rate and systolic blood pressure; simultaneously, dynamically measure the volume of the suspected hematoma area. The volume measurement is based on three-dimensional mask voxel counting and combined with pixel spacing and layer thickness for volume conversion, calculate the hematoma volume change rate, and extract the maximum volume change rate; the volume change rate is obtained through phase... The volume difference between adjacent assessment periods is divided by the volume of the previous period to obtain the maximum rate of change, which is the maximum value of the rate of change across multiple periods. The coefficient of variation of heart rate is incremented by one, the natural logarithm is taken, and then multiplied by the maximum rate of change of volume. The product is divided by the maximum rate of change of volume and incremented by one to obtain the dynamic change adjustment term. During the calculation, the logarithmic input is constrained to a positive value to avoid numerical overflow. The cerebral blood volume is divided by the cerebral blood flow and incremented by one. The ratio is then added to the ratio of prothrombin time to reference prothrombin time to obtain the circulatory coagulation composite term. The product of blood oxygen saturation and respiratory rate is added to body temperature, and the square root of the result is taken to obtain the physiological baseline term. The hemorrhage characteristic assessment value is divided by the physiological baseline term to obtain the physiological correction term. This correction term is used to correct for the basic physiological differences between different individuals to improve the individualized adaptability of risk assessment. The dynamic change adjustment term, the circulatory coagulation composite term, and the physiological correction term are multiplied sequentially to obtain the cerebral hemorrhage risk assessment value. The calculation result is assigned a unique timestamp and written into the risk assessment sequence as the input basis for subsequent risk level determination and dynamic threshold control.

[0048] The specific formula for calculating the risk assessment value for cerebral hemorrhage is as follows:

[0049] ;

[0050] In the formula, This indicates the risk assessment value for cerebral hemorrhage. This represents the maximum rate of change of volume. This represents the coefficient of variation of heart rate. This represents the coefficient of variation of systolic blood pressure. Indicates cerebral blood volume. Indicates cerebral blood flow. Indicates prothrombin time. This indicates the reference prothrombin time. Indicates the bleeding feature assessment value, Indicates the blood oxygen saturation, Indicates the respiratory rate, Indicates the body temperature.

[0051] In this embodiment, by introducing a dynamic calculation mechanism based on a time window, the heart rate variability coefficient, systolic blood pressure variability coefficient, and hematoma volume change rate can reflect the temporal fluctuation law of the bleeding focus; through the calculation of the circulating coagulation composite term, the change characteristics of cerebral blood volume, cerebral blood flow, and prothrombin time are jointly incorporated into the risk calculation process to characterize the coupling relationship between hemodynamics and coagulation status; by jointly introducing the physiological reference term and physiological correction term, the influence of differences in physiological indicators such as blood oxygen saturation, respiratory rate, and body temperature on the risk assessment result is fully corrected to improve the individual accuracy of the intracerebral hemorrhage risk assessment value; through the comprehensive calculation of the dynamic change adjustment term, circulating coagulation composite term, and physiological correction term, a risk quantification model driven by multi-dimensional parameters is established to achieve a fine classification of the risk degree of the suspected hematoma area, providing data support for the dynamic risk identification and closed-loop control of intracerebral hemorrhage.

[0052] Optionally, the specific steps for determining the risk level based on the evaluation result and generating a hierarchical control strategy are as follows: As Figure 4 shown, the intracerebral hemorrhage risk assessment value S is compared with the multi-level risk thresholds S1 and S2 in real time to determine the risk level and generate a hierarchical control strategy. At the end of the same evaluation cycle, the time stamp index is used, and the boundary rules of the closed interval and open interval are used for numerical comparison. S1 and S2 are fixed thresholds calibrated by clinical statistics and satisfy S1 < S2; when S ≤ S1, the computer device executes a data processing program to determine it as a first-level risk, automatically shortens the evaluation cycle and continuously detects, reduces the evaluation cycle length at the parameter level, and records the corresponding relationship between the intracerebral hemorrhage risk assessment value of this cycle and the binary mask matrix; when S1 < S < S2, the computer device automatically determines it as a second-level risk, executes the key attention processing logic, generates a structured risk prompt message based on the spatial position, volume, and bleeding feature assessment value of the suspected hematoma area in the binary mask matrix, and the processing result is synchronously output to the attending doctor management terminal through the data interface for medical staff to refer to; when S ≥ S2, the computer device executes the risk control instruction process, automatically determines it as a third-level risk, performs a spatial expansion simulation on the suspected hematoma area, generates a prediction voxel set using an isotropic structuring element and a fixed number of iteration steps, combines the original CT image and the binary mask matrix to form a bleeding prediction report, the report includes the spatial position, boundary coordinates, and predicted volume of the suspected hematoma area and is accompanied by a time stamp, and the data is output to the attending doctor terminal and the nursing station terminal through the network communication channel, and at the same time, an alarm prompt signal is automatically triggered. The output result of this information processing method is only used as an auxiliary reference for the intracerebral hemorrhage risk classification assessment, providing a decision-making basis for medical staff and does not constitute a medical diagnosis conclusion.

[0053] This implementation plan ensures the stability and consistency of risk level classification by comparing the intracerebral hemorrhage risk assessment value with multi-level risk thresholds in real time, thus maintaining the traceability of assessment results within a continuous monitoring period. Through a periodic shortening mechanism for Level 1 risk, high-frequency detection in low-risk states is achieved, enhancing early risk capture capabilities. A focused attention mechanism for Level 2 risk enables structured output of the spatial location, volume, and hemorrhage characteristic assessment values ​​of suspected hematoma areas, providing medical staff with quantitative references based on spatiotemporal parameters. Through spatial expansion simulation and hemorrhage prediction report generation for Level 3 risk, the potential hemorrhage expansion trend is quantified, providing a visual basis for subsequent intervention. Overall, this method achieves fully automated execution of the entire process from intracerebral hemorrhage risk assessment to grading determination and strategy generation, improving the real-time performance and accuracy of risk identification while maintaining the safety and standardization of assessment results as auxiliary decision-making for medical staff.

[0054] Optionally, based on the cerebral hemorrhage risk assessment results of the current assessment period and historical assessment periods, the specific steps for quantitatively assessing the fluctuation of the cerebral hemorrhage risk level are as follows: After each assessment period ends, the computer device automatically retrieves the cerebral hemorrhage risk assessment values ​​of the current assessment period and the previous N assessment periods to construct a time-series-based set of historical cerebral hemorrhage risk assessment values. An index matrix is ​​established according to the timestamp order to ensure the continuity and traceability of the risk data; where N is a positive integer; the retrieved historical cerebral hemorrhage risk assessment values ​​are summed and averaged using an arithmetic mean algorithm to obtain the historical risk assessment mean; the current cerebral hemorrhage risk assessment value is incremented by one and the natural logarithm is taken to obtain the current period's mean. The logarithmic risk value is calculated as follows: The logarithmic risk value of the historical mean is obtained by adding one to the historical risk assessment mean and then taking the natural logarithm. The logarithmic risk value of the current period is subtracted from the logarithmic risk value of the historical mean, and the difference is used as the numerator to characterize the instantaneous shift in the risk level of cerebral hemorrhage. The square of the current cerebral hemorrhage risk assessment value, the square of the historical risk assessment mean, and a constant 1 are added sequentially, and the square root of the sum is used as the denominator to suppress the numerical shift caused by extreme fluctuations. Finally, the numerator is divided by the denominator to obtain the risk adjustment value. The risk adjustment value serves as a quantitative indicator of the cerebral hemorrhage risk fluctuation state and is used for the adaptive updating of the risk threshold and the dynamic closed-loop control of the assessment logic.

[0055] The specific formula for calculating the risk adjustment value is as follows:

[0056] ;

[0057] In the formula, Indicates the risk adjustment value. This indicates the current risk assessment value for cerebral hemorrhage. This represents the historical average risk assessment.

[0058] In this implementation scheme, by calculating the logarithmic difference between the current cerebral hemorrhage risk assessment value and the historical average risk assessment value, subtle fluctuations in risk status can be effectively captured, thus reflecting the dynamic shift characteristics of risk level over time. By introducing square terms and square root normalization operations during the calculation process, the risk adjustment value is kept balanced in numerical distribution, avoiding misjudgments caused by single outliers. Based on the statistical analysis of multi-period data, this method achieves time-series comparison and standardized measurement of risk status, providing a stable mathematical basis for the subsequent adaptive update of the cerebral hemorrhage risk threshold, and improving the timeliness and computational reliability of the risk monitoring process.

[0059] Optionally, the specific steps for dynamically updating the risk threshold to achieve a closed loop of cerebral hemorrhage risk prediction and control are as follows: The risk threshold is adjusted based on the positive or negative trend of the risk adjustment value. After each assessment cycle, the computer device calls the risk adjustment value and the historical cerebral hemorrhage risk assessment value sequence, performing trend analysis and direction determination according to the timestamp order. When the risk adjustment value is positive, it indicates that the cerebral hemorrhage risk assessment value is within a stable fluctuation range. The calculation logic keeps the current risk threshold unchanged and records the threshold state in the parameter cache for the next cycle. When the risk adjustment value is negative, it indicates that the cerebral hemorrhage risk assessment value has a downward trend or fluctuation deviation. The calculation logic combines the changing trend of historical cerebral hemorrhage risk assessment values ​​and dynamically corrects the risk threshold based on the linear fitting results and historical mean square deviation. Smoothing constraints are executed during the numerical update process to prevent sudden changes. After the threshold update is completed, the updated risk threshold is rewritten into the risk judgment parameter set and called during the cerebral hemorrhage risk level judgment process in the next assessment cycle. This achieves adaptive iteration of the risk threshold and closed-loop control of the assessment logic, enabling a continuous feedback mechanism for risk assessment in the time dimension.

[0060] In this implementation scheme, by combining the trend analysis of risk adjustment values ​​with historical intracerebral hemorrhage risk assessment values, the adjustment of risk thresholds no longer relies on fixed empirical parameters, but is based on a data-driven dynamic calibration mechanism. This automatically corrects the judgment boundary when the risk status changes. Linear fitting and smoothing constraint methods are used to control the magnitude of threshold changes, ensuring the stability and continuity of the update process and avoiding risk level jumps caused by abrupt threshold changes. This method achieves closed-loop control of the risk assessment logic, making intracerebral hemorrhage risk prediction more sensitive and reliable, and maintaining consistent risk response accuracy in clinical monitoring.

[0061] On the other hand, a multimodal learning-based system for predicting the risk of cerebral hemorrhage is provided, such as... Figure 2The diagram shows a structural schematic of a brain hemorrhage risk prediction system based on multimodal learning. This system is applied to a multimodal learning-based brain hemorrhage risk prediction method and includes: a data acquisition and preprocessing module for acquiring multimodal brain data and performing time alignment, noise suppression, feature enhancement, continuous correction, and normalization on the data to obtain preprocessed multimodal brain data; a hemorrhage feature recognition module for identifying high-grayscale suspicious regions based on the preprocessed multimodal brain data, calculating grayscale feature parameters of these regions, quantifying the hemorrhage features of the corresponding regions, marking suspected hematoma regions, and generating a binary mask matrix for these regions; a risk assessment and determination module for calling the binary mask matrix of the suspected hematoma regions and the temporal multimodal data to assess the degree of brain hemorrhage risk in these regions, determining the risk level based on the assessment results, and generating a graded control strategy; and a dynamic threshold control module for quantitatively assessing the fluctuation state of brain hemorrhage risk levels based on the brain hemorrhage risk assessment results of the current and historical assessment periods, dynamically updating the risk threshold, and realizing a closed loop of brain hemorrhage risk prediction and control.

[0062] This implementation scheme achieves information fusion between cranial imaging features and physiological signal features within a unified temporal dimension through the organic coordination of four stages: data acquisition and preprocessing, hemorrhage feature identification, risk assessment and determination, and dynamic threshold adjustment. This forms a closed-loop computational path from data input to risk output. By accurately identifying high-grayscale suspicious areas and generating a binary mask matrix, the system can effectively distinguish the spatial distribution characteristics of suspected hematoma areas. Through real-time comparison of cerebral hemorrhage risk assessment values ​​with multi-level risk thresholds, the system can automatically determine the risk level. Through adaptive dynamic updates of risk thresholds, the system can maintain assessment accuracy and timeliness under continuous monitoring conditions. The system has a complete structure and clear logic, enabling real-time identification, graded early warning, and dynamic adjustment of cerebral hemorrhage risk in clinical decision support, providing reliable data support and decision-making basis for medical personnel.

[0063] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0064] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0065] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0066] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0067] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0068] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0069] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0070] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0071] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0072] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the risk of cerebral hemorrhage based on multimodal learning, characterized in that, The method includes: S1. Collect cranial multimodal data and perform time alignment, noise suppression, feature enhancement, continuous correction and normalization on the cranial multimodal data to obtain preprocessed cranial multimodal data. S2, based on the preprocessed cranial multimodal data, identify high grayscale suspicious regions, calculate the grayscale feature parameters of high grayscale suspicious regions, quantify the bleeding features of the corresponding regions, mark the suspected hematoma regions, and generate a binary mask matrix of the suspected hematoma regions; S3 calls the binary mask matrix and temporal multimodal data of the suspected hematoma area to assess the risk of cerebral hemorrhage in the suspected hematoma area, and determines the risk level based on the assessment results and generates a graded control strategy. S4, based on the risk assessment results of cerebral hemorrhage in the current assessment cycle and the historical assessment cycle, quantifies the fluctuation of the risk level of cerebral hemorrhage, dynamically updates the risk threshold, and realizes a closed loop of cerebral hemorrhage risk prediction and regulation.

2. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 1, characterized in that, The specific steps for acquiring multimodal brain data and performing time alignment, noise suppression, feature enhancement, continuous correction, and normalization on the preprocessed multimodal brain data are as follows: Using the anterior commissure of the patient as the origin, a three-dimensional spatial reconstruction method based on the DICOM world coordinate system was used to establish a cranial spatial coordinate system, and real-time multimodal data of the cranium at various locations in the patient's cranium were collected. The cranial multimodal data included CT pixel gray values, cerebral blood flow, cerebral blood volume, systolic blood pressure, diastolic blood pressure, heart rate, blood oxygen saturation, respiratory rate, body temperature, and prothrombin time. For the acquired cranial multimodal data, a time synchronization calibration algorithm was used to achieve unified time alignment; an adaptive anisotropic filtering algorithm was used to denoise and smooth the cranial image data while preserving structural fidelity, suppressing random noise and maintaining hematoma boundary information; a sliding window averaging and outlier detection algorithm was used to remove artifacts and smooth trends; a Kalman filter combined with a dynamic state estimation algorithm was used to continuously correct the cranial multimodal data; finally, a Z-Score normalization and min-max normalization algorithm were used to unify the scale and eliminate dimensions of the cranial multimodal data.

3. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 1, characterized in that, The specific steps for identifying high-grayscale suspicious regions based on preprocessed cranial multimodal data are as follows: Preprocessed cranial multimodal data were extracted, and statistical analysis of CT pixel grayscale values ​​was performed over the entire brain. The grayscale frequency distribution vector was calculated, and an adaptive threshold segmentation algorithm was used to determine the grayscale segmentation threshold based on the grayscale frequency distribution vector. High grayscale pixel regions with CT pixel grayscale values ​​higher than the grayscale segmentation threshold were then selected.

4. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 3, characterized in that, The specific steps for calculating the grayscale feature parameters of the high-grayscale suspicious region are as follows: Three-dimensional connectivity analysis is performed on high grayscale pixel regions to identify continuous high-density regions and calculate the volume of each high-density region. Isolated small regions with volumes smaller than the volume threshold are removed. The mean grayscale value of CT pixels and the mean gradient value of CT edges are calculated for the retained high-density regions. The current high-density region is mirrored along the brain midline to the contralateral brain region, and the grayscale values ​​of CT pixels at symmetrical positions are extracted and the regional standard deviation is calculated to obtain the control grayscale standard deviation.

5. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 1, characterized in that, The specific steps for quantifying the bleeding characteristics of the corresponding region, marking suspected hematoma areas, and generating a binary mask matrix for the suspected hematoma areas are as follows: The grayscale ratio term is obtained by dividing the mean grayscale value of CT pixels by the sum of the mean grayscale value of CT pixels and the standard deviation of the grayscale value of the control. The edge gradient regulation term is obtained by calculating the absolute value of the mean CT edge gradient and dividing it by the sum of the absolute value of the mean CT edge gradient and a constant 1. The density structure factor is obtained by multiplying the grayscale ratio term and the edge gradient regulation term. The cerebral blood volume is divided by the sum of the cerebral blood flow and a constant 1, and the ratio is increased by 1 and the natural logarithm is taken to obtain the cerebral circulation regulation factor. The coagulation status factor is obtained by multiplying the ratio of prothrombin time to reference prothrombin time by the cerebral circulation regulation factor. The blood pressure response factor is obtained by dividing the difference between systolic blood pressure and diastolic blood pressure by the sum of the two and taking the hyperbolic tangent of the comparison value. The hemorrhage characteristic assessment value is obtained by adding the density structure factor, the coagulation status factor, and the blood pressure response factor in sequence. The system compares the bleeding feature assessment value and the bleeding threshold in real time. When the bleeding feature assessment value is less than or equal to the bleeding threshold, the corresponding high-density area is marked as a normal area. When the bleeding feature assessment value is greater than the bleeding threshold, the corresponding high-density area is marked as a suspected hematoma area. The system performs a three-dimensional morphological closing operation on the suspected hematoma area to smooth the area boundary and generate a binary mask matrix with the same spatial size as the original CT image.

6. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 1, characterized in that, The specific steps for assessing the risk of cerebral hemorrhage in the suspected hematoma area by calling the binary mask matrix and temporal multimodal data of the suspected hematoma area are as follows: A binary mask matrix is ​​invoked to extract the spatial locations of all suspected hematoma regions and obtain the corresponding intracranial multimodal data. A fixed time window is set as an evaluation period. Within each evaluation period, the mean and standard deviation of heart rate and systolic blood pressure are calculated, and the ratio of the corresponding standard deviation to the mean is calculated as the coefficient of variation of heart rate and the coefficient of variation of systolic blood pressure. At the same time, the volume of suspected hematoma regions is dynamically measured, the rate of change of hematoma volume is calculated, and the maximum rate of change of volume is extracted. Add one to the coefficient of variation of heart rate, take the natural logarithm, multiply by the maximum rate of change of volume, divide the product by the maximum rate of change of volume and add one to obtain the dynamic change regulation term; divide cerebral blood volume by cerebral blood flow and add one, add the ratio of prothrombin time to reference prothrombin time to obtain the circulatory coagulation complex term; add the product of blood oxygen saturation and respiratory rate to body temperature, take the square root of the result to obtain the physiological baseline term; divide the bleeding characteristic assessment value by the physiological baseline term to obtain the physiological correction term. The risk assessment value for cerebral hemorrhage is obtained by multiplying the dynamic change adjustment term, the circulatory coagulation complex term, and the physiological correction term in sequence.

7. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 6, characterized in that, The specific steps for determining the risk level and generating a tiered control strategy based on the assessment results are as follows: Real-time comparison of the intracerebral hemorrhage risk assessment value S with multi-level risk thresholds S1 and S2 to determine the risk level and generate a graded control strategy: When S≤S1, it is determined to be a Level 1 risk, the assessment cycle is shortened, and continuous monitoring is carried out; When S1 < S < S2, it is determined as a secondary risk, and the key attention mechanism is activated. Based on the spatial position, volume, and bleeding feature evaluation value of the suspected hematoma area in the binary mask matrix, a structured risk prompt message is generated and synchronized to the attending doctor's management terminal; When S ≥ S2, it is determined as a tertiary risk, and the risk control process is initiated. Spatial expansion simulation is performed on the suspected hematoma area, and a bleeding prediction report is generated by combining the original CT image and the binary mask matrix, which is pushed to the attending doctor and the nursing station, and the alarm mechanism is triggered simultaneously.

8. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 1, characterized in that, The specific steps for quantitatively evaluating the fluctuation state of the cerebral hemorrhage risk level based on the cerebral hemorrhage risk assessment results of the current evaluation period and the historical evaluation period are as follows: After each evaluation period ends, the cerebral hemorrhage risk assessment values of the current evaluation period and the previous N evaluation periods are extracted, a historical cerebral hemorrhage risk assessment value sequence is constructed, and the historical risk assessment mean value is calculated; the natural logarithm of the current cerebral hemorrhage risk assessment value plus one is subtracted from the natural logarithm of the historical risk assessment mean value plus one, and the obtained difference is used as the numerator; the square of the current cerebral hemorrhage risk assessment value, the square of the historical risk assessment mean value, and the constant 1 are added in sequence, and the square root of the sum value is used as the denominator; The numerator is divided by the denominator to obtain the risk adjustment value.

9. The method for predicting the risk of cerebral hemorrhage based on multimodal learning according to claim 8, characterized in that, The specific steps for dynamically updating the risk threshold to achieve a closed-loop of cerebral hemorrhage risk prediction and regulation are as follows: The risk threshold is adjusted according to the positive or negative trend of the risk adjustment value. When the risk adjustment value is positive, the current risk threshold remains unchanged; When the risk adjustment value is negative, the risk threshold is dynamically adjusted in combination with the change trend of the historical cerebral hemorrhage risk assessment value; The updated risk threshold is reapplied to the determination of the cerebral hemorrhage risk level in the next evaluation period to achieve closed-loop control of the evaluation logic.

10. A multimodal learning-based system for predicting the risk of intracerebral hemorrhage, wherein the multimodal learning-based system for predicting the risk of intracerebral hemorrhage is used to implement the multimodal learning-based method for predicting the risk of intracerebral hemorrhage as described in any one of claims 1-9, characterized in that, The system includes: A data acquisition and preprocessing module, which is used to acquire multi-modal brain data and perform time alignment, noise suppression, feature enhancement, continuous correction, and normalization processing on the multi-modal brain data to obtain preprocessed multi-modal brain data; A bleeding feature recognition module, which is used to identify high-gray suspicious areas based on the preprocessed multi-modal brain data, calculate the gray feature parameters of the high-gray suspicious areas, quantify the bleeding features of the corresponding areas, mark the suspected hematoma areas, and generate a binary mask matrix of the suspected hematoma areas; A risk assessment and determination module, which is used to call the binary mask matrix of the suspected hematoma area and the time-series multi-modal data, evaluate the cerebral hemorrhage risk degree of the suspected hematoma area, determine the risk level based on the evaluation result, and generate a hierarchical control strategy; A dynamic threshold control module, which is used to quantitatively evaluate the fluctuation state of the cerebral hemorrhage risk level based on the cerebral hemorrhage risk assessment results of the current evaluation period and the historical evaluation period, dynamically update the risk threshold, and achieve a closed-loop of cerebral hemorrhage risk prediction and regulation.

Images