Hemorrhagic brain injury intracranial pressure prediction and early warning system based on CT image

Abstract

The invention provides an hemorrhagic brain injury intracranial pressure prediction and early warning system based on a CT image, and the system comprises a data collection module which is used for collecting the original data of a target data source of a target patient; the hybrid model comprises a 3DCNN model used for extracting hematoma spatial features of the head CT image sequence; the LSTM model is used for extracting hematoma time sequence characteristics of the physiological signals; the multi-modal fusion module is used for fusing the hematoma space features and the hematoma time sequence features and outputting hematoma fusion features; the prediction module is used for predicting an ICP sequence of the target patient in a future preset time period based on the hematoma fusion features; the self-adaptive early warning module is used for outputting a visual result based on the predicted ICP sequence in the future preset time period; and through a dynamic threshold adjustment mechanism, adjusting an early warning threshold of the target patient in combination with the case information, and triggering graded early warning in real time. Through personalized data processing and dynamic early warning, accurate, efficient and real-time patient risk management is realized.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence and neurocritical care technology, and in particular to a CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury. Background Art

[0002] Among current medical treatments, computed tomography (CT) of the head is the preferred diagnostic tool for hemorrhagic brain injury, clearly displaying key parameters such as hematoma volume, mass effect, ventricular morphology, and midline shift. Studies have shown that these imaging features are significantly correlated with changes in intracranial pressure. For example, when the hematoma volume exceeds 30ml or the midline shift is ≥5mm, the risk of uncontrolled intracranial pressure increases dramatically.

[0003] However, traditional imaging features cannot capture the dynamic evolution of pathological processes and are difficult to integrate with real-time physiological data, making it easy to miss subtle changes. Analysis is time-consuming and error-prone due to reliance on physician experience, limiting the development of accurate early warning and comprehensive predictive models.

[0004] Therefore, the present invention proposes a CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury. Summary of the Invention

[0005] The present invention provides a CT image-based intracranial pressure prediction and warning system for hemorrhagic brain injury, which is used to analyze patient data in real time and dynamically adjust warning thresholds through data acquisition and processing, spatiotemporal and temporal feature extraction, multimodal fusion and prediction modules, combined with an adaptive warning mechanism, to provide accurate graded warning signals and realize personalized and intelligent risk management.

[0006] The present invention provides a CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury, comprising: Data acquisition module: used to collect case information, physiological signals and continuously shot head CT image sequences of target patients from multiple target data sources; The hybrid model includes: a three-dimensional convolutional neural network (3DCNN) model for extracting hematoma spatial features from head CT image sequences; a long short-term memory (LSTM) model for extracting hematoma temporal features from physiological signals; a multimodal fusion module for fusing hematoma spatial features with hematoma temporal features to output hematoma fusion features; and a prediction module for predicting the target patient's ICP sequence within a preset future time period based on the hematoma fusion features. Adaptive early warning module: used to output visualization results based on the predicted ICP sequence within a preset future time period; and through a dynamic threshold adjustment mechanism, adjust the warning threshold of the target patient based on the target patient's case information and predicted ICP value, and trigger graded warnings in real time.

[0007] Preferably, the data acquisition module includes: Case information data collection unit: used to collect case information from the target patient's electronic medical record based on the target patient's unique identifier; Dynamic CT image sequence data acquisition unit: used to continuously acquire head CT at preset time intervals after the target patient is admitted to the hospital to obtain the target patient's original CT image sequence; Physiological signal acquisition unit: used to collect the original physiological signals of the target patient according to the unique identifier of the target patient; Preprocessing unit: used for preprocessing the original CT image sequence and original physiological signal to generate a temporally and spatially consistent and structured CT image sequence and physiological signal of the target patient.

[0008] Preferably, the 3DCNN model is specifically used for: By using the time-series slice splicing technology, the original CT image sequence of continuous time points in the CT image sequence is constructed into four-dimensional space-time CT data; The convolutional layer in the 3DCNN model uses underlying convolution kernels of different sizes to perform sliding scans in the voxel space of the 4D spatiotemporal CT data, dividing the 4D spatiotemporal CT data into multiple regions and detecting the local texture spatial features of each region; The fully connected layer in the 3DCNN model uses multi-scale fusion technology to integrate the local texture spatial features of all regions and output the hematoma spatial features of the original CT image sequence of the target patient.

[0009] Preferably, the third convolution block of the 3DCNN model further includes a spatial attention module, which is used to pay spatial attention to the local texture spatial features of each region through a dynamic feature clipping strategy, remove the region smaller than the clipping threshold, and retain the hemorrhagic brain tissue region; wherein the formula for calculating the clipping threshold T based on spatial attention is: ; Formula (1) In formula (1), M represents the attention weight map generated by the spatial attention map, , the spatial attention map is mapped to the attention weight map through 3DCNN , represents the mean of the attention weight map, represents the standard deviation of the attention weight map, Represents an adjustable hyperparameter.

[0010] Preferably, a multi-layer cascaded LSTM unit is used within the LSTM model, including: Input the physiological signal into a multi-layer LSTM unit. Each LSTM unit includes a cell state, a hidden state, and a gating mechanism. The gating mechanism processes feature input and output and forgetting. The cell state captures the long-term trend characteristics of the original physiological signal, and the hidden state captures the short-term fluctuation characteristics of the original physiological signal. The long-term trend characteristics and short-term fluctuation characteristics output by each LSTM unit constitute the unit time series characteristics output by the LSTM unit. Among them, an attention mechanism is added between the connections of adjacent layers of LSTM units. The attention mechanism adjusts the weights of the preset key time points of the unit temporal features output by each layer of LSTM units; Finally, the hematoma time series characteristics of the target patient's physiological signals are output.

[0011] Preferably, in the multimodal fusion module, the hematoma temporal features and the hematoma spatial features are gated and fused through an adaptive weighting mechanism. The specific formula is as follows: ; (2) In formula (2), Indicates the characteristics of hematoma fusion, represents the Sigmoid function, represents the hematoma spatial feature weight matrix, represents the hematoma time series feature weight matrix, Indicates the spatial characteristics of hematoma, Indicates the temporal characteristics of hematoma, Represents element-wise product.

[0012] Preferably, the prediction module in the hybrid model includes a Transformer model, which is used to add a position code to each time node of the hematoma fusion feature, and then map the hematoma fusion feature into an ICP sequence of a future preset time period.

[0013] Preferably, the training of the hybrid model includes: Collecting a multimodal dataset of historical medical imaging data and historical physiological signal data of multiple patients as a sample dataset, and dividing the sample dataset into a training set and a validation set; Build an initial hybrid model including 3DCNN model, LSTM model, multimodal fusion module, and prediction module; The initial hybrid model is trained using the training set and validation set, and the trend loss and fluctuation loss are combined to form a multi-scale loss function; The model parameters are optimized through a multi-scale loss function. When the model parameters of the initial hybrid model meet the stability requirements, the training is stopped to obtain the hybrid model.

[0014] Preferably, the adaptive early warning module includes: The visualization unit is used to determine the risk interval corresponding to each ICP value at each time point in a preset future time period based on four preset risk levels, and color-code it to generate an ICP trend graph containing the color-coded risk intervals. It also combines the ICP sequence and hematoma fusion features and visualizes them as an ICP change curve and a hematoma diffusion probability heat map. Threshold adjustment unit: used to adjust the preset warning threshold based on the case information of the target patient collected by the case information collection subunit to obtain the final warning threshold; Multi-level warning unit: used to divide the risk level into four levels for graded alarm based on the multi-level warning mechanism; when the ICP value obtained based on the hybrid model exceeds the final warning threshold, the corresponding level of warning is triggered according to the graded alarm, and a graded warning signal is generated for the target patient in real time, and clinical treatment recommendations and warning reports containing visual results are generated for the target patient based on the graded warning signal.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a CT image-based intracranial pressure prediction and warning system for hemorrhagic brain injury, which is used to analyze patient data in real time and dynamically adjust warning thresholds through data acquisition and processing, spatiotemporal and temporal feature extraction, multimodal fusion and prediction modules, combined with an adaptive warning mechanism, to provide accurate graded warning signals and realize personalized and intelligent risk management. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 Schematic diagram of the structure of a CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury provided by an embodiment of the present invention; Figure 2 This is a table of definitions of four levels of population-specific gradient chromatograms provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0019] Example 1: like Figure 1 As shown, the embodiment of the present invention provides a CT image-based intracranial pressure prediction and warning system for hemorrhagic brain injury, including: Data acquisition module: used to collect case information, physiological signals and continuously shot head CT image sequences of target patients from multiple target data sources; The hybrid model includes: a three-dimensional convolutional neural network (3DCNN) model for extracting hematoma spatial features from head CT image sequences; a long short-term memory (LSTM) model for extracting hematoma temporal features from physiological signals; a multimodal fusion module for fusing hematoma spatial features with hematoma temporal features to output hematoma fusion features; and a prediction module for predicting the target patient's ICP sequence within a preset future time period based on the hematoma fusion features. Adaptive early warning module: used to output visualization results based on the predicted ICP sequence within a preset future time period; and through a dynamic threshold adjustment mechanism, adjust the warning threshold of the target patient based on the target patient's case information and predicted ICP value, and trigger graded warnings in real time.

[0020] In this embodiment, the target patient refers to the specific individual that the system focuses on during the data collection, processing and prediction process.

[0021] In this embodiment, the target data source refers to all raw data sources related to the target patient for collection and analysis, including: dynamic CT image sequences, physiological signal data, case information data, etc.

[0022] In this embodiment, the case information is the patient's basic health information: age, medical history.

[0023] In this embodiment, the physiological signal refers to the patient's real-time vital sign signals: intracranial pressure (ICP), arterial blood pressure (ABP), and heart rate (HR).

[0024] In this embodiment, the head CT image sequence is a medical imaging sequence that uses a combination of X-rays and computer technology to generate cross-sectional images of the patient's body. The CT scanner collects X-ray data from different angles by rotating the X-ray source and detector. The computer processes this data to generate cross-sectional images of different layers within the patient's body.

[0025] In this embodiment, the hematoma spatial feature represents spatial information related to the hematoma, and may include spatial features such as the position, size, and shape of the hematoma in the image.

[0026] In this embodiment, the three-dimensional convolutional neural network (3DCNN) model is a model that extends the traditional two-dimensional convolutional neural network (2D CNN). During the convolution operation, it not only considers the information of the two-dimensional space (such as the width and height of the image), but also can simultaneously capture the third dimension of the image or data (depth or time series).

[0027] In this embodiment, the Long Short-Term Memory (LSTM) model is a special recurrent neural network (RNN) used to process and predict time series data. By introducing a "memory unit," LSTM effectively addresses the vanishing and exploding gradient problems encountered by traditional RNNs when learning long sequences, making it particularly effective at handling long-term dependencies.

[0028] In this embodiment, the hematoma time series feature refers to the time dynamic change feature related to the hematoma obtained after the original physiological signal is processed by a multi-layer LSTM unit.

[0029] In this embodiment, the multimodal fusion module is a module for integrating and processing multiple data modes from different data sources.

[0030] In this embodiment, the prediction module generates a prediction result of the future health status of the target patient based on the feature information extracted from the data acquisition module and the hybrid model.

[0031] In this embodiment, the ICP series refers to time series data of intracranial pressure prediction.

[0032] In this embodiment, the dynamic threshold adjustment mechanism refers to automatically adjusting the threshold used to trigger the warning signal in the early warning system based on the real-time health status and predicted results of the target patient. Suppose there is a 50-year-old male patient who suffered severe craniocerebral injury due to a car accident. His GCS score was 8 points when he was admitted to the hospital. The CT scan showed a right frontotemporal subdural hematoma and a midline structure shift of about 10mm. The patient underwent intracranial pressure monitoring after admission, and the ICP value continued to fluctuate between 25-30mmHg, accompanied by frequent intracranial pressure events. Because the patient was older and had a history of hypertension, the system lowered the ICP warning threshold to 20mmHg. This enables the system to detect the patient's intracranial pressure events earlier, which buys more time for medical staff to intervene.

[0033] In this embodiment, the warning threshold refers to the threshold used in the health monitoring system to determine whether a specific health indicator or prediction result has reached a standard value that requires attention or intervention. When certain indicators exceed or fall below the set threshold, the system will issue an alarm or warning signal, indicating a possible health risk.

[0034] In this embodiment, graded warning means that the system uses different levels or categories to indicate the urgency of health risks based on the patient's health status, prediction results, and set warning thresholds.

[0035] The working principle and beneficial effects of the above technical solution are as follows: through data collection and processing, feature extraction, multimodal fusion and prediction modules, combined with an adaptive early warning mechanism, dynamic monitoring and accurate prediction of patient status are achieved. By adjusting the early warning threshold, the accuracy and personalization of early warnings are improved, and risk management capabilities are enhanced.

[0036] Example 2: Based on the above embodiment, the data acquisition module includes: Case information data collection unit: used to collect case information from the target patient's electronic medical record based on the target patient's unique identifier; Dynamic CT image sequence data acquisition unit: used to continuously acquire head CT at preset time intervals after the target patient is admitted to the hospital to obtain the target patient's original CT image sequence; Physiological signal acquisition unit: used to collect the original physiological signals of the target patient according to the unique identifier of the target patient; Preprocessing unit: used for preprocessing the original CT image sequence and original physiological signal to generate a temporally and spatially consistent and structured CT image sequence and physiological signal of the target patient.

[0037] In this embodiment, the unique identifier refers to a unique identifier that can uniquely identify the patient's identity in the system, which is used to distinguish different patients and ensure that all medical data and records are bound to a specific patient to avoid confusion or misprocessing.

[0038] In this embodiment, the preset time interval refers to a time interval set in the dynamic CT image sequence data acquisition unit, which is used to control the frequency of each CT image acquisition.

[0039] In this embodiment, the original CT image sequence refers to unprocessed continuous image data obtained through CT scanning.

[0040] In this embodiment, the raw physiological signal refers to unprocessed physiological data collected from the patient, such as intracranial pressure (ICP), arterial blood pressure (ABP), heart rate (HR), etc.

[0041] In this embodiment, preprocessing refers to performing a series of operations on the original data to improve data quality, reduce noise, and enhance signal characteristics. Preprocessing can be performed using statistical parameters such as mean, standard deviation, etc.

[0042] In this embodiment, spatiotemporal consistency refers to ensuring temporal and spatial consistency and alignment of multiple types of data (such as CT images and physiological signals) when processing these data. Assuming that CT images are collected at regular intervals (e.g., every hour), while physiological signals (such as electrocardiograms and blood pressure) are collected in real time, there may be multiple data points between the CT images. To achieve temporal consistency, an interpolation method can be used to interpolate the physiological signal data at the time points of the CT images, aligning the acquisition times of the physiological signal and the CT images. For example, linear interpolation can be used to interpolate the time points between the CT images into the physiological signal data, ensuring that each CT image data point has corresponding physiological signal data.

[0043] In this embodiment, structuring refers to organizing the original data according to certain standards and formats to facilitate subsequent processing and analysis.

[0044] The working principle and beneficial effects of the above technical solution are as follows: through the continuous acquisition of patient information, CT image sequences, and physiological signals, and the generation of structured data through a preprocessing unit to ensure the temporal and spatial consistency of the data, comprehensive and accurate patient data collection is achieved, providing high-quality raw data support for subsequent analysis and prediction.

[0045] Example 3: Based on the above embodiment, the 3DCNN model in the hybrid model is specifically used for: By using the time-series slice splicing technology, the original CT image sequence of continuous time points in the CT image sequence is constructed into four-dimensional space-time CT data; The convolutional layer in the 3DCNN model uses underlying convolution kernels of different sizes to perform sliding scans in the voxel space of the 4D spatiotemporal CT data, dividing the 4D spatiotemporal CT data into multiple regions and detecting the local texture spatial features of each region; The fully connected layer in the 3DCNN model uses multi-scale fusion technology to integrate the local texture spatial features of all regions and output the hematoma spatial features of the original CT image sequence of the target patient.

[0046] In this embodiment, time-series slice stitching technology is used to stitch together a sequence of raw CT images at consecutive time points into four-dimensional spatiotemporal CT data. Each CT image frame is typically a three-dimensional data set (x, y, z). CT images acquired at consecutive time points form a time series. The CT image at each time point has corresponding spatial information (voxel information). Time-series slice stitching technology stitches the CT image data at these time points into a four-dimensional dataset, where the first three dimensions are spatial dimensions (x, y, z) and the fourth dimension is time (t).

[0047] In this embodiment, four-dimensional spatiotemporal CT data is a type of CT image data that combines the time dimension and the space dimension, and is used to describe the dynamic process of changes in body tissues or lesions (such as bleeding, tumors, etc.) over time at different time points.

[0048] In this example, the underlying convolution kernels refer to the initial convolution kernels in a convolutional neural network, which are used to capture the basic features of the input data. For a three-dimensional convolutional neural network (3DCNN), the underlying convolution kernels slide in three-dimensional space and extract local spatial features from the four-dimensional space-time CT image.

[0049] In this embodiment, sliding scanning is a common technique used in convolutional neural networks to perform local processing on input data. It slides a fixed-size window across each location in the data to extract local features. In medical image processing (such as CT images), sliding scanning can help extract local spatial features of the image.

[0050] In this embodiment, the region refers to a small area in the voxel space divided after sliding scanning of four-dimensional spatiotemporal CT data through a three-dimensional convolutional neural network (3DCNN).

[0051] In this embodiment, voxel space refers to the four-dimensional space-time CT data, where each "voxel" represents a small three-dimensional area or volume unit. In medical imaging, CT images are usually viewed as a three-dimensional grid, and each voxel corresponds to a volume element in the image, similar to a pixel in a two-dimensional image (each pixel represents a position and grayscale value on a plane).

[0052] In this embodiment, the local texture spatial feature refers to the spatial information of the local area of ​​the image extracted from the four-dimensional spatiotemporal CT data by a three-dimensional convolutional neural network (3DCNN).

[0053] The working principle and beneficial effects of this technical solution are as follows: By splicing time-series slices and using a 3DCNN dynamic feature cropping strategy, local texture spatial features are extracted, unimportant regions are eliminated, and hematoma areas are accurately identified. By optimizing feature extraction through spatial attention, the accuracy and effectiveness of hematoma detection are improved, providing high-quality data support for lesion analysis.

[0054] Example 4: Based on the above embodiment, the 3DCNN model further includes a spatial attention module after the third convolution block, which is used to use a dynamic feature clipping strategy to focus on the spatial attention of the local texture spatial features of each region, remove the region smaller than the clipping threshold, and retain the brain tissue region with bleeding; wherein the formula for calculating the clipping threshold T based on spatial attention is: ; Formula (1) In formula (1), M represents the attention weight map generated by the spatial attention map, , where H represents height, W represents width, and D represents depth; the spatial attention map is mapped to the attention weight map through 3DCNN , represents the mean of the attention weight map, represents the standard deviation of the attention weight map, Represents an adjustable hyperparameter.

[0055] In this embodiment, the dynamic feature clipping strategy is a screening method for local texture space features, which aims to improve processing efficiency and accuracy by weighting the spatial attention of each region and eliminating unimportant or irrelevant regions.

[0056] In this embodiment, spatial attention refers to assigning different importance weights to regions at different spatial locations through an attention mechanism when processing four-dimensional spatiotemporal CT data. Its purpose is to help the model focus on more important regions, thereby improving processing efficiency and accuracy.

[0057] In this embodiment, the clipping threshold is a standard used to determine which areas of local texture space features should be removed in the 3DCNN dynamic feature clipping unit, and the removed areas are brain tissue areas without bleeding.

[0058] In this embodiment, the attention weight map is a graph generated by the spatial attention mechanism to represent the importance of each region in the input data. By calculating the spatial attention map, the 3DCNN determines which regions are more important for the current task (e.g., bleeding area detection) and focuses on these regions.

[0059] In this example, hemorrhagic brain tissue regions are identified and located using a local texture spatial feature extraction unit and a 3D CNN dynamic feature cropping unit. By processing spatiotemporal data, the model is able to detect areas that exhibit significant changes compared to normal tissue (such as hematomas or hemorrhages), which are hallmarks of hemorrhage.

[0060] The working principle and beneficial effects of this technical solution are as follows: By introducing a spatial attention module and utilizing a dynamic feature cropping strategy to optimize the 3DCNN model, non-bleeding areas are eliminated while retaining hematoma-related areas, improving hematoma detection accuracy. Adjusting the cropping threshold enhances the model's ability to identify hemorrhagic brain tissue.

[0061] Example 5: Based on the above embodiment, a multi-layer cascaded LSTM unit is used within the LSTM model, including: Input the physiological signal into a multi-layer LSTM unit. Each LSTM unit includes a cell state, a hidden state, and a gating mechanism. The gating mechanism processes feature input and output and forgetting. The cell state captures the long-term trend characteristics of the original physiological signal, and the hidden state captures the short-term fluctuation characteristics of the original physiological signal. The long-term trend characteristics and short-term fluctuation characteristics output by each LSTM unit constitute the unit time series characteristics output by the LSTM unit. Among them, an attention mechanism is added between the connections of adjacent layers of LSTM units. The attention mechanism adjusts the weights of the preset key time points of the unit temporal features output by each layer of LSTM units; Finally, the hematoma time series characteristics of the target patient's physiological signals are output.

[0062] In this embodiment, a multi-layer LSTM stacks multiple LSTM units to form a network architecture. The output of each layer's LSTM subunits serves as the input to the next layer, processing data layer by layer.

[0063] In this embodiment, the gating mechanism includes an input gate, a forget gate, and an output gate. The input gate determines which new information is added to the cell state, the forget gate determines which old information is forgotten from the cell state, and the output gate determines which information is output based on the cell state.

[0064] In this embodiment, the cell state stores long-term information from past time steps to the current moment. By transferring this information between time steps, the network can continuously update and remember key information from earlier in the sequence. For physiological signal processing, the cell state helps capture long-term trends in signals, such as persistent fluctuations or chronic changes in human physiological signals.

[0065] In this embodiment, long-term trend features describe slow, long-term changes in physiological signals. For example, in medical data, these features may reflect the progression of chronic diseases, changes in the effects of long-term treatments, or gradual changes in patient vital signs.

[0066] In this embodiment, the hidden state is used to capture short-term changes and fluctuations in the input data. In physiological signals, this feature may reflect instantaneous changes in the data, such as short-term fluctuations in heart rate or rapid fluctuations in blood pressure. These fluctuations are typically small changes in amplitude over time and have short time dependencies.

[0067] In this embodiment, the short-term fluctuation feature refers to the changes and fluctuations of the original physiological signal on a short time scale. For example, in the patient's real-time vital sign signals: intracranial pressure (ICP), arterial blood pressure (ABP), heart rate (HR), the short-term fluctuation feature may be a sharp change in heart rate, rapid fluctuations in blood pressure, etc.

[0068] In this embodiment, the unit temporal feature refers to the time-related feature representation output by the LSTM unit, which combines the output of each layer of LSTM sub-units, including long-term trend features and short-term fluctuation features.

[0069] In this embodiment, the attention mechanism assigns weights to the output features of each LSTM layer, especially giving higher weights to those time points that are most important to the task. For example, in physiological signal analysis, certain sudden physiological changes (such as a sudden increase in heart rate, a sharp change in blood pressure, etc.) may be more important to the diagnosis results. The attention mechanism can ensure that the model pays more attention to these time points.

[0070] In this embodiment, the preset key time point refers to a time node in the time series data that is determined in advance to be most critical to the task.

[0071] The working principle and beneficial effects of this technical solution are as follows: It uses multi-layer LSTM units to capture the long-term trends and short-term fluctuations of physiological signals, and introduces an attention mechanism to enhance the feature weights of key time points. By accurately extracting the temporal characteristics of hematomas, it optimizes physiological signal analysis and improves the temporal accuracy and effectiveness of hematoma prediction.

[0072] Example 6: On the basis of the above embodiment, the hematoma temporal features and hematoma spatial features are gated and fused in the multimodal fusion module through an adaptive weighting mechanism. The specific formula is as follows: ; (2) In formula (2), Indicates the characteristics of hematoma fusion, represents the Sigmoid function, represents the hematoma spatial feature weight matrix, represents the hematoma time series feature weight matrix, Indicates the spatial characteristics of hematoma, Indicates the temporal characteristics of hematoma, Represents element-wise product.

[0073] In this embodiment, the sigmoid function compresses the input value into the (0, 1) interval, generating a "gating signal." If the σ output is close to 1, the model focuses more on imaging features (such as the volume of the hemorrhage). If the σ output is close to 0, the model relies more on time series features (such as the ICP fluctuation trend).

[0074] In this embodiment, the hematoma spatial feature weight matrix is ​​a parameter matrix that weights the hematoma spatial features according to an adaptive weighting mechanism when fusing the hematoma temporal features and the hematoma spatial features. It is used to control the importance or weight of the hematoma spatial features in the final fusion process.

[0075] In this embodiment, the hematoma time series feature weight matrix is ​​a matrix used in an adaptive weighting mechanism to control the importance or weight of the hematoma time series feature in the fusion process.

[0076] In this embodiment, element-by-element multiplication multiplies the gated signal by the original fused features, achieving dynamic feature weighting. This weight distribution suppresses irrelevant features and enhances key features. For example, when ventriculomegaly (an imaging feature) is detected, σ increases the weight of the imaging feature.

[0077] When the ICP exhibits periodic fluctuations (time series characteristics), σ will enhance the contribution of the time series characteristics.

[0078] In this embodiment, the formula avoids manually setting fusion rules through a learnable weight matrix and gating mechanism, highlights the complementary key information in the two modalities, suppresses redundancy or noise, and can intuitively reflect the contribution of different features.

[0079] The working principle and beneficial effects of this technical solution are as follows: An adaptive weighting mechanism is used to gate and fuse the temporal and spatial features of hematomas, and a sigmoid function is used to adjust the weights of each feature to optimize the fusion effect. The multimodal fusion module improves the comprehensive representation of hematoma features, enhancing the accuracy and stability of hematoma detection and prediction.

[0080] Example 7: Based on the above embodiment, the prediction module in the hybrid model includes a Transformer model, which is used to add position coding to each time node of the hematoma fusion feature, and then map the hematoma fusion feature into an ICP sequence of a future preset time period.

[0081] In this embodiment, the Transformer model consists of an encoder and a decoder. The encoder obtains global information by calculating the relationship between each element in the input sequence and other elements.

[0082] In this example, positional encoding is a key component of the Transformer model, providing the model with information about the position of elements in a sequence. Because the Transformer architecture lacks a recursive structure like RNNs or LSTMs, it lacks built-in sequential information, so positional information must be passed to the model. This is where positional encoding comes in: it ensures the model understands the relative and absolute position of each element in the input sequence.

[0083] In this embodiment, the future preset time period refers to a time range that is defined in advance and needs to be predicted during the model prediction process, and a commonly used time period is 6 hours.

[0084] In this embodiment, the ICP sequence refers to the change trend sequence of the patient's ICP within a period of time in the future predicted by the model. The ICP value at each time node represents the patient's intracranial pressure status at that moment.

[0085] The working principle and beneficial effects of this technical solution are as follows: A Transformer model is used to process hematoma fusion features in conjunction with position encoding to predict future ICP sequences. This is then color-coded based on pre-set risk intervals to visually display different risk levels. This effectively improves the accuracy of ICP prediction and the visualization of risk identification.

[0086] Example 8: Based on the above embodiment, the training of the hybrid model includes: Collecting a multimodal dataset of historical medical imaging data and historical physiological signal data of multiple patients as a sample dataset, and dividing the sample dataset into a training set and a validation set; Build an initial hybrid model including 3DCNN model, LSTM model, multimodal fusion module, and prediction module; The initial hybrid model is trained using the training set and validation set, and the trend loss and fluctuation loss are combined to form a multi-scale loss function; The model parameters are optimized through a multi-scale loss function. When the model parameters of the initial hybrid model meet the stability requirements, the training is stopped to obtain the hybrid model.

[0087] In this embodiment, historical medical imaging data refers to the medical imaging examination results that the patient has undergone in the past, and the data covers patients with different disease severity and individual differences to improve the generalization ability of the model.

[0088] In this embodiment, historical physiological signal data refers to physiological test data that the patient has received in the past, and the data covers patients with different disease severity and individual differences to improve the generalization ability of the model.

[0089] In this embodiment, the sample data set refers to the original data set used to train, verify and test the model.

[0090] In this embodiment, the training set is used for model training.

[0091] In this embodiment, the validation set is used for model tuning and to prevent overfitting.

[0092] In this embodiment, the algorithm base models refer to the two main models used to construct the hybrid model: the 3DCNN (3D Convolutional Neural Network) model and the LSTM (Long Short-Term Memory Network) model. These two models are used to extract spatiotemporal features and temporal features, respectively, and serve as the core components of the hybrid model.

[0093] In this example, trend loss is used to measure the difference between the model output and the long-term trend of the actual data. In time series forecasting tasks, trend loss helps optimize the model's prediction results to better fit the long-term trend of the data.

[0094] In this example, the volatility loss measures the difference between the predicted value and the actual data at the level of short-term or local fluctuations. In time series forecasting tasks, the volatility loss focuses on instantaneous changes or local fluctuations in the data, such as short-term fluctuations or frequent ups and downs in the signal, rather than the overall trend. It helps the model learn the details of the data, especially in tasks with high volatility.

[0095] In this embodiment, the loss function is a function used to measure the gap between the model prediction value and the actual value. In this system, trend loss and volatility loss are combined to form a multi-scale loss function, which can be achieved by weighted summing the two loss functions.

[0096] In this embodiment, the stability requirement is met by training the model using the training set and monitoring indicators such as loss and accuracy during training to evaluate the model's performance. Training can be stopped when the loss no longer decreases significantly or the accuracy reaches a stable level.

[0097] The working principle and beneficial effects of the above technical solution are: by combining trend loss and fluctuation loss, a multi-scale loss function is constructed, the parameters of the 3DCNN and LSTM hybrid model are optimized, the model's prediction accuracy for patient physiological signals and imaging data is improved, and the accurate capture of long-term trends and short-term fluctuations is ensured.

[0098] Example 9: Based on the above embodiment, the adaptive warning module includes: The visualization unit is used to determine the risk interval corresponding to each ICP value at each time point in a preset future time period based on four preset risk levels, and color-code it to generate an ICP trend graph containing the color-coded risk intervals. It also combines the ICP sequence and hematoma fusion features and visualizes them as an ICP change curve and a hematoma diffusion probability heat map. Threshold adjustment unit: used to adjust the preset warning threshold based on the case information of the target patient collected by the case information collection subunit to obtain the final warning threshold; Multi-level warning unit: used to divide the risk level into four levels for graded alarm based on the multi-level warning mechanism; when the ICP value obtained based on the hybrid model exceeds the final warning threshold, the corresponding level of warning is triggered according to the graded alarm, and a graded warning signal is generated for the target patient in real time, and clinical treatment recommendations and warning reports containing visual results are generated for the target patient based on the graded warning signal.

[0099] In this embodiment, the four levels of preset risk intervals refer to the classification of ICP (intracranial pressure) values ​​into four risk levels according to different ranges. Each level corresponds to a color label to help quickly identify the patient's ICP status and risk, such as Figure 2 shown.

[0100] In this embodiment, real-time condition information refers to data collected and reflecting the patient's physiological status in real time during the patient monitoring process. These data are usually obtained by various physiological signal acquisition devices.

[0101] In this embodiment, the preset warning threshold refers to one or more standard values ​​pre-set during system design based on medical knowledge, clinical experience, historical data analysis, and other factors, and is used to monitor changes in the patient's condition.

[0102] In this embodiment, the final warning threshold is a warning standard value that is dynamically adjusted based on the patient's real-time condition information and specific data of the target patient.

[0103] In this embodiment, the ICP change curve is a graph used to represent changes in the patient's intracranial pressure (ICP) over time.

[0104] In this embodiment, the hematoma spread probability heat map is a visual chart that is primarily used to represent the spread probability of a hematoma (e.g., intracranial hemorrhage) within a patient. The chart reflects the hematoma spread risk in different areas through color changes.

[0105] In this embodiment, the visualization results include an ICP change curve and a hematoma diffusion probability heat map.

[0106] In this embodiment, the multi-level warning mechanism is a warning system that is graded according to the degree of disease risk. Risk is usually divided into several levels, and each level represents a different degree of severity of the disease.

[0107] In this embodiment, the graded alarm is achieved by dividing the intracranial pressure risk into four levels: normal, slightly elevated, moderately elevated, and severely elevated, and by color coding (green / yellow / orange / red).

[0108] In this embodiment, clinical treatment recommendations refer to treatment recommendations automatically generated in the medical early warning system based on the patient's condition, early warning signals, and risk level.

[0109] The working principle and beneficial effects of the above technical solution are: through a multi-level warning mechanism, threshold adjustment and visualization unit, combined with ICP value and hematoma diffusion characteristics, real-time risk assessment and graded alarm are achieved, clinical treatment recommendations and warning reports are generated, and the risk management and decision support capabilities of target patients are improved.

[0110] The workflow of the above-mentioned CT image-based hemorrhagic brain injury intracranial pressure prediction and warning system is described below with reference to a specific example. However, it is worth noting that this specific example is only for better illustrating the present invention and does not constitute an improper limitation to the present invention.

[0111] Example 10: Patient Li, a 45-year-old male, was rushed to the hospital due to head trauma caused by a car accident. When admitted to the hospital, the patient was unconscious and had symptoms such as nausea and vomiting. The medical staff immediately conducted a comprehensive multimodal data acquisition on the patient. First, the dynamic change data of the patient's intracranial hemorrhage were obtained through a series of continuously taken head CT images. These CT images are presented in the form of three-dimensional voxels, which can clearly show the location, volume and morphological changes of the hematoma. At the same time, the medical staff also recorded the patient's real-time vital signs, including key indicators such as heart rate and blood pressure, and important information such as age and medical history.

[0112] The acquired CT image sequence is input into the 3DCNN spatiotemporal feature extraction module. This module uses a three-dimensional convolutional neural network architecture that can simultaneously capture the morphological characteristics of hematomas in the spatial dimension and their dynamic evolution in the temporal dimension. Through a combination of multiple convolutional layers and pooling layers, the module gradually extracts multi-scale features from microscopic to macroscopic. The underlying convolution kernel detects tiny intracranial bleeding points in the patient, while the high-level network integrates global information such as hematoma volume, edge diffusion speed, and the degree of midline structural displacement. In addition, the module also tracks the dynamic trajectory of hematoma volume changes through feature splicing in the temporal dimension, providing key input for subsequent time series modeling.

[0113] The spatiotemporal feature sequences output by the 3DCNN module are input into the LSTM temporal dynamic modeling module. This module utilizes a long short-term memory (LSTM) network to construct a temporal dependency model of intracranial pressure changes. Through its two-layer LSTM architecture, the module captures short-term fluctuations, such as minute-level pressure changes, while also enabling long-term trend memory through cell state transfer. In particular, the module incorporates an attention mechanism, assigning greater weight to critical time windows, such as periods of rapid hematoma expansion. For example, between the second and fourth hours after the patient's admission, the LSTM module detected a significant increase in hematoma volume and, based on this, predicted a rising trend in intracranial pressure.

[0114] The outputs of the 3DCNN and LSTM modules are input into the multimodal fusion and prediction module. This module fuses 3D structural information with temporal dynamic parameters through feature concatenation, followed by nonlinear mapping using a fully connected network. The prediction results include two dimensional outputs: a continuous value prediction, which generates an intracranial pressure curve for the next six hours using a sliding window; and a risk stratification, which uses a four-color warning system (green / yellow / orange / red) to correspond to different intervention priorities. For example, five hours after the patient's admission, the module predicted that the intracranial pressure might exceed 25 mmHg within the next two hours, triggering a red alert.

[0115] The output of the multimodal fusion and prediction module is fed into the adaptive early warning module. This module dynamically adjusts the early warning strategy based on individual patient differences and real-time condition. Using a built-in library of patient-specific parameters (such as age and underlying diseases), the module calibrates the early warning threshold. For example: Consider a 50-year-old male patient who suffered a severe craniocerebral injury in a car accident. His GCS score was 8 upon admission, and a CT scan showed a right frontotemporal subdural hematoma with a midline structure shift of approximately 10 mm. Upon admission, the patient underwent intracranial pressure monitoring, and the ICP value continued to fluctuate between 25 and 30 mmHg, accompanied by frequent episodes of increased intracranial pressure. Because the patient was older and had a history of hypertension, the system lowered the ICP warning threshold to 20 mmHg. This enabled the system to detect increased intracranial pressure events earlier, buying more time for medical staff to intervene.

[0116] During monitoring, the system detected a continuous rise in the patient's ICP, accompanied by physiological abnormalities such as a decrease in ABP and an increase in HR. Based on these changes, the system dynamically raised the alert level from yellow to red. This attracted the attention of medical staff, who promptly implemented treatment measures to reduce intracranial pressure, effectively controlling the patient's condition.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury, characterized by: include: A data acquisition module is used to collect case information, physiological signals, and continuously shot head CT image sequences of target patients from multiple target data sources; The hybrid model includes: a three-dimensional convolutional neural network (3DCNN) model to extract the spatial features of hematomas from head CT image sequences; The long short-term memory network (LSTM) model is used to extract the hematoma temporal features of physiological signals; the multimodal fusion module is used to fuse the hematoma spatial features with the hematoma temporal features to output the hematoma fusion features; A prediction module, used to predict the ICP sequence of the target patient within a preset future time period based on the hematoma fusion feature; Adaptive early warning module: used to output visualization results based on the predicted ICP sequence within a preset future period; And through the dynamic threshold adjustment mechanism, the warning threshold of the target patient is adjusted in combination with the case information of the target patient and the predicted ICP value, and the graded warning is triggered in real time.

2. The CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury according to claim 1, wherein the data acquisition module comprises: Case information data collection unit: used to collect case information from the target patient's electronic medical record based on the target patient's unique identifier; Dynamic CT image sequence data acquisition unit: used to continuously acquire head CT at preset time intervals after the target patient is admitted to the hospital to obtain the target patient's original CT image sequence; Physiological signal acquisition unit: used to collect the original physiological signals of the target patient according to the unique identifier of the target patient; Preprocessing unit: used for preprocessing the original CT image sequence and original physiological signal to generate a temporally and spatially consistent and structured CT image sequence and physiological signal of the target patient.

3. According to the CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury of claim 1, the 3DCNN model is specifically used for: By using the time-series slice splicing technology, the original CT image sequence of continuous time points in the CT image sequence is constructed into four-dimensional space-time CT data; The convolutional layer in the 3DCNN model uses underlying convolution kernels of different sizes to perform sliding scans in the voxel space of the 4D spatiotemporal CT data, dividing the 4D spatiotemporal CT data into multiple regions and detecting the local texture spatial features of each region; The fully connected layer in the 3DCNN model uses multi-scale fusion technology to integrate the local texture spatial features of all regions and output the hematoma spatial features of the original CT image sequence of the target patient.

4. According to the CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury of claim 3, the 3DCNN model further includes a spatial attention module after the third convolutional block, which is used to use a dynamic feature clipping strategy to focus on the spatial characteristics of the local texture space of each region, remove the region smaller than the clipping threshold, and retain the hemorrhagic brain tissue region; wherein, The formula for calculating the clipping threshold T based on spatial attention is: ; Formula (1) In formula (1), M represents the attention weight map generated by the spatial attention map, , where H represents height, W represents width, and D represents depth; the spatial attention map is mapped to the attention weight map through 3DCNN , represents the mean of the attention weight map, represents the standard deviation of the attention weight map, Represents an adjustable hyperparameter.

5. The CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury according to claim 1, wherein a multi-layer cascaded LSTM unit is used within the LSTM model, including: Input the physiological signal into a multi-layer LSTM unit. Each LSTM unit includes a cell state, a hidden state, and a gating mechanism. The gating mechanism processes feature input and output and forgetting. The cell state captures the long-term trend characteristics of the original physiological signal, and the hidden state captures the short-term fluctuation characteristics of the original physiological signal. The long-term trend characteristics and short-term fluctuation characteristics output by each LSTM unit constitute the unit time series characteristics output by the LSTM unit. Among them, an attention mechanism is added between the connections of adjacent layers of LSTM units. The attention mechanism adjusts the weights of the preset key time points of the unit temporal features output by each layer of LSTM units; Finally, the hematoma time series characteristics of the target patient's physiological signals are output.

6. According to the CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury of claim 1, the multimodal fusion module uses an adaptive weighting mechanism to perform gated fusion of the hematoma temporal characteristics and the hematoma spatial characteristics. The specific formula is as follows: ； （2） In formula (2), Indicates the characteristics of hematoma fusion, represents the Sigmoid function, represents the hematoma spatial feature weight matrix, represents the hematoma time series feature weight matrix, Indicates the spatial characteristics of hematoma, Indicates the temporal characteristics of hematoma, Represents element-wise product.

7. According to the CT image-based intracranial pressure prediction and warning system for hemorrhagic brain injury according to claim 1, the prediction module in the hybrid model includes a Transformer model, which is used to add a position code to each time node of the hematoma fusion feature, and then map the hematoma fusion feature into an ICP sequence for a preset time period in the future.

8. According to the CT image-based intracranial pressure prediction and early warning system for hemorrhagic brain injury of claim 1, the training of the hybrid model comprises: Collecting a multimodal dataset of historical medical imaging data and historical physiological signal data of multiple patients as a sample dataset, and dividing the sample dataset into a training set and a validation set; Build an initial hybrid model including 3DCNN model, LSTM model, multimodal fusion module, and prediction module; The initial hybrid model is trained using the training set and validation set, and the trend loss and fluctuation loss are combined to form a multi-scale loss function; The model parameters are optimized through a multi-scale loss function. When the model parameters of the initial hybrid model meet the stability requirements, the training is stopped to obtain the hybrid model.

9. The CT image-based intracranial pressure prediction and warning system for hemorrhagic brain injury according to claim 1, wherein the adaptive warning module comprises: The visualization unit is used to determine the risk interval corresponding to each ICP value at each time point in a preset future time period based on four preset risk levels, and color-code it to generate an ICP trend graph containing the color-coded risk intervals. It also combines the ICP sequence and hematoma fusion features and visualizes them as an ICP change curve and a hematoma diffusion probability heat map. Threshold adjustment unit: used to adjust the preset warning threshold based on the case information of the target patient collected by the case information collection subunit to obtain the final warning threshold; Multi-level warning unit: used to divide the risk level into four levels for graded alarm based on the multi-level warning mechanism; when the ICP value obtained based on the hybrid model exceeds the final warning threshold, the corresponding level of warning is triggered according to the graded alarm, and a graded warning signal is generated for the target patient in real time, and clinical treatment recommendations and warning reports containing visual results are generated for the target patient based on the graded warning signal.

Images