Heart magnetic resonance image intelligent processing method and system

Abstract

The invention relates to the technical field of image processing, in particular to a heart magnetic resonance image intelligent processing method and system, and the method comprises the steps: obtaining an original heart magnetic resonance four-dimensional image sequence; performing phase dislocation correction processing on the four-dimensional image sequence to generate a space-time alignment image sequence; dynamically segmenting the cardiac muscle tissue on the basis of the space-time alignment image sequence; performing multi-dimensional feature extraction on the three-dimensional dynamic mask sequence to generate a dynamic feature map including a displacement field, a strain rate and a blood flow propagation velocity; inputting the dynamic characteristic chromatogram into a hybrid analysis model, and outputting myocardial movement abnormality region and microcirculation disturbance quantization parameters; generating a dynamic visual diagnosis report based on the myocardial movement abnormality region and the microcirculation disturbance quantization parameters; according to the heart magnetic resonance image processing method, by introducing the hybrid analysis model of the space-time diagram convolutional network and the physical constraint differential equation, the heart magnetic resonance image processing precision and efficiency are remarkably improved, and comprehensive analysis of heart functions is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for intelligent processing of cardiac magnetic resonance images. Background Art

[0002] Heart disease is one of the diseases with the highest mortality rate worldwide, and accurate cardiac function assessment is crucial for the early diagnosis and treatment of heart disease. Cardiac Magnetic Resonance Imaging (CMR), as a non-invasive, high-resolution imaging technology, provides rich information on cardiac anatomy, motor function, and blood flow dynamics. Through four-dimensional cardiac magnetic resonance images (including three-dimensional spatial coordinates and time dimensions), the movement state of the heart at all levels and the condition of microcirculation can be comprehensively evaluated. However, existing cardiac magnetic resonance image processing methods still face multiple technical challenges, especially in the spatiotemporal alignment of images, myocardial tissue segmentation, dynamic feature extraction, and visualization.

[0003] The heart undergoes complex movements during each beat, and physiological factors such as breathing can also cause phase misalignment of the image, resulting in temporal and spatial misalignment between image sequences at different time points. Existing image processing methods are difficult to accurately correct these motion artifacts, which affects the accuracy of subsequent analysis. In addition, the myocardial boundary in cardiac magnetic resonance images is complex and fuzzy, and existing segmentation methods are usually unable to effectively handle these dynamic changes, resulting in difficulty in accurately segmenting the endocardium, myocardium, and epicardium. In addition, how to accurately extract dynamic features such as displacement field, strain rate, and blood flow velocity from image sequences, and comprehensively consider physiological constraints to evaluate microcirculation, remains a difficult problem that needs to be solved. Most existing technologies focus on a single task and lack a comprehensive analysis of the overall function of the heart. Summary of the invention

[0004] Based on the above purpose, the present invention provides a method for intelligent processing of cardiac magnetic resonance images.

[0005] A cardiac magnetic resonance image intelligent processing method comprises the following steps:

[0006] S1: Acquire an original cardiac magnetic resonance four-dimensional image sequence, wherein the four-dimensional image sequence includes a spatial three-dimensional coordinate and a time dimension;

[0007] S2: performing phase misalignment correction processing on the four-dimensional image sequence to generate a spatiotemporal aligned image sequence, wherein the phase misalignment correction includes a respiratory motion compensation layer and an ECG gating synchronization layer;

[0008] S3: Perform dynamic segmentation of the myocardial tissue based on the spatio-temporal aligned image sequence, and output a three-dimensional dynamic mask sequence including the endocardium, myocardium, and epicardium;

[0009] S4: Extract multi-dimensional features from the three-dimensional dynamic mask sequence to generate a dynamic feature map including a displacement field, strain rate, and blood flow propagation velocity;

[0010] S5: Input the dynamic feature map into a hybrid analysis model, where the hybrid analysis model includes a spatio-temporal graph convolutional network branch and a physical constraint differential equation branch, and output the myocardial motion abnormal region and microcirculation disorder quantification parameters;

[0011] S6: Generate a dynamic visualization diagnosis report based on the myocardial motion abnormal region and microcirculation disorder quantification parameters, where the dynamic visualization diagnosis report includes a myocardial anisotropic motion trajectory cloud map and a microcirculation parameter heat map.

[0012] Further, the four-dimensional image sequence in S1 is represented as {I(x, y, z, t)}, including spatial three-dimensional coordinates (x, y, z) and time dimension t; the time interval of the four-dimensional image sequence is Δt, and the resolution of the four-dimensional image sequence is D res =(D x , D y , D z ), where D x , D y , D z respectively represent the resolutions of the image in the three spatial dimensions of x, y, and z, and Δt is the time interval between two consecutive frames.

[0013] Further, S2 includes:

[0014] S21: Perform respiratory motion compensation processing on the four-dimensional image sequence. Based on the time dimension t in the four-dimensional image sequence, by extracting the thoracic displacement information of the cardiac image in each frame of the image, use a non-rigid deformation model to correct the respiratory motion. The non-rigid deformation model is expressed as:

[0015]

[0016] where I corr (x, y, z, t) is the image after respiratory motion compensation, I(x, y, z, t) is the four-dimensional image sequence, represents the convolution operation, is the respiratory motion compensation field, indicating the displacement field at the spatial position (x, y, z) and time frame t;

[0017] S22: After respiratory motion compensation, perform electrocardiogram (ECG) gating synchronization correction. By analyzing the ECG signal, synchronize and align the four-dimensional image sequence with the ECG signal. The ECG gating synchronization correction includes:

[0018] Obtain the ECG signal, analyze the QRS peak positions of the ECG signal to obtain the cardiac cycle T ECG ;

[0019] Based on the QRS peak positions of the ECG signal, determine the acquisition time of each image frame and map it to the corresponding cardiac cycle phase;

[0020] Resample the four-dimensional image sequence to make it evenly distributed in the phase of the cardiac cycle, thereby obtaining the four-dimensional image sequence I ECG (x, y, z, t′), where t′ is the synchronized time dimension;

[0021] S23: Based on the obtained four-dimensional image sequence I corr (x, y, z, t) after respiratory motion compensation and the image sequence I ECG (x, y, z, t′) after ECG gating synchronization, perform fusion through a spatio-temporal transformation model to generate a spatio-temporally aligned image sequence I aligned (x, y, z, t”), and the spatio-temporal transformation model is expressed as:

[0022]

[0023] where, I aligned (x, y, z, t”) is the spatio-temporally aligned image sequence, t” is the synchronized time dimension, and Δt1 represents the time alignment offset.

[0024] Further, the S3 includes:

[0025] S31: Input the spatio-temporally aligned image sequence {I aligned (x, y, z, t”)};

[0026] S32: Segment the myocardial tissue of the spatio-temporally aligned image sequence, and use an image segmentation model to automatically segment the heart region. The image segmentation model extracts the boundaries of the endocardium, myocardium, and epicardium by comparing the gray information, texture features, and edge information in the spatio-temporally aligned image sequence;

[0027] S33: During the segmentation process, each layer of the heart (endocardium, myocardium, epicardium) is processed through a series of convolutional operations and pooling operations, and finally the masks of the three tissues are output through a fully connected layer; for each frame t”, generate the mask of the endocardium through the boundaries of the endocardium, myocardium, and epicardium segmented and output by the image segmentation model Mask of the myocardium and the mask of the epicardium

[0028] S34: The mask of the endocardium of each time frame t” Mask of the myocardium and the mask of the epicardium are combined to obtain a three-dimensional dynamic mask sequence containing the dynamic changes of the endocardium, myocardium, and epicardium in all time frames

[0029] Furthermore, the step of automatically segmenting the heart region by the image segmentation model in S32 includes:

[0030] S321, Input image: Input the spatio-temporal aligned image sequence {I aligned (x,y,z,t”)};

[0031] S322, Encoder part: The feature map F1(x,y,z,t”) after the first layer of 3D convolution of the input spatio-temporal aligned image sequence {I aligned (x,y,z,t”)} is expressed as:

[0032] F1(x,y,z,t″) = ReLU(Conv3(I aligned (x,y,z,t″)));

[0033] where Conv3 is the 3D convolution operation and ReLU is the activation function;

[0034] After the k-th layer of convolution and pooling, the output feature map is F k (x,y,z,t”) and is expressed as:

[0035] F k (x,y,z,t″) = ReLU(Conv3(MaxPool3(F k-1 (x,y,z,t″))));

[0036] where MaxPool3 represents the 3D max pooling operation and F k-1 (x,y,z,y”) is the feature map of k - 1;

[0037] S323, Decoder part: The feature map F k (x,y,z,t”) of the encoder part is input to the k-th layer of the decoder. After deconvolution and the activation function, the decoder output D k ′(x,y,z,t”) is obtained and expressed as:

[0038] D k′(x, y, z, t”) = ReLU(Deconv3(F k (x, y, z, t”)));

[0039] Where Deconv3 is a 3D transposed convolution operation and ReLU is an activation function;

[0040] S324, skip connection: The skip connection concatenates the feature map F k (x, y, z, t”) of the k-th layer of the encoder with the input feature map D k ′(x, y, z, t”) of the k-th layer of the decoder to obtain the final decoder output D k (x, y, z, t”), expressed as:

[0041]

[0042] Where: represents the concatenation operation;

[0043] S325, image output: The final decoder output is passed through the Softmax function for output to obtain the output image M(x, y, z, t”), expressed as:

[0044] O(x, y, z, t″) = Softmax(Conv1(D k (x, y, z, t”)));

[0045] Where Conv1 is a 1×1×1 convolution operation and Softmax is a normalization operation.

[0046] Furthermore, in S33, it is processed through a series of convolution operations and pooling operations, and finally the masks of three tissues are output through a fully connected layer, including:

[0047] S331, convolution operation: The input is the spatio-temporally aligned image sequence I aligned (x, y, z, t”) with a convolution kernel size of 3×3×3, that is, convolution processing is performed on the three-dimensional space, and the output feature map is F1(x, y, z, t”). After the convolution operation, the output feature map is obtained:

[0048]

[0049] Where w ijk is the convolution kernel weight, representing the convolution weight at the spatial position (i, j, k) with the input image I aligned (x, y, z, t”), b is the bias term, and ReLU is the activation function, which acts on the convolution result to ensure the non-linear characteristics of the output. In a deep convolutional network, the feature map F k(x, y, z, t”) is processed through several layers of convolution and calculated as:

[0050] F k (x, y, z, t″) = ReLU(Conv3(MaxPool3(F k-1 (x, y, z, t″)))) ;

[0051] Among them, Conv3 represents the 3D convolution operation, and MaxPool3 is the 3D max pooling operation;

[0052] S332, pooling operation: The input feature map is F k-1 (x, y, z, t”), after the 3D max pooling operation, the output feature map is F k (x, y, z, t”), which is expressed as:

[0053] F k (x, y, z, ty”) = max(F k-1 (x + i, y + j, z + k, t”)), i, j, k ∈ {0, 1, 2} ;

[0054] S333, output result (mask): Each layer of the output mask is expressed as:

[0055]

[0056] Among them, are the masks of the endocardium, myocardium, and epicardium of the output.

[0057] Furthermore, the S4 includes:

[0058] S41, displacement field extraction: Based on the three-dimensional dynamic mask sequence, by calculating the position changes of each tissue region of the heart between adjacent time frames, a displacement field sequence is generated. For adjacent frames at time points t” and t” + 1, the displacement vector field is calculated which is expressed as:

[0059]

[0060] Among them, x t” , y t” , z t” and x t”+1 , y t”+1 , z t”+1 are the three-dimensional spatial coordinates of each tissue of the heart at time t” and t” + 1 respectively;

[0061] S42, strain rate extraction: By performing differential calculations on the displacement field sequence, the strain rate of each voxel is obtained, and ∈(x, y, z, t”) is expressed as:

[0062]

[0063] Among them, Δt is the time interval, that is, the time difference between adjacent time frames, is the displacement field the change amount at time t”;

[0064] S43, blood flow propagation speed extraction: Using the spatial position changes of each point in the mask sequence and combining with the movement trajectory of the heart, calculate the blood flow propagation speed v(x, y, z, t”) expressed as:

[0065] Among them, is the modulus of the displacement vector, indicating the distance of blood flow propagation within the time interval Δt;

[0066] S44, generating a dynamic feature map: Synthesize the displacement field, strain rate, and blood flow propagation speed to generate a dynamic feature map F(x, y, z, y”) containing the displacement field, strain rate, and blood flow propagation speed, expressed as:

[0067]

[0068] Among them, represents the displacement field, ∈(x, y, z, t”) represents the strain rate, and v(x, y, z, y”) represents the blood flow propagation speed.

[0069] Furthermore, the above S5 includes:

[0070] S51, inputting the dynamic feature map: Use the dynamic feature map F(x, y, z, t”) as the model input and transfer it to the two branches of the hybrid analysis model. The branches include the spatio-temporal graph convolutional network branch and the physical constraint differential equation branch;

[0071] S52, spatio-temporal graph convolutional network branch: In the spatio-temporal graph convolutional network branch, capture the feature relationships at different time points and spatial positions in the spatio-temporal domain through convolution operations, expressed as:

[0072] H k = σ(A · H k-1 · W k );

[0073] Among them, H k is the feature representation of the k-th layer, A is the adjacency matrix of the graph, representing the connection relationship between different spatial points, W k is the convolutional kernel weight of the k-th layer, and σ(·) is the activation function, used to introduce non-linearity;

[0074] S53, Physical Constraint Differential Equation Branch: Based on the blood flow propagation velocity v(x, y, z, t”) and the strain rate ∈(x, y, z, t”), a physical constraint differential equation model is introduced to simulate hemodynamics and myocardial motion. The physical constraint differential equation model is expressed as:

[0075]

[0076] where is the displacement field, representing the motion of the myocardium, ∈ is the strain rate, representing the deformation rate of the myocardium, v is the blood flow velocity, representing the flow rate of blood in the heart, F(·) is a differential equation function derived from the principles of fluid mechanics and elasticity, used to describe the physical behavior of the heart, λ and μ are the Lame constants, describing the elastic properties of the material, tr(∈) is the trace of the strain tensor, I is the identity matrix, η is the viscosity coefficient of the blood, representing the mechanical action of blood flow on tissues, represents the gradient of the blood flow velocity, capturing the influence of blood flow on tissues;

[0077] S54, Fusion and Output: Based on the output results of the spatio-temporal graph convolutional network branch and the physical constraint differential equation branch, they are combined through a fusion mechanism to obtain the abnormal myocardial motion region and the quantitative parameters of microcirculation disorders, expressed as:

[0078] Y = λ1·H k + λ2·Y PDE ;

[0079]

[0080] where H k is the output feature of the spatio-temporal graph convolutional network, representing the abnormal myocardial motion region, and Y PDE is the output feature of the physical constraint differential equation branch, representing the quantitative parameters of microcirculation disorders, describes the displacement field, strain rate, and blood flow velocity of the heart, and λ1 and λ2 are weighting coefficients, representing the contribution degrees of the outputs of different branches, and λ1, λ2 ∈ [0, 1].

[0081] Furthermore, the S6 includes:

[0082] S61, Input Quantitative Parameters: Input the quantitative parameters from the hybrid analysis model. The quantitative parameters include the abnormal myocardial motion region and the quantitative parameters of microcirculation disorders;

[0083] S62, Generate the Anisotropic Motion Trajectory Cloud Map of the Myocardium: Based on the abnormal myocardial motion region and its anisotropic information, generate the motion trajectory cloud map;

[0084] S63, Generate the Heat Map of Microcirculation Parameters: Based on the quantitative parameters of microcirculation disorders, generate the heat map;

[0085] S64, Dynamic display and interaction function: Generate a dynamic visualization report through a graphical user interface. The report includes a myocardial anisotropic motion trajectory cloud map and a microcirculation parameter heat map, and provides an interaction function;

[0086] S65, Generate a final diagnosis report: Based on the myocardial anisotropic motion trajectory cloud map and the microcirculation parameter heat map, generate a final diagnosis report, and provide specific descriptions of regional annotation, motion abnormalities, and microcirculation disorders.

[0087] A cardiac magnetic resonance image intelligent processing system for implementing the above-mentioned cardiac magnetic resonance image intelligent processing method, including the following modules:

[0088] Four-dimensional image acquisition module: Acquire the original cardiac magnetic resonance four-dimensional image sequence, and the image sequence includes spatial three-dimensional coordinates and time dimension;

[0089] Phase misalignment correction module: Perform phase misalignment correction processing on the four-dimensional image sequence to generate a spatio-temporal alignment image sequence;

[0090] Myocardial tissue dynamic segmentation module: Based on the spatio-temporal alignment image sequence, dynamically segment the myocardial tissue, and output a three-dimensional dynamic mask sequence including the endocardium, myocardial layer, and epicardium;

[0091] Feature extraction module: Perform multi-dimensional feature extraction on the three-dimensional dynamic mask sequence to generate a dynamic feature map including displacement field, strain rate, and blood flow propagation velocity;

[0092] Hybrid analysis model module: Input the dynamic feature map into the hybrid analysis model, and the hybrid analysis model includes a spatio-temporal graph convolutional network branch and a physical constraint differential equation branch, and output the myocardial motion abnormal region and the microcirculation disorder quantification parameter;

[0093] Visualization report generation module: Generate a dynamic visualization diagnosis report, and the dynamic visualization diagnosis report includes a myocardial anisotropic motion trajectory cloud map and a microcirculation parameter heat map.

[0094] Advantages of the present invention:

[0095] In the present invention, by introducing a hybrid analysis model of spatio-temporal graph convolutional network and physically constrained differential equations, the accuracy and efficiency of cardiac magnetic resonance image processing are significantly improved. In terms of image spatio-temporal alignment, a phase misalignment correction technique is adopted, combined with respiratory motion compensation and electrocardiogram gating synchronization, which can effectively eliminate motion artifacts caused by respiration and heartbeat, and achieve precise spatio-temporal alignment of four-dimensional image sequences. This method not only improves the image quality, but also provides high-quality input data for subsequent myocardial tissue segmentation and feature extraction. Through precise image alignment, it can ensure the consistency of cardiac image data at different time points, thus effectively reducing the error between images and avoiding diagnostic deviations caused by image misalignment.

[0096] In the present invention, through dynamic segmentation of myocardial tissue and multi-dimensional feature extraction, a comprehensive analysis of cardiac function is achieved. Advanced image segmentation techniques such as 3D U-Net are used to accurately divide the regions of the endocardium, myocardium, and epicardium, laying a foundation for subsequent myocardial motion analysis and microcirculation assessment. At the same time, the generation of dynamic feature maps covers multiple indicators such as the displacement field, strain rate, and blood flow propagation velocity of the myocardium, enabling a comprehensive analysis of cardiac motion abnormalities and microcirculation disorders from multiple dimensions. Finally, by generating a dynamic visualization diagnostic report, a cloud map of myocardial anisotropic motion trajectories and a heat map of microcirculation parameters are provided. This not only makes the diagnostic results more intuitive, but also provides dynamic display and interaction functions, helping clinicians comprehensively evaluate cardiac health status from different angles and time points, greatly improving the diagnostic accuracy of heart diseases and the support for clinical decision-making. Brief Description of the Drawings

[0097] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings described below are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0098] Figure 1 It is a schematic flowchart of the method according to an embodiment of the present invention;

[0099] Figure 2 It is a schematic diagram of the system module according to an embodiment of the present invention. Detailed Embodiments

[0100] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the following further details the present invention in combination with specific embodiments.

[0101] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention pertains. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "comprising" or "including" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0102] As Figure 1 shown, an intelligent processing method for cardiac magnetic resonance images includes the following steps:

[0103] S1: Obtain the original four-dimensional cardiac magnetic resonance image sequence, and the four-dimensional image sequence includes three-dimensional spatial coordinates and the time dimension;

[0104] S2: Perform phase misalignment correction processing on the four-dimensional image sequence to generate a spatio-temporal alignment image sequence, and the phase misalignment correction includes a respiratory motion compensation layer and an electrocardiogram gating synchronization layer;

[0105] S3: Perform dynamic segmentation of myocardial tissue based on the spatio-temporal alignment image sequence, and output a three-dimensional dynamic mask sequence including the endocardium, myocardial layer, and epicardium;

[0106] S4: Extract multi-dimensional features from the three-dimensional dynamic mask sequence to generate a dynamic feature map including a displacement field, strain rate, and blood flow propagation velocity;

[0107] S5: Input the dynamic feature map into a hybrid analysis model, and the hybrid analysis model includes a spatio-temporal graph convolutional network branch and a physical constraint differential equation branch, and output the myocardial motion abnormal region and the microcirculation disorder quantification parameters;

[0108] S6: Generate a dynamic visualization diagnosis report based on the myocardial motion abnormal region and the microcirculation disorder quantification parameters, and the dynamic visualization diagnosis report includes a myocardial anisotropic motion trajectory cloud map and a microcirculation parameter heat map.

[0109] The four-dimensional image sequence in S1 is represented as {I(x, y, z, t)}, including the three-dimensional spatial coordinates (x, y, z) and the time dimension t. Here, I(x, y, z, t) is the image pixel value at the position (x, y, z) at time t, and t is the time dimension, representing the time frame index in this image sequence; the time interval of the four-dimensional image sequence is Δt, and the resolution of the four-dimensional image sequence is D res =(D x , D y , D z ), where D x , D y , D z respectively represent the resolutions of the image in the three spatial dimensions of x, y, and z. Δt is the time interval between two consecutive frames, and the unit is seconds.

[0110] S2 includes:

[0111] S21: Perform respiratory motion compensation processing on the four-dimensional image sequence. Based on the time dimension t in the four-dimensional image sequence, by extracting the thoracic displacement information of the cardiac image in each frame of the image, use a non-rigid deformation model to correct the respiratory motion. The non-rigid deformation model is expressed as:

[0112]

[0113] where I corr (x, y, z, t) is the image after respiratory motion compensation, and I(x, y, z, t) is the four-dimensional image sequence, represents the convolution operation, is the respiratory motion compensation field, representing the displacement field at the spatial position (x, y, z) and the time frame t;

[0114] S22: After respiratory motion compensation, perform electrocardiogram gating synchronization correction. By analyzing the electrocardiogram (ECG) signal, synchronize and align the four-dimensional image sequence with the electrocardiogram signal. The electrocardiogram gating synchronization correction includes:

[0115] Obtain the electrocardiogram signal, analyze the QRS peak position of the electrocardiogram signal, and obtain the electrocardiogram cycle T ECG , perform QRS wave detection on the ECG signal to obtain the QRS peak time series t QRS ={t1, t2,..., t n}), where t n represents the start time of the nth electrocardiogram cycle. By the QRS peak position, calculate the time interval between two adjacent QRS peaks, and this time interval is the electrocardiogram cycle T ECG , for adjacent QRS peak times t i and t i+1 , the electrocardiogram cycle T ECGCalculated as: T ECG = t i+1 - t i This cardiac cycle represents a complete cycle of the heart, and the cardiac cycle will be used as the time reference for subsequent synchronization and image resampling;

[0116] According to the QRS peak positions of the electrocardiogram signal, determine the acquisition time of each image frame, map it to the corresponding cardiac cycle phase. In the four-dimensional image sequence, each frame of image corresponds to a time point t i ′, representing the acquisition time of this image. In order to align the image sequence with the cardiac cycle, it is necessary to find the cardiac cycle phase corresponding to each frame of image. For each frame of image, find its corresponding QRS peak time point t j , such that the image acquisition time t i ′ is between two adjacent QRS peaks, that is: t j ≤ t i ′ ≤ t j+1 . In this way, ensure that the acquisition time of each frame of image can correspond to the start and end of the cardiac cycle; resample the four-dimensional image sequence so that it is evenly distributed in the phase of the cardiac cycle, thereby obtaining the four-dimensional image sequence I ECG (x, y, z, t′), where t′ is the synchronized time dimension, perform linear interpolation processing on the original image sequence, and resample each image frame I(x, y, z, t′) to a new time point ′t i ”, thereby generating the synchronized image sequence I ECG (x, y, z, t”), where t” is the resampled time point, is the phase of the image acquisition time t i ′ within the cardiac cycle, and the range is

[0117] Through this process, ensure that there is an image frame corresponding to each phase of the cardiac cycle in the image sequence, so that the image sequence is completely synchronized with the cardiac cycle in time;

[0118] S23: Based on the obtained four-dimensional image sequence I corr (x, y, z, t) after respiratory motion compensation and the image sequence I ECG (x, y, z, t′) synchronized by electrocardiogram gating, fuse them through a spatio-temporal transformation model to generate a spatio-temporally aligned image sequence I aligned (x, y, z, t”), and the spatio-temporal transformation model is expressed as:

[0119]

[0120] Among them, Ialigned (x, y, z, t”) is the image sequence after spatio-temporal alignment, t” is the synchronized time dimension, and Δt1 represents the time alignment offset, which is optimized by minimizing the differences between image sequences.

[0121] S3 includes:

[0122] S31: Input the spatio-temporally aligned image sequence {I aligned (x, y, z, t”)};

[0123] S32: Segment the myocardial tissue from the spatio-temporally aligned image sequence. Use an image segmentation model (3DU-Net model) to automatically segment the heart region. The image segmentation model extracts the boundaries of the endocardium, myocardium, and epicardium by comparing the gray-scale information, texture features, and edge information in the spatio-temporally aligned image sequence;

[0124] S33: During the segmentation process, each layer of the heart (endocardium, myocardium, epicardium) is processed through a series of convolutional operations and pooling operations, and finally, the masks of the three tissues are output through a fully connected layer; each layer of mask is classified according to different gray-scale and structural features, so as to obtain the precise positions of the endocardium, myocardium, and epicardium on each frame of image; for each frame t”, the boundaries of the endocardium, myocardium, and epicardium segmented and output by the image segmentation model are used to generate the mask of the endocardium The mask of the myocardium and the mask of the epicardium The mask of each tissue is a three-dimensional matrix, representing the tissue region in space (x, y, z), where the value of 1 indicates that the position belongs to the tissue, and the value of 0 indicates that the position does not belong to the tissue;

[0125] S34: Combine the masks of the endocardium The mask of the myocardium and the mask of the epicardium for each time frame t” to obtain a three-dimensional dynamic mask sequence containing the dynamic changes of the endocardium, myocardium, and epicardium in all time frames It can accurately represent the spatio-temporal dynamics of myocardial tissue.

[0126] The steps of automatically segmenting the heart region by the image segmentation model in S32 include:

[0127] S321, Input image: Input the spatio-temporally aligned image sequence {I aligned (x, y, z, t”)};

[0128] S322, Encoder Part: The purpose of the encoder part is to extract the features of the input image and gradually downsample the spatial resolution of the image through multi-layer convolutional operations. A max pooling layer follows each convolutional layer for downsampling. The input spatio-temporally aligned image sequence {I aligned (x, y, z, t”)} passes through the first 3D convolutional layer to obtain the feature map F1(x, y, z, t”), which is expressed as:

[0129] F1(x, y, z, t″) = ReLU(Conv3(I aligned (x, y, z, t″)));

[0130] where Conv3 is the 3D convolutional operation acting on the input image I aligned (x, y, z, t”), and ReLU is the activation function applied to the result of the convolutional operation. By continuously stacking convolutional layers and pooling layers, the encoder extracts increasingly abstract image features;

[0131] After passing through the k-th layer of convolution and pooling, the output feature map is F k (x, y, z, t”), which is expressed as:

[0132] F k (x, y, z, t″) = ReLU(Conv3(MaxPool3(F k-1 (x, y, z, t″))));

[0133] where MaxPool3 represents the 3D max pooling operation acting on the feature map F k-1 (x, y, z, t”), and F k-1 (x, y, z, t”) is the feature map of k - 1;

[0134] S323, Decoder Part: The decoder part gradually restores the feature map to the original resolution through deconvolution operations. Each layer of the decoder contains a 3D deconvolution layer for the purpose of restoring the spatial resolution. The feature map F k (x, y, z, t”) of the encoder part is input into the k-th layer of the decoder. After deconvolution and the activation function, the decoder output D k ′(x, y, z, t”) is obtained, which is expressed as:

[0135] D k ′(x, y, z, t”) = ReLU(Deconv3(F k (x, y, z, t”)));

[0136] where Deconv3 is the 3D deconvolution operation and ReLU is the activation function acting on the deconvolution result;

[0137] S324, Skip Connection: The skip connection is used to connect the shallow features of the encoder to the corresponding layer of the decoder to retain more spatial details. Specifically, the skip connection concatenates the feature map F k (x, y, z, t”) of the k-th layer of the encoder with the input feature map of the k-th layer of the decoder and inputs it into the deconvolution operation. The skip connection concatenates the feature map F k (k, y, z, t”) of the k-th layer of the encoder with the input feature map D k ′(x, y, z, t”) of the k-th layer of the decoder, and the final decoder output D k (x, y, z, t”) is obtained, which is expressed as:

[0138]

[0139] Where: represents the concatenation operation;

[0140] S325, Image Output: The final decoder output is output through the Softmax function to obtain the output image M(x, y, z, t”), which is expressed as:

[0141] O(x, y, z, t″) = Softmax(Conv1(D k (x, y, z, t″)));

[0142] Where Conv1 is a 1×1×1 convolution operation that maps the decoded feature map D k (x, y, z, t”) to three-class outputs. Softmax is a normalization operation that converts the value of each voxel into a probability, representing the probability that the voxel belongs to the endocardium, myocardium, or epicardium.

[0143] In S33, it is processed through a series of convolution operations and pooling operations, and finally the masks of three tissues are output through the fully connected layer, including:

[0144] S331, Convolution Operation: The input is the spatio-temporally aligned image sequence I aligned (x, y, z, t”) with a convolution kernel size of 3×3×3, that is, convolution processing is performed on the three-dimensional space, and the output feature map F1(x, y, z, t”) is obtained. After the convolution operation, the output feature map (also called the convolution layer output) is obtained:

[0145]

[0146] Where w ijk is the convolution kernel weight, representing the convolution with the input image I alignedThe convolution weights at the spatial position (i, j, k) are (x, y, z, t”), b is the bias term, and ReLU is the activation function that acts on the convolution result to ensure the non - linear characteristics of the output. This convolution operation extracts local features in the image, such as edges, textures, etc. In a deep convolutional network, the feature map F after the input passes through the k - th layer of convolution is k (x, y, z, t”) is processed through several layers of convolution and is calculated as:

[0147] F k (x, y, z, t") = ReLU(Conv3(MaxPool3(Fk -1 (x, y, z, t″)))) ;

[0148] Among them, Conv3 represents the 3D convolution operation, and MaxPool3 is the 3D max - pooling operation. It downsamples the feature map of the previous layer to reduce the spatial dimension;

[0149] S332, pooling operation: It extracts the maximum value of each small local area, thereby reducing the spatial resolution and retaining important spatial information. The input feature map is F k-1 (x, y, z, t”), after passing through the 3D max - pooling operation, the output feature map is F k (x, y, z, t”), which is expressed as:

[0150] F k (x, y, z, t”) = max(F k-1 (x + i, y + j, z + k, t”)), i, j, k ∈ {0, 1, 2} ;

[0151] S333, output result (mask): Each layer of the output mask is expressed as:

[0152]

[0153] Among them, are the masks of the endocardium, myocardium, and epicardium of the output. The output of each mask represents the probability that each voxel belongs to this category. The final mask output is a three - dimensional dynamic mask sequence, representing the dynamic changes of each tissue region of the heart in the entire time series.

[0154] S4 includes:

[0155] S41, displacement field extraction: Based on the three - dimensional dynamic mask sequence, by calculating the position changes of each tissue region of the heart (including the endocardium, myocardium, epicardium) between adjacent time frames, a displacement field sequence is generated. For adjacent frames at time points t” and t” + 1, the displacement vector field is calculated which is expressed as:

[0156]

[0157] Among them, x t” , y t” , z t” and x t”+1 , y t”+1 , z t”+1 are respectively the three-dimensional spatial coordinates of each tissue of the heart at times t” and t” + 1;

[0158] S42, Strain rate extraction: By performing a difference calculation on the displacement field sequence, the strain rate of each voxel is obtained. ∈(x, y, z, t”) is expressed as:

[0159]

[0160] Among them, Δt is the time interval, that is, the time difference between adjacent time frames, is the change amount of the displacement field at time t”;

[0161] S43, Blood flow propagation speed extraction: Using the cardiac blood flow information in the three-dimensional mask sequence, by calculating the movement rate of each part of the heart in the time dimension, the blood flow propagation speed is inferred. Using the spatial position change of each point in the mask sequence and combining with the movement trajectory of the heart, the blood flow propagation speed v(x, y, z, t”) is calculated and expressed as:

[0162] Among them, is the modulus of the displacement vector, indicating the distance of blood flow propagation within the time interval Δt;

[0163] S44, Generate dynamic feature map: Synthesize the displacement field, strain rate, and blood flow propagation speed to generate a dynamic feature map F(x, y, z, t”) containing the displacement field, strain rate, and blood flow propagation speed, which is expressed as:

[0164]

[0165] Among them, represents the displacement field, ∈(x, y, z, t”) represents the strain rate, and v(x, y, z, t”) represents the blood flow propagation speed.

[0166] S5 includes:

[0167] S51, Input dynamic feature map: Use the dynamic feature map F(x, y, z, t”) as the model input and pass it to the two branches of the hybrid analysis model. The branches include the spatio-temporal graph convolutional network branch and the physical constraint differential equation branch;

[0168] S52, Spatiotemporal Graph Convolutional Network Branch: In the spatiotemporal graph convolutional network branch, the feature relationships at different time points and spatial positions are captured in the spatiotemporal domain through convolutional operations, expressed as:

[0169] H k = σ(A · H k-1 · W k );

[0170] where H k is the feature representation of the k-th layer, A is the adjacency matrix of the graph, representing the connection relationships between different spatial points, W k is the convolutional kernel weight of the k-th layer, and σ(·) is the activation function (Sigmoid), used to introduce non-linearity;

[0171] S53, Physical Constraint Differential Equation Branch: Based on the blood flow propagation velocity v(x, y, z, t'') and the strain rate ∈(x, y, z, t''), a physical constraint differential equation model is introduced to simulate hemodynamics and myocardial motion. The physical constraint differential equation model is expressed as:

[0172]

[0173] where is the displacement field, representing the motion of the myocardium, ∈ is the strain rate, representing the deformation rate of the myocardium, v is the blood flow velocity, representing the flow rate of blood in the heart, F(·) is a differential equation function derived from the principles of fluid mechanics and elasticity, used to describe the physical behavior of the heart, λ and μ are the Lame constants, describing the elastic properties of the material, tr(∈) is the trace of the strain tensor (i.e., the volume strain), I is the identity matrix, and η is the viscosity coefficient of the blood, representing the mechanical effect of blood flow on tissues, represents the gradient of the blood flow velocity, capturing the influence of blood flow on tissues;

[0174] S54, Fusion and Output: Based on the output results of the spatiotemporal graph convolutional network branch and the physical constraint differential equation branch, they are combined through a fusion mechanism to obtain the abnormal myocardial motion region and the quantitative parameters of microcirculation disorders, expressed as:

[0175] Y = λ1 · H k + λ2 · Y PDE ;

[0176]

[0177] where H k is the output feature of the spatiotemporal graph convolutional network, representing the abnormal myocardial motion region, and Y PDE is the output feature of the physical constraint differential equation branch, representing the quantitative parameters of microcirculation disorders, Describes the displacement field, strain rate, and blood flow velocity of the heart. λ1 and λ2 are weighting coefficients representing the contribution degrees of different branch outputs, where λ1, λ2 ∈ [0, 1].

[0178] S6 includes:

[0179] S61, input quantization parameters: Input the quantization parameters from the hybrid analysis model. The quantization parameters include the quantification parameters of myocardial motion abnormality regions and microcirculation disorders;

[0180] S62, generate myocardial anisotropic motion trajectory cloud map: Based on the myocardial motion abnormality regions and their anisotropic information, generate a motion trajectory cloud map;

[0181] The steps to generate the motion trajectory cloud map include:

[0182] Data extraction: Extract data on myocardial anisotropic motion regions from the quantization parameters, especially the direction and amplitude of myocardial anisotropic motion;

[0183] Motion trajectory modeling: Use interpolation algorithms to smooth and fit the myocardial motion trajectories to construct dynamic myocardial motion trajectories;

[0184] Trajectory visualization: Render the anisotropic motion trajectories in three-dimensional space, using color mapping and transparency changes to highlight the intensity and abnormality degree of myocardial motion, and generate a dynamic myocardial anisotropic motion trajectory cloud map;

[0185] S63, generate a heat map of microcirculation parameters: Based on the quantization parameters of microcirculation disorders, generate a heat map;

[0186] The steps to generate the heat map include:

[0187] Data standardization: Normalize the quantified microcirculation disorder parameters to ensure that the color mapping of the heat map meets the visualization requirements and highlights the regions of microcirculation abnormalities;

[0188] Heat map plotting: Associate the microcirculation parameters with the three-dimensional spatial coordinates of the heart and generate corresponding color values at each spatial point. Visualize the degree of microcirculation abnormalities in different regions using a heat map. The common color scheme is a gradient color from cold tones (low values) to warm tones (high values) to represent different degrees of microcirculation disorders;

[0189] S64, dynamic display and interaction functions: Generate a dynamic visualization report through a graphical user interface. The report includes the myocardial anisotropic motion trajectory cloud map and the heat map of microcirculation parameters, and provides interaction functions such as rotation, zooming, and perspective switching;

[0190] S65. Generate a final diagnosis report: Based on the myocardial anisotropic motion trajectory cloud map and the microcirculation parameter heat map, generate a final diagnosis report, and provide specific descriptions of regional annotation, abnormal motion, and microcirculation disorders.

[0191] As Figure 2 shown, a cardiac magnetic resonance image intelligent processing system for implementing the above-mentioned cardiac magnetic resonance image intelligent processing method includes the following modules:

[0192] Four-dimensional image acquisition module: Acquire the original cardiac magnetic resonance four-dimensional image sequence, and the image sequence includes spatial three-dimensional coordinates and time dimension;

[0193] Phase misalignment correction module: Perform phase misalignment correction processing on the four-dimensional image sequence to generate a spatio-temporally aligned image sequence;

[0194] Myocardial tissue dynamic segmentation module: Based on the spatio-temporally aligned image sequence, perform dynamic segmentation on the myocardial tissue, and output a three-dimensional dynamic mask sequence including the endocardium, myocardial layer, and epicardium;

[0195] Feature extraction module: Perform multi-dimensional feature extraction on the three-dimensional dynamic mask sequence to generate a dynamic feature map including displacement field, strain rate, and blood flow propagation velocity;

[0196] Hybrid analysis model module: Input the dynamic feature map into the hybrid analysis model, and the hybrid analysis model includes a spatio-temporal graph convolutional network branch and a physical constraint differential equation branch, and output the myocardial motion abnormal region and the microcirculation disorder quantification parameter;

[0197] Visualization report generation module: Generate a dynamic visualization diagnosis report, and the dynamic visualization diagnosis report includes a myocardial anisotropic motion trajectory cloud map and a microcirculation parameter heat map.

[0198] Those of ordinary skill in the art should understand that the discussion of any above embodiment is only exemplary, and is not intended to imply that the scope of the present invention is limited to these examples; under the idea of the present invention, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of the present invention as described above, and they are not provided in detail for the sake of brevity.

Claims

1. A cardiac magnetic resonance image intelligent processing method, characterized in that: The following steps are involved: S1: Acquire an original cardiac magnetic resonance four-dimensional image sequence, wherein the four-dimensional image sequence includes a spatial three-dimensional coordinate and a time dimension; S2: performing phase misalignment correction processing on the four-dimensional image sequence to generate a spatiotemporal aligned image sequence, wherein the phase misalignment correction includes a respiratory motion compensation layer and an ECG gating synchronization layer; S3: performing dynamic segmentation of myocardial tissue based on the spatiotemporal aligned image sequence, and outputting a three-dimensional dynamic mask sequence including the endocardium, myocardium, and epicardium; S4: performing multi-dimensional feature extraction on the three-dimensional dynamic mask sequence to generate a dynamic feature map including displacement field, strain rate, and blood flow propagation velocity; S5: inputting the dynamic characteristic map into a hybrid analysis model, wherein the hybrid analysis model comprises a spatiotemporal graph convolutional network branch and a physical constraint differential equation branch, and outputting quantitative parameters of abnormal myocardial motion areas and microcirculatory disorders; S6: Generate a dynamic visualization diagnosis report based on the abnormal myocardial motion area and the quantitative parameters of microcirculation disorder, wherein the dynamic visualization diagnosis report includes a cloud map of myocardial anisotropic motion trajectory and a heat map of microcirculation parameters.

2. The method for intelligent processing of cardiac magnetic resonance images according to claim 1, characterized in that: The four-dimensional image sequence in S1 is represented as {I(x, y, z, t)}, including three-dimensional spatial coordinates (x, y, z) and time dimension t; the time interval of the four-dimensional image sequence is Δt, and the resolution of the four-dimensional image sequence is D res =(D x ,D y ,D z ), where D x ,D y ,D z They represent the resolution of the image in the three spatial dimensions of x, y, and z respectively, and Δt is the time interval between two consecutive frames.

3. A cardiac magnetic resonance image intelligent processing method according to claim 2, characterized in that: The S2 includes: S21: performing respiratory motion compensation processing on the four-dimensional image sequence, based on the time dimension t in the four-dimensional image sequence, by extracting the chest displacement information of the heart image in each frame image, and using a non-rigid deformation model to correct the respiratory motion, the non-rigid deformation model is expressed as: Among them, I coor (x, y, z, t) is the image after respiratory motion compensation, I(x, y, z, t) is the four-dimensional image sequence, represents the convolution operation, is the respiratory motion compensation field, which represents the displacement field at the spatial position (x, y, z) and time frame t; S22: After respiratory motion compensation, ECG gating synchronization correction is performed, and the four-dimensional image sequence is synchronously aligned with the ECG signal by analyzing the ECG signal. The ECG gating synchronization correction includes: Get the ECG signal, analyze the QRS peak position of the ECG signal, and get the ECG cycle T ECG ; According to the QRS peak position of the ECG signal, the acquisition time of each image frame is determined and mapped to the corresponding ECG cycle phase; The four-dimensional image sequence is resampled so that it is evenly distributed in the phase of the ECG cycle, thereby obtaining a four-dimensional image sequence I after ECG gating synchronization. ECG (x, y, z, t′), where t′ is the time dimension after synchronization; S23: Based on the obtained respiratory motion compensated four-dimensional image sequence I corr (x,y,z,t) image sequence after synchronization with ECG gating I ECG (x, y, z, t′), fused through the spatiotemporal transformation model to generate a spatiotemporal aligned image sequence I aligned (x,y,z,t”).

4. The method for intelligent processing of cardiac magnetic resonance images according to claim 3, characterized in that: The S3 includes: S31: Input the spatiotemporal aligned image sequence {I aligned (x,y,z,t')}; S32: performing myocardial tissue segmentation on the spatiotemporal aligned image sequence, and automatically segmenting the heart region using an image segmentation model, wherein the image segmentation model extracts the boundaries of the endocardium, myocardium, and epicardium by comparing grayscale information, texture features, and edge information in the spatiotemporal aligned image sequence; S33: During the segmentation process, each layer of the heart is processed through a series of convolution operations and pooling operations, and finally the masks of the three tissues are output through the fully connected layer; for each frame t", the endocardium, myocardium and epicardium boundaries are segmented and output by the image segmentation model to generate the endocardial mask Myocardial mask and epicardial mask S34: The endocardial mask of each time frame t' is Myocardial mask and epicardial mask Combined, we get a 3D dynamic mask sequence that contains the dynamic changes of the endocardium, myocardium, and epicardium in all time frames:

5. The method for intelligent processing of cardiac magnetic resonance images according to claim 4, characterized in that: The step of automatically segmenting the heart region using the image segmentation model in S32 includes: S321, input image: input spatiotemporal aligned image sequence {I aligned (x,y,z,t')}; S322, encoder part: input spatiotemporal aligned image sequence {I aligned The feature map F1(x,y,z,t") after the first layer of 3D convolution is expressed as: F1(x,y,z,t″)=ReLU(Conv3(I aligned (x,y,z,t″)))? Among them, Conv3 is a 3D convolution operation and ReLU is an activation function; After the kth layer of convolution and pooling, the output feature map is F k (x,y,z,t”), expressed as: F k (x,y,z,t″)=ReLU(Conv3(MaxPool3(F k-1 (x,y,z,t″))))? Among them, MaxPool3 represents the 3D maximum pooling operation, F k-1 (x,y,z,t”) is the k-1 feature map; S323, decoder part: feature map F of the encoder part k (x, y, z, t”) is input to the kth layer of the decoder, and after deconvolution and activation function, the decoder output D is obtained. k ′(x,y,z,t”), expressed as: D k ′(x,y,z,t”)=ReLU(Deconv3(F k (x,y,z,t”)))? Among them, Deconv3 is a 3D deconvolution operation, and ReLU is an activation function; S324, skip connection: The skip connection connects the feature map F of the kth layer of the encoder k (x,y,z,t') and the input feature map D of the kth layer of the decoder k ′(x,y,z,t”) to get the final decoder output D k (x,y,z,t'); S325, image output: The final decoder output is output through the Softmax function to obtain an output image M (x, y, z, t”).

6. The method for intelligent processing of cardiac magnetic resonance images according to claim 4, characterized in that: In S33, a series of convolution operations and pooling operations are performed, and finally the masks of three tissues are output through the fully connected layer, including: S331, convolution operation: input is a spatiotemporal aligned image sequence I aligned (x, y, z, t”), the convolution kernel size is 3×3×3, that is, convolution processing is performed on the three-dimensional space, and the feature map F1(x, y, z, t”) is output. In the convolution network, the feature map F after the k-th layer of convolution is input. k (x,y,z,t”) is processed through several layers of convolution and calculated as: F k (x,y,z,t″)=ReLU(Conv3(MaxPool3(F k-1 (x,y,z,t″))))? Among them, Conv3 represents the 3D convolution operation, and MaxPool3 is the 3D maximum pooling operation; S332, pooling operation: input feature map is F k-1 (x,y,z,t”), after 3D maximum pooling operation, the output feature map F k (x,y,z,t”), expressed as: F k (x,y,z,t”)=max(F k-1 (x+i,y+j,z+k,t”)),i,j,k∈{0,1,2}; S333, outputs the mask representations of the endocardium, myocardium and epicardium, which are:

7. A cardiac magnetic resonance image intelligent processing method according to claim 6, characterized in that: The S4 includes: S41, displacement field extraction: Based on the three-dimensional dynamic mask sequence, the displacement field sequence is generated by calculating the position changes of various cardiac tissue regions between adjacent time frames. For adjacent frames at time points t” and t”+1, the displacement vector field is calculated. It is expressed as: Among them, x t” ,y t” ,z t” and x t”+1 ,y t”+1 ,z t”+1 are the three-dimensional spatial coordinates of each cardiac tissue at time t” and t”+1 respectively; S42, strain rate extraction: By performing differential calculation on the displacement field sequence, the strain rate of each voxel is obtained, ∈(x,y,z,t”) expressed as: Among them, Δt is the time interval, that is, the time difference between adjacent time frames, is the displacement field The amount of change at time t'; S43, blood flow propagation velocity extraction: Using the spatial position changes of each point in the mask sequence and the motion trajectory of the heart, the blood flow propagation velocity v(x, y, z, t”) is calculated as: in, is the modulus of the displacement vector, which represents the distance the blood flow travels in the time interval Δt; S44, generating a dynamic characteristic spectrum: synthesizing the displacement field, strain rate and blood flow propagation velocity to generate a dynamic characteristic spectrum F(x, y, z, t”) including the displacement field, strain rate and blood flow propagation velocity.

8. The method for intelligent processing of cardiac magnetic resonance images according to claim 1, characterized in that: The S5 includes: S51, input dynamic feature map: take the dynamic feature map F(x, y, z, t”) as model input and pass it to two branches of the hybrid analysis model, including a spatiotemporal graph convolutional network branch and a physical constraint differential equation branch; S52, spatiotemporal graph convolutional network branch: In the spatiotemporal graph convolutional network branch, the feature relationship between different time points and spatial positions is captured in the spatiotemporal domain through convolution operations, which is expressed as: H k =σ(A·H k-1 ·W k ); Among them, H k is the feature representation of the kth layer, A is the adjacency matrix of the graph, which represents the connection relationship between different spatial points, and W k is the weight of the k-th convolution kernel, σ(·) is the activation function used to introduce nonlinearity; S53, physical constraint differential equation branch: Based on the blood flow propagation velocity v(x,y,z,t”) and strain rate ∈(x,y,z,t”), a physical constraint differential equation model is introduced to simulate hemodynamics and myocardial motion; S54, fusion and output: Based on the output results of the spatiotemporal graph convolutional network branch and the physical constraint differential equation branch, the quantitative parameters of myocardial motion abnormality area and microcirculatory disorder are obtained through the fusion mechanism, which is expressed as: Y=λ1·H k +λ2·Y PDE ; Among them, H k is the output feature of the spatiotemporal graph convolutional network, indicating the area of ​​abnormal myocardial motion, Y PDE is the output feature of the physical constraint differential equation branch, representing the quantitative parameter of microcirculatory disorder, Describes the displacement field, strain rate and blood flow velocity of the heart, is the displacement field, representing the movement of the myocardium, ∈ is the strain rate, representing the deformation rate of the myocardium, v is the blood flow velocity, representing the flow rate of blood in the heart, λ1 and λ2 are weighting coefficients, representing the contribution of different branch outputs, λ1,λ2∈[0,1].

9. The method for intelligent processing of cardiac magnetic resonance images according to claim 1, characterized in that: The S6 includes: S61, inputting quantitative parameters: inputting quantitative parameters from the hybrid analysis model, wherein the quantitative parameters include quantitative parameters of abnormal myocardial motion area and microcirculatory disorder; S62, generating a myocardial anisotropic motion trajectory cloud map: generating a motion trajectory cloud map based on the abnormal motion region of the myocardium and its anisotropy information; S63, generating a microcirculatory parameter heat map: generating a heat map based on the quantitative parameters of microcirculatory disorders; S64, dynamic display and interactive function: Generate dynamic visualization reports through the graphical user interface, including myocardial anisotropic motion trajectory cloud map and microcirculation parameter thermal map, and provide interactive functions; S65, generate final diagnosis report: based on the myocardial anisotropic motion trajectory cloud map and microcirculation parameter heat map, generate the final diagnosis report and provide specific descriptions of regional annotations, motion abnormalities and microcirculation disorders.

10. A cardiac magnetic resonance image intelligent processing system, used to implement a cardiac magnetic resonance image intelligent processing method according to any one of claims 1 to 9, characterized in that: Includes the following modules: Four-dimensional image acquisition module: acquires original cardiac magnetic resonance four-dimensional image sequence, wherein the image sequence includes spatial three-dimensional coordinates and time dimension; Phase misalignment correction module: performs phase misalignment correction on the four-dimensional image sequence to generate a time-space aligned image sequence; Myocardial tissue dynamic segmentation module: Based on the spatiotemporal aligned image sequence, the myocardial tissue is dynamically segmented and a three-dimensional dynamic mask sequence including the endocardium, myocardium and epicardium is output; Feature extraction module: performs multi-dimensional feature extraction on the three-dimensional dynamic mask sequence to generate dynamic feature maps including displacement field, strain rate, and blood flow propagation velocity; Hybrid analysis model module: input the dynamic feature map into the hybrid analysis model, the hybrid analysis model includes a spatiotemporal graph convolutional network branch and a physical constraint differential equation branch, and outputs quantitative parameters of abnormal myocardial motion area and microcirculatory disorder; Visual report generation module: generates a dynamic visual diagnosis report, which includes a myocardial anisotropic motion trajectory cloud map and a microcirculation parameter thermal map.

Images