A Delay Prediction and Decision-Making Method for Vehicle Headset CTA

By using a data flow model that combines baseline delay prediction and physiological residual correction, the problem of individualized delay prediction in vehicle-mounted head-mounted CTA was solved, reducing contrast agent waste and duplicate scans, and improving the safety and operability of the vehicle-mounted CTA process.

CN123058764APending Publication Date: 2026-11-10XUANWU HOSPITAL OF CAPITAL UNIV OF MEDICAL SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611463304.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-20
Publication Date
2026-11-10

Smart Images

  • Figure CN123058764A_ABST
    Figure CN123058764A_ABST
Patent Text Reader

Abstract

This invention relates to a delay time prediction and decision-making method for mobile head-mounted CTA, belonging to the field of mobile CT scan control technology. It solves the problems of existing mobile stroke unit mobile head-mounted CTA methods, which rely on test-bolus analysis, cannot individually predict delay times, and lack comprehensive decision-making based on the mobile status before injection. The method includes acquiring clinical parameters and mobile CT scan parameters; preprocessing the clinical parameters to obtain a standardized feature vector; inputting the standardized feature vector and mobile CT scan parameters into a trained prediction model to obtain an individualized delay time prediction value; acquiring the current mobile status information of the mobile stroke unit and determining whether the current mobile status information meets the readiness conditions; if so, outputting an injection feasibility signal; and controlling the mobile CT to perform head-mounted CTA scanning based on the injection feasibility signal and the individualized delay time prediction value. This achieves individualized delay time prediction and comprehensive decision-making for mobile head-mounted CTA.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of vehicle-mounted CT scan control technology, and in particular to a delay time prediction and decision-making method for vehicle-mounted head-mounted CTA. Background Technology

[0002] For acute ischemic stroke, especially in patients suspected of having large vessel occlusion, it is crucial to determine the status of intracranial major arteries, the Circle of Willis, and related feeding vessels as early as possible to assist in decisions regarding intravenous thrombolysis, bridging thrombectomy, and transport. Mobile stroke units bring mobile CT scanners, emergency equipment, remote communication systems, and stroke teams to the pre-hospital setting, which can shorten the treatment process.

[0003] However, the mobile stroke units are typically equipped with 16-slice or miniaturized CT scanners, which differ from in-hospital high-slice fixed CT scanners in terms of scanning speed, detector coverage, vehicle space, patient fixation conditions, and operator configuration. In-hospital methods such as aortic arch level bolus injection tracking or low-dose test bolus may require additional injections, scans, and ROI curve interpretation steps in a mobile environment, hindering rapid pre-hospital decision-making.

[0004] While fixed-delay scanning is simple to perform, it cannot fully reflect individual differences in patient circulatory status, blood volume, height and weight, blood pressure, heart rate, vascular elasticity, and the time it takes for contrast agent to reach the target vessels in the head. For 16-slice mobile head CTA, if the scanning phase is too early, insufficient arterial enhancement may occur; if the scanning phase is too late, venous contamination may occur, affecting the evaluation of distal branches and collateral circulation.

[0005] Furthermore, mobile stroke units may be in motion during patient transport. Studies have shown that the incidence of motion artifacts is significantly increased when performing 16-slice mobile CT head scans while the ambulance is in motion. Therefore, even if the delay time is accurately predicted, if injections are performed under conditions such as the CT scanner not being ready, the vehicle being subjected to severe vibrations, or the infusion pump not being prepared, it may still lead to waste of contrast agent, duplicate scans, or the acquisition of clinically unusable images.

[0006] Therefore, there is a need for an intelligent prediction of delayed scanning time and pre-injection decision support method for 16-slice mobile CT head-mounted CTA, which can output individualized delayed time prediction values ​​suitable for 16-slice mobile CT countdown triggering without performing additional test bolus or real-time ROI swarm tracking for each case, and output injection suggestion signals based on the overall mobile situation before injection for the operator's decision reference. Summary of the Invention

[0007] Based on the above analysis, the embodiments of the present invention aim to provide a delay time prediction and decision-making method for vehicle-mounted head CTA, in order to solve the problems of existing mobile stroke unit vehicle-mounted head CTA relying on case-by-case test bolus, being unable to predict delay time individually, and lacking comprehensive decision-making based on vehicle status before injection.

[0008] This invention provides a latency prediction and decision-making method for vehicle-mounted head unit (CTA), comprising the following steps: S1. Obtain the patient's clinical parameters and vehicle-mounted CT scan parameters, and preprocess the clinical parameters to obtain a standardized feature vector; S2. Input the standardized feature vector and vehicle-mounted CT scan parameters into the trained prediction model to obtain the individualized delay time prediction value. The individualized delay time prediction value is used to determine the time interval between the start of contrast agent injection and the start of CT scan. S3. Obtain the current vehicle status information of the mobile stroke unit, including the vehicle CT status, vehicle motion status, and infusion pump status. S4. Determine whether the current vehicle status information meets the preset ready conditions. If yes, output an injection feasible signal; otherwise, output an injection infeasible signal and return to step S3 after the set collection interval to continue execution. S5. Based on the injection feasibility signal and the individualized delay time prediction value, control the on-board CT execution head CTA scan.

[0009] Furthermore, the prediction model includes: The baseline delay prediction branch is used to output the baseline delay scan time based on the first set of parameters in the standardized feature vector; The physiological residual correction branch is used to output a nonlinear physiological residual correction amount based on the second set of parameters in the standardized feature vector. The time-delay fusion module is used to add the basic delayed scan time to the nonlinear physiological residual correction amount to obtain the patient-level physiological prediction delay time; the patient-level physiological prediction delay time is used to determine the individualized delay time prediction value.

[0010] Furthermore, the prediction model also includes a CT timing correction branch, which is used to output the scan trigger time based on the patient-level physiological prediction delay time and the on-board CT scan parameters; the scan trigger time is used to determine the individualized delay time prediction value.

[0011] Furthermore, the scan trigger time Represented as: ; in, ; In the formula, This indicates the delay in physiological prediction at the patient level. Indicates timing correction amount, This indicates a device trigger response delay. Indicates the scan duration. This represents the target blood vessel location coefficient.

[0012] Furthermore, the individualized delay time prediction value is determined based on the scan trigger time in the following manner. : ; In the formula, This indicates the preset fine-tuning amount. , These represent the minimum and maximum allowed times, respectively. Represents the amplitude limiting function. This represents the floor function.

[0013] Furthermore, the input to the baseline delay prediction branch is the first set of parameters in the standardized feature vector, including age, heart rate, mean arterial pressure, pulse pressure, and BMI; the baseline delay prediction branch uses a linear regression, ridge regression, or Lasso regression model, and its prediction expression is: ; In the formula, Indicates the baseline delayed scan time. Represents the intercept term. , , , , Indicates the parameters of each model. , , , , These represent the values ​​of age, heart rate, mean arterial pressure, pulse pressure, and BMI in the standardized feature vector, respectively.

[0014] Further, the input to the physiological residual correction branch is the second set of parameters in the standardized feature vector, including heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure and pulse pressure, height, weight, BMI, sex, cardiac rhythm status, contrast agent injection flow rate, contrast agent dosage, and saline flushing parameters; the physiological residual correction branch includes: The parameter segmentation layer is used to obtain the circulatory dynamics feature group, body size-blood volume feature group, and injection protocol feature group based on the second set of parameters; A cyclic dynamics coding layer is used to convert the received cyclic dynamics feature set into a cyclic dynamics embedding vector; A body size-blood volume encoding layer is used to convert the received body size-blood volume feature set into a body size embedding vector; An injection scheme encoding layer is used to convert the received injection scheme feature set into an injection scheme embedding vector; The fusion layer is used to fuse the cyclic dynamics embedding vector, body shape embedding vector, and injection scheme embedding vector to obtain fused features; A residual prediction network layer is used to obtain a nonlinear physiological residual correction amount based on the fusion features.

[0015] Furthermore, the training of the prediction model employs a composite loss function, which includes one or more of the following: a time prediction error loss term, a range constraint loss term, a physiological trend constraint loss term, and an input perturbation smoothing loss term. The range constraint loss term is used to penalize when the model output is below the minimum allowed time or above the maximum allowed time. The physiological trend constraint loss term is used to penalize when the model output violates a preset physiological monotonicity. The input perturbation smoothing loss term is used to constrain the change in the model output to not exceed a preset threshold when the input variable experiences a perturbation of a preset magnitude.

[0016] Furthermore, after outputting the injection feasibility signal, if no injection initiation event is detected, at least one of the following is performed: If the injection initiation event is not detected within the preset waiting time, return to step S3; The vehicle status information is continuously monitored. If the vehicle status information no longer meets the preset ready conditions, the injection feasibility signal is revoked and the process returns to step S3.

[0017] Furthermore, the control of the vehicle-mounted CT to perform head CTA scanning based on the injection feasibility signal and individualized delay time prediction includes: Based on the injection feasibility signal, an injection prompt message is output, which is used to prompt the operator to perform the contrast agent injection. In response to an injection initiation event generated by the operator performing the injection, a countdown is started; wherein the countdown duration is the individualized delay time prediction value; After the countdown reaches zero, the onboard CT scanner will perform a CTA scan, either triggered or delayed, depending on the vehicle's motion status.

[0018] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: This invention provides a delay time prediction and decision-making method for mobile head-mounted CTA. Through a hierarchical data flow model of baseline delay prediction, physiological residual correction, and CT timing correction, it outputs individualized delay time prediction values ​​without performing additional test bolus or real-time ROI bolus tracking for each case. Before injection, it outputs an injection feasibility signal by comprehensively considering the mobile CT status, vehicle motion status, and infusion pump status, avoiding blind injection under conditions of equipment incomplete preparation or vehicle vibration. Simultaneously, by responding to the injection initiation event to start a countdown, and after the countdown reaches zero, it triggers or delays the mobile CT head-mounted CTA scan based on the vehicle motion status, reducing contrast agent waste and the risk of repeated scans, and improving the standardization, safety, and operability of the mobile stroke unit mobile head-mounted CTA process.

[0019] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0020] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0021] Figure 1 This is a flowchart illustrating the delay prediction and decision-making method for vehicle-mounted head unit (CTA) provided in an embodiment of the present invention. Detailed Implementation

[0022] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0023] A specific embodiment of the present invention discloses a latency prediction and decision-making method for vehicle head unit CTA, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the patient's clinical parameters and vehicle-mounted CT scan parameters, and preprocess the clinical parameters to obtain a standardized feature vector; S2. Input the standardized feature vector and vehicle-mounted CT scan parameters into the trained prediction model to obtain the individualized delay time prediction value. The individualized delay time prediction value is used to determine the time interval between the start of contrast agent injection and the start of CT scan. S3. Obtain the current vehicle status information of the mobile stroke unit, including the vehicle CT status, vehicle motion status, and infusion pump status. S4. Determine whether the current vehicle status information meets the preset ready conditions. If yes, output an injection feasible signal; otherwise, output an injection infeasible signal and return to step S3 after the set collection interval to continue execution. S5. Based on the injection feasibility signal and the individualized delay time prediction value, control the on-board CT execution head CTA scan.

[0024] During implementation, in step S1, clinical parameters include age, heart rate, blood pressure, and body type, as well as gender, cardiac rhythm, contrast agent injection flow rate, contrast agent dosage, and saline flushing parameters. Among these, age is used to characterize age-related differences in vascular elasticity, circulation time, and degree of arteriosclerosis; heart rate is used to characterize the speed at which the contrast agent reaches the target blood vessels in the head via the central circulation; blood pressure parameters are used to characterize the patient's circulatory perfusion status and hemodynamic level, including systolic and diastolic blood pressure; and body type parameters are used to characterize circulation path length, blood volume, and contrast agent dilution, including height and weight.

[0025] During implementation, the parameters for vehicle-mounted CT scanning include scan duration, device trigger response delay, and target blood vessel location coefficient.

[0026] In practice, clinical parameters are preprocessed, including: unit standardization, outlier identification, missing value imputation, missing indicator feature generation, continuous variable standardization, and categorical variable coding.

[0027] It should be noted that the mobile CT scan parameters are not standardized but are saved directly in physical quantity units for subsequent calculation of CT timing correction branches.

[0028] Specifically, the units are uniformly as follows: age in years, heart rate in beats per minute, blood pressure in mmHg, height in centimeters or meters, and weight in kilograms.

[0029] Specifically, missing value imputation is performed using the overall median of the training set, the median stratified by gender, or the median stratified by age.

[0030] Specifically, the missing indicator feature is generated to record whether one or more of the following parameters are missing: blood pressure, height, and weight.

[0031] Specifically, continuous variable standardization refers to performing Z-score standardization or min-max normalization on continuous variables.

[0032] Specifically, categorical variable encoding involves binary encoding or one-hot encoding of categorical variables.

[0033] It should be noted that during the training phase, the standardized parameters and imputation rules for continuous variables are calculated and saved from the training set; during the application phase, the standardized parameters and imputation rules saved during the training phase are used to preprocess the clinical parameters to ensure that the input distribution in the application phase is consistent with that in the training phase.

[0034] Specifically, outlier values ​​are judged based on the following physiologically reasonable ranges: heart rate is 20–250 beats / min; systolic blood pressure is 50–300 mmHg; diastolic blood pressure is 20–200 mmHg; height is 100–250 cm; and weight is 20–300 kg. Values ​​exceeding the above reasonable ranges are marked as abnormal and corrected or removed.

[0035] In practice, the preprocessing also includes calculating derived features based on the clinical parameters. These derived features include one or more of mean arterial pressure (MAP), pulse pressure (PP), and body mass index (BMI). Specifically, mean arterial pressure (MAP) = DBP + 1 / 3 × (SBP) DBP), pulse pressure difference PP=SBP DBP, BMI = Weight / Height²; SBP is systolic blood pressure, DBP is diastolic blood pressure, Weight is body weight, and Height is height. When calculating BMI, Height is in meters and Weight is in kilograms.

[0036] Specifically, the preprocessing process includes: standardizing the units of the parameters and identifying outliers; calculating mean arterial pressure, pulse pressure, and BMI based on the standardized parameters; imputing missing values ​​and simultaneously generating missing value indication features; standardizing continuous variables and encoding categorical variables; and forming a standardized feature vector. The standardized feature vector includes age, gender, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse pressure, height, weight, BMI, and missing value indication features. Age, heart rate, systolic blood pressure, diastolic blood pressure, height, and weight are the standardized values ​​of directly obtained parameters; mean arterial pressure, pulse pressure, and BMI are the standardized values ​​of derived features; and gender is the encoded value of the categorical variable.

[0037] In implementation, step S2, the prediction model includes: The baseline delay prediction branch is used to output the baseline delay scan time based on the first set of parameters in the standardized feature vector; The physiological residual correction branch is used to output a nonlinear physiological residual correction amount based on the second set of parameters in the standardized feature vector. The time-delay fusion module is used to add the basic delayed scan time to the nonlinear physiological residual correction amount to obtain the patient-level physiological prediction delay time; the patient-level physiological prediction delay time is used to determine the individualized delay time prediction value.

[0038] In practice, the input to the baseline time delay prediction branch is the first set of parameters in the standardized feature vector, including age, heart rate, mean arterial pressure, pulse pressure, and BMI. The baseline time delay prediction branch employs a linear regression, ridge regression, or Lasso regression model, and its prediction expression is as follows: ; In the formula, Indicates the baseline delayed scan time. Represents the intercept term. , , , , Indicates the parameters of each model. , , , , These represent the values ​​of age, heart rate, mean arterial pressure, pulse pressure, and BMI in the standardized feature vector, respectively.

[0039] The model parameters and intercept terms are obtained from the training set through linear regression, ridge regression, or Lasso regression.

[0040] In practice, the input to the physiological residual correction branch is the second set of parameters in the standardized feature vector, including heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure and pulse pressure, height, weight, BMI, sex, cardiac rhythm status, contrast agent injection flow rate, contrast agent dosage, and saline flushing parameters. The physiological residual correction branch includes: The parameter segmentation layer is used to obtain the circulatory dynamics feature group, body size-blood volume feature group, and injection protocol feature group based on the second set of parameters; A cyclic dynamics coding layer is used to convert the received cyclic dynamics feature set into a cyclic dynamics embedding vector; A body size-blood volume encoding layer is used to convert the received body size-blood volume feature set into a body size embedding vector; An injection scheme encoding layer is used to convert the received injection scheme feature set into an injection scheme embedding vector; The fusion layer is used to fuse the cyclic dynamics embedding vector, body shape embedding vector, and injection scheme embedding vector to obtain fused features; A residual prediction network layer is used to obtain a nonlinear physiological residual correction amount based on the fusion features.

[0041] Specifically, the circulatory dynamics characteristic group includes heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse pressure, and cardiac rhythm status, used to characterize the patient's circulatory velocity, perfusion pressure, pulsation status, and hemodynamic level; the body size-blood volume characteristic group includes height, weight, BMI, and sex, used to characterize circulatory path length, blood volume, contrast agent dilution, and individual body size differences; the injection protocol characteristic group includes contrast agent injection flow rate, contrast agent dosage, and saline flushing parameters.

[0042] Specifically, the circulatory dynamics coding layer, the body size-blood volume coding layer, and the injection protocol coding layer are each composed of one or more of the following: fully connected layers, batch normalized layers, ReLU or LeakyReLU activation functions, and Dropout layers.

[0043] Understandably, the circulatory dynamics encoding layer extracts nonlinear feature representations of the patient's circulatory velocity, perfusion pressure, pulsation state, and hemodynamic level, outputting a circulatory dynamics embedding vector; the body size-blood volume encoding layer extracts feature representations of the patient's circulatory path length, blood volume, contrast agent dilution, and individual body size differences, outputting a body size embedding vector; and the injection protocol encoding layer extracts feature representations of the initial conditions and propulsion speed of the contrast agent entering the body, outputting an injection protocol embedding vector. The circulatory dynamics encoding layer, body size-blood volume encoding layer, and injection protocol encoding layer each map their respective received feature sets into low-dimensional, dense embedding vectors, enabling subsequent fusion layers to fuse information from different dimensions within a unified feature space.

[0044] Specifically, the fusion layer uses one or more of the following methods: splicing, weighted summation, gated fusion, or attention fusion, to fuse the cyclic dynamics embedding vector, body shape embedding vector, and injection scheme embedding vector to obtain fused features.

[0045] Preferably, the input to the physiological residual correction branch also includes missing indicator features. The fusion layer is used to fuse the circulatory dynamics embedding vector, body shape embedding vector, injection protocol embedding vector, and missing indicator features to obtain fused features. It is understood that missing indicator features enable the model to distinguish between true collected values ​​and imputed values, allowing it to output relatively reliable predictions even when parameters such as blood pressure, height, and weight are missing in pre-hospital emergency scenarios, avoiding information loss and prediction bias caused by simple imputation.

[0046] Specifically, the residual prediction network layer includes two fully connected layers. The first layer is used for feature fusion, and the second layer is used to output a single continuous value, namely the nonlinear physiological residual correction amount. The number of hidden layer nodes can be set to 8, 16, 32 or 64.

[0047] In practice, the prediction model is trained on the baseline delay prediction branch and the physiological residual correction branch in the following manner: Construct training samples, each sample including a standardized feature vector and a supervised learning label. The supervised learning label is the actual delayed scan time of head CTA or the peak time of head target vessels obtained by the test bolus method. The baseline latency prediction branch is trained, using supervised learning labels as labels, and outputs the baseline latency scan time. The difference between the supervised learning label and the baseline delayed scan time is used as the residual label; the residual label is used to represent individual differences that the baseline delay prediction branch fails to explain. The physiological residual correction branch is trained, with the residual label as the learning target, and outputs a nonlinear physiological residual correction quantity. The patient-level physiological prediction delay time is obtained by adding the baseline delayed scan time to the nonlinear physiological residual correction.

[0048] It should be noted that the construction of standardized feature vectors for each sample is based on the preprocessing method in step S1. Specifically, the standardized parameters in continuous variable standardization are calculated and saved from the training set; the imputation rules in missing value imputation are calculated and saved from the training set. In the application phase, the standardized parameters and imputation rules saved in the training phase are used to preprocess the clinical parameters to ensure that the input distribution in the application phase is consistent with that in the training phase.

[0049] Specifically, when training the baseline delay prediction branch, if it is linear regression, the least squares method is used to solve the model parameters; if it is ridge regression, the least squares method with L2 regularization is used to solve the model parameters; if it is Lasso regression, the least squares method with L1 regularization is used to solve the model parameters.

[0050] In implementation, the baseline delay prediction branch and the physiological residual correction branch are jointly trained and optimized end-to-end using a composite loss function. The composite loss function includes one or more of the following: time prediction error loss term, range constraint loss term, physiological trend constraint loss term, and input perturbation smoothing loss term. The range constraint loss term is used to penalize when the model output is lower than the minimum allowable time or higher than the maximum allowable time. The physiological trend constraint loss term is used to penalize when the model output violates the preset physiological monotonicity. The input perturbation smoothing loss term is used to constrain the change in model output to not exceed a preset threshold when the input variable undergoes a perturbation of a preset magnitude.

[0051] In practical implementation, the composite loss function Represented as: ; In the formula, , , , These represent the time prediction error loss term, range constraint loss term, physiological trend constraint loss term, and input disturbance smoothing loss term, respectively. , , , These represent the weight coefficients of the time prediction error loss term, the range constraint loss term, the physiological trend constraint loss term, and the input disturbance smoothing loss term, respectively.

[0052] Specifically, weighting coefficients , , , Based on the safety priority settings in the vehicle-mounted head unit (CTA) scenario, the weight of the range constraint loss term is determined. The weight of the time prediction error loss term is higher than that of the time prediction error loss term. To ensure that the model output prioritizes meeting the safety boundary; the weights of the physiological trend constraint loss term. The physiological trend constraint weight is determined based on the measurement reliability of each physiological variable in the vehicle-mounted monitoring equipment; the lower the measurement reliability of a variable, the higher its physiological trend constraint weight. For example, , , , Set them to 0.5, 1, 0.3, and 0.1 respectively.

[0053] Specifically, the time prediction error loss term is expressed as: ; In the formula, For the first Patient-level physiological prediction delay time for each sample For the first The actual delayed scan time for each sample This represents the number of training samples.

[0054] Specifically, the scope constraint loss term is expressed as: ; In the formula, , These represent the minimum and maximum allowed times, respectively.

[0055] Specifically, the physiological trend constraint loss term is expressed as: ; In the formula, This represents a pre-defined set of physiological variables; For the first Physiological variables in a sample The value; Indicates the first Patient-level physiological prediction delay time for each sample Physiological variables The partial derivatives; Physiological variables The constraint direction sign has a value of +1 indicating expected positive correlation and a value of -1 indicating expected negative correlation. Representing physiological variables The weighting coefficients.

[0056] Specifically, the set of physiological variables includes one or more of heart rate, BMI, MAP, and PP; among them, the constraint direction of heart rate is negative correlation, the constraint direction of BMI is positive correlation, the constraint direction of MAP is negative correlation, and the constraint direction of PP is positive correlation.

[0057] Specifically, physiological variables Weighting coefficients The weighting factor is determined based on the clarity of the physiological relationship between the physiological variable and the latency, the data quality of the physiological variable in the training set, the degree of influence of the physiological variable on the latency, and the balance of magnitude among the loss terms. For example, the weighting factor for heart rate is set to 0.3–0.5, the weighting factor for BMI is set to 0.1–0.3, the weighting factor for MAP is set to 0.1–0.3, and the weighting factor for PP is set to 0.1–0.2.

[0058] Specifically, the input perturbation smoothing loss term is expressed as: ; In the formula, For the first Standardized feature vectors in each sample A preset amplitude perturbation applied to the standardized feature vector; For the first Standardized feature vectors in each sample Preset amplitude disturbance applied Post-patient physiological prediction delay time; For the first Standardized feature vectors in each sample The corresponding patient-level physiological prediction delay time, Represents the norm.

[0059] Specifically, the amplitude perturbation in the input perturbation smoothing loss term The amplitude perturbation is determined based on one or more of the following factors: the measurement error range of the vehicle-mounted monitoring equipment, the standard deviation of the input variables, the clinically acceptable range of predictive fluctuations, and the training sample size. For example, the amplitude perturbation... The amplitude disturbance is ±5% to ±10% of the standard deviation of each continuous variable. It can be applied to all continuous variables in the standardized feature vector, or only to variables with large measurement errors such as heart rate, systolic blood pressure, and diastolic blood pressure.

[0060] Specifically, under small sample conditions, the prediction model is trained using one or more strategies among K-fold cross-validation, leave-one-out cross-validation, Bootstrap ensemble, early stopping, and regularization.

[0061] Preferably, after the predictive model is deployed, one or more of the following are recorded: actual usage time, target vessel CT value, image quality score, vein contamination grade, and whether the diagnosis is met. A weakly supervised calibration target is constructed based on the recorded information to update the predictive model. Specifically, the actual usage time is automatically recorded by the vehicle-mounted CT control terminal, representing the delay time for final confirmation by the operator; the target vessel CT value is automatically read by the system after the operator selects the region of interest (ROI) of the target vessel on the vehicle-mounted CT image; the image quality score is assigned by the operator according to a preset scoring standard; the vein contamination grade is assigned by the operator according to a preset grading standard; and whether the diagnosis is met is determined by the operator based on whether the image meets the diagnostic requirements.

[0062] Specifically, weakly supervised calibration targets are constructed based on the recorded information, including: If the CT value of the target blood vessel is less than the preset enhancement threshold, the calibration target time is set to the actual trigger time plus the first correction amount; where the first correction amount is 1 to 3 seconds. If the venous contamination level reaches the preset level, the calibration target time is set to the actual trigger time minus the second correction amount, which is 1 to 3 seconds. If the image quality score reaches the preset threshold and meets the diagnostic requirements, the actual trigger time will be used as the calibration target. If a case fails to meet diagnostic criteria due to severe motion artifacts, injection failure, or device alarms, its weight is reduced or it is excluded from model updates.

[0063] It should be noted that the actual time taken, the target vessel CT value, the image quality score, the venous contamination grade, and whether the diagnosis is met, as well as the failure to meet the diagnosis due to severe motion artifacts, injection failure, or equipment alarms, can all be directly obtained.

[0064] More specifically, the image quality score is based on a scale of 1 to 5, where 1 is not diagnosable and 5 is excellent; the venous contamination is graded on a scale of 0 to 3, where 0 is no venous contamination and 3 is severe venous contamination.

[0065] Specifically, the calibration target is used to update the prediction model, including: using the calibration target as a weak supervision label to fine-tune the physiological residual correction branch parameters.

[0066] In practice, the prediction model also includes a CT timing correction branch, which is used to output the scan trigger time based on the patient-level physiological prediction delay time and the on-board CT scan parameters; the scan trigger time is used to determine the individualized delay time prediction value.

[0067] Specifically, the scan trigger time Represented as: ; In the formula, This indicates the delay in physiological prediction at the patient level. This indicates the timing correction amount.

[0068] More specifically, timing correction amount Represented as: ; In the formula, This indicates a device trigger response delay. Indicates the scan duration. This represents the target blood vessel location coefficient.

[0069] Furthermore, the target blood vessel location coefficient This value, representing the relative position of the target vessel at the appropriate enhancement point throughout the scan acquisition process, can be set to a value between 0.3 and 0.7 depending on the target vessel's position in the scan sequence. When it is desired that the scan acquisition center is close to the peak enhancement of the large intracranial arteries, It can be set to 0.5; when the target vessel is located in an area acquired earlier in the scan sequence, The size can be appropriately reduced; when the target blood vessel is located in a region acquired later in the scanning sequence, It can be increased appropriately.

[0070] In practice, the individualized delay time prediction value is determined based on the scan trigger time using the following method. : ; In the formula, This indicates the preset fine-tuning amount. , These represent the minimum and maximum allowed times, respectively. Represents the amplitude limiting function. This represents the floor function.

[0071] Specifically, preset fine-tuning amount The limited manual adjustment amount, which can be made by the operator under special circumstances, is input through the input interface of the vehicle control terminal; for example, the preset fine-tuning amount is provided. The interval should not exceed ±3 seconds to avoid excessive adjustments by the operator that could cause deviation in the scanning phase.

[0072] Specifically, minimum allowable time and maximum allowed time The settings are based on clinical experience with head CTA scans and the characteristics of the mobile CT equipment. For example... It lasts for 10 seconds. It is 30 seconds. If Below or higher ,but The function limits the output to or .

[0073] Specifically, The function rounds the input according to the countdown granularity of the vehicle-mounted CT control terminal. For example, if the countdown granularity is 1 second, Round will round the input to the nearest integer seconds; if the countdown granularity is 0.5 seconds, Round will round the input to the nearest 0.5 seconds.

[0074] It should be noted that if there is no CT timing correction branch, then the personalized delay time prediction value is obtained based on the patient-level physiological prediction delay time. : .

[0075] Understandably, during the training phase, the baseline delay prediction branch is trained using supervised learning labels, and the physiological residual correction branch is trained using residual labels, without going through the CT timing correction branch. During the application phase, the baseline delayed scan time is added to the nonlinear physiological residual correction amount to obtain the patient-level physiological prediction delay time, which is then input into the CT timing correction branch to obtain the scan trigger time. Output constraints are applied to the scan trigger time to obtain the individualized delay time prediction value.

[0076] Preferably, if the prediction model includes a CT timing correction branch, the composite loss function for training the baseline delay prediction branch and the physiological residual correction branch also includes a timing consistency loss term, which is used to constrain the deviation between the result of the patient-level physiological prediction delay time after CT timing correction and the actual scan trigger time recorded in the training samples.

[0077] Specifically, the time-series consistency loss term Represented as: ; In the formula, Indicates the first The time-series correction amount corresponding to each sample For the first The actual scan trigger time for each sample.

[0078] Specifically, during the forward computation in the training phase, the timing correction branch is invoked to compute... The timing consistency loss term is calculated through the timing correction branch. The result after time-series correction, and compared with Comparison. During backpropagation in the training phase, the gradient changes from... The baseline delay prediction branch and the physiological residual correction branch are returned and their parameters are updated. The device trigger response delay, scan duration, and target vessel position coefficient in the timing correction branch are device parameters or set coefficients and do not participate in the parameter update of backpropagation.

[0079] Understandably, the temporal consistency loss term enables the model to perceive the existence of the CT temporal correction branch during the training phase, constraining the result of the patient-level physiological prediction delay time after temporal correction to be consistent with the actual scan trigger time, thereby compensating for the systematic deviation between the training labels and the application target, and improving the model's generalization ability under different devices and scanning protocols.

[0080] Preferably, after the prediction model training is completed, one or more of the following should be saved: model version number, training sample size, input variable list, standardized parameters, missing value imputation rules, model parameters, output upper and lower limits, CT time-series correction parameters, and update time. Internal validation should be performed before deployment after each model update. If the updated model output exceeds the preset allowable range, prediction fluctuations are too large, or image quality feedback is unstable, then this model version should not be deployed, and the previous stable version should be retained.

[0081] During implementation, in step S3, the vehicle-mounted CT status includes one or more of the following: X-ray tube ready signal, X-ray tube rotation stabilization signal, high voltage generator charging completion signal, and detector temperature stabilization signal; the infusion pump status includes one or more of the following: infusion pump ready signal, contrast agent dosage confirmation signal, tubing venting completion signal, and power supply status signal; and the vehicle motion status includes one or more of the following: triaxial acceleration, triaxial angular velocity, vehicle speed, and parking status signal.

[0082] During implementation, in step S4, the preset ready conditions include: the vehicle-mounted CT status meets the requirements of tube X-ray ready, tube rotation stable, and high-voltage generator charging completed; the injection pump status meets the requirements of injection pump ready, pipeline venting completed, and power supply sufficient; and the vehicle movement status meets the preset stable conditions.

[0083] Specifically, the stability conditions include: the absolute values ​​of all three-axis accelerations are less than preset acceleration thresholds, and the absolute values ​​of all three-axis angular velocities are less than preset angular velocity thresholds. For example, the acceleration threshold is 0.05g, and the angular velocity threshold is 2° / s.

[0084] Understandably, since the contrast agent injection begins an irreversible countdown process, it is necessary to comprehensively assess whether the mobile CT scanner, the infusion pump, and the vehicle's movement status all meet the readiness conditions before injection. If it is discovered after injection that the equipment is not ready or the vehicle is moving, it will result in wasted contrast agent or scan failure.

[0085] Preferably, the generation of the injection feasibility signal can also be achieved using a learnable decision model. The inputs of the decision model include individualized delay time prediction, on-board CT readiness status, injection pump status, and vehicle motion status. The output of the decision model is an injection suggestion signal, which includes three categories: suggested injection, suggested waiting before injection, and not suggested injection.

[0086] Specifically, the decision model can be implemented using one or more of the following: multilayer perceptron, gradient boosting tree, random forest, or logistic regression. For example, the decision model includes an input layer, two fully connected hidden layers, and an output layer. The output layer uses a Softmax function to output the probability distributions of the three types of injection proposal signals.

[0087] Specifically, the decision model and the prediction model are jointly trained using a multi-task learning architecture. This architecture includes a shared bottom-level feature extraction layer and two task output heads. The first task output head outputs the individualized latency prediction value, and the second task output head outputs the injection proposal signal. The shared bottom-level feature extraction layer consists of a fully connected layer, a batch normalization layer, and an activation function.

[0088] Specifically, the training labels for the decision-making model are derived from actual clinical records, including one or more of the following: whether an injection was performed at the time, whether the post-injection scan was successful, and whether the image quality met the diagnostic requirements. For example, the training labels are determined according to the following rules: if an injection was performed and the scan was successful with image quality meeting the diagnostic requirements, the label is "Injection Recommended"; if an injection was performed but the scan failed or the image quality did not meet the diagnostic requirements, the label is "Injection Not Recommended"; if an injection was not performed at the time, the label is "Injection Recommended After Waiting".

[0089] Specifically, the training loss function of the multi-task learning architecture is a weighted sum of the time prediction error loss term and the classification loss term of the decision model. The classification loss term can be the cross-entropy loss.

[0090] In specific implementation, in step S4, if no injection initiation event is detected after outputting the injection feasibility signal, at least one of the following is executed: If the injection initiation event is not detected within the preset waiting time, return to step S3; The vehicle status information is continuously monitored. If the vehicle status information no longer meets the preset ready conditions, the injection feasibility signal is revoked and the process returns to step S3.

[0091] Specifically, the preset waiting time is set according to the clinical operation rhythm and the time requirements of the vehicle-mounted emergency rescue scenario, and is exemplarily set to 10 to 30 seconds, preferably 15 seconds.

[0092] During implementation, step S5, which involves performing a vehicle-mounted CT head CTA scan based on the injection feasibility signal and the individualized delay time prediction value, includes: Based on the injection feasibility signal, an injection prompt message is output, which is used to prompt the operator to perform the contrast agent injection. In response to an injection initiation event generated by the operator performing the injection, a countdown is started; wherein the countdown duration is the individualized delay time prediction value; After the countdown reaches zero, the onboard CT scanner will perform a CTA scan, either triggered or delayed, depending on the vehicle's motion status.

[0093] In practice, after the countdown reaches zero, the onboard CT scanner is triggered or delayed to perform a CTA scan based on the vehicle's motion status, including: After the countdown reaches zero, the vehicle's motion status is obtained; If the vehicle's motion meets the preset stable conditions, the onboard CT scanner will be triggered to perform a CTA scan. If the vehicle's motion state does not meet the preset stable conditions, it enters a waiting state and continuously monitors the vehicle's motion state for the preset maximum waiting time. If the vehicle returns to stable within the maximum waiting time, a scan is triggered; otherwise, the operator is prompted to make a selection.

[0094] Specifically, the longest waiting time The plateau phase width of arterial enhancement in head CTA is dynamically set to meet the following requirements: .

[0095] In the formula, The plateau width of enhancement in the target artery of the head. This is the preset maximum waiting time.

[0096] It should be noted that the predicted delay time and injection feasibility signal obtained in this application are intermediate information processing results, intended for the operator's (physician's) reference and decision-making. The injection action is performed autonomously by the operator, who only responds to the injection initiation event and executes subsequent device control logic.

[0097] Compared with existing technologies, this embodiment provides a delay time prediction and decision-making method for mobile head-mounted CTA. Through a hierarchical data flow model of baseline delay prediction, physiological residual correction, and CT timing correction, it outputs individualized delay time prediction values ​​without performing additional test bolus or real-time ROI bolus tracking for each case. Before injection, it outputs an injection feasibility signal by comprehensively considering the mobile CT status, vehicle motion status, and infusion pump status, avoiding blind injection under conditions of equipment incomplete preparation or vehicle vibration. Simultaneously, by responding to the injection initiation event to start a countdown, and after the countdown reaches zero, it triggers or delays the mobile CT head-mounted CTA scan based on the vehicle motion status, reducing contrast agent waste and the risk of repeated scans, and improving the standardization, safety, and operability of the mobile stroke unit mobile head-mounted CTA process.

[0098] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0099] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A latency prediction and decision-making method for vehicle-mounted head unit (CTA), characterized in that, Includes the following steps: S1. Obtain the patient's clinical parameters and vehicle-mounted CT scan parameters, and preprocess the clinical parameters to obtain a standardized feature vector; S2. Input the standardized feature vector and vehicle-mounted CT scan parameters into the trained prediction model to obtain the individualized delay time prediction value. The individualized delay time prediction value is used to determine the time interval between the start of contrast agent injection and the start of CT scan. S3. Obtain the current vehicle status information of the mobile stroke unit, including the vehicle CT status, vehicle motion status, and infusion pump status. S4. Determine whether the current vehicle status information meets the preset ready conditions. If yes, output an injection feasible signal; otherwise, output an injection infeasible signal and return to step S3 after the set collection interval to continue execution. S5. Based on the injection feasibility signal and the individualized delay time prediction value, control the on-board CT execution head CTA scan.

2. The latency prediction and decision-making method for vehicle-mounted head-up camera (CTA) according to claim 1, characterized in that, The prediction model includes: The baseline delay prediction branch is used to output the baseline delay scan time based on the first set of parameters in the standardized feature vector; The physiological residual correction branch is used to output a nonlinear physiological residual correction amount based on the second set of parameters in the standardized feature vector. The time-delay fusion module is used to add the basic delayed scan time to the nonlinear physiological residual correction amount to obtain the patient-level physiological prediction delay time; the patient-level physiological prediction delay time is used to determine the individualized delay time prediction value.

3. The latency prediction and decision-making method for vehicle-mounted head unit (CTA) according to claim 2, characterized in that, The prediction model also includes a CT timing correction branch, which is used to output the scan trigger time based on the patient-level physiological prediction delay time and the on-board CT scan parameters; the scan trigger time is used to determine the individualized delay time prediction value.

4. The latency prediction and decision-making method for vehicle-mounted head unit (CTA) according to claim 3, characterized in that, The scan trigger time Represented as: ; in, ; In the formula, This indicates the delay in physiological prediction at the patient level. Indicates timing correction amount, This indicates a device trigger response delay. Indicates the scan duration. This represents the target blood vessel location coefficient.

5. The latency prediction and decision-making method for vehicle-mounted head-up camera (CTA) according to claim 4, characterized in that, The individualized delay time prediction value is determined based on the scan trigger time using the following method. : ; In the formula, This indicates the preset fine-tuning amount. , These represent the minimum and maximum allowed times, respectively. Represents the amplitude limiting function. This represents the floor function.

6. The latency prediction and decision-making method for vehicle-mounted head-up camera (CTA) according to claim 2, characterized in that, The input to the baseline delay prediction branch is the first set of parameters in the standardized feature vector, including age, heart rate, mean arterial pressure, pulse pressure, and BMI; the baseline delay prediction branch uses a linear regression, ridge regression, or Lasso regression model, and its prediction expression is: ; In the formula, Indicates the baseline delayed scan time. Represents the intercept term. , , , , Indicates the parameters of each model. , , , , These represent the values ​​of age, heart rate, mean arterial pressure, pulse pressure, and BMI in the standardized feature vector, respectively.

7. The latency prediction and decision-making method for vehicle-mounted head-up camera (CTA) according to claim 2, characterized in that, The input to the physiological residual correction branch is the second set of parameters in the standardized feature vector, including heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure and pulse pressure, height, weight, BMI, gender, heart rhythm status, contrast agent injection flow rate, contrast agent dosage and saline flushing parameters. The physiological residual correction branch includes: The parameter segmentation layer is used to obtain the circulatory dynamics feature group, body size-blood volume feature group, and injection protocol feature group based on the second set of parameters; A cyclic dynamics coding layer is used to convert the received cyclic dynamics feature set into a cyclic dynamics embedding vector; A body size-blood volume encoding layer is used to convert the received body size-blood volume feature set into a body size embedding vector; An injection scheme encoding layer is used to convert the received injection scheme feature set into an injection scheme embedding vector; The fusion layer is used to fuse the cyclic dynamics embedding vector, body shape embedding vector, and injection scheme embedding vector to obtain fused features; A residual prediction network layer is used to obtain a nonlinear physiological residual correction amount based on the fusion features.

8. The latency prediction and decision-making method for vehicle-mounted head-up CTA according to claim 1, characterized in that, The prediction model is trained using a composite loss function, which includes one or more of the following: time prediction error loss term, range constraint loss term, physiological trend constraint loss term, and input perturbation smoothing loss term. The range constraint loss term penalizes the model output when it is below the minimum allowed time or above the maximum allowed time. The physiological trend constraint loss term penalizes the model output when it violates a preset physiological monotonicity. The input perturbation smoothing loss term constrains the change in model output to not exceed a preset threshold when the input variable experiences a perturbation of a preset magnitude.

9. The latency prediction and decision-making method for vehicle-mounted head-up camera (CTA) according to claim 1, characterized in that, After outputting the injection feasible signal, if no injection initiation event is detected to generate an injection, then at least one of the following is performed: If the injection initiation event is not detected within the preset waiting time, return to step S3; The vehicle status information is continuously monitored. If the vehicle status information no longer meets the preset ready conditions, the injection feasibility signal is revoked and the process returns to step S3.

10. The delay prediction and decision-making method for vehicle-mounted head-up camera (CTA) according to claim 1, characterized in that, The method of controlling the onboard CT to perform head CTA scanning based on injection feasibility signals and individualized delay time predictions includes: Based on the injection feasibility signal, an injection prompt message is output, which is used to prompt the operator to perform the contrast agent injection. In response to an injection initiation event generated by the operator performing the injection, a countdown is started; wherein the countdown duration is the individualized delay time prediction value; After the countdown reaches zero, the onboard CT scanner will perform a CTA scan, either triggered or delayed, depending on the vehicle's motion status.