Cardiac magnetic resonance dynamic imaging analysis method and system

Abstract

The application relates to a cardiac magnetic resonance dynamic image analysis method and system. Cardiac magnetic resonance dynamic image data is acquired; a plurality of types of anchor point markers are established in initial cardiac magnetic resonance dynamic image data; analysis and detection are performed on the cardiac magnetic resonance dynamic image data in which the plurality of types of anchor point markers are established; when it is detected that an anchor point marker has an abnormality, an anchor point correction process is triggered to correct the anchor point marker with the abnormality. By establishing the plurality of types of anchor point markers including an anatomical anchor point marker, a functional anchor point marker and a texture anchor point marker, the cardiac magnetic resonance dynamic image data can be analyzed based on a multi-type anchor point cooperative tracking mechanism. When body motion interference occurs during the analysis of the cardiac magnetic resonance dynamic image data, the offset can be automatically corrected, the anatomical position accuracy of myocardial segment markers is ensured, the need for manual re-marking is avoided, and the efficiency and reliability of cardiac function quantification analysis are improved.

Description

Technical Field

[0001] This application relates to the field of intelligent medical image analysis technology, and in particular to a method and system for dynamic analysis of cardiac magnetic resonance images. Background Technology

[0002] Dynamic imaging analysis of cardiac magnetic resonance imaging (MRI) is a core tool for clinical assessment of cardiac structure and function. The accuracy of myocardial segment tracking directly determines the reliability of key quantitative indicators such as left ventricular ejection fraction and myocardial thickening rate. However, dynamic imaging analysis of cardiac MRI still faces significant technical limitations. On the one hand, the acquired dynamic cardiac MRI images may contain anatomical shifts between adjacent frames due to slight body movements. Current techniques for analyzing dynamic cardiac MRI images mostly rely on manually marking keyframes to address these shifts. When tracking is lost, manual remarking is required. This manual marking method suffers from problems such as high labor costs, low analysis efficiency, and inconsistent analysis results due to differences in the analysts. Summary of the Invention

[0003] Therefore, it is necessary to address the above-mentioned problems by providing a method and system for dynamic imaging analysis of cardiac magnetic resonance imaging.

[0004] In a first aspect, this application provides a method for dynamic analysis of cardiac magnetic resonance imaging, the method comprising:

[0005] Acquire dynamic cardiac magnetic resonance imaging data;

[0006] Multiple types of anchor markers are established in the initial cardiac magnetic resonance dynamic imaging data, including anatomical anchor markers, functional anchor markers, and texture anchor markers;

[0007] The dynamic cardiac magnetic resonance imaging data with multiple anchor point markers were analyzed and detected. When an anomaly was detected in the anchor point marker, an anchor point correction process was triggered to correct the abnormal anchor point marker. The correction process includes: identifying the type of abnormal anchor point marker; if the anatomical anchor point marker is abnormal, correcting the abnormal anatomical anchor point marker based on a thin-plate spline deformation model; if the functional anchor point marker is abnormal, correcting the abnormal functional anchor point marker based on an adaptive Kalman filter algorithm; and if the texture anchor point marker is abnormal, correcting the abnormal texture anchor point marker based on an improved RANSAC-SIFT algorithm.

[0008] In some embodiments, anatomical anchor point markers are established in the initial cardiac magnetic resonance imaging data, including:

[0009] The cardiac magnetic resonance imaging data is divided into data blocks of the required volume.

[0010] The cardiac chamber structure in the data block is segmented using a 3D U-Net network to obtain a valve annulus segmentation probability map;

[0011] Based on the region of interest in the lobe ring segmentation probability map, a ResNet-50 network is used for fine localization to obtain an anchor point probability heatmap.

[0012] Based on the anchor point probability heatmap, the precise coordinates of the anatomical anchor points are calculated through subpixel interpolation; the anatomical anchor points include the center point of the aortic valve annulus and the anterolateral junction of the mitral valve annulus.

[0013] In some embodiments, establishing functional anchor markers in the initial cardiac magnetic resonance imaging data includes:

[0014] Extract the left ventricular apex coordinate sequence from the first N frames of the initial cardiac magnetic resonance dynamic imaging data;

[0015] A quadratic polynomial prediction model is constructed based on the left ventricular apex coordinate sequence of the first N frames;

[0016] Prediction is performed based on the quadratic polynomial prediction model to obtain the coordinates of the left ventricular apex in the next frame.

[0017] The motion trajectory of the midpoint of the interventricular septum was obtained by optical flow tracking and verified by changes in myocardial thickness to obtain the coordinates of the midpoint of the interventricular septum. The coordinates of the midpoint of the interventricular septum and the coordinates of the left ventricular apex in the next frame are the coordinates of the functional anchor point marker.

[0018] In some embodiments, texture anchor markers are established in the initial cardiac magnetic resonance imaging data, including:

[0019] An improved SIFT algorithm was used to extract multi-scale feature points at the myocardial-pericardial junction;

[0020] Construct a local texture feature library, which includes radial gradient histogram features and local binary pattern features;

[0021] A bidirectional matching strategy is used to match the texture feature points of the reference frame and the current frame, and the matching error is calculated based on the dynamic weight error function to obtain texture anchor point markers with a matching error less than a preset threshold.

[0022] In some embodiments, the criteria for determining whether the anchor point marking is abnormal include satisfying any one of the following conditions:

[0023] The displacement mutation value is greater than the preset displacement mutation threshold;

[0024] The deviation of the motion trajectory exceeds the 3σ range;

[0025] The matching confidence level is greater than the adaptive threshold.

[0026] In some embodiments, the adaptive Kalman filter algorithm includes:

[0027]

[0028] in, Here is the state transition equation; For the first The parameter state at any given time; This is the state transition matrix; The input matrix; For external input; This is process noise; The observation equation; In order to be in Observation results at any given time; The observation matrix; To observe noise; Weights for texture motion trends; This refers to the motion trend term extracted from the texture matching results. For adaptive weights.

[0029] In some embodiments, the adaptive weights The expressions include:

[0030]

[0031] in, The variance of the state error predicted by the quadratic polynomial prediction model is... The variance of the motion trend error in the texture matching results.

[0032] In some embodiments, the expression for the deformation model of the thin plate spline includes:

[0033]

[0034] in, For point The value after deformation; It is a linear term; These are radial basis function terms; for To the control point The Euclidean distance; The weights of the control points.

[0035] In some embodiments, the improved RANSAC-SIFT algorithm includes motion consistency constraints, the expressions of which include:

[0036]

[0037] in, This is the feature matching error term; It is the texture anchor point of frame t; It is the corresponding point matched in frame t+1; To predict the Euclidean distance between the location and the actual matching point; These are the weighting coefficients; For motion consistency constraints.

[0038] Secondly, this application also provides a cardiac magnetic resonance dynamic imaging analysis system, the cardiac magnetic resonance dynamic imaging analysis system comprising:

[0039] The data loading module is used to acquire dynamic cardiac magnetic resonance imaging data.

[0040] A multi-anchor mark module is used to establish multiple types of anchor marks in the initial cardiac magnetic resonance dynamic imaging data, including anatomical anchor marks, functional anchor marks, and texture anchor marks;

[0041] The analysis and detection module is used to analyze and detect the dynamic cardiac magnetic resonance imaging data with multiple anchor point markers and to determine whether there are any abnormalities in the multiple anchor point markers.

[0042] A collaborative correction module is used to trigger an anchor point correction process when the analysis and detection module detects anomalies in anchor point markers, in order to correct the anomaly-causing anchor point markers. The correction process includes: identifying the type of the anomaly-causing anchor point markers; if the anatomical anchor point markers are anomalies, correcting them based on a deformation model of a thin-plate spline; if the functional anchor point markers are anomalies, correcting them based on an adaptive Kalman filter algorithm; and if the texture anchor point markers are anomalies, correcting them based on an improved RANSAC-SIFT algorithm.

[0043] The cardiac magnetic resonance dynamic image analysis method and system of this application establishes multiple types of anchor marks, including anatomical anchor marks, functional anchor marks, and texture anchor marks. Based on the multi-type anchor mark collaborative tracking mechanism, it can automatically correct the offset when body motion interference occurs during the analysis of cardiac magnetic resonance dynamic image data, ensuring the accuracy of the anatomical position of myocardial segment marks, thereby avoiding the need for manual remarking and improving the efficiency and reliability of cardiac function quantitative analysis. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a cardiac magnetic resonance dynamic image analysis method provided in one embodiment of this application;

[0046] Figure 2 This is a structural block diagram of a cardiac magnetic resonance dynamic imaging analysis system provided in another embodiment of this application;

[0047] Figure labeling: 10, Data loading module; 20, Multi-anchor point marking module; 30, Analysis and detection module; 40, Collaborative correction module. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] In one embodiment, see Figure 1 This application provides a method for dynamic analysis of cardiac magnetic resonance imaging, which includes the following steps: S11~S13.

[0050] S11: Acquire dynamic cardiac magnetic resonance imaging data.

[0051] S12: Establish multiple types of anchor marks in the initial cardiac magnetic resonance dynamic imaging data, including anatomical anchor marks, functional anchor marks, and texture anchor marks.

[0052] S13: Analyze and detect cardiac magnetic resonance imaging data with multiple anchor point markers. When an abnormality is detected in an anchor point marker, trigger an anchor point correction process to correct the abnormal anchor point marker. The correction process includes: identifying the type of abnormal anchor point marker; if the anatomical anchor point marker is abnormal, correcting it based on a thin-plate spline deformation model; if the functional anchor point marker is abnormal, correcting it based on an adaptive Kalman filter algorithm; and if the texture anchor point marker is abnormal, correcting it based on an improved RANSAC-SIFT algorithm.

[0053] The cardiac magnetic resonance dynamic image analysis method of this application establishes multiple types of anchor marks, including anatomical anchor marks, functional anchor marks, and texture anchor marks. Based on the multi-type anchor mark collaborative tracking mechanism, it can automatically correct the offset when body motion interference occurs during the analysis of cardiac magnetic resonance dynamic image data, ensuring the accuracy of the anatomical position of myocardial segment marks, thereby avoiding the need for manual remarking and improving the efficiency and reliability of cardiac function quantitative analysis.

[0054] In step S11, please refer to Figure 1 Step S11 is used to acquire dynamic cardiac magnetic resonance imaging data.

[0055] Specifically, the acquisition module can acquire pre-stored dynamic cardiac magnetic resonance images, and the scanning area corresponding to the dynamic cardiac magnetic resonance images includes key structures such as the aortic valve annulus, mitral valve annulus, and left ventricular apex.

[0056] In step S12, please refer to Figure 1 In step S12, multiple types of anchor markers are established in the initial cardiac magnetic resonance dynamic imaging data. These multiple types of anchor markers include anatomical anchor markers, functional anchor markers, and texture anchor markers.

[0057] As an example, establishing anatomical anchor point markers in the initial cardiac magnetic resonance imaging data may include the following:

[0058] The cardiac magnetic resonance imaging data is divided into data blocks of the required volume.

[0059] The cardiac chamber structure in the data block is segmented using a 3D U-Net network to obtain a valve annulus segmentation probability map;

[0060] Based on the region of interest in the lobe ring segmentation probability map, a ResNet-50 network is used for fine localization to obtain an anchor point probability heatmap.

[0061] Based on the anchor point probability heatmap, the precise coordinates of the anatomical anchor points are calculated through subpixel interpolation; the anatomical anchor points include the center point of the aortic valve annulus and the anterolateral junction of the mitral valve annulus.

[0062] Specifically, firstly, the initial frame data of the cardiac magnetic resonance imaging (MRI) in DICOM format is loaded. The initial frame image can be cropped into a volumetric data block of 128×128×16 to adapt to the network input requirements. The data block is then input into a 3D U-Net network. The encoder-decoder structure of the 3D U-Net network extracts the three-dimensional spatial features of the heart, segments the cardiac chamber structure, and outputs a valve annulus segmentation probability map with the same size as the data block. The value (0~1) of each voxel in the valve annulus segmentation probability map represents the confidence that the location belongs to the aortic valve annulus or mitral valve annulus structure. The higher the value, the greater the probability of the corresponding target structure. Subsequently, high-confidence voxels with a probability value greater than 0.9 can be screened from the valve annulus segmentation probability map to minimize the interference of noise, blood vessels, and other similar structures, ensuring that the screened area is a high-confidence valve annulus core region. After removing noisy voxels through connected component analysis, the region of interest (ROI) is delineated with the complete aortic valve annulus and mitral valve annulus structures as the core.

[0063] Specifically, the ROI data is input into the ResNet-50 network. The ResNet-50 network captures subtle anatomical features of the valve annulus (such as the annular contour of the aortic valve annulus and the edge morphology of the anterolateral junction of the mitral valve annulus) through the Skip Connection structure to perform fine anchor point localization and generate an anchor point probability heatmap. The probability peak area in the heatmap corresponds to the target anchor point location.

[0064] Specifically, based on the anchor point probability heatmap, for the aortic annulus center point, the geometric center coordinates of the minimum circumscribed ellipse in the high-confidence region can be calculated, which is the precise coordinate of the aortic annulus center point. For the anterolateral junction of the mitral annulus, subpixel interpolation can be performed on the peak region of the probability heatmap. By fitting the probability distribution of adjacent pixels, the coordinate accuracy is improved from the pixel level to the subpixel level (accuracy up to 0.1 pixel unit), outputting the precise coordinates of the anterolateral junction of the mitral annulus. The precise coordinates of the aortic annulus center point and the anterolateral junction of the mitral annulus are then combined to complete the initial frame anatomical anchor point marking.

[0065] As an example, establishing functional anchor point markers in the initial cardiac magnetic resonance imaging data may include the following:

[0066] Extract the left ventricular apex coordinate sequence from the first N frames of the initial cardiac magnetic resonance dynamic imaging data;

[0067] A quadratic polynomial prediction model is constructed based on the left ventricular apex coordinate sequence of the first N frames;

[0068] Prediction is performed based on the quadratic polynomial prediction model to obtain the coordinates of the left ventricular apex in the next frame.

[0069] The motion trajectory of the midpoint of the interventricular septum was obtained by optical flow tracking and verified by changes in myocardial thickness to obtain the coordinates of the midpoint of the interventricular septum. The coordinates of the midpoint of the interventricular septum and the coordinates of the left ventricular apex in the next frame are the coordinates of the functional anchor point marker.

[0070] Specifically, functional anchor point tracking focuses on key points with regular characteristics in cardiac motion. It achieves precise localization of myocardial segments by dynamically tracking the motion trajectory of these key points. These key points mainly include two types of markers: the left ventricular apex and the midpoint of the interventricular septum.

[0071] Specifically, regarding the left ventricular apex, whose motion is periodic and predictable, the actual coordinate sequence of the left ventricular apex from the first N frames (e.g., the first 5 frames) of the image (e.g., (120,85), (122,86), (123,86.5), (124,87), (125,87.5)) can be extracted as training samples. Six model parameters (a0, a1, a2, b0, b1, b2) are calculated by fitting a quadratic polynomial prediction model using the least squares method. The coordinate prediction formula of the quadratic polynomial prediction model can be:

[0072]

[0073]

[0074] in, Represents the coordinates of the t-th frame; This is the time frame number.

[0075] Subsequently, the time sequence number of the next frame is substituted into the quadratic polynomial prediction model to obtain the predicted coordinates of the left ventricular apex (for example, the prediction for the 6th frame is (126.2, 88.1)).

[0076] Specifically, regarding the midpoint of the interventricular septum, located in the middle of the septum between the left and right ventricles, its trajectory is affected by the contraction / diastole of both ventricles. The initial position of the midpoint can be determined in the initial frame. Optical flow is used to calculate the pixel displacement vector between consecutive frames to track the trajectory of this point (optical flow can capture subtle deformations of myocardial tissue). Simultaneously, cross-validation is performed using changes in myocardial thickness. The interventricular septum thickness changes regularly during the cardiac cycle. If the optical flow tracking result does not match the pattern of myocardial thickness changes, an anomaly detection mechanism is triggered. Multi-feature fusion is used to improve tracking robustness, thus obtaining the accurate coordinates of the midpoint of the interventricular septum. The coordinates of the left ventricular apex predicted by the quadratic polynomial model and the coordinates of the midpoint of the interventricular septum are used together as the coordinates of the functional anchor point marker.

[0077] As an example, establishing texture anchor point markers in the initial cardiac magnetic resonance imaging data may include the following:

[0078] An improved SIFT algorithm was used to extract multi-scale feature points at the myocardial-pericardial junction;

[0079] Construct a local texture feature library, which includes radial gradient histogram features and local binary pattern features;

[0080] A bidirectional matching strategy is used to match the texture feature points of the reference frame and the current frame, and the matching error is calculated based on the dynamic weight error function to obtain texture anchor point markers with a matching error less than a preset threshold.

[0081] Specifically, firstly, focusing on the myocardial-pericardial junction, where texture differences are significant and feature point stability is high, an improved 256-dimensional SIFT algorithm can be used to extract multi-scale feature points. By optimizing the scale space construction of the difference-of-Gaussian pyramid, the adaptability to low-contrast and high-noise features of dynamic cardiac magnetic resonance images is enhanced, resulting in the extraction of 12 candidate feature points. The improved higher-dimensional descriptor of the 256-dimensional SIFT algorithm can contain richer local texture information (such as detail gradients and orientation distribution) to reduce feature ambiguity and improve matching discrimination (especially in complex texture scenarios, such as cardiac ultrasound or CT images).

[0082] Specifically, a local texture feature library can be constructed for each candidate feature point. A radial gradient histogram (HOG) centered on the feature point is calculated to capture the gradient direction distribution pattern. Simultaneously, local binary pattern (LBP) is extracted to characterize the gray-level contrast pattern of neighboring pixels. The HOG and LBP features are then fused to form a more dimensional texture descriptor. Subsequently, using the initial frame as a reference frame, texture feature point matching is performed on subsequent frames (e.g., the second frame). A bidirectional matching strategy is adopted, where feature points in the reference frame (ref) and the current frame (curr) are matched mutually, retaining only the feature points that successfully match bidirectionally. A dynamic weighted error function is introduced to calculate the matching error. The formula for calculating the matching error function can be:

[0083]

[0084] in, This represents the final matching error; The total number of dimensions of the feature descriptors; For the first Weights of each feature dimension; It is the descriptor of the i-th matching feature point in the reference frame; It is the descriptor of the i-th matching feature point in the current frame; finally, from the initially extracted candidate points, three feature points with matching errors less than a preset threshold (e.g., 0.3) are selected (average coordinates such as (89,105), (92,110), (90,108), etc.) as the coordinates of the texture anchor point marker.

[0085] In step S13, please refer to Figure 1 In step S13, the dynamic cardiac magnetic resonance imaging data with multiple anchor point markers are analyzed and detected. When an abnormality is detected in an anchor point marker, an anchor point correction process is triggered to correct the abnormal anchor point marker. The correction process includes: identifying the type of abnormal anchor point marker; if the anatomical anchor point marker is abnormal, correcting the abnormal anatomical anchor point marker based on the deformation model of a thin plate spline; if the functional anchor point marker is abnormal, correcting the abnormal functional anchor point marker based on an adaptive Kalman filter algorithm; and if the texture anchor point marker is abnormal, correcting the abnormal texture anchor point marker based on an improved RANSAC-SIFT algorithm.

[0086] As an example, the criteria for determining whether anchor point markings are abnormal may include satisfying any of the following conditions:

[0087] The displacement mutation value is greater than the preset displacement mutation threshold;

[0088] The deviation of the motion trajectory exceeds the 3σ range;

[0089] The matching confidence level is greater than the adaptive threshold.

[0090] Specifically, displacement abrupt change detection: the displacement difference between the current frame and the previous frame exceeds a threshold. The formula for calculating the displacement mutation value can be:

[0091]

[0092] in, Let be the coordinates of the anchor point in frame t; Let be the coordinates of the anchor point in frame t-1.

[0093] Can Set to 2mm, when > At that time, a displacement mutation anomaly was triggered.

[0094] Specifically, the trajectory deviation is the Mahalanobis distance between the actual position and the predicted position. The calculation formula can be:

[0095]

[0096] in, The vectors that are actually observed; The vector predicted by the model; Let be the covariance matrix of the data.

[0097] If the Mahalanobis distance (MD) is too large, it indicates that the anchor point movement deviates from the predicted pattern, requiring a correction. When the Mahalanobis distance between the actual and predicted positions exceeds [a certain value], [the correction is needed]. The range triggers an abnormal deviation from the motion trajectory; The standard deviation of the historical motion trajectory data of the anchor point is used to quantify the degree of fluctuation in the anchor point's motion trajectory; Based on the statistical law of normal distribution: if the normal motion trajectory data of the anchor point approximately follows a normal distribution, about 99.73% of the trajectory data will... Within the defined range; the deviation of the motion trajectory exceeds The range can be used to determine whether the trajectory of the anchor point deviates from the normal pattern formed in history.

[0098] Specifically, the texture matching confidence score is defined as the feature matching error exceeding an adaptive threshold. The formula for expressing the texture matching confidence score can be:

[0099]

[0100] in, This represents the current matching error; This represents the mean of the matching error; This represents the standard deviation of the matching error.

[0101] when When the current matching error is significantly higher than the historical average (exceeding 2 standard deviations), texture matching confidence anomaly is triggered.

[0102] Specifically, and Based on the dynamic calculation of the matching error of the first N frames, it can adapt to the dynamic changes of the scene (such as the difference in texture change patterns of different cardiac cycles in heart movement), avoid misjudgment caused by fixed thresholds (such as error fluctuations caused by normal movement being misjudged as unreliable); it can also update the abnormal standard of matching error with historical data, and improve the robustness of confidence judgment.

[0103] As an example, the expression for the deformation model of the thin plate spline may include:

[0104]

[0105] in, For point The value after deformation; It is a linear term; These are radial basis function terms; for To the control point The Euclidean distance; The weights of the control points.

[0106] Specifically, firstly, based on the anomaly detection criteria, abnormal anatomical anchor points are identified. Then, anatomical anchor points in adjacent normal frames (e.g., the two frames before and one frame after the abnormal frame, and whose motion state has been verified to be normal) are selected as control vertices for the TPS deformation model. For example, if the anatomical anchor point in frame 23 is abnormal, the stable aortic annulus center point and the anterolateral junction of the mitral annulus in frames 21 and 22, which conform to anatomical constraints, are selected, and their coordinates are recorded. (i = 1,2,...,n, where n is the number of control vertices); then the original coordinates of the anomaly dissection anchor points are... Substituting the TPS deformation function, the corrected coordinates are calculated. .

[0107] As an example, the adaptive Kalman filter algorithm may include:

[0108]

[0109] in, Here is the state transition equation; For the first The parameter state at any given time; This is the state transition matrix; The input matrix; For external input; This is process noise; The observation equation; In order to be in Observation results at any given time; The observation matrix; To observe noise; Weights for texture motion trends; This refers to the motion trend term extracted from the texture matching results. For adaptive weights.

[0110] As an example, the adaptive weights The expression can include:

[0111]

[0112] in, The variance of the state error predicted by the quadratic polynomial prediction model is... The variance of the motion trend error in the texture matching results.

[0113] Specifically, firstly, based on abnormality criteria, abnormal markers at the left ventricular apex or the midpoint of the interventricular septum are identified, and then the state transition equation is used to predict the next abnormality. Functional anchor state at any moment Subsequently, the motion trend term of the texture anchor point is extracted from the texture matching results of the current frame and the reference frame. The state error variance of the quadratic polynomial prediction model was statistically obtained. (Statistical results of the deviation between historical predicted values ​​and actual values) and the variance of motion trend error in texture matching results ; Calculate adaptive weights When the polynomial prediction error is large ( Large value) and reliable texture matching ( When the value is small, Approaching 1, enhancing The influence of polynomial prediction depends more on the texture motion trend correction state; when the polynomial prediction is reliable ( Small value) while large texture matching error ( When the value is large, Approaching 0, weakening The impact depends more on system model predictions.

[0114] Specifically, the collection of the first Observation results of time-function anchor points The texture motion trend term will be determined through the observation equation. With adaptive weights The data is integrated into the observations to obtain observations with fused texture trends; based on Kalman filtering, it is combined with the predicted state. Update the observations of the fused texture trend. Estimation of the functional anchor state at any given time.

[0115] As an example, the improved RANSAC-SIFT algorithm includes motion consistency constraints, the expressions of which may include:

[0116]

[0117] in, This is the feature matching error term; It is the texture anchor point of frame t; It is the corresponding point matched in frame t+1; To predict the Euclidean distance between the location and the actual matching point; These are the weighting coefficients; For motion consistency constraints.

[0118] Specifically, the set of texture anchor points at the myocardial-pericardial junction in frame t is first extracted using an improved SIFT algorithm. And based on the SIFT algorithm, the first matching is obtained. In-frame and Corresponding candidate matching points The objective function is minimized using an iterative optimization method (such as gradient descent); during the iteration, the matching relationship of texture anchor points and the motion vector field are adjusted synchronously. The goal is to minimize the feature matching error and keep the divergence of the motion vector field within a small range (to satisfy the consistency of heart motion). After the objective function converges, the texture anchor point matching pairs that minimize the feature matching error and motion consistency constraints are selected, and erroneous matches that do not conform to motion consistency (such as erroneous SIFT matches caused by image artifacts or noise) are eliminated to obtain reliable texture anchor point matching results that satisfy the motion laws.

[0119] As an example, in step S13, the anchor point detection and correction uses an appropriate fusion algorithm to correct anchor points of different types, such as anatomical, functional, and textured anchor points, thereby achieving multimodal collaboration and significantly improving the stability of anchor points in complex scenes.

[0120] In another embodiment, this application also provides a cardiac magnetic resonance imaging (MRI) dynamic image analysis system, which may include: a data loading module 10, a multi-anchor mark module 20, an analysis and detection module 30, and a collaborative correction module 40; wherein the data loading module 10 is used to acquire cardiac MRI dynamic image data; the multi-anchor mark module 20 is used to establish multiple types of anchor marks in the initial cardiac MRI dynamic image data, the multiple types of anchor marks including anatomical anchor marks, functional anchor marks, and texture anchor marks; the analysis and detection module 30 is used to analyze and detect the cardiac MRI dynamic image data with established multiple types of anchor marks, and determine the multiple anchor marks. The analysis and detection module 40 is used to trigger an anchor point correction process when the analysis and detection module detects that an anchor point is abnormal, in order to correct the abnormal anchor point. The correction process includes: identifying the type of the abnormal anchor point; if the anatomical anchor point is abnormal, correcting the abnormal anatomical anchor point based on the deformation model of the thin plate spline; if the functional anchor point is abnormal, correcting the abnormal functional anchor point based on the adaptive Kalman filter algorithm; if the texture anchor point is abnormal, correcting the abnormal texture anchor point based on the improved RANSAC-SIFT algorithm.

[0121] In the cardiac magnetic resonance dynamic image analysis system of this application, through the data loading module 10, the multi-anchor mark module 20, the analysis and detection module 30, and the collaborative correction module 40, by establishing multiple types of anchor marks including anatomical anchor marks, functional anchor marks, and texture anchor marks, the system can automatically correct the offset when body motion interference occurs during the analysis of cardiac magnetic resonance dynamic image data based on the multi-type anchor mark collaborative tracking mechanism. This ensures the accuracy of the anatomical position of the myocardial segment marks, thereby avoiding the need for manual remarking and improving the efficiency and reliability of cardiac function quantitative analysis.

[0122] As an example, the cardiac magnetic resonance dynamic imaging analysis system of this application can be used to perform, for example... Figure 1 And the method for dynamic imaging analysis of cardiac magnetic resonance in related embodiments.

[0123] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features of the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0124] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method of cardiac magnetic resonance dynamic image analysis, characterized by, The method comprises: acquiring initial cardiac magnetic resonance dynamic image data; establishing multi-class anchor point markers in the initial cardiac magnetic resonance dynamic image data, wherein the multi-class anchor point markers comprise an anatomical anchor point marker, a functional anchor point marker and a texture anchor point marker; wherein the establishment of the anatomical anchor point marker comprises: dividing the cardiac magnetic resonance dynamic image data into data blocks of a required volume size; segmenting the cardiac chamber structure in the data block by using a 3D U-Net network to obtain an annulus segmentation probability map; performing fine positioning by using a ResNet-50 network based on the region of interest of the annulus segmentation probability map to obtain an anchor point probability heat map; and calculating the accurate coordinates of the anatomical anchor point by sub-pixel interpolation based on the anchor point probability heat map; the anatomical anchor point comprises a aortic annulus center point and an anterior-lateral commissure point of the mitral annulus; the establishment of the functional anchor point marker comprises: extracting a left ventricular apex point coordinate sequence of the first N frames in the initial cardiac magnetic resonance dynamic image data; constructing a quadratic polynomial prediction model based on the left ventricular apex point coordinate sequence of the first N frames; predicting the coordinates of the left ventricular apex point in the next frame based on the quadratic polynomial prediction model; obtaining the motion trajectory of the interventricular septum midpoint by using an optical flow method and verifying it by using the change of myocardial thickness to obtain the coordinates of the interventricular septum midpoint; and the coordinates of the interventricular septum midpoint and the coordinates of the left ventricular apex point in the next frame obtained are the coordinates of the functional anchor point marker; the establishment of the texture anchor point marker comprises: extracting multi-scale feature points at the myocardium-pericardium junction by using an improved SIFT algorithm; constructing a local texture feature library, wherein the local texture feature library comprises a radial gradient histogram feature and a local binary pattern feature; matching the texture feature points of the reference frame and the current frame by using a bidirectional matching strategy, and calculating the matching error based on a dynamic weight error function to obtain a texture anchor point marker with a matching error less than a preset threshold; analyzing and detecting the cardiac magnetic resonance dynamic image data with the established multi-class anchor point markers, triggering an anchor point correction process when an anchor point marker is detected to have an abnormality, to correct the anchor point marker with the abnormality; the anchor point correction process comprises: identifying the type of the anchor point marker with the abnormality; if the anatomical anchor point marker has an abnormality, correcting the anatomical anchor point marker with the abnormality based on a thin plate spline deformation model; if the functional anchor point marker has an abnormality, correcting the functional anchor point marker with the abnormality based on an adaptive Kalman filtering algorithm; and if the texture anchor point marker has an abnormality, correcting the texture anchor point marker with the abnormality based on an improved RANSAC-SIFT algorithm.

2. The method of claim 1, wherein, The judgment criteria for the anchor point marker having an abnormality include any one of the following conditions: a displacement mutation value greater than a preset displacement mutation threshold; a motion trajectory deviation degree exceeding a 3σ range; wherein σ is the standard deviation of the historical motion trajectory data of the anchor point; a matching confidence greater than an adaptive threshold.

3. The method of claim 1, wherein, The adaptive Kalman filtering algorithm comprises: ； wherein, is a state transition equation; is a state transition matrix; is a parameter state at time is a state transition matrix; is an input matrix; is an external input; is a process noise; is an observation equation; is an observation matrix; is an observation result at time is an observation matrix; is an observation noise; is a texture motion trend weight; is a motion trend item extracted from a texture matching result; is an adaptive weight.

4. The method of claim 3, wherein, The adaptive weight The expression of the adaptive weight includes: ； wherein, a state error variance predicted for the quadratic polynomial prediction model; a motion trend error variance for the texture matching result.

5. The method of claim 1, wherein, The expression of the thin plate spline deformation model comprises: ； wherein, is a point the value after deformation; is a linear term; is a radial basis function term; is the Euclidean distance to the control point ; is the weight of the control point.

6. The method of claim 1, wherein, The improved RANSAC-SIFT algorithm comprises a motion consistency constraint, and an expression of the motion consistency constraint comprises: ； wherein, is a feature matching error term; is a texture anchor point of the t-th frame; is a corresponding point matched in the t+1-th frame; is the Euclidean distance between the predicted position and the actual matched point; is a weight coefficient; is a motion consistency constraint.

7. A cardiac magnetic resonance dynamic imaging analysis system, characterized by The cardiac magnetic resonance dynamic image analysis system comprises: A data loading module is configured to acquire initial cardiac magnetic resonance dynamic image data. A multi-anchor labeling module is configured to establish multi-type anchor labels in the initial cardiac magnetic resonance dynamic image data, wherein the multi-type anchor labels comprise an anatomical anchor label, a functional anchor label, and a texture anchor label. An analysis and detection module is configured to perform analysis and detection on the cardiac magnetic resonance dynamic image data with the multi-type anchor labels established, and to determine whether the multi-type anchor labels are abnormal. A cooperative correction module is configured to trigger an anchor correction process to correct the anchor labels that are abnormal when the analysis and detection module detects that the anchor labels are abnormal, wherein the anchor correction process comprises: identifying the type of the anchor labels that are abnormal; correcting the anatomical anchor labels that are abnormal based on a thin plate spline deformation model; correcting the functional anchor labels that are abnormal based on an adaptive Kalman filtering algorithm; and correcting the texture anchor labels that are abnormal based on an improved RANSAC-SIFT algorithm.

Images