A method and system for tube flow monitoring for ECMO

Abstract

The present application belongs to the technical field of medical monitoring, and particularly relates to a pipe diameter flow monitoring method and system for ECMO, which comprises the following steps: a video image of an ECMO pipe is acquired, a dynamic contour model containing a statistical shape model and a Kalman filter is constructed to adapt to non-rigid deformation of the pipe and track the non-rigid deformation; macro tracking features such as model fitting residuals and kinematic deviations, and micro features such as textures and optical flows inside the pipe are combined to calculate a multi-dimensional tracking confidence, which is used to adaptively adjust parameters of the Kalman filter, so that optimal compensation for pipe vibration and deformation is realized; furthermore, the system can accurately extract the pipe diameter and flow rate, and calculate real-time instantaneous flow. The accuracy, robustness and intelligent level of ECMO monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology. More specifically, this invention relates to a method and system for monitoring the diameter and flow rate of ECMO tubes. Background Technology

[0002] As a critical life support technology, the stability and patency of blood flow within ECMO tubing are essential for patient safety. Therefore, real-time and accurate monitoring of ECMO tubing diameter and internal blood flow status through non-contact methods such as video image analysis is of great significance for preventing complications such as thrombosis.

[0003] In clinical applications of ECMO, the tubing not only experiences high-frequency vibrations due to pump operation but may also undergo non-rigid deformations such as twisting and rotation due to changes in patient positioning or nursing procedures. This complex dynamic characteristic presents a significant challenge to accurate vision-based tracking. To address these issues, existing technologies typically employ dynamic contour models for target tracking. These models usually combine a statistical shape model to describe non-rigid deformations with a Kalman filter to predict the object's motion. The statistical shape model effectively handles tubing bending and torsion, while the Kalman filter predicts and smooths rigid movements such as translation and rotation.

[0004] However, traditional Kalman filter-based dynamic contour tracking methods have inherent limitations when applied to ECMO pipeline monitoring. These methods typically employ fixed noise parameters and cannot dynamically adjust their trust in model predictions and actual observations based on real-time fluctuations in observation quality caused by factors such as motion blur, partial occlusion, or changes in lighting in the video image. When image quality deteriorates or observation data becomes unreliable, fixed parameter settings can cause the filter to over-rely on noisy data, leading to tracking deviations or even target loss. Conversely, when observation quality is good, over-reliance on the prediction model may prevent timely responses to changes in the pipeline's actual condition, impacting the accuracy and robustness of monitoring. Summary of the Invention

[0005] To address the aforementioned technical problems of insufficient accuracy and robustness in ECMO pipe diameter and flow monitoring, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for monitoring the pipe diameter and flow rate of an ECMO, comprising:

[0007] Acquire and track video image sequences of the ECMO tubing; the tracking includes: constructing a dynamic contour model comprising a statistical shape model and a Kalman filter, wherein the statistical shape model is used to handle non-rigid deformation of the tubing contour; acquiring the observation parameters of the current frame; calculating the Mahalanobis distance between the observation parameters and the predicted parameters of the Kalman filter to obtain the kinematic bias; calculating the mean distance between the model contour and the edge in the current frame to obtain the model fitting residual; extracting the Region of Interest (ROI) within the model contour and acquiring the texture features and optical flow features within the ROI; calculating the multidimensional tracking confidence score, wherein the multidimensional tracking confidence score is negatively correlated with the model fitting residual and kinematic bias, and positively correlated with the health status of the texture features and optical flow features; adaptively adjusting the measurement noise covariance of the Kalman filter using the multidimensional tracking confidence score; extracting the precise tubing diameter based on the posterior state of the adjusted Kalman filter; obtaining the flow velocity based on the optical flow features; and calculating the instantaneous flow rate based on the precise tubing diameter and flow velocity.

[0008] This invention constructs a dynamic contour model incorporating a statistical shape model and a Kalman filter, enabling effective tracking and adaptation to the non-rigid deformation of pipelines. It introduces texture and optical flow features within the contour and combines model fitting residuals and kinematic bias to construct a multidimensional tracking confidence level. This confidence level not only assesses the accuracy of macroscopic tracking but also monitors microscopic health conditions within the blood, such as potential thrombosis risk. Finally, by adaptively adjusting the measurement noise of the Kalman filter using this confidence level, dynamic optimization is achieved, suppressing noise when observation quality is poor and improving accuracy when observation quality is good. This results in more accurate, robust, and information-rich pipe diameter and flow monitoring results.

[0009] Preferably, the construction of the statistical shape model includes:

[0010] Collect multiple pipeline profile samples containing non-rigid deformation; obtain the coordinates of key points from each profile sample;

[0011] Align the coordinates of all key points and apply principal component analysis to obtain the mean shape vector and deformation basis matrix;

[0012] The statistical shape model is composed of a mean shape vector, a deformation basis matrix, and a shape parameter vector. The shape parameter vector is used to control the deformation of the contour along the deformation basis direction.

[0013] This invention collects various pipeline contour samples and uses principal component analysis to construct a statistical shape model, which can describe and constrain the reasonable non-rigid deformation space of ECMO pipelines with low-dimensional parameters. This allows the model to not only adapt to the bending and torsion of the pipeline, but also ensure the rationality of the deformation, avoiding abnormal contours that do not conform to physical constraints during the tracking process. This improves the accuracy and stability of the dynamic contour model when dealing with complex deformations.

[0014] Preferably, the calculation of the multidimensional tracking confidence includes:

[0015] The model fitting residuals and kinematic biases together constitute the macroscopic tracking features; the texture features and optical flow features together constitute the microscopic internal features.

[0016] The multidimensional tracking confidence level is the product of the macroscopic tracking feature confidence level and the microscopic internal feature confidence level.

[0017] Preferably, the calculation of the confidence level of the microscopic internal features includes:

[0018] Texture confidence is obtained based on the texture features, and optical flow confidence is obtained based on the optical flow features;

[0019] The geometric mean of the texture confidence and optical flow confidence is used as the confidence of the microscopic internal features.

[0020] This invention calculates the confidence level of microscopic internal features by geometrically averaging the confidence levels of texture and optical flow. This allows for more sensitive detection of deterioration in any microscopic dimension. When any abnormality in texture or flow field causes a sharp decrease in confidence level, the geometric mean is more significantly affected. This improves the system's sensitivity to detecting early pathological features such as thrombosis and turbulence, and avoids the problem of normal indicators masking abnormalities in other indicators.

[0021] Preferably, the texture feature includes the LBP histogram of the image within the ROI; the texture confidence is negatively correlated with the Bach distance between the current LBP histogram and the initial LBP histogram.

[0022] Preferably, the optical flow feature includes the directional distribution of multiple optical flow vectors within the ROI; the optical flow confidence is negatively correlated with the entropy value of the directional distribution.

[0023] Preferably, the measurement noise covariance of the adaptively adjusted Kalman filter includes:

[0024] ;

[0025] in, The measurement noise covariance at the current moment; The reference noise covariance obtained from calibration; The preset maximum noise covariance; The confidence level for multidimensional tracking at the current moment.

[0026] This invention uses multidimensional tracking confidence as a dynamic adjustment factor to realize real-time adjustment of the filter's confidence in the observed data. When the confidence is high, the observation is trusted, and when the confidence is low, the prediction is trusted. This mechanism enhances the robustness of the filter in the face of video noise, partial occlusion, or internal flow field anomalies and the smoothness of the output results.

[0027] Preferably, the extraction of the precise pipe diameter includes:

[0028] Based on the shape and pose parameters of the posterior state, a corrected contour is generated;

[0029] Calculate the distance between multiple pairs of corresponding points along the centerline normal direction of the corrected profile, and use the average distance as the accurate pipe diameter.

[0030] This invention uses the results of dynamic model smoothing and correction, which can effectively filter out instantaneous high-frequency noise caused by pump vibration, etc., making the pipe diameter measurement results more stable and accurate, and providing a reliable basis for subsequent accurate flow calculation.

[0031] Preferably, the method further includes preprocessing the video image sequence, including:

[0032] Convert each frame of the image to a grayscale image;

[0033] A multi-layer image pyramid is constructed for each grayscale image, and the dynamic contour model is tracked starting from the top layer of the pyramid.

[0034] Secondly, the present invention provides a pipe diameter flow monitoring system for ECMO, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned pipe diameter flow monitoring method for ECMO is implemented.

[0035] By adopting the above technical solution, a computer program for monitoring pipe diameter and flow rate of ECMO is generated and stored in a memory so that it can be loaded and executed by a processor. A terminal device can then be made based on the memory and processor for convenient use.

[0036] The beneficial effects of this invention are as follows:

[0037] (1) This invention avoids the interference of non-rigid deformation of pipeline on tracking by constructing a dynamic model that combines statistical shape model and Kalman filter;

[0038] (2) This invention integrates macroscopic motion tracking information with microscopic texture and optical flow characteristics inside the pipeline to construct a multidimensional confidence assessment system. This system can not only be used to adaptively optimize the filtering process and improve monitoring accuracy and robustness, but also assess the blood health status in real time and provide early warning of risks such as thrombosis for clinical use. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a method for monitoring pipe diameter flow in ECMO according to the present invention;

[0040] Figure 2 This is a schematic diagram illustrating the change in the final instantaneous flow rate. Detailed Implementation

[0041] This invention discloses a method for monitoring pipe diameter and flow rate in ECMO, referring to... Figure 1 This includes steps S1-S4:

[0042] S1: Based on the video image sequence of the ECMO tubing, a structured dynamic contour model is constructed and initialized through statistical shape analysis.

[0043] It should be noted that in the actual clinical environment of ECMO monitoring, the tubing not only experiences high-frequency translation and vibration due to the pump, but may also undergo slight twisting and rotation due to changes in patient position and nursing procedures. Traditional rigid template matching methods based on pixel grayscale values ​​are quite sensitive to these types of deformations; once non-rigid deformation occurs, the matching score decreases, leading to tracking failure. Considering that although ECMO tubing undergoes non-rigid deformation, its deformation is not completely random but follows the constraints of its physical structure, possessing a reasonable deformation space, this invention uses principal component analysis to learn and establish a statistical shape model from training samples that can describe its legal deformation space with low-dimensional parameters, enabling the model to adapt to non-rigid deformations.

[0044] Specifically, based on the video image sequence of the ECMO tubing, a structured dynamic contour model is constructed and initialized through statistical shape analysis, including:

[0045] First, using an industrial camera fixed to the ECMO device or bedside, real-time video streams of specific monitoring segments of the ECMO tubing are acquired at a preset frame rate to obtain a raw image sequence. For example, the preset frame rate is 30fps. To improve the efficiency and robustness of subsequent processing, each frame in the sequence is preprocessed, including converting it to grayscale and constructing an L-layer image pyramid. For example, L is set to 3, containing three layers of images: original resolution, 1 / 2 resolution, and 1 / 4 resolution.

[0046] Secondly, a statistical shape model is constructed: a predetermined number of tubing contour samples representing different ECMO operating states are collected, such as vibrations at different pump speeds and bending and torsion caused by changes in patient position. For example, the predetermined number is 50. M key points are uniformly sampled for each tubing contour sample to form a coordinate matrix. For example, the number of key points is 100. The coordinate matrices of all tubing contour samples are aligned to eliminate differences in translation, rotation, and scaling, thus obtaining shape information. Principal component analysis is applied to the hypervector formed by the coordinates of all aligned key points to calculate the average shape vector and covariance matrix. Eigenvalue decomposition is performed on the covariance matrix, and the principal component eigenvectors corresponding to the k largest eigenvalues ​​are selected to form the deformation basis matrix. For example, k=5. The tubing contour samples are automatically extracted from the initial video frames using methods such as Canny edge detection and Hough transform. Canny edge detection, Hough transform, and principal component analysis are existing technologies and will not be elaborated upon here.

[0047] Therefore, the shape information of any pipeline profile sample satisfies the expression:

[0048] ;

[0049] In the formula, It is a 2M-dimensional column vector containing the coordinates of all key points; It is the average shape vector; It is from the front The principal component eigenvectors corresponding to the largest eigenvalues ​​constitute the structure. matrix; It is a k-dimensional shape parameter vector. It should be noted that... The column vectors define the main legal deformation modes of the shape, such as bending and twisting.

[0050] In the formula, Decompose a high-dimensional, complex contour shape into stable average shapes. and consisting of k parameters Controlled linear deformation combination, by adjusting the parameter vector By using the learned components, various reasonable pipeline profiles can be generated within the legal deformation space, thus enabling the model to describe and adapt to non-rigid deformations.

[0051] Finally, the dynamic model is initialized: a Kalman filter is used to model and predict the dynamic process of the pipeline, and its state vector is defined as follows: This includes the shape parameters of the target. Global translation Scaling Rotation and its corresponding first rate of change , , , Initialize the state vector based on the detection results of the first frame image. and its covariance matrix.

[0052] Thus, the image pyramid of the original image sequence and the structured dynamic contour model containing the statistical shape model and the initial Kalman filter were obtained.

[0053] S2: Based on multi-scale images and dynamic contour models, preliminary structural and motion features are obtained through model-driven matching and kinematic prediction.

[0054] It should be noted that after obtaining the deformable structured dynamic contour model, it needs to be matched with the newly generated image. Independent model fitting processes may result in geometrically plausible but kinematically abnormal performance due to local image noise or occlusion; therefore, results obtained solely through geometric fitting are unreliable. Considering that the pump's drive makes the pipeline's motion continuous and predictable, reliable observation results should not only closely match the image edges geometrically in the current frame but also conform to the motion patterns predicted from historical states. Therefore, this invention not only calculates the geometric fitting residuals between the model and the image but also introduces Mahalanobis distance to calculate the statistical deviation between the observed state and the kinematically predicted state in a high-dimensional parameter space, thereby simultaneously evaluating both geometric fit and kinematic conformance.

[0055] Specifically, based on multi-scale images and dynamic contour models, preliminary structural and motion features are obtained through model-driven matching and kinematic prediction, including:

[0056] Model-driven contour matching is performed on the t-th frame of the original image sequence, and the model fitting residual is obtained: Starting from the top layer of the pyramid with the lowest resolution, the shape predicted by the Kalman filter is used as the initial contour; for each keypoint of the current contour, the pixel with the strongest gradient is found as the new candidate position of the keypoint along its normal direction in a one-dimensional search line; after collecting the new candidate positions of all keypoints, a new set of shape and pose parameters is found by solving a least squares problem, such that the sum of the distances between the contour generated by these parameters and all candidate positions is minimized; the fitted contour is projected onto the next layer of the pyramid until matching is completed at the original resolution layer; at the original resolution layer, the average Euclidean distance between the model-fitted contour and the real edge of the image is calculated and denoted as the model fitting residual.

[0057] Kinematic prediction and bias quantization are performed on the t-th frame of the original image sequence to obtain the kinematic bias: using the state transition equation of the Kalman filter, based on the optimal posterior state at time t-1... Predict the prior state at time t and its covariance matrix The new shape and pose parameters are used to construct an observation parameter vector. ;Calculate the Mahalanobis distance between the observed parameter vector and the prior prediction vector as the kinematic bias.

[0058] Kinematic deviations satisfy the expression:

[0059] ;

[0060] In the formula, This represents the kinematic deviation at time t; This represents the vector of observation parameters at time t; Let represent the prior parameter vector at time t; Let represent the prediction covariance matrix at time t; This indicates that the vector is transposed.

[0061] In the formula, The observation parameter vector was measured in a high-dimensional parameter space. Deviation from prior parameter vector The statistical distance, and by Normalization was performed, taking into account the correlation and uncertainty between various parameters. Not only were deviations in position and size calculated, but deviations in shape parameters were also considered, thus enabling the detection of whether non-rigid deformation conforms to its historical motion law.

[0062] Thus, the model fitting residuals describing geometric fit and the kinematic bias describing the fit of motion laws were obtained.

[0063] S3: Based on the preliminary structure and motion characteristics, obtain multidimensional tracking confidence through internal feature mining and hierarchical confidence modeling.

[0064] It should be noted that the model fitting residuals and kinematic biases primarily analyze the macroscopic geometry and motion state of the target. However, these macroscopic features are insensitive to changes in the patient's physiological state, such as thrombosis, leading to changes in internal microscopic texture and internal flow field structure. Even if the tubing is correctly tracked macroscopically, lesions indicating potential risks may have already occurred inside. Considering that healthy blood flow has relatively stable microscopic texture and consistent flow patterns within the tubing, this invention constructs a hierarchical confidence model. First, it calculates the structural tracking confidence based on macroscopic features as a foundation. Then, it calculates the internal integrity score based on internal microscopic features. Finally, it uses the internal integrity score to verify and modulate the structural tracking confidence, forming a comprehensive confidence model that considers both macroscopic and microscopic aspects and is logically progressive.

[0065] Specifically, based on the preliminary structure and motion characteristics, multi-dimensional tracking confidence is obtained through internal feature mining and hierarchical confidence modeling, including:

[0066] First, internal texture and flow field features are extracted to obtain texture drift and flow field structure entropy: Based on the precise contour, the blood region image within it is extracted as the region of interest (ROI), and the local binary pattern histogram of the ROI is calculated. The texture drift is obtained by calculating the Bach distance between the current local binary pattern histogram and the template local binary pattern histogram stored during initialization. Within the ROI, a sparse optical flow algorithm is used to calculate the motion vectors of a set of feature points, and the directions of these motion vectors are statistically analyzed to construct an angular distribution histogram. Its information entropy is then calculated to obtain the flow field structure entropy. It should be noted that under healthy laminar flow conditions, the optical flow direction is unidirectional, and the entropy value is low; when turbulence or thrombosis occurs, the flow field becomes chaotic, and the entropy value increases significantly.

[0067] Secondly, a hierarchical multidimensional confidence model is constructed, including:

[0068] Geometric and kinematic confidence scores are obtained, and their product is denoted as the structural tracking confidence score. These scores are derived from the model fitting residuals and kinematic biases through online statistical learning of their mean and standard deviation during the stable tracking phase, and then transformed using a chi-square cumulative distribution function. It should be noted that the structural tracking confidence score integrates two macroscopic tracking indicators, jointly reflecting whether the target as a whole is being correctly tracked.

[0069] The internal integrity score is obtained by square rooting the product of texture confidence and flow field confidence. Texture confidence and flow field confidence are derived through online statistical normalization and probability transformation based on texture drift and flow field structure entropy, respectively. It should be noted that a geometric mean is used to fuse the two internal micro-features. Compared to the arithmetic mean, the geometric mean more severely penalizes any significant deterioration in a single dimension, avoiding the risk of one normal indicator masking another failing indicator. Compared to direct multiplication, the square root operation, while preserving logical rigor, restores the fused score to a scale close to the input value, enhancing the model's numerical stability and interpretability.

[0070] The product of structural tracking confidence and internal integrity score is denoted as multidimensional tracking confidence. This hierarchical model reflects a logical progression. The final multidimensional tracking confidence is based on the macroscopic structural tracking confidence and is verified and modulated by the microscopic internal integrity score. A tracking result that is good in geometry and motion will lead to a decrease in internal integrity score if its internal texture or flow field is abnormal. The multidimensional tracking confidence will also decrease, thus making the entire decision-making system more robust and intelligent.

[0071] Thus, a hierarchical multidimensional tracking confidence score was obtained, capable of comprehensively assessing the health of non-rigid deformation, motion, texture, and flow field.

[0072] S4: Based on the multidimensional tracking confidence, an accurate instantaneous flow sequence is obtained through adaptive dynamic compensation and flow calculation.

[0073] It should be noted that traditional Kalman filters use fixed noise parameters and cannot dynamically adjust their confidence in predictions and observations based on real-time changes in observation quality. This makes them susceptible to noise interference when observation quality is poor, and inaccurate when observation quality is good. Considering that the confidence level of multidimensional tracking is directly related to the reliability of observation data, this invention uses the multidimensional tracking confidence level as a feedback signal to dynamically adjust the measurement noise covariance of the Kalman filter. This allows the filter to trust observations more when confidence is high and rely more on predictions when confidence is low, thereby achieving optimal compensation for vibration and deformation.

[0074] Specifically, based on multidimensional tracking confidence, an accurate instantaneous traffic sequence is obtained through adaptive dynamic compensation and traffic calculation, including:

[0075] First, a confidence-adaptive Kalman filter update is performed, and the dynamically adjusted measurement noise covariance matrix is ​​obtained. Satisfying the expression:

[0076] ;

[0077] In the formula, Let be the measurement noise covariance matrix at the current moment; The reference noise covariance is obtained through system calibration; The preset maximum noise covariance; The confidence level for multidimensional tracking at the current moment.

[0078] In the formula, when the confidence level of multidimensional tracking is large, Approaching , near This makes the filter more confident in the current observation; when the confidence level... When very low, Approaching , near This allows the filter to almost completely ignore the current observation and rely more on its own kinematic predictions, thus ensuring the smoothness of the output.

[0079] The instantaneous observation parameters and the dynamically adjusted measurement noise covariance are input into the Kalman filter, and the posterior state estimate at the current moment is output.

[0080] Secondly, precise pipe diameter extraction and flow rate calculation are performed: using the shape and pose parameters in the optimal posterior state estimation, a precise, vibration-free pipe profile that accurately reflects the current non-rigid deformation is generated; on the pipe profile, along the direction of its centerline normal, the distance between multiple pairs of corresponding points is calculated, and the average value is taken to obtain the precise instantaneous pipe diameter; from the optical flow field calculation in S3, the average blood flow velocity within the ROI is obtained.

[0081] The final instantaneous flow rate satisfies the expression:

[0082] ;

[0083] In the formula, Indicates the final instantaneous flow rate; Indicates average blood flow velocity; Indicates the precise pipe diameter; It represents pi (π).

[0084] It should be noted that, as Figure 2 The diagram shows the change in the final instantaneous flow rate. In the gray area after 100 frames, the instantaneous flow rate drops significantly, indicating the presence of thrombosis.

[0085] Thus, the Kalman filter was adaptively improved using multidimensional tracking confidence, achieving optimal compensation for vibration and non-rigid deformation, and based on this, the final accurate instantaneous flow sequence was obtained.

[0086] This invention also discloses a pipe diameter flow monitoring system for ECMO, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a pipe diameter flow monitoring method for ECMO according to the present invention.

[0087] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0088] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for monitoring pipe diameter flow rate in ECMO, characterized in that, include: Acquire and track video image sequences of the ECMO tubing; the tracking includes: A dynamic profile model is constructed, which includes a statistical shape model and a Kalman filter. The statistical shape model is used to handle the non-rigid deformation of the pipeline profile. Obtain the observation parameters of the current frame; calculate the Mahalanobis distance between the observation parameters and the Kalman filter prediction parameters to obtain the kinematic bias; calculate the mean distance between the model contour and the edge in the current frame to obtain the model fitting residual; Extract the Region of Interest (ROI) within the model contour, and obtain the texture features and optical flow features within the ROI; calculate the multidimensional tracking confidence score, which is negatively correlated with the model fitting residual and kinematic bias, and positively correlated with the health status of the texture features and optical flow features; the calculation of the multidimensional tracking confidence score includes: The model fitting residuals and kinematic biases together constitute the macroscopic tracking features; the texture features and optical flow features together constitute the microscopic internal features. The multidimensional tracking confidence level is the product of the macroscopic tracking feature confidence level and the microscopic internal feature confidence level; The measurement noise covariance of the Kalman filter is adaptively adjusted using the multidimensional tracking confidence; the precise pipe diameter is extracted based on the posterior state of the adjusted Kalman filter; and the flow velocity is obtained based on the optical flow characteristics. The instantaneous flow rate is calculated based on the precise pipe diameter and flow velocity.

2. The method for monitoring pipe diameter and flow rate for ECMO according to claim 1, characterized in that, The construction of the statistical shape model includes: Collect multiple pipeline profile samples containing non-rigid deformation; obtain the coordinates of key points from each profile sample; Align the coordinates of all key points and apply principal component analysis to obtain the mean shape vector and deformation basis matrix; The statistical shape model is composed of a mean shape vector, a deformation basis matrix, and a shape parameter vector. The shape parameter vector is used to control the deformation of the contour along the deformation basis direction.

3. The method for monitoring pipe diameter and flow rate for ECMO according to claim 1, characterized in that, The calculation of the confidence level of the microscopic internal features includes: Texture confidence is obtained based on the texture features, and optical flow confidence is obtained based on the optical flow features; The geometric mean of the texture confidence and optical flow confidence is used as the confidence of the microscopic internal features.

4. The method for monitoring pipe diameter and flow rate for ECMO according to claim 3, characterized in that, The texture features include the LBP histogram of the image within the ROI; the texture confidence is negatively correlated with the Bach distance between the current LBP histogram and the initial LBP histogram.

5. The method for monitoring pipe diameter and flow rate for ECMO according to claim 3, characterized in that, The optical flow features include the directional distribution of multiple optical flow vectors within the ROI; the optical flow confidence is negatively correlated with the entropy value of the directional distribution.

6. The method for monitoring pipe diameter and flow rate for ECMO according to claim 1, characterized in that, The measurement noise covariance of the adaptively adjusted Kalman filter includes: ； in, The measurement noise covariance at the current moment; The reference noise covariance obtained from calibration; The preset maximum noise covariance; The confidence level for multidimensional tracking at the current moment.

7. The method for monitoring pipe diameter and flow rate for ECMO according to claim 1, characterized in that, The extraction of the precise pipe diameter includes: Based on the shape and pose parameters of the posterior state, a corrected contour is generated; Calculate the distance between multiple pairs of corresponding points along the centerline normal direction of the corrected profile, and use the average distance as the accurate pipe diameter.

8. The method for monitoring pipe diameter and flow rate for ECMO according to claim 1, characterized in that, It also includes preprocessing the video image sequence, including: Convert each frame of the image to a grayscale image; A multi-layer image pyramid is constructed for each grayscale image, and the dynamic contour model is tracked starting from the top layer of the pyramid.

9. A pipe diameter flow monitoring system for ECMO, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a pipe diameter flow monitoring method for ECMO according to any one of claims 1-8.

Images