Grading early warning method and system for MRI scene safety management

Abstract

The invention discloses a hierarchical early warning method and system for MRI scene safety management, and relates to the technical field of medical safety management, and the method comprises the following steps: constructing a scene multi-source information set, carrying out the data fusion processing of the multi-source information set, and extracting a state vector of a current time period; inputting the state vector into a pre-trained security risk assessment model, and outputting a risk grade score of the current scene; according to the risk grade score and a set grading early warning threshold value, determining a risk grade category to which the current MRI scene belongs; matching an early warning strategy according to the risk level category, and pushing the early warning strategy to a safety management system; according to the method, the multi-source information set is constructed, cross-modal data fusion is carried out, and security risk assessment based on classification, regression and fusion judgment models is combined, so that accurate grading early warning of the MRI scene is realized, the problems of difficulty in fusion of multi-heterogeneous data, inaccurate risk grade judgment and difficulty in threshold setting are solved, and the accuracy and the reliability of the MRI scene are improved. And the accuracy and the intelligent level of risk identification are improved.

Description

Technical Field

[0001] This invention relates to the field of medical safety management technology, and more specifically, to a graded early warning method and system for safety management in MRI scenarios. Background Technology

[0002] With the rapid development of medical imaging technology, magnetic resonance imaging (MRI) has become an indispensable tool in clinical diagnosis due to its advantages such as being non-invasive, high-resolution, and free of ionizing radiation. However, MRI equipment poses certain safety risks during operation, mainly due to its strong magnetic field, electromagnetic radiation, and high-power radiofrequency excitation. These risks not only pose a potential threat to the lives of patients but also affect the health of medical staff and caregivers. Especially in the MRI examination environment, due to the complex operating conditions of the equipment and the diverse personnel flow, effectively monitoring and assessing safety risks in the scenario and achieving dynamic early warning has become an important issue in ensuring medical safety.

[0003] Currently, risk assessment methods that integrate multiple models are gaining increasing attention. By integrating multi-class classification and regression models and employing ensemble learning algorithms such as gradient boosting trees to train the fused features through nonlinear mapping, the potential correlations within the data can be fully explored, improving the accuracy and robustness of risk level determination. Furthermore, by combining mathematical tools such as fuzzy membership functions and Bayesian probabilistic inference, a probabilistic expression of risk scores and dynamic threshold adjustment can be achieved, enabling the early warning system to possess stronger adaptive capabilities and adapt to the complex and ever-changing MRI operating environment.

[0004] Based on the aforementioned needs and current situation, there is an urgent need for an intelligent hierarchical early warning method and system capable of integrating multi-source information, deeply fusing it, extracting dynamic states, and accurately assessing risks in MRI scenarios. This system can not only reflect the safety status of the MRI examination environment in real time, but also provide hierarchical early warnings and strategy recommendations based on risk levels. This assists medical administrators in responding promptly to potential safety hazards, protecting the health and lives of patients and staff, and improving the safe operation of medical equipment.

[0005] The above-disclosed technical solutions have at least the following technical problems:

[0006] Traditional MRI safety management often relies on a single data source (such as equipment status or environmental monitoring alone), lacking comprehensive utilization of multimodal information on personnel, equipment, and environment. This results in incomplete risk assessment information and affects accuracy. Furthermore, traditional safety risk prediction models struggle to balance the discrete classification of risk levels with the refined expression of continuous risk scores, reducing the overall performance of risk assessment.

[0007] To address the above problems, this invention proposes a solution. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of existing technologies, embodiments of the present invention provide a graded early warning method and system for MRI scene safety management. By constructing a multi-source information set for the MRI scene, fusing multi-modal data on personnel, equipment, and the environment, and extracting vectors representing the current time period, a comprehensive perception and real-time modeling of the MRI examination environment is achieved. Based on a pre-trained safety risk assessment model, combined with the output of classification and regression models, a fusion judgment model is used to non-linearly correct the risk score, avoiding misjudgments and missed judgments caused by traditional threshold settings. Simultaneously, fuzzy membership functions and Bayesian inference methods are employed to dynamically determine the risk level of the MRI scene, achieving more accurate graded early warning. The system can match and push corresponding early warning strategies in real time, improving the intelligence and reliability of MRI examination room safety management, and effectively solving the technical problems of complex multi-source heterogeneous data fusion, inaccurate risk assessment, and delayed early warning response.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A hierarchical early warning method for MRI scene safety management includes the following steps: constructing a multi-source information set for the scene, performing data fusion processing on the multi-source information set, and extracting the state vector for the current time period; inputting the state vector into a pre-trained safety risk assessment model to output the risk level score of the current scene; determining the risk level category of the current MRI scene based on the risk level score and the set hierarchical early warning threshold; matching an early warning strategy according to the risk level category and pushing it to the safety management system.

[0011] In a preferred embodiment, the construction of the multi-source information set for the scenario is specifically as follows: acquiring personnel information, equipment operating status, and environmental data of the area to be tested; integrating the collected data into structured records according to time windows, and then splicing them together through time windows to form a time series dataset, i.e., the multi-source information set.

[0012] In a preferred embodiment, the data fusion processing of the multi-source information set is specifically as follows: according to the type of the multi-source information set, features are extracted by encoders of different modalities, including: personnel modality, equipment modality and environment modality; the feature representations encoded by each modality are input into the attention mechanism module, cross-modal correlation is calculated, and weighted fusion is performed to generate a fusion representation vector.

[0013] In a preferred embodiment, the extraction of the state vector for the current time period is specifically as follows: a feedforward neural network module is constructed using two fully connected layers and an activation function; the output of the feedforward neural network module is added to the fusion representation vector to form an intermediate representation with residuals; the intermediate representation with residuals is subjected to layer normalization to obtain the final enhanced fusion representation vector; the enhanced fusion representation vector is input into a multilayer perceptron for dimensionality reduction and nonlinear mapping to output a fixed-dimensional state vector, i.e., the state vector for the current time period.

[0014] In a preferred embodiment, the step of inputting the state vector into a pre-trained security risk assessment model and outputting a risk level score for the current scenario is as follows: The security risk assessment model is trained based on historical labeled data. The security risk assessment model includes a classification model and a regression model. The state vector extracted within the current time window is obtained and input into the classification model and regression model respectively, outputting an initial risk level and a risk score. The risk score is concatenated with the state vector to form a new fusion feature vector. The label of the initial risk level is used as a supervision signal to construct a training dataset. The training dataset and the fusion feature vector are input into a gradient boosting tree model for training to obtain a fusion decision model. The risk score is corrected using the fusion decision model to obtain the final corrected risk level score.

[0015] In a preferred embodiment, the step of correcting the risk score through a fusion decision model to obtain the final corrected risk level score is as follows: the fusion feature vector is input into a pre-trained fusion decision model, which outputs a confidence probability distribution for each risk level based on the nonlinear mapping relationship learned from historical training samples; a weight value is predefined for each risk level, forming a weight vector; the confidence probability distribution is multiplied by the weight vector to output the final corrected risk level score.

[0016] In a preferred embodiment, the step of determining the risk level category of the current MRI scene based on the risk level score and the set graded warning threshold is as follows: Multiple fuzzy membership functions for risk levels are designed based on expert experience and historical risk score data; the real-time calculated risk level scores are input into the fuzzy membership functions respectively, their corresponding membership degrees are calculated, and a membership degree vector is generated; the membership degree vector is used as a fuzzy prior to weight and correct the prior probabilities of historical risk levels, resulting in a weighted prior probability; the posterior probability is output based on the Bayesian formula according to the weighted prior probability; and the risk level with the highest posterior probability is selected as the risk level determination result for the current MRI scene.

[0017] A graded early warning system for MRI scene safety management includes an extraction module, an output module, a judgment module, and a push module, with interconnections between the modules. The extraction module constructs a multi-source information set for the scene, performs data fusion processing on the multi-source information set, and extracts the state vector for the current time period. The output module inputs the state vector into a pre-trained safety risk assessment model and outputs a risk level score for the current scene. The judgment module determines the risk level category of the current MRI scene based on the risk level score and a set graded early warning threshold. The push module matches an early warning strategy according to the risk level category and pushes it to the safety management system.

[0018] The technical effects and advantages of the graded early warning method and system for MRI scene safety management disclosed in this invention are as follows:

[0019] 1. This invention employs an attention-based feature fusion method that automatically learns and weights the intrinsic relationships between different modalities (such as the correlation between a specific person carrying a metal object near the scanning room and environmental magnetic field disturbances), generating a fusion representation vector richer in semantic information. This overcomes the limitations of traditional methods that simply splice or average fusion, significantly improving the representational power of features and laying a solid foundation for accurate risk assessment.

[0020] 2. This invention addresses the uncertainty of risk scoring boundaries (such as scores falling precisely near two level thresholds) by introducing fuzzy membership functions. It calculates the membership degree to each risk level, better aligning with human experts' understanding of the "gradual" nature of risk levels. Using the fuzzy membership degree as a prior correction factor, combined with historical risk level distributions (prior probabilities), the posterior probability is calculated using Bayes' theorem. This method dynamically integrates real-time assessment results with historical experience, effectively reducing the risk of misjudgment caused by fluctuations or noise in single real-time scores. This makes the final risk level determination more statistically robust, especially in handling cases with ambiguous boundaries more stably. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a graded early warning method for MRI scene safety management according to the present invention.

[0022] Figure 2 This is a schematic diagram of a graded early warning system for MRI scene safety management according to the present invention. Detailed Implementation

[0023] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1, Figure 1 This invention provides a graded early warning method for safety management in MRI scenes, comprising:

[0025] S1. Construct a multi-source information set for the scene, perform data fusion processing on the multi-source information set, and extract the state vector for the current time period.

[0026] In this embodiment, a multi-source information set for the scene is constructed as follows:

[0027] By deploying positioning devices, identification sensors, and identity verification terminals in the MRI examination room, personnel information of those entering the room is collected in real time. The personnel information includes identity category (such as medical staff, patients, caregivers, etc.), three-dimensional position of the personnel, identification of the personnel's action status (such as walking, standing still, approaching the magnet area), physiological status (such as heart rate, body temperature), and other attribute data.

[0028] The current equipment operating status is collected in real time through the communication interface with the MRI master control system (such as DICOM Worklist service, vendor open API or SCPI command set), including parameters such as operating mode, magnet status (such as whether it is in superconducting maintenance state), radio frequency excitation power, gradient magnetic field switching frequency, etc.

[0029] Environmental sensing equipment is installed in the inspection room and its auxiliary areas to collect environmental data, including air temperature and humidity, electromagnetic radiation intensity, oxygen concentration, and combustible gas concentration.

[0030] The various heterogeneous data collected above are integrated into structured records (such as JSON format, Protocol Buffer format, etc.) according to time windows, and then spliced ​​together by time windows to form a time series dataset, i.e. a multi-source information set, which serves as the input for subsequent processing.

[0031] In this embodiment, data fusion processing is performed on the multi-source information set, and the state vector for the current time period is extracted, as follows:

[0032] Based on the type of multi-source information set, feature extraction is performed using encoders of different modalities, including:

[0033] Personnel modality encoder: It adopts bidirectional LSTM (Bi-LSTM), and the input is a sequence of personnel position (x,y,z), activity state, physiological parameters, etc., and the output is a context hidden vector;

[0034] Device modal encoder: Using a unidirectional LSTM or 1D-CNN+LSTM structure, the device operating status (such as magnet current, RF power, and scanning mode) is encoded in a time sequence, and the device operating status sequence is output.

[0035] Environmental modal encoder: The field sensors (temperature, humidity, electromagnetic field, gas concentration, etc.) are deployed to form a map structure, and spatial correlation features are extracted using graph convolutional networks (GCN);

[0036] The feature representations encoded by each modality are input into the attention mechanism module. The person modality is used as the query, and the device and environment are used as the key and value. Cross-modal correlation is calculated as the matching weight between different modalities, and the weighted fusion is used to generate a fusion representation vector.

[0037] A feedforward neural network module is constructed by using two fully connected layers and activation functions. The output of the feedforward neural network module is added to the fused representation vector to form an intermediate representation with residuals.

[0038] Layer normalization is performed on the intermediate representation with residuals to obtain the final enhanced fused representation vector;

[0039] The enhanced fusion representation vector is input into a multilayer perceptron (MLP) for dimensionality reduction and nonlinear mapping, and the output is a fixed-dimensional state vector, that is, the state vector of the current time period.

[0040] The fusion representation vector is as follows:

[0041]

[0042] The feedforward neural network module is specifically as follows:

[0043] FFN=W2×σ×(W1×H f +b1)+b2

[0044] Where: H f It is a fused representation vector. Softmax is an operation that transforms each row into a probability distribution. Q is the Query matrix obtained by linearly mapping the modal features of the people at the current time step. K T It is the result of concatenating the modal features of the device and the environment, then multiplying by a linear mapping to obtain the Key matrix, d. kσ is the scaling factor, V is the Value matrix obtained by linear mapping of device and environment modal features, FFN is the output of the feedforward neural network, σ is the activation function (e.g., ReLU or GELU), b1 and b2 are the corresponding bias vectors, W1 is the weight matrix of the first layer, and W2 is the weight matrix of the second layer.

[0045] It should be noted that an attention mechanism is introduced, using personnel modalities as queries and devices and environments as keys / values, to achieve adaptive matching and weighted fusion of different information. Compared to simple concatenation or averaging, this solution can dynamically allocate attention weights based on real-time scene changes, highlighting key modal signals and enhancing the sensitivity and targeting of the fused representation to potential risks.

[0046] Furthermore, based on the fusion representation, features are nonlinearly mapped through a two-layer feedforward network and residual connections. Layer normalization is then applied to effectively suppress gradient vanishing and exploding problems, improving the model's convergence speed and generalization ability, and ensuring long-term stable operation after online deployment. A multilayer perceptron (MLP) is used to reduce the dimensionality of the enhanced fusion vector, outputting a fixed-dimensional state vector that can be directly used as input to subsequent risk assessment models. This lightweight representation retains multimodal feature information while reducing computational overhead, making it suitable for real-time embedded or edge computing environments and meeting the stringent timeliness requirements of MRI scenarios.

[0047] S2 inputs the state vector into a pre-trained safety risk assessment model and outputs a risk level score for the current scenario.

[0048] In this embodiment, the state vector is input into a pre-trained security risk assessment model, which outputs a risk level score for the current scenario, as follows:

[0049] The security risk assessment model is trained based on historical labeled data. The security risk assessment model includes a first model and a second model. The first model is a multi-class classification model (such as random forest, LightGBM or deep neural network), and its output is an initial risk level (such as "normal", "warning", "alarm"). The second model is a regression model (such as XGBoost regression, SVR or deep regression network), and its output is a risk score [0, 1].

[0050] Obtain the state vector extracted within the current time window and input it into the classification model and regression model respectively, and output the initial risk level and risk score;

[0051] The risk score is concatenated with the state vector to form a new fused feature vector;

[0052] The initial risk level labels are used as supervision signals to construct a training dataset;

[0053] The training dataset and the fused feature vector are input into the gradient boosting tree model for training. The non-linear mapping relationship between the fused feature vector and the risk level label is learned to obtain the fused judgment model. This fused model can automatically determine the risk level by comprehensively utilizing continuous risk scores and original state information without the need for manually setting thresholds.

[0054] The risk score is corrected by fusion judgment model to obtain the final corrected risk level score.

[0055] The risk score is corrected using a fusion judgment model to obtain the final corrected risk level score, as detailed below:

[0056] The fused feature vector is input into the pre-trained fusion decision model. The fusion decision model outputs the confidence probability distribution for each risk level based on the nonlinear mapping relationship learned from historical training samples.

[0057] To achieve a continuous mapping of risk level scores, a weight value is predefined for each risk level, forming a weight vector. This weight value reflects the severity and priority of the risk level, for example:

[0058] Normal: w1 = 0;

[0059] Low risk (early warning): w2 = 0.3;

[0060] Medium risk (alarm): w3 = 0.7;

[0061] High risk (critical alert): w4 = 1.0;

[0062] The confidence probability distribution is multiplied by the weight vector, and the weighted average is calculated as the final corrected risk level score.

[0063] It should be noted that a multi-class classification model is used to initially determine the discrete risk level, and then a regression model outputs a continuous risk score; the two complement each other. The classification model excels at distinguishing different risk categories, ensuring the basic accuracy of the warning level; the regression model can characterize the continuity of risk changes, providing rich numerical references for subsequent integration and avoiding the limitation of a single model only providing a rough classification.

[0064] Furthermore, traditional methods often require manually preset fixed thresholds to distinguish risk levels. Threshold selection is heavily influenced by expert experience and historical data distribution, easily leading to false positives or false negatives. In this solution, ensemble learning algorithms such as gradient boosting trees are used to train the classification output, regression score, and state vector through nonlinear mapping. This automatically learns the intrinsic relationships between various elements, eliminating the need for manually setting thresholds and achieving adaptive risk level determination, thus enhancing the model's robustness and maintainability.

[0065] S3 determines the risk level category of the current MRI scene based on the risk level score and the set graded warning threshold.

[0066] In this embodiment, the risk level category of the current MRI scene is determined based on the risk level score and the set graded early warning threshold, as follows:

[0067] Based on expert experience and historical risk scoring data, multiple fuzzy membership functions for risk levels are designed. The risk levels include "normal", "low risk", "medium risk" and "high risk". Each fuzzy membership function adopts the form of a triangle or Gaussian function, and its center value and width parameter are defined.

[0068] The membership function parameters are stored in a preset parameter library for subsequent fuzzy membership degree calculation of risk scoring;

[0069] The risk level scores obtained in real time are input into the fuzzy membership function to calculate their corresponding membership values ​​and generate a membership vector. Each membership value is between 0 and 1, reflecting the degree of belonging to the risk level corresponding to the risk score.

[0070] By using the membership vector as a fuzzy prior, the prior probability of historical risk level is weighted and corrected to obtain the weighted prior probability.

[0071] Output the posterior probability based on the weighted prior probability and Bayes' theorem;

[0072] Based on the posterior probability, the risk level with the highest posterior probability is selected as the risk level determination result for the current MRI scene. If the maximum value is lower than the set confidence threshold, a review or manual confirmation process is triggered.

[0073] The real-time judgment level is compared with the subsequent actual feedback to generate the judgment error;

[0074] The membership function parameters and conditional probabilities are adjusted based on the error, including: the membership function automatically optimizes the membership curve through gradient descent; and the Bayesian conditional probabilities are incrementally updated.

[0075] The weighted prior probabilities are as follows:

[0076]

[0077] The specific output of the posterior probability based on Bayes' theorem is as follows:

[0078]

[0079] In the formula: P i ′ is the weighted prior probability of the i-th risk level, μ i P is the membership value of the risk rating score under the i-th risk level (e.g., "normal", "low risk", "medium risk", "high risk"). i μ is the historical prior probability of the i-th risk level. j P is the membership value of the risk level score under the j-th risk level. j P(R) is the historical prior probability of the j-th risk level, where j is the number of risk levels. i |L) is the posterior probability of the i-th risk level, P(L|R) i P(L|R) is the conditional probability density of the risk score L at the i-th risk level. This density can be obtained by fitting a probability distribution (such as a Gaussian distribution) to historical labeled samples or by kernel density estimation. j ) is the conditional probability density of the risk score L under the i-th risk level.

[0080] S4 matches early warning strategies based on risk level categories and pushes them to the security management system.

[0081] In this embodiment, an early warning strategy is matched according to the risk level category and pushed to the security management system, as follows:

[0082] Based on the risk level category, the corresponding early warning strategy template is called from the preset strategy library. The template includes fields such as: early warning level identifier, suggested handling process, emergency response time, responsible person information, and linkage control instructions.

[0083] If the strategy template contains linkage control requirements (such as frequency reduction vibration, starting bypass air duct, etc.), the control module interface is called to generate a list of control instructions and write them into the queue of instructions to be executed.

[0084] The matched early warning strategy template and its accompanying control instructions are packaged into an early warning response information package, which is then pushed to the upper-level safety management system through the device communication gateway module.

[0085] In the safety management system, the graphical warning interface is triggered to update, and alarm information is sent to preset terminals (such as mobile APP, duty room screen, remote operation and maintenance platform) according to the responsible person identifier in the policy template.

[0086] Record the early warning strategy push behavior and response status to the early warning log system for subsequent statistical analysis and model optimization.

[0087] Example 2, Figure 2 The present invention provides a graded early warning system for safety management in MRI scenes, comprising an extraction module, an output module, a judgment module, and a push module, with connections between the modules;

[0088] The extraction module is used to construct a multi-source information set of the scene, perform data fusion processing on the multi-source information set, and extract the state vector of the current time period;

[0089] The output module is used to input the state vector into the pre-trained safety risk assessment model and output the risk level score of the current scene;

[0090] The judgment module is used to determine the risk level category of the current MRI scene based on the risk level score and the set graded warning threshold;

[0091] The push module is used to match early warning strategies based on risk level categories and push them to the security management system.

[0092] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0093] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0094] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0095] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0097] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A graded early warning method for safety management in MRI scenarios, characterized in that, Includes the following steps: Construct a multi-source information set for the scene, perform data fusion processing on the multi-source information set, and extract the state vector for the current time period; The state vector is input into a pre-trained safety risk assessment model, which outputs a risk level score for the current scenario. Based on the risk level score and the set graded early warning threshold, determine the risk level category of the current MRI scene; Based on the risk level category, match the early warning strategy and push it to the security management system.

2. The graded early warning method for MRI scene safety management according to claim 1, characterized in that, The construction of the multi-source information set for the scenario is as follows: Acquire personnel information, equipment operating status, and environmental data for the area to be tested; The collected data is integrated into structured records according to time windows, and then spliced ​​together by time windows to form a time series dataset, i.e., a multi-source information set.

3. The graded early warning method for MRI scene safety management according to claim 2, characterized in that, The data fusion processing of the multi-source information set is as follows: Based on the type of multi-source information set, features are extracted using encoders of different modalities, including: personnel modality, equipment modality, and environmental modality; The feature representations encoded by each modality are input into the attention mechanism module to calculate cross-modal correlations and then weighted and fused to generate a fused representation vector.

4. The graded early warning method for MRI scene safety management according to claim 3, characterized in that, The extraction of the state vector for the current time period is as follows: A feedforward neural network module is constructed by using two fully connected layers and activation functions. The output of the feedforward neural network module is added to the fused representation vector to form an intermediate representation with residuals. Layer normalization is performed on the intermediate representation with residuals to obtain the final enhanced fused representation vector; The enhanced fusion representation vector is input into a multilayer perceptron for dimensionality reduction and nonlinear mapping, and the output is a fixed-dimensional state vector, that is, the state vector of the current time period.

5. The graded early warning method for MRI scene safety management according to claim 4, characterized in that, The process of inputting the state vector into a pre-trained security risk assessment model and outputting a risk level score for the current scenario is as follows: A security risk assessment model is trained based on historical labeled data, and the security risk assessment model includes a classification model and a regression model; Obtain the state vector extracted within the current time window and input it into the classification model and regression model respectively, and output the initial risk level and risk score; The risk score is concatenated with the state vector to form a new fused feature vector; The initial risk level labels are used as supervision signals to construct a training dataset; The training dataset and the fused feature vector are input into the gradient boosting tree model for training to obtain the fusion decision model. The risk score is corrected by fusion judgment model to obtain the final corrected risk level score.

6. The graded early warning method for MRI scene safety management according to claim 5, characterized in that, The risk score is corrected using a fusion judgment model to obtain the final corrected risk level score, as detailed below: The fused feature vector is input into the pre-trained fusion decision model. The fusion decision model outputs the confidence probability distribution for each risk level based on the nonlinear mapping relationship learned from historical training samples. A weight value is predefined for each risk level, and a weight vector is constructed accordingly; Multiply the confidence probability distribution by the weight vector to output the final corrected risk level score.

7. The graded early warning method for MRI scene safety management according to claim 6, characterized in that, The risk level category of the current MRI scene is determined based on the risk level score and the set graded early warning threshold, as follows: Based on expert experience and historical risk scoring data, multiple fuzzy membership functions for risk levels are designed. The risk level scores obtained in real time are input into the fuzzy membership function to calculate their corresponding membership values ​​and generate a membership vector. By using the membership vector as a fuzzy prior, the prior probability of historical risk level is weighted and corrected to obtain the weighted prior probability. Output the posterior probability based on the weighted prior probability and Bayes' theorem; The risk level with the highest posterior probability is selected as the risk level determination result for the current MRI scene based on the posterior probability.

8. A graded early warning system for MRI scene safety management as described in any one of claims 1-7, characterized in that, It includes an extraction module, an output module, a judgment module, and a push module, and the modules are interconnected. The extraction module is used to construct a multi-source information set of the scene, perform data fusion processing on the multi-source information set, and extract the state vector of the current time period; The output module is used to input the state vector into the pre-trained safety risk assessment model and output the risk level score of the current scene; The judgment module is used to determine the risk level category of the current MRI scene based on the risk level score and the set graded warning threshold; The push module is used to match early warning strategies based on risk level categories and push them to the security management system.

Images