Data management method based on CT blood vessel image and anaphylactic reaction

Abstract

The invention relates to the technical field of medical data management, and discloses a data management method based on CT blood vessel images and anaphylaxis. The method comprises the following steps: acquiring CT blood vessel image sequence data of a patient, performing multi-scale feature extraction to obtain a blood vessel structure feature set, and synchronously acquiring allergen detection report data to extract an allergy type and a reaction level; performing space-time alignment matching on the two types of data to generate a blood vessel-allergy associated feature matrix; and constructing a dynamic probability graph model based on the matrix, calculating early warning probability values of allergy-related lesions in different blood vessel sections, and performing three-level layered marking on the blood vessel image according to the early warning probability values. The method also receives a real-time medication record to dynamically adjust model parameters, and triggers a local rescanning instruction of the blood vessel image when detecting that a cross reaction exists between a newly-ingested allergen and a historical allergen; and through differential feature analysis of the re-scanning area, an early warning mark is updated.

Description

Technical Field

[0001] This invention relates to the field of medical data management technology, specifically a data management method based on CT vascular imaging and allergic reactions. Background Technology

[0002] Allergic diseases, as a common abnormal immune system reaction, present with diverse clinical manifestations, traditionally focusing primarily on local symptoms such as those affecting the skin, respiratory tract, and digestive system. However, with in-depth clinical observation and research, the potential impact of systemic allergic reactions on the cardiovascular system has gradually gained attention. Some severe allergic reactions can be accompanied by angioedema, blood pressure fluctuations, and even shock, while long-term, chronic allergic inflammatory states are also considered to potentially participate in pathological processes such as atherosclerosis and vascular endothelial dysfunction. This cross-systemic correlation brings new challenges and opportunities to clinical diagnosis and treatment. Currently, in routine medical practice, allergen detection and cardiovascular imaging assessment are usually handled by different specialties, with data management and analysis being independent. Allergists primarily rely on in vitro specific IgE testing or skin prick tests to determine the patient's sensitization status and allergen type, and record the reaction grade; while radiologists and cardiologists use techniques such as computed tomography angiography to focus on assessing morphological changes such as the degree of vascular stenosis, plaque characteristics, and vessel wall calcification.

[0003] This data isolation has led to a significant cognitive gap: when a patient has both a confirmed allergic constitution and vascular abnormalities detected by CTA, clinicians struggle to determine whether there is an intrinsic pathophysiological link between these two findings, or whether the allergic reaction is a risk factor that induces or exacerbates the vascular lesions. Current diagnostic procedures lack effective tools to correlate dynamically changing immunological information with detailed vascular anatomy imaging, making it difficult to quantify, assess, and provide early warnings of potential "allergy-vascular" association risks.

[0004] A patient's allergic state is not static. New allergens may emerge at any time, cross-reactions may exist between known allergens, and the use of drugs such as antihistamines and glucocorticoids can also modulate the body's immune response. How these dynamic changes affect the state of the vascular system is a blind spot in current medical data management. Routine CTA examinations are usually scheduled based on fixed time intervals or significant clinical symptoms, failing to capture the short-term or immediate impact of these dynamic changes in immune status on blood vessels. Therefore, existing diagnostic and treatment models are lagging and one-sided in assessing the vascular health risks of allergic patients, and cannot meet the clinical needs for early warning, dynamic monitoring, and individualized risk management of allergy-related vascular lesions. There is an urgent need for a data management method that can integrate multi-source heterogeneous medical data, especially breaking down the barriers between allergy immunology information and vascular imaging information, and can adapt to dynamic data changes, thereby achieving quantitative assessment and prediction of associated risks. Summary of the Invention

[0005] The purpose of this invention is to provide a data management method based on CT vascular imaging and allergic reactions to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a data management method based on CT vascular imaging and allergic reactions, the method comprising: The patient's CT vascular image sequence data was acquired, and multi-scale feature extraction was performed on the image data to obtain a vascular structure feature set; Simultaneously collect patient allergen test report data and extract allergen type and allergy reaction level information; Spatiotemporal alignment matching of vascular structure feature set with allergic reaction data is performed to generate vascular-allergy association feature matrix; A dynamic probabilistic graphical model is constructed based on the correlation feature matrix to calculate the early warning probability value of allergy-related lesions in different vascular segments; Based on the early warning probability value, the vascular images are stratified and labeled to generate a three-level early warning labeling result; Receive real-time updated patient medication records and dynamically adjust the calculation parameters for the early warning probability value; When a new allergen cross-reacts with a historical allergen, a local rescan instruction for vascular imaging is triggered. Differential feature analysis is performed on the vascular image regions obtained from rescanning to update the warning labeling results of the corresponding segments; Establish a mapping model between allergen exposure time axis and vascular lesion development; The output includes a prediction report on the development of vascular lesions that varies over time.

[0007] Preferably, the multi-scale feature extraction of the image data includes: The pyramid decomposition algorithm was used to process CT vascular image sequences to separate macroscopic vascular orientation features and microscopic vessel wall texture features. Anisotropic diffusion filtering is applied to the features at each scale to preserve the continuity of blood vessel boundaries; Calculate the topological similarity between features at different scales and construct a cross-scale vascular feature association map; Feature subsets with pathological significance were selected based on the association map.

[0008] Preferably, the spatiotemporal alignment matching includes: Establish time window matching rules between vascular image acquisition timestamps and allergic reaction occurrence times; Spatial registration of vascular features and allergic reaction data within the same time window; A dynamic time warping algorithm is used to compensate for the difference between the frequency of vascular image acquisition and the frequency of allergy detection. Generate a vascular-allergy association feature matrix with time synchronization identifiers.

[0009] Preferably, the construction of the dynamic probabilistic graphical model includes: An initial graph structure is constructed using vascular anatomy regions as nodes and allergen transmission paths as edges; The edge weight update function of the graph neural network is trained based on historical case data; Real-time injection of newly occurring allergic reaction event data triggers dynamic restructuring of the graph structure; Output the probability distribution of anomalies for each blood vessel segment in the current graph structure.

[0010] Preferably, the layer markers include: Establish a three-level early warning classification standard based on probability thresholds; Vascular segments exceeding the first-level threshold are marked with red dynamic outline markers; Vascular segments within the secondary threshold range are marked with a flashing yellow marker; For vascular segments below the warning threshold, the original image display is maintained.

[0011] Preferably, the calculation parameters for dynamically adjusting the early warning probability value include: Identify antihistamine use in medication records; The effective duration of allergen action is recalculated based on the drug's half-life; Correct the time decay coefficient in the vascular-allergy association feature matrix; Update the edge weight calculation formula for the dynamic probabilistic graphical model.

[0012] Preferably, the triggering command for local rescanning of vascular images includes: The protein structure similarity between new allergens and historical allergens is detected. When the similarity exceeds the cross-reactivity threshold, the vascular region corresponding to the previous allergic reaction is located, a priority scanning command containing the coordinates of the target vascular segment is generated, and the CT equipment is controlled to perform a low-dose local rapid scan.

[0013] Preferably, the differential feature analysis includes: Align the vascular spatial coordinate system between the old and new images; Extract the rate of change of vessel wall thickness at the same anatomical location; Calculate the difference in contrast agent filling rate; The changing feature vector is input into the early warning probability update model.

[0014] Preferably, the model for establishing the mapping relationship between the allergen exposure timeline and the development of vascular lesions includes: Allergen exposure events are encoded into discrete time points in chronological order. Morphological change parameters of vascular images before and after each time point are extracted. A temporal convolutional network is trained to predict the trajectory of vascular lesion development and generate a heat map of vascular state evolution containing time markers.

[0015] Preferably, the output includes a vascular lesion development prediction report with a time dimension variation, comprising: By comparing the development trajectory of vascular lesions with standard growth curves, abnormal time intervals that deviate from the normal development rate are marked, and allergen exposure records within the corresponding time intervals are associated to generate a multimodal data association report sorted by time axis.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention provides an innovative tool for exploring and monitoring the risk of allergy-related vascular lesions by deeply integrating CT vascular imaging data with clinical data on allergic reactions. Multi-scale feature extraction from CT vascular images comprehensively quantifies vascular status from macroscopic morphology to microscopic structure. The extracted vascular features are spatiotemporally aligned and matched with specific allergen types and reaction levels to construct a correlation feature matrix. This organically links data from different disciplines, laying the foundation for discovering potential patterns. A dynamic probabilistic graphical model built based on this matrix can quantitatively assess the risk probability of lesions caused by allergic factors in different vascular segments, achieving a leap from qualitative correlation to quantitative risk assessment.

[0017] Based on the probability values ​​of early warning, vascular images are stratified and labeled to generate a three-level early warning result. This risk assessment is presented to physicians in an intuitive and visual manner, assisting them in focusing on high-risk areas. The method incorporates a dynamic update mechanism. By receiving real-time medication records and adjusting model parameters, the risk assessment can be adjusted according to changes in treatment, making it more timely. When a new allergen or cross-reaction occurs, a local rescan instruction for vascular images is triggered, and differential feature analysis is performed on the old and new images. This design allows the system to proactively capture the potential acute impact of specific immune events on vascular structures, achieving proactive monitoring of dynamic risks. Establishing a mapping model between the allergen exposure timeline and the development of vascular lesions explores the correlation between the two from a longitudinal perspective. The output prediction report, which includes changes over time, not only focuses on the current state but also assesses future trends. This method promotes the cross-integration of allergy and cardiovascular imaging, and is expected to improve the understanding and management of allergy-related vascular complications. Attached Figure Description

[0018] Figure 1 A diagram showing the relationship between allergen exposure and vascular lesions; Figure 2 This is a flowchart for multi-scale feature extraction; Figure 3 A flowchart for constructing a dynamic probabilistic graphical model; Figure 4 A weighted heatmap connecting nodes in a dynamic probabilistic graphical model. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides a data management method based on CT vascular imaging and allergic reactions. The method integrates medical imaging and allergic reaction information to achieve early warning and dynamic monitoring of vascular lesions. The method involves acquiring patient CT vascular image sequence data, extracting multi-scale features, and using image processing algorithms to separate vascular features at different scales, including filtering and enhancement operations. Simultaneously, patient allergen detection report data is collected, and allergen type and allergic reaction level information are extracted and quantified. The vascular structure feature set is spatiotemporally aligned and matched with allergic reaction data, considering timestamps and spatial coordinates, to generate a vascular-allergy association feature matrix. Based on this matrix, a dynamic probabilistic graphical model is constructed, using vascular anatomical regions as nodes, and a graph neural network is used to calculate the probability value of allergy-related lesions in each vascular segment. Vascular images are stratified and labeled according to the probability values. Real-time updated patient medication record data is received, and the early warning probability calculation parameters are dynamically adjusted based on pharmacokinetics. When a newly ingested allergen cross-reacts with a historical allergen and the protein sequence similarity exceeds a threshold, a low-dose rescan of a specific vascular region is triggered. Differential feature analysis is performed on the rescanned images, and the early warning labeling results are updated. A model mapping the relationship between allergen exposure time axis and vascular lesion development was established to predict the evolution of vascular status and output a vascular lesion development prediction report with time dimension changes, and the lesion trajectory was displayed with a heat map.

[0021] Example 1: See Figure 2 Multi-scale feature extraction employs a pyramid decomposition algorithm to process CT vascular image sequences. This algorithm is based on a Gaussian pyramid structure, which generates multi-resolution image sequences through layer-by-layer downsampling. The bottom layer represents the original resolution image, with the resolution halved for each subsequent layer. When separating macroscopic vascular orientation features from microscopic vessel wall texture features, macroscopic feature extraction uses a vessel centerline tracing algorithm. This algorithm obtains the vessel centerline path based on distance transformation and skeletonization methods, calculating the path's curvature, twist, and bifurcation angle. Microscopic feature extraction uses gray-level co-occurrence matrix analysis to analyze vessel wall texture, calculating the contrast, correlation, and entropy of texture features. Anisotropic diffusion filtering is applied to features at each scale. This filtering is based on the Perona-Malik model, where the diffusion coefficient is adjusted according to the image gradient magnitude, resulting in strong smoothing in uniform regions and maintaining smoothness intensity in edge regions. Preserving vessel boundary continuity employs an edge enhancement algorithm. This algorithm extracts boundary points using a Canny edge detector and connects discontinuous boundaries using a polynomial curve fitting method.

[0022] The topological similarity between features at different scales is calculated using a persistent cohomology method. This method constructs a simple complex structure of the vascular network, calculates the variation of the Betti number at each scale, and generates persistent barcodes to represent the topological features. A graph theory approach is used to construct a cross-scale vascular feature association map. Graph nodes represent feature points at each scale, and edge weights are calculated based on the Euclidean distance between features, with smaller distances resulting in larger weights. A random forest feature importance assessment method is used to select feature subsets with pathological significance based on the association map. The Gini index decrease value for each feature in the classification task is calculated, and features with an index decrease value greater than a threshold are selected. Spatiotemporal alignment matching includes establishing time window matching rules between vascular image acquisition timestamps and allergic reaction occurrence times. The time window matching rules are set based on the allergic reaction type: a 2-hour time window is used for immediate allergic reactions, and a 24-hour time window is used for delayed allergic reactions. Spatial registration is performed on vascular features and allergic reaction data within the same time window. Spatial registration uses a feature point-based registration algorithm, extracting vascular bifurcation points as feature points, and using an iterative nearest-point algorithm to calculate the rigid body transformation matrix. A dynamic time warping algorithm is employed to compensate for the difference between the frequency of vascular image acquisition and the frequency of allergy detection. The dynamic time warping algorithm finds the optimal curved path by constructing a cost matrix, and the cumulative cost of the path minimizes the distance between time series. A vascular-allergy correlation feature matrix with time synchronization identifier is generated. The matrix structure is a three-dimensional tensor, with the three dimensions representing the vascular segment, time point, and feature type, respectively. Each element stores the feature value and timestamp information.

[0023] In the spatial registration of spatiotemporal alignment matching, the SIFT algorithm is used for feature point extraction to detect scale-invariant features of vascular bifurcation points. The iterative nearest-point algorithm incorporates a RANSAC mechanism to eliminate mismatched point pairs. The cost function for dynamic time warping uses Euclidean distance, and the window size for path constraints is set to 10% of the time series length. The time synchronization identifier for the vascular-allergy association feature matrix uses Unix timestamp format, accurate to the millisecond level. The integration of multi-scale feature extraction and spatiotemporal alignment matching is achieved through a message middleware; the feature extraction module publishes results to a RabbitMQ message queue, and the spatiotemporal alignment module subscribes to and processes the messages. The inter-level scale factor for pyramid decomposition is set to 2, and the gradient threshold for anisotropic diffusion filtering is automatically calculated using the Otsu method. The persistent coherence parameter for topological similarity calculation is simplified to a threshold of 0.5, with a maximum dimension of 2. The edge weights of the association graph are calculated using a Gaussian kernel function, with the kernel width adjusted according to the feature distribution. The time window sliding step size for spatiotemporal alignment matching is set to 50% of the time window length to achieve overlapping window analysis. The coordinate transformation for spatial registration uses an affine transformation model including rotation, translation, and scaling parameters. The path search for dynamic time warping employs dynamic programming, achieving a time complexity of O(nm) for filling the cost matrix. The index construction of the vascular-allergy association feature matrix uses a B+ tree structure to accelerate time-range queries.

[0024] Parallelization of multi-scale feature extraction is achieved using OpenMP multithreaded programming, with different threads allocated to each pyramid layer. GPU acceleration of anisotropic diffusion filtering utilizes CUDA programming, assigning a thread block to each pixel. Distributed computation for topological similarity calculation employs the Apache Spark framework, with persistent cohomology computation distributed across multiple nodes. The association map is stored using the Neo4j graph database, supporting the Cypher query language. Stream processing for spatiotemporal alignment matching uses the Apache Flink framework, with time window operations based on event-temporal semantics. Accelerated spatial registration uses KD-tree nearest neighbor search to reduce feature point matching complexity. The online algorithm for dynamic time warping employs a sliding window approach, maintaining a fixed-size cost matrix. Compression of the vascular-allergy association feature matrix uses the Zstandard algorithm to reduce storage space usage. Quality control for multi-scale feature extraction includes feature integrity checks and outlier detection. The root mean square error (RMSE) is calculated for pyramid decomposition reconstruction errors to ensure information loss remains within acceptable limits. Convergence of anisotropic diffusion filtering is monitored by the rate of change of image entropy; iteration stops when the rate of change is less than 0.1%. The stability of topological similarity calculation is evaluated using the bootstrap method, and confidence intervals are calculated through repeated sampling. Connectivity checks of the association graph are performed using a depth-first search algorithm to ensure that the graph contains no isolated nodes.

[0025] The accuracy of spatiotemporal alignment matching was validated using manually annotated gold-standard data. The accuracy of temporal window matching was evaluated using F1 scores, and the spatial registration error was calculated based on the target registration error. The alignment quality of dynamic temporal warping was measured using a regularized path distance metric, and the completeness of the vascular-allergy association feature matrix was checked for missing data ratios. Parameter optimization for multi-scale feature extraction employed a grid search method, evaluating feature quality under different parameter combinations. The optimal number of pyramid decomposition layers was determined through cross-validation, and the optimal number of iterations for anisotropic diffusion filtering was based on image quality metrics. The optimal persistence threshold for topological similarity calculation was selected based on clustering performance, and the optimal community resolution for the association map was optimized using modularity metrics. Parameter tuning for spatiotemporal alignment matching was conducted using simulated data experiments. The optimal value for the temporal window size was determined using receiver operating characteristic curves, and the optimal transformation model for spatial registration was selected based on the Bayesian information criterion. The optimal constraint window for dynamic temporal warping was determined by minimizing the alignment error, and the optimal dimension of the vascular-allergy association feature matrix was determined using principal component analysis to explain variance.

[0026] The scalable design of multi-scale feature extraction supports the addition of new feature types. The pyramid decomposition algorithm supports non-integer scale factors, and anisotropic diffusion filtering can be extended to 3D image processing. Topological similarity calculation supports higher-dimensional topological features, and the association map can integrate multimodal features. The scalability of spatiotemporal alignment matching supports the access of new data sources. The time window matching rule can be configured with multiple time window types, and the spatial registration algorithm supports non-rigid transformation models. Dynamic time warping can be extended to multivariate time series, and the vascular-allergy association feature matrix supports dynamic dimensional expansion. The system implementation adopts a modular architecture. The multi-scale feature extraction module includes image preprocessing, feature calculation, and feature selection sub-modules. The spatiotemporal alignment matching module includes time processing, spatial registration, and data fusion sub-modules. The modules communicate with each other through a standard interface, exchanging data in JSON format. The pyramid decomposition module takes DICOM format images as input and outputs a multi-scale feature array. The anisotropic diffusion filtering module takes the original image as input and outputs the filtered image. The topological similarity calculation module takes multi-scale features as input and outputs a similarity matrix. The association map construction module takes feature similarity as input and outputs graph structure data.

[0027] The spatiotemporal alignment and matching module takes time-series data as input and outputs time alignment results. The spatial registration module takes vascular features and allergy data as input and outputs registration parameters. The dynamic time warping module takes unequal-length sequences as input and outputs alignment paths. The vascular-allergy association feature matrix generation module takes alignment data as input and outputs a 3D tensor. The system deployment uses Docker containerization technology, with each module running independently in a container. Container orchestration is managed using Kubernetes to achieve elastic scaling and fault recovery. The monitoring system collects performance metrics, including processing latency and resource utilization. The logging system records operation trajectories, supporting fault diagnosis and audit trails. The multi-scale feature extraction module verifies the input image format and resolution, and verifies the output feature dimensions. The spatiotemporal alignment and matching module verifies the input timestamp format and output alignment accuracy. The parameter range check of pyramid decomposition ensures the scale factor is greater than 1, and the parameter check of anisotropic diffusion filtering ensures the time step meets stability conditions. The input verification for topological similarity calculation checks feature dimension consistency, and the construction of the association graph verifies graph connectivity.

[0028] The time window parameter checks for spatiotemporal alignment matching ensure the window length is positive, and the input validation for spatial registration checks the number of feature points. Dynamic time warping parameter checks ensure the sequence is not empty, and the output validation for the vascular-allergy association feature matrix checks the tensor shape. Error handling mechanisms include exception handling and retry logic, with network timeout settings for connection retries. Data backup strategies include full and incremental backups, with backup frequency set according to the data update cycle. Security measures include data transmission encryption and access control, and authentication uses the OAuth 2.0 protocol. Performance optimization includes caching frequently used data and pre-computing intermediate results, with cache invalidation based on timestamps and data versions. Visualization of multi-scale feature extraction results uses the Matplotlib library to generate feature distribution maps, and visualization of spatiotemporal alignment matching uses the Plotly library to generate time series alignment maps. Pyramid decomposition visualization displays images at each layer, and anisotropic diffusion filtering visualization shows a comparison before and after filtering. Topological similarity calculation visualization displays persistent barcodes, and association map visualization uses Gephi software to generate force-directed maps. The time window visualization for spatiotemporal alignment matching displays window partitioning, and spatial registration visualization displays feature point matching. The visualization of the curved path is based on dynamic time regularization, and the heatmap is used to visualize the vascular-allergy association feature matrix.

[0029] System testing includes unit testing, integration testing, and system testing. Unit tests for multi-scale feature extraction cover all function branches, while integration tests for spatiotemporal alignment matching simulate real data flow. Boundary tests for pyramid decomposition handle minima and maxima images, and convergence tests for anisotropic diffusion filtering verify numerical stability. Accuracy tests for topological similarity calculation use simulated complexes, and performance tests for association maps stress-test large-scale graph data. Reliability tests for spatiotemporal alignment matching run over long periods, and accuracy tests for spatial registration use phantom data. Correctness tests for dynamic time warping verify path optimality, and integrity tests for the vascular-allergy association feature matrix check data consistency. System maintenance includes regularly updating dependency library versions and timely application of security patches. Performance monitoring sets threshold alerts, and automatic scaling occurs when resources are insufficient. User support provides detailed documentation and training materials, and a problem feedback mechanism collects improvement suggestions. Version management uses Git to record changes, and continuous integration automates test builds. Deployment scripts automate the installation process, and configuration management unifies parameter settings. Disaster recovery plans define backup and recovery procedures, and business continuity ensures high system availability.

[0030] Example 2: See Figure 3The dynamic probabilistic graphical model is constructed using vascular anatomical regions as nodes. These regions follow the American Heart Association's 17-segment coronary artery classification, with each segment corresponding to a unique coded identifier. Node attributes include geometric features such as segment length, diameter, and wall thickness, as well as functional parameters such as fractional flow reserve. The node feature vector is 50-dimensional, with the first 20 dimensions storing morphological features extracted from CT images, the middle 20 dimensions storing hemodynamic parameters, and the last 10 dimensions storing allergen binding affinity data. An initial graph structure is constructed using allergen propagation paths as edges, with edge directions strictly following the blood flow direction. Initial edge weights are calculated based on fluid dynamics formulas, considering pressure gradients and flow resistance between vascular segments. Edge attributes include propagation time delay parameters and allergen attenuation coefficients. The initial graph structure contains approximately 200 nodes and 500 directed edges, stored using an adjacency list data structure.

[0031] The edge weight update function of the graph neural network was trained using historical case data. The historical case database contained 3000 complete clinical records, each including 12 consecutive months of CT angiography sequences, weekly allergen testing reports, and lesion localization data verified by intravascular ultrasound. The graph neural network adopted a gated graph attention network architecture, with input layer nodes having a feature dimension of 50, processed through three layers of graph attention convolutional layers, each with 8 attention heads, and a hidden layer dimension of 128. The edge weight update function calculated the influence weights between nodes through a multi-head attention mechanism, with the weight values ​​dynamically adjusted based on feature similarity and spatiotemporal proximity. The training process used the Adam optimizer, with an initial learning rate of 0.001, adjusted using a cosine annealing strategy, and a batch size of 16 patient data. The loss function used was the focal loss function, with the focal parameter γ set to 2, and the class weights were inversely weighted based on the lesion incidence rate. The training cycle was 500 epochs, with model performance evaluated on the validation set after each epoch. An early stopping mechanism was triggered after 20 consecutive epochs without improvement.

[0032] Real-time injection of newly occurring allergic reaction event data triggers dynamic reorganization of the graph structure. Event data is received via the HL7 message interface, and parsed to obtain allergen IgE concentration, exposure time, and affected anatomical regions. During the dynamic graph structure reorganization process, temporary event nodes are added, and weighted connections are established with relevant vascular nodes. Edge weights are calculated based on allergen concentration and vascular segment sensitivity coefficients. The reorganized graph structure undergoes real-time inference, employing subgraph sampling technology to update only the node states of affected regions. The abnormal probability distribution of each vascular segment in the current graph structure is output. Probability calculations are normalized using a softmax function, generating 17-dimensional probability vectors corresponding to each coronary artery segment. The probability distribution is stored in a time-series database, supporting sliding window analysis and trend prediction. The probability value update frequency is synchronized with the CT image acquisition cycle, with a minimum time interval of 1 hour. Hierarchical labeling includes setting a three-level warning classification standard based on probability thresholds. The thresholds are determined based on baseline data statistics from thousands of healthy individuals. The first-level threshold corresponds to a probability value of 0.85 (95th percentile), and the second-level threshold corresponds to a probability value of 0.70 (80th percentile). The threshold parameter can be dynamically adjusted through the management interface, with an adjustment step of 0.01.

[0033] For vascular segments exceeding the first-level threshold, a red dynamic contour marker is used. The contour is generated automatically using an active contour model to delineate the vessel boundaries, with a contour line width of 4 pixels and a standard red color (RGB255,0,0). The dynamic effect is achieved by linearly interpolating the contour transparency, which cycles between 0.6 and 1.0 at a frequency of 3Hz. The fusion of the contour markers and DICOM images uses alpha blending technology to preserve the details of the underlying image. For vascular segments within the second-level threshold range, a yellow flashing marker is used. The marked area is determined using a region growing algorithm, filled with standard yellow (RGB255,255,0), and the base transparency is set to 0.4. The flashing effect is controlled by a square wave signal with a frequency of 2Hz and a duty cycle of 60%. The marked area supports mouse hover interaction, displaying the probability value and confidence interval when hovering. For vascular segments below the warning threshold, the original image display is maintained, with display parameters following the DICOM grayscale display standard, and window width and window level automatically optimized according to tissue characteristics. Image rendering uses multi-texture binding technology to ensure a real-time display frame rate of no less than 30fps.

[0034] The integration of the dynamic probabilistic graphical model and hierarchical labeling is achieved through a medical image browser developed based on the VTK framework, supporting 2D multi-planar reconstruction and 3D volumetric rendering. Spatial registration of probabilistic data and image data uses an affine transformation matrix, with a registration error of less than 0.5 mm. The graph structure visualization module employs a force-guided layout algorithm, arranging node positions according to vascular anatomy. Training data preprocessing for the graph neural network includes outlier detection and feature normalization; continuous features are normalized using Z-scores, and categorical features are encoded using one-hot encoding. Patient-level separation is used to divide the training and test sets, ensuring that data from the same patient appears in only one set. Model evaluation uses five-fold cross-validation, reporting precision, recall, and F1 score. The real-time inference engine is optimized using TensorRT, with inference latency controlled within 100 milliseconds. Probabilistic data is stored in the InfluxDB time-series database, supporting retrospective analysis using SQL-like query language. Label information management uses Redis caching to accelerate the reading of frequently accessed data. Feature updates for vascular anatomy partition nodes use an exponentially weighted moving average algorithm, with recent feature changes having higher weights. The edge weight update of the allergen transmission path incorporates a time decay factor, with the decay coefficient set according to the allergen's half-life. The computational complexity of dynamic graph restructuring is controlled by an incremental update algorithm, updating only the affected subgraphs.

[0035] The layered marking color management system supports ICC color configuration, ensuring consistency across different display devices. Marker transparency blending employs a pre-multiplied alpha synthesis method to avoid color distortion. Contour anti-aliasing utilizes multi-sampling technology with a sampling multiplier of 4x. Marker depth testing is synchronized with the image depth buffer, and the depth comparison function is set to LEQUAL. The user interface provides a real-time probability threshold adjustment slider; changes in the slider value immediately trigger marker updates. Display presets support various clinical scenario configurations, such as emergency mode, screening mode, and follow-up mode. User configuration information is stored in a locally encrypted file and automatically synchronized to the cloud. Abnormal probability distribution calibration uses a temperature scaling method, with temperature parameters learned through maximum likelihood estimation. The calibration model is updated monthly using the latest accumulated clinical data. Probability distribution visualization provides a contour display mode with contour line spacing of 0.1 probability units.

[0036] Interpretive analysis of the graph neural network employs the integral gradient method to calculate the contribution of each feature to the final probability. The interpretation results are displayed as a superimposed heatmap, with important feature regions highlighted. The feature contribution ranking algorithm supports multi-dimensional analysis to identify key risk factors. System performance monitoring uses Prometheus to collect metrics, including inference latency, memory usage, and GPU utilization. Anomaly detection rules are based on the 3σ principle, triggering an alarm when three consecutive time points exceed the specified range. Operation logs are recorded in a structured format for easy ELK stack analysis. The dynamic splitting and merging algorithm for vascular anatomy partition nodes is based on region growth results, with a splitting threshold set to 10 mm of vessel length. Node merging conditions are based on feature similarity; adjacent nodes with a cosine similarity greater than 0.9 are automatically merged. Partition granularity adjustment adaptively changes according to image resolution, supporting sub-millimeter level partitioning. Edge weight learning for allergen propagation paths introduces physical constraints, forcing edge weights that violate blood flow direction to zero. Path validation is based on vascular connectivity detection, using a depth-first search algorithm to verify path validity. Edge weight normalization employs L2 norm constraints to prevent gradient explosion. The rendering optimization for layered markers employs an instantiation drawing technique, allowing batch processing of markers of the same type. Outline wireframe generation utilizes a geometry shader program, dynamically adjusting the level of detail within the wireframe. Marker state management employs a finite state machine model, with state transitions based on probability-based events.

[0037] User access management is based on the RBAC model, defining three roles: physician, technician, and administrator. Data access auditing records all query operations, with audit logs retained for 10 years. System authentication uses two-factor authentication, and passwords are mandatory to be changed every 90 days. Incremental learning of the dynamic probabilistic graphical model employs an elastic weight consolidation algorithm, with importance weights calculated based on the Fisher information matrix. Model version management uses semantic version numbers, with major updates incrementing the major version number. Model performance is compared using an A / B testing framework, with new models validated in shadow mode. System integration test cases cover 17 typical allergy scenarios, including pollen allergies, drug allergies, and food allergies. Compatibility testing covers image data formats from the three major CT equipment manufacturers. Stress testing simulates 100 allergy events per second to verify system stability. Technical documentation includes system architecture diagrams, data flow diagrams, and interface description diagrams. API documentation uses the OpenAPI specification and provides online debugging tools. The user manual includes operation videos and troubleshooting guides, supporting multiple languages. The maintenance plan specifies daily full backups and hourly incremental backups, with backup data stored off-site. Model training is performed automatically monthly, with training reports sent to the administrator's email address. Software updates utilize a blue-green deployment strategy to ensure uninterrupted service. Security controls include TLS encryption for data transmission and AES-256 encryption for static data. Access control is based on an attribute-based encryption scheme, providing fine-grained control over data access. Security audits are performed quarterly, and vulnerability scans run automatically weekly.

[0038] See Figure 4 This figure visualizes the connection weights between nodes in a dynamic probabilistic graphical model constructed from CT angiography and allergic reaction data. The figure shows the connection weight values ​​between nodes in different vascular segments, with colors ranging from light yellow to dark red corresponding to a gradient from 0 to close to 1. The color bars on the right clearly indicate the numerical range of the connection weights. The value of each cell represents the strength of the connection weight between the corresponding two vascular nodes. A higher weight indicates a greater probability that the allergen will affect the corresponding vascular segment through this path, indicating a stronger allergen transmission association between these node pairs. This is a key association path for assessing the risk of allergy-related vascular lesions. This heatmap intuitively presents the topology and weight distribution of the dynamic probabilistic graphical model, providing core data support for subsequent calculation of the early warning probability of allergy-related lesions in vascular segments and the generation of hierarchical early warning markers. It helps clinicians accurately identify high-risk paths associated with allergies and blood vessels, demonstrating the technical value of multi-source medical data fusion in the quantitative analysis of pathological associations.

[0039] Example 3: The calculation parameters for dynamically adjusting the early warning probability value include identifying the use of antihistamines in medication records. These records are derived from the hospital information system's drug treatment record table, which includes fields for patient ID, drug code, dosage, administration timestamp, and route of administration. Antihistamine identification utilizes the Anatomical Therapeutic Chemochemistry Classification (ATC) system from the World Health Organization's Drug Statistics Methods Integration Database, specifically matching antihistamines coded as R06A series, including specific drug components such as astemizole, acetaminophen, and atazatidine.

[0040] The time series of medication records was aligned with the timeline of allergic events using a timestamp-accurate matching algorithm, with a maximum allowed time deviation of 5 minutes. Drug usage frequency statistics employed a time window counting method, with the window size scientifically determined based on the drug's half-life. Short-acting antihistamines such as chlorpheniramine used a 4-hour statistical window, while long-acting antihistamines such as desloratadine used a 24-hour statistical window. Drug dosage standardization adopted the daily dose unit defined by the World Health Organization. Differences in bioavailability across different routes of administration were adjusted using conversion factors: the bioavailability factor for oral administration was set at 1.0, for intramuscular injection at 1.1, and for intravenous injection at 1.2.

[0041] The effective duration of action of the allergen was recalculated based on the drug's half-life. The drug half-life data was derived from the elimination half-life parameters published in the drug's package insert. A professional database of drug half-lives containing the half-life values ​​of 200 antihistamines was established. The effective duration of action was calculated using a one-compartment pharmacokinetic model, expressed by the following formula:

[0042] in: The plasma concentration of the drug at time t (unit: mg / L) Indicates the initial drug concentration (unit: mg / L). Indicates the dosage (unit: mg). Indicates bioavailability (unitless). Represents apparent volume distribution (unit: L). Represents the elimination rate constant (unit: h) -1 ), Represents the absorption rate constant (unit: h). -1 ), The term represents the exponential decay of the drug elimination process. The term represents the exponential decay of drug absorption. The effective duration of action is defined as the time point at which the drug concentration drops below the minimum effective concentration, which is determined based on clinical pharmacology trial data.

[0043] The time decay coefficient in the vascular-allergy association feature matrix is ​​corrected. The initial value of the time decay coefficient is scientifically set based on the type of allergic reaction: 0.8 per hour for immediate-type allergic reactions and 0.2 per hour for delayed-type allergic reactions. The correction method introduces a drug influence factor, calculated based on the dose-response curve of drug plasma concentration and receptor occupancy, fitted using the Hill equation. The corrected time decay coefficient is negatively correlated with drug plasma concentration; the higher the concentration, the smaller the time decay coefficient, effectively prolonging the duration of allergen action. The vascular-allergy association feature matrix is ​​updated using an element-level correction algorithm. The algorithm traverses each element in the matrix, retrieves the corresponding drug usage record, and calculates the new time decay coefficient value. The matrix update frequency is synchronized with the medication record update, with a minimum update interval set to 15 minutes to ensure real-time reflection of the latest medication information. Row-locking is used during the update process to ensure data consistency and avoid concurrent write conflicts. The edge weight calculation formula of the dynamic probabilistic graphical model is updated. The original edge weight calculation formula is based on the anatomical distance between vascular segments and hemodynamic parameters. The revised formula introduces a drug regulation term, which takes the form of an sigmoid function. The input to the function is the ratio of drug concentration to standard therapeutic concentration. The recalculated edge weight formula consists of three components: basic anatomical weight, hemodynamic weight, and drug regulation weight. These three weights are summed to obtain the final edge weight value, and the weight coefficients are assigned based on clinical importance.

[0044] The calculation of drug regulation weights requires individualized pharmacokinetic parameters for each patient, including individual differences such as liver metabolic enzyme activity and renal clearance. These individualized parameters are estimated using a population pharmacokinetic model, with patient age, weight, and liver function indicators as covariates as inputs. Edge weight updates are triggered by three conditions: new medication record entry, arrival of the dosing time, and significant changes in drug concentration. The system monitors these conditions in real time. The dynamic probabilistic graphical model uses incremental learning for parameter updates, updating only the model parameters corresponding to the affected vascular segment at a time. The incremental learning algorithm employs stochastic gradient descent, with the learning rate adaptively adjusted based on the parameter update frequency; frequently updated parameters use a smaller learning rate to ensure stability. Model parameter version management uses snapshot technology to save historical parameter states for rollback and comparative analysis. The system implementation adopts a microservice architecture. The medication identification service is deployed independently, receiving medication records via a RESTful interface. The pharmacokinetic calculation service is deployed as a Docker container, with container resource configuration dynamically adjusted based on computational complexity. The message queue uses Kafka to process real-time data streams, ensuring temporal consistency between medication records and allergic events. Data quality checks for medication records include format verification, range checks, and logical checks. Format verification ensures timestamps conform to ISO 8601 standards, range checks verify that dosage values ​​are within reasonable ranges, and logical checks identify issues such as inconsistent dosing times. The data anomaly handling mechanism includes three strategies: automatic correction, manual review, and data discarding, with the appropriate strategy selected based on the severity of the anomaly.

[0045] A version control mechanism is established for updating and maintaining drug half-life data, ensuring timely synchronization of the latest data when new drugs are launched or drug instructions are updated. Data sources are prioritized based on the authority of the data, including drug regulatory agency approval documents, clinical trial reports, and academic literature. Uncertainty in half-life values ​​is represented by confidence intervals, calculated based on sample size and the coefficient of variation. The vascular-allergy association feature matrix is ​​stored using a sparse matrix compression format to reduce storage space. The matrix access interface supports both time-range and vascular segment queries, with query response times optimized to milliseconds. Matrix version management retains historical version data, supporting retrospective analysis of parameter adjustment effects. The edge weight calculation service for the dynamic probabilistic graphical model is deployed on a GPU server, utilizing parallel computing to accelerate matrix operations. Numerical stability checks are incorporated into the weight calculation process to prevent division by zero anomalies and numerical overflow. Weights are constrained to a range of 0 to 1, with out-of-range weights automatically pruned to boundary values. System performance monitoring includes metrics such as drug identification accuracy, half-life query response time, and edge weight calculation latency. Monitoring data is displayed in real-time on the operations dashboard, triggering alarm notifications for abnormal situations. The system log records the complete parameter tuning pipeline, including input data, calculation process, and output results.

[0046] The user interface provides a visual display of parameter adjustments, showing the correlation between drug concentration change curves and time decay coefficients. Doctors can manually adjust parameter weights, and manual adjustment records are saved in the operation log. Parameter adjustment history generates trend charts to help analyze the adjustment effects. A regular synchronization mechanism is established for updating and maintaining the drug dictionary, with weekly incremental synchronization with authoritative drug databases. New drug additions require review by pharmaceutical experts to ensure classification accuracy. Drug alias management establishes a synonym mapping table to improve recognition recall. Pharmacokinetic model validation uses cross-validation, dividing patient data into training and testing sets. Model prediction accuracy is evaluated using the mean absolute percentage error (MAE) metric; models with errors exceeding 20% ​​require retraining. The model is updated every six months, incorporating new clinical research data. System integration testing includes drug interaction scenarios to test the effect of parameter adjustments when multiple drugs are used in combination. Performance testing simulates medication record input during peak hours to verify the system's processing capacity. Security testing focuses on checking medication data privacy protection measures to prevent information leakage.

[0047] The documentation includes pharmacokinetic parameter tables, a medication identification rule base, and explanations of parameter adjustment algorithms. The interface documentation details data formats and error codes, and provides call examples. User training materials include operation guidelines for typical medication scenarios to help doctors understand parameter adjustment logic. The system maintenance plan includes regular updates to the drug dictionary, retraining of the pharmacokinetic model, and expansion of the computing service. The backup strategy provides dual backups of medication records and adjustment parameters, with encrypted storage of backup data. The disaster recovery plan ensures rapid switchover to a backup node in the event of system failure. Security controls include encrypted transmission of medication data, hierarchical access control, and operational audit trails. Data encryption uses national cryptographic algorithms, and the key is changed regularly. Access control is role-based, with different roles having different parameter adjustment permissions. Audit logs record all parameter modification operations, meeting medical regulatory requirements.

[0048] Example 4: The triggering of a local rescan command for vascular imaging includes detecting the protein structural similarity between new and historical allergens. Protein structural similarity calculation uses the Smith-Waterman local sequence alignment algorithm. Algorithm parameters are set as follows: a matching score of +2 points, a mismatch penalty of -1 point, a gap opening penalty of -2 points, and a gap extension penalty of -0.5 points. Matrix filling uses a dynamic programming method, and the backtracking path selects the path with the highest score. The protein sequences of new allergens are obtained from the Allergen Database of the International Federation of Immunological Societies (IFIS), version 2023.1, containing 987 validated allergen protein sequences. Historical allergen sequences are stored in a local patient profile database managed by a MySQL relational database management system. When the similarity exceeds the cross-reactivity threshold, the system triggers a rescan mechanism. The cross-reactivity threshold is set based on allergen family classification, and the threshold varies for different allergen families. See Table 1 for the cross-reactivity threshold parameter settings for major allergen families. Table 1: Allergen Cross-Reactivity Threshold Parameters

[0049] Locating the vascular region corresponding to past allergic reactions is achieved by querying a historical warning database, which records the coordinates of the vascular segment corresponding to each allergic reaction. The vascular segment coordinates are represented in a three-dimensional spatial coordinate system, with the origin at the center of the sternal angle, the X-axis pointing to the patient's right side, the Y-axis pointing towards the foot, and the Z-axis pointing towards the back. The coordinate accuracy reaches 0.1 mm, ensuring accurate location of the target vascular region. Coordinate transformation uses a rigid body transformation matrix, and the matrix parameters are calibrated using a calibration phantom. A priority scan command containing the coordinates of the target vascular segment is generated, and the command format follows the DICOM supplemental acquisition protocol standard. The command content includes fields such as patient identifier, target vascular center point coordinates, scan radius, and suggested scan parameters. The scan radius is dynamically calculated based on the vascular diameter, using a formula of three times the average vascular diameter to ensure complete coverage of the target area. Command transmission uses the DICOMMWL protocol and is sent to the CT equipment via the DIMSE message service.

[0050] The CT scanner was controlled to perform low-dose local rapid scanning. Scanning parameters were set as follows: tube voltage 80kV, tube current 50mA, slice thickness 0.5mm, pitch 1.5, and an iterative reconstruction algorithm was used. The low-dose scanning scheme was optimized based on noise simulation, controlling the radiation dose to below 30% of conventional scanning while ensuring image quality. The rapid scanning mode employed partial scan reconstruction technology, reducing scan time to 40% of the standard protocol, with a reconstruction matrix size of 512×512 pixels. Differential feature analysis included aligning the vascular spatial coordinate system of the old and new images. The alignment process used a feature point-based registration algorithm. Vessel bifurcation points were selected as natural markers, and their coordinates were extracted using a 3D SIFT feature detector. The feature descriptor dimension was 128. The registration transformation matrix was calculated using singular value decomposition to solve for the rotation matrix and translation vector, achieving spatial alignment of the old and new images with a registration accuracy requirement of less than 0.5mm. Vessel wall thickness change rate data at the same anatomical location were extracted, and the vessel wall thickness was measured using a fully automated boundary detection algorithm. The algorithm is based on an active contour model. The initial contour is set at the vascular intima boundary, and contour evolution is driven by a gradient vector flow field. The energy function includes internal and external energy terms. The thickness change rate is calculated using a relative percentage formula, which is the ratio of the difference between the newly measured thickness and the baseline thickness to the baseline thickness. The result is rounded to two decimal places. The difference in contrast agent filling rate is calculated, and the filling rate is obtained through time-density curve analysis. The curve acquisition points are set at three measurement time points: the arterial phase, the venous phase, and the delayed phase. Density values ​​are measured in Heinz units, and the sampling frequency is 1 frame per second. The difference in filling rate is calculated using the area under the curve comparison method, comparing the difference in area under the curve between the new and old scans. The integration method uses the compound Simpson's rule.

[0051] The change feature vector is input into the early warning probability update model. This vector includes 10 feature dimensions, such as the rate of change in vessel wall thickness, the difference in filling speed, and the change in vascular elasticity coefficient. The early warning probability update model uses a random forest classifier with 100 decision trees and a maximum tree depth of 10 layers. The Gini impurity criterion is used for node splitting. The model output is the updated early warning probability value, ranging from 0 to 1, representing the risk level of allergy-related lesions in the vascular segment. The triggering logic for local rescanning commands is based on a real-time monitoring mechanism, with the system continuously listening for new allergen detection events. Event processing uses an asynchronous message queue model, implemented using RabbitMQ. Messages are persistently stored to prevent data loss, and the queue capacity is set to 10,000 messages. The consumer thread pool size is dynamically adjusted according to system load, with a maximum concurrency of 50 threads. The protein structure similarity calculation service is deployed on a high-performance computing cluster with 128 CPU cores, 256GB of memory, and an MPI parallel computing framework. Similarity calculation tasks are processed in a distributed manner, with each computing node handling one sequence alignment task, and the results aggregated to the master node. The computational speedup is linear, processing 1000 sequence alignments in less than 5 minutes. The dynamic adjustment mechanism for the cross-reactivity threshold is based on clinical feedback data; the system records the consistency between each rescan result and the final diagnosis. When the prediction accuracy for a particular allergen family consistently falls below 85%, a threshold recalibration process is automatically triggered. The calibration algorithm employs a logistic regression model, adjusting threshold parameters based on the true positive rate and false positive rate. The model training data comes from clinical records accumulated over the past 365 days.

[0052] The accuracy of vascular region localization was verified using a phantom experiment method, employing a vascular biomimetic phantom for repeated scanning tests. The phantom material used was tissue-equivalent, achieving a vascular structure accuracy of 0.1 mm. The localization error was calculated as the Euclidean distance between the actual and theoretical coordinates; the system required a localization error of less than 1 mm, triggering a calibration process when the error exceeded the threshold. The scan command generation service was integrated into the hospital's radiology information system. Command transmission used the HL7 message protocol, with message type ORM_O01. The message structure included fields such as basic patient information, scan site description, and suggested scan parameters, with field validation using HL7 standard validation rules. Commands were electronically signed by a radiologist before being sent to ensure medical safety. The CT equipment control interface was implemented based on the DICOMModalityWorklist protocol. Worklist items included attributes such as basic patient information, examination instance UID, and planned scan time. Equipment status monitoring was achieved through DICOM status messages, tracking scan progress in real time. An automatic retry mechanism was triggered in case of abnormalities, with a maximum of 3 retries.

[0053] Spatial registration accuracy for differential feature analysis was assessed using reprojection error, calculated using the least squares method. The registration algorithm employed a multi-resolution strategy with a three-layer pyramid, and the bottom layer image was downsampled to 1 / 8 of the original image. The registration results were visualized as a registration error distribution map, with error colors ranging from blue to red indicating increasing error. The tube wall thickness measurement algorithm was certified by the Medical Image Analysis Association, achieving a measurement accuracy of 0.01 mm. The statistical significance of the thickness change rate was tested using a t-test with a significance level of 0.05 and a power requirement of 0.9. Trend analysis employed an autoregressive integral moving average model, with model parameters determined through maximum likelihood estimation. Temporal registration for filling velocity analysis was based on ECG gating technology, with an ECG signal sampling frequency of 1000 Hz. Density calibration was performed periodically using a standard phantom containing five contrast agent samples of different concentrations. The area under the curve was calculated using the trapezoidal rule, with the integration interval covering the entire enhancement cycle and an integration step size of 0.1 seconds. The training data for the early warning probability update model comes from a multi-center clinical trial, with the dataset containing 5,000 validation cases. Model performance is evaluated using 10-fold cross-validation, with evaluation metrics including accuracy, recall, and F1 score. The model is updated quarterly, and new models must be validated through regression testing before deployment. System integration testing covers CT equipment from different vendors, with test cases including 200 normal scenarios and 50 abnormal scenarios. Performance testing uses a load generation tool to simulate peak load, requiring a system response time of less than 2 seconds and a throughput of 100 requests per second. Security testing employs penetration testing methods, with vulnerability scanning covering the OWASP Top Ten Security Risks.

[0054] The user interface provides a visual analysis of rescanning decisions, employing a three-column layout. The left side displays protein alignment results and cross-reaction risk assessment, the middle shows a vascular region location map, and the right side displays historical warning records. Interactive operations support drag-and-drop, zooming, and filtering, with a view update response time of less than 100 milliseconds. A quality control system establishes a regular audit mechanism, with audits conducted monthly. Audit content includes assessments of the rationality and effectiveness of rescanning instructions, and quality indicators include parameters such as rescanning positivity rate, radiation dose control level, and diagnostic accuracy rate. A continuous improvement mechanism optimizes system parameters and algorithms based on quality audit results, with improvements recorded in quality reports. The documentation system includes a system operation manual, technical white papers, and troubleshooting guides, using Markdown format. Training materials include video tutorials and interactive demonstrations, with a video resolution of 1080p and a frame rate of 30fps. Technical support establishes a multi-level response mechanism, with first-line support response time of less than 15 minutes and second-line support response time of less than 2 hours. The system maintenance plan includes daily monitoring, regular backups, and disaster recovery drills. Monitoring metrics include system availability, data integrity, and performance metrics, with monitoring data retained for 365 days. The backup strategy combines full and incremental backups, with full backups performed weekly and incremental backups performed daily. The disaster recovery plan specifies a recovery time target of 4 hours and a service recovery point target of 1 hour.

[0055] Example 5: The Vascular Remodeling Trend Prediction Report comprises three core components: multi-timepoint data alignment, predictive model building, and visualization report generation. Multi-timepoint data alignment integrates all vascular imaging data of the patient over the past 36 months. This data originates from DICOM format files stored in the hospital's image archiving system, containing scan images from Siemens SOMATOM Force, Philips Brilliance iCT, and GE Revolution CT devices. Each timepoint image data includes three-phase enhanced scans of the arterial, venous, and delayed phases, with slice thicknesses of 0.5 mm, 0.625 mm, and 1.0 mm, and pixel pitch values ​​between 0.3 mm and 0.5 mm. In the data preprocessing stage, anisotropic diffusion filtering is used for image denoising. The filter's transfer function is designed based on the Perona-Malik model, employing a larger diffusion coefficient in uniform regions with small image gradients to achieve strong smoothing, and a smaller diffusion coefficient in vascular edge regions with large image gradients to preserve detail. Spatial registration employs a B-spline-based free deformation transformation model, with the grid node spacing set to 10 mm. The registration algorithm maximizes mutual information to achieve accurate alignment of images from different time points, and the mutual information is calculated using the Parzen window density estimation method. Intensity standardization utilizes histogram matching technology, using the intensity distribution of the image from the initial examination as a reference standard. The matching process employs a histogram specification algorithm to eliminate intensity differences caused by different scanning parameters.

[0056] Temporal resampling transforms irregularly acquired time series into equally spaced sequences, with a fixed sampling interval of ninety days. Missing time points are filled using cubic spline interpolation, with interpolation nodes selected from known time point measurements. Vessel segmentation employs a level set method based on region growing. Initial seed points are automatically set at the center of the vessel lumen. The evolution rate function combines image gradient information and prior knowledge of vessel shape, iterating 500 times or until the contour change is less than 0.01 mm. Segmentation results are independently verified by two radiologists, with an accuracy requirement of over 95%. Unsatisfactory segmentation results are returned for reprocessing. Establishing a temporal prediction model for vessel morphology parameters requires extracting quantitative features for each time point. These features include 28 morphological indicators such as vessel lumen area, wall thickness, vessel tortuosity, and bifurcation angle. The feature extraction algorithm is based on distance transformation and skeletonization methods. First, a thinning algorithm is used to calculate the vessel centerline. Then, a measurement cross-section is set every 0.5 mm along the centerline, and morphological parameters are calculated on each cross-section. Outlier detection employed the Isolation Forest algorithm, with a contamination parameter set to 0.1 to remove outlier data points caused by measurement errors. Time series forecasting used a seasonal autoregressive integral moving average model, with model parameters determined using the Akaike information criterion. The optimal model structure was ARIMA(2,1,1)(1,0,1)12. Model training employed a rolling forecasting method, using data from the past thirty-six months to predict the trend for the next twelve months. The prediction interval was calculated using the Bootstrap method, with 1000 repeated samplings to obtain a 95% confidence interval. Residual analysis employed the Ljung-Box test with a lag order of 12 to ensure the residual series was free of autocorrelation.

[0057] Predicting trends in vascular function parameters requires analyzing contrast agent filling kinetics, including hemodynamic parameters such as peak enhancement time, area under the curve, and mean transit time. Filling curve fitting employs a gamma-variable function, with parameters estimated using the Levenberg-Marquardt optimization algorithm at 1000 iterations. Functional parameter prediction utilizes a vector autoregressive model, considering the interactions between different parameters. Model stability is validated using the Augmented Dickey-Fuller unit root test, and unstable time series are stabilized through first-order differencing. Visualized reports are generated using the D3.js JavaScript library to create interactive charts. The report layout employs a responsive design, using CSS media queries to adapt to various screen sizes from desktops to tablets. Trend charts display the changes in morphological and functional parameters over the past thirty-six months and the predicted range for the next twelve months. Curve colors are color-coded: blue for historical data and red for predicted data. Confidence intervals are displayed using semi-transparent bands with a transparency of 30%. The report's interactive functionality supports time window zooming; users can drag a slider to select the time range of interest, zooming from one month to thirty-six months. Key time points are marked to indicate important clinical events, including the onset time of allergic reactions, the start time of drug treatment, and the time of dose adjustment. The marked information is displayed in a floating window, including the event type, specific date, and clinical notes. A vessel region selector allows users to focus on specific anatomical sites, such as the descending coronary arteries, the left anterior descending coronary artery, or the circumflex coronary artery. Risk level assessment is based on the deviation of predicted values ​​from baseline values; a deviation exceeding 20% ​​is marked as a yellow alert, and a deviation exceeding 40% is marked as a red alert. The alert rule considers the rate of parameter change; even if the current value is within the normal range, a rate of change exceeding two standard deviations will trigger an alert. The risk assessment model integrates morphological and functional parameters, using logistic regression to calculate a comprehensive risk score, ranging from 0 to 100 points.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data management method based on CT angiography and allergic reaction, characterized in that, The method comprises the following steps: Obtain CT angiography sequence data of a patient, and perform multi-scale feature extraction on the image data to obtain a blood vessel structure feature set; Synchronously collect patient allergen detection report data, and extract allergen type and allergic reaction level information; Align and match the blood vessel structure feature set and the allergic reaction data in space-time to generate a blood vessel-allergy association feature matrix; Based on the association feature matrix, a dynamic probabilistic graph model is constructed to calculate the early warning probability value of different blood vessel segments for allergic related lesions; According to the early warning probability value, the blood vessel image is marked in layers to generate a three-level early warning marking result; Receive real-time updated patient medication record data, and dynamically adjust the calculation parameters of the early warning probability value; When there is a cross reaction between the newly ingested allergen and the historical allergen, a local re-scanning instruction of the blood vessel image is triggered; Perform differential feature analysis on the blood vessel image region obtained by re-scanning, and update the early warning marking result of the corresponding segment; Establish a mapping relationship model between the allergen exposure time axis and the development of blood vessel lesions; Output a blood vessel lesion development prediction report containing time dimension changes.

2. The data management method based on CT angiography and allergic reaction according to claim 1, characterized in that, The multi-scale feature extraction on the image data comprises: Pyramid decomposition algorithm is used to process the CT angiography sequence to separate the macroscopic blood vessel direction features and the microscopic tube wall texture features; Anisotropic diffusion filtering is performed on the features at each scale to preserve the continuity of the blood vessel boundary; The topological similarity between different scale features is calculated to construct a cross-scale blood vessel feature association graph; According to the association graph, a feature subset with pathological significance is screened.

3. The data management method based on CT angiography and allergic reaction according to claim 1, characterized in that, The space-time alignment and matching comprises: Establish a time window matching rule between the blood vessel image acquisition timestamp and the allergic reaction occurrence time; Spatially register the blood vessel features and the allergic reaction data within the same time window; Use dynamic time warping algorithm to compensate for the difference between the blood vessel image acquisition frequency and the allergic detection frequency; Generate a blood vessel-allergy association feature matrix with time synchronization identification.

4. The data management method based on CT angiography and allergic reaction according to claim 1, characterized in that, The construction of the dynamic probabilistic graph model comprises: An initial graph structure is constructed with blood vessel anatomical partitions as nodes and allergen transmission paths as edges; The edge weight update function of the graph neural network is trained according to historical case data; Real-time injection of newly occurring allergic reaction event data triggers dynamic reorganization of the graph structure; Output the abnormal probability distribution of each blood vessel segment in the current graph structure.

5. The data management method based on CT angiography and allergic reaction according to claim 4, characterized in that, The hierarchical marking comprises: Set a three-level early warning division standard based on the probability threshold; Mark the blood vessel segments exceeding the first threshold with a red dynamic contour marker; Mark the blood vessel segments within the second threshold interval with a yellow flashing marker; Maintain the original image display for the blood vessel segments below the early warning threshold.

6. The data management method based on CT angiography and allergic reaction according to claim 1, characterized in that, The dynamic adjustment of the calculation parameters of the early warning probability value comprises: Identify the use of antihistamine drugs in the medication record; Recalculate the effective action time of the allergen according to the drug half-life; Correct the time decay coefficient in the blood vessel-allergy association feature matrix; Update the edge weight calculation formula of the dynamic probabilistic graph model.

7. The data management method based on CT angiography and allergic reaction according to claim 1, characterized in that, The local re-scanning instruction of the blood vessel image comprises: Detecting the protein structure similarity between the new allergen and the historical allergen, positioning the blood vessel region corresponding to the past allergic reaction when the similarity exceeds the cross-reaction threshold, generating priority scanning instructions containing the target blood vessel segment coordinates, and controlling the CT device to perform a low-dose local fast scan.

8. The data management method based on CT angiography and allergic reaction according to claim 7, characterized in that, The differential feature analysis includes: Aligning the blood vessel spatial coordinate systems of the new and old images; Extracting the tube wall thickness change rate data of the same anatomical position; Calculating the difference value of the angiographic contrast agent filling speed; Inputting the change feature vector into the early warning probability updating model.

9. The data management method based on CT angiography and allergic reaction according to claim 1, characterized in that, The mapping relationship model between the allergen exposure timeline and the development of vascular lesions includes: Encoding the allergen contact events into discrete time points in chronological order, extracting the morphological change parameters of the blood vessel images before and after each time point, training the time series convolution network to predict the development trajectory of the vascular lesions, and generating a blood vessel state evolution heat map containing time labels.

10. The data management method based on CT angiography and allergic reaction according to claim 9, characterized in that, The output of the vascular lesion development prediction report containing the time dimension change includes: Comparing the vascular lesion development trajectory with the standard growth curve, labeling the abnormal time interval deviating from the normal development rate, correlating the allergen exposure records in the corresponding time interval, and generating a multi-modal data correlation report sorted by time axis.

Images