Myocardial perfusion and survivability evaluation method based on cardiac ultrasound contrast intelligent analysis

Abstract

The invention relates to the technical field of cardiovascular disease diagnosis, and discloses a myocardial perfusion and survivability evaluation method based on cardiac ultrasound contrast intelligent analysis. The method comprises the following steps: firstly, acquiring cardiac ultrasound contrast dynamic image data and pre-processing the data, then performing myocardial region segmentation by adopting an improved U-Net network in combination with an attention mechanism and a dynamic region growing algorithm, and constructing a myocardial perfusion zoning model. Then, myocardial perfusion dynamic characteristics are extracted, key time sequence characteristics are screened, and a multi-modal fusion evaluation model is constructed to generate a myocardial survivability probability distribution diagram. And finally, performing quantitative evaluation based on the probability distribution diagram, outputting a myocardial perfusion defect index and a survivability score, and generating a visual three-dimensional evaluation report. The accuracy of myocardial perfusion and survivability evaluation is improved, multi-modal information is integrated, quantitative evaluation and visualization are achieved, model robustness and clinical working efficiency are improved, and powerful support is provided for diagnosis and treatment of cardiovascular diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of cardiovascular disease diagnosis, and specifically provides a method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac contrast echocardiography. Background Art

[0002] Cardiovascular diseases pose a serious threat to human health and are one of the main causes of death and disability globally. Accurately evaluating myocardial perfusion and viability is crucial for the diagnosis, treatment decision-making, and prognosis judgment of cardiovascular diseases. Currently, there are numerous methods for evaluating myocardial perfusion and viability in clinical practice, but each method has certain limitations.

[0003] Traditional coronary angiography, regarded as the "gold standard" for diagnosing coronary heart disease, can clearly show the anatomical structure of the coronary arteries. However, it is an invasive examination with certain risks and high costs, making it impossible to be widely used in large-scale screening. Meanwhile, coronary angiography can only reflect the morphological changes of the coronary arteries and cannot directly evaluate the myocardial perfusion and viability.

[0004] Radionuclide myocardial perfusion imaging has high accuracy in evaluating myocardial perfusion. However, this method requires the use of radioactive nuclides, posing a radiation hazard and having a potential impact on the health of patients and medical staff. In addition, the examination equipment is expensive and the examination process is complex, restricting its popularization in primary medical institutions.

[0005] Cardiac magnetic resonance imaging (MRI) can also be used for evaluating myocardial perfusion and viability. It has good soft tissue resolution and can provide rich myocardial information. However, the MRI examination takes a long time, requires a high degree of patient cooperation, and is not suitable for special patient groups with metal implants in the body. Moreover, the cost of MRI equipment is high and the examination fee is expensive, making it difficult to promote clinically.

[0006] Cardiac contrast echocardiography technology has been widely used in clinical practice due to its advantages such as simple operation, relatively low price, and no radiation. However, the current analysis of cardiac contrast echocardiography images mainly relies on doctors' subjective experience, resulting in poor accuracy and repeatability. When observing images, doctors are easily affected by factors such as their own professional level, fatigue degree, and image quality, leading to deviations in the evaluation of myocardial perfusion and viability. In addition, manually analyzing contrast echocardiography images is time-consuming and laborious, with low efficiency, and it is difficult to meet the rapid diagnosis needs of a large number of clinical cases.

[0007] With the rapid development of artificial intelligence technology, its application in the field of medical image processing has become a research hotspot. However, there are still many deficiencies in the existing artificial intelligence-based methods for myocardial perfusion and viability assessment. Some methods only analyze data from a single modality and cannot fully integrate multi-source information, resulting in incomplete and inaccurate assessment results. There are also some methods that are too simple in model construction and cannot effectively extract the complex features of the myocardium, making it difficult to adapt to the differences in the conditions of different patients. Moreover, in key links such as image preprocessing, region segmentation, and feature selection, there is a lack of targeted optimization strategies, which affects the final assessment accuracy. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for assessing myocardial perfusion and viability based on intelligent analysis of cardiac contrast echocardiography to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: A method for assessing myocardial perfusion and viability based on intelligent analysis of cardiac contrast echocardiography, the method comprising:

[0010] Step 1: Obtain dynamic imaging data of cardiac contrast echocardiography, including myocardial perfusion sequence images at multiple angles and multiple time phases; further preprocess the sequence images;

[0011] Step 2: Perform myocardial region segmentation on the preprocessed temporal images. An improved U-Net network combined with an attention mechanism is used to automatically extract the inner and outer myocardial contours and divide the myocardial segments with the contrast characteristics of the left ventricular cavity as the boundary constraint; at the same time, a dynamic region growing algorithm is introduced to adaptively segment the regions with abnormal contrast agent filling to construct a myocardial perfusion partition model;

[0012] Step 3: Extract dynamic features of myocardial perfusion based on the segmentation results. A three-dimensional convolutional neural network is used to construct a spatio-temporal feature encoder to extract the contrast agent filling rate, peak intensity decay curve, and spatial heterogeneity parameters of each myocardial segment; combined with the reinforcement learning method, the feature selection process is optimized through the Q-learning strategy to screen the key temporal features related to myocardial viability;

[0013] Step 4: Construct a multi-modal fusion assessment model to perform cross-modal alignment of dynamic features and clinical parameters; use a graph neural network to establish an association map of myocardial perfusion and viability, and fuse spatio-temporal features and topological relationships through node embedding technology to generate a myocardial viability probability distribution map;

[0014] Step 5: Perform quantitative assessment based on the probability distribution map. Design an adaptive threshold segmentation algorithm, and dynamically divide the ischemic core area and the marginal area in combination with fuzzy logic reasoning; use the Bayesian optimization method to iteratively adjust the assessment parameters, output the myocardial perfusion defect index and viability score, and generate a visual three-dimensional assessment report.

[0015] Preferably, in the step 1, the method for preprocessing the sequence images includes:

[0016] Aiming at the problems of noise interference and motion artifacts in the original images, an image denoising and motion compensation algorithm based on a generative adversarial network is adopted, and the local motion trajectory of the myocardium is extracted by combining the time series optical flow method to generate high signal-to-noise ratio myocardial contrast time series images;

[0017] The image denoising algorithm of the generative adversarial network adopts a dual-path discriminator structure, where the first discriminator focuses on spatial resolution restoration, and the second discriminator optimizes the temporal continuity based on the cycle consistency loss function; the motion compensation algorithm combines optical flow field estimation and deformation field interpolation to achieve non-rigid registration of local myocardial motion.

[0018] Preferably, in the step 2, the improved U-Net network introduces a multi-scale dilated convolution module in the encoder part to capture multi-resolution features of the myocardial boundary; the attention mechanism adopts a channel-spatial dual attention module to dynamically weight the contrast intensity information of different regions.

[0019] Preferably, in the step 3, the three-dimensional convolution kernel of the spatio-temporal feature encoder adopts an asymmetric structure, and a long short-term memory unit is used in the time dimension to capture the dynamic evolution law of the contrast agent filling; the Q-learning strategy designs a reward function to preferentially retain the feature parameters with a correlation higher than 0.8 with the clinical survival gold standard.

[0020] Preferably, in the step 4, the nodes of the graph neural network represent myocardial segments, and the edge weights are calculated by the hemodynamic similarity between segments; the topological relationship fusion adopts a graph attention mechanism to dynamically adjust the contribution degree of different modality features in the node embedding.

[0021] Preferably, in the step 5, the adaptive threshold segmentation algorithm adopts a level set method combined with a region competition model, and the rule base of fuzzy logic reasoning is driven by expert knowledge, including the contrast agent retention time, the intensity gradient change rate, and the spatial continuity constraint.

[0022] Preferably, the training of the dual-path discriminator adopts a progressive learning strategy. In the first stage, only the spatial resolution is optimized. In the second stage, the temporal consistency loss is jointly optimized. In the final stage, adversarial perturbations are introduced to enhance the robustness of the model.

[0023] Preferably, the seed points of the dynamic region growing algorithm are selected based on the contrast agent first-pass time map, and the growth criterion combines local intensity similarity and the contrast agent diffusion kinetic model. The smoothness of the segmentation boundary is optimized through a Markov random field, where the region growth criterion is expressed as:

[0024]

[0025] Among them, G(p) is the growth probability of pixel point p, I(p) is the intensity value of pixel point p, I(s) is the average intensity of seed point s, σ is the standard deviation of intensity distribution, D(p) is the contrast agent diffusion coefficient, is the time gradient vector, and α and β are the spatial and temporal weight coefficients respectively.

[0026] Preferably, the design of the reward function includes a feature stability index and a redundancy penalty term. The stability index calculates the reciprocal of the variance of feature parameters through Bootstrap resampling, and the redundancy penalty term quantifies the correlation between features based on the mutual information theory. The reward function is expressed as:

[0027] R = γ·(1 / Var(X)) + η·(1 - MI(X,Y))

[0028] Among them, R is the final reward value, Var(X) is the variance of feature parameters, MI(X,Y) is the mutual information between features, γ is the stability weight coefficient, and η is the redundancy penalty coefficient.

[0029] Preferably, the surrogate model of the Bayesian optimization method adopts a deep kernel learning framework, combines Gaussian processes and deep neural networks, and accelerates the hyperparameter search through variational inference; the evaluation parameters include the weight of the ischemic area, the survival probability threshold, and the spatial continuity penalty coefficient, and the objective function is expressed as:

[0030]

[0031] Among them, θ is the parameter combination to be optimized, A is the area of the ischemic area, P is the survival probability, P0 is the probability normalization constant, C is the spatial continuity metric value, and w1, w2, and w3 are weight coefficients.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] The present invention uses an improved U-Net network combined with an attention mechanism for myocardial region segmentation, which can more accurately extract the inner and outer membrane contours of the myocardium and divide myocardial segments. For example, in practical applications, for patients with myocardial ischemia, the improved network can clearly identify the boundaries of the ischemic regions, and the segmentation error is significantly reduced compared with traditional segmentation methods. At the same time, a dynamic region growing algorithm is introduced to adaptively segment the regions with abnormal contrast agent filling, and the constructed myocardial perfusion partition model is more in line with the actual situation. When extracting the dynamic features of myocardial perfusion, a three-dimensional convolutional neural network is used to construct a spatio-temporal feature encoder, and a reinforcement learning method is combined to screen key temporal features, greatly improving the extraction accuracy of features related to myocardial viability. Verified by clinical data, the correlation between evaluating myocardial viability using the method of the present invention and the clinical gold standard has been significantly improved, providing a more reliable diagnostic basis for doctors.

[0034] A multi-modal fusion evaluation model is constructed to perform cross-modal alignment of dynamic features and clinical parameters. An association map of myocardial perfusion and viability is established through a graph neural network, fusing spatio-temporal features and topological relationships, and making full use of multiple information sources. For example, by combining clinical parameters such as the patient's age, blood pressure, and blood lipids, as well as dynamic features such as the contrast agent filling rate and peak intensity decay curve extracted from contrast-enhanced ultrasound images, the myocardial condition can be evaluated more comprehensively. This multi-modal fusion method makes up for the deficiencies of single-information evaluation, making the evaluation results more accurate and comprehensive, helping doctors to understand the patient's condition more deeply and formulate more targeted treatment plans.

[0035] Quantitative evaluation is carried out based on the probability distribution map. The designed adaptive threshold segmentation algorithm combines fuzzy logic reasoning to dynamically divide the ischemic core area and the edge area, and the Bayesian optimization method is used to iteratively adjust the evaluation parameters, outputting the myocardial perfusion defect index and viability score, and generating a visual three-dimensional evaluation report. This makes the evaluation results more intuitive and clear, facilitating doctors to quickly understand and analyze the patient's myocardial condition. For example, doctors can visually see the location, scope, and severity of the myocardial ischemic region through the visual three-dimensional evaluation report, greatly improving the diagnostic efficiency. At the same time, the quantitative evaluation indicators are also beneficial for tracking and comparing the patient's condition and timely adjusting the treatment plan.

[0036] During the image preprocessing process, an image denoising and motion compensation algorithm based on a generative adversarial network is adopted, effectively solving the problems of noise interference and motion artifacts in the original images. The training of the dual-path discriminator adopts a progressive learning strategy, enhancing the robustness of the model. During the feature selection process, the design of the reward function takes into account the feature stability index and the redundancy penalty term, improving the adaptability of the model to different data. For example, when processing contrast-enhanced ultrasound images collected by different devices, the method of the present invention can still stably extract effective features, accurately evaluate myocardial perfusion and viability, and reduce the evaluation errors caused by data differences.

[0037] The present invention realizes an automated process from image acquisition to the generation of evaluation reports, greatly reducing the time and workload of doctors for manual image analysis. Traditional manual analysis methods require doctors to spend a lot of time observing and measuring images, while the method of the present invention can quickly process and analyze images and output evaluation results in a short time. This not only improves the efficiency of clinical diagnosis but also enables doctors to have more time and energy to focus on the treatment and management of patients, which is of great significance for improving the overall diagnosis and treatment level of medical institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is the working principle diagram of the method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac contrast-enhanced ultrasound according to the present invention;

[0039] Figure 2 is the working flow chart of the preprocessing of sequential images;

[0040] Figure 3 is the working flow chart of the extraction of dynamic features of myocardial perfusion and the screening of key timing features;

[0041] Figure 4 is the working flow chart of the construction of a multi-modal fusion evaluation model and the generation of a probability distribution map of myocardial viability. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] Please refer to Figures 1-4 , the present invention provides a method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac contrast-enhanced ultrasound, and its overall implementation solution is as follows:

[0044] Step 1: Obtain and preprocess dynamic myocardial contrast echocardiography image data: Collect dynamic myocardial contrast echocardiography image data through professional contrast echocardiography equipment to ensure the acquisition of myocardial perfusion sequence images from multiple angles and at multiple time phases. These images can comprehensively reflect the perfusion of the myocardium at different angles and times, providing rich information for subsequent analysis. For the problems of noise interference and motion artifacts in the original images, an image denoising and motion compensation algorithm based on a generative adversarial network is adopted. Combine the time-series optical flow method to extract the local motion trajectories of the myocardium, and then generate high-signal-to-noise ratio myocardial contrast time-series images. This process can effectively remove noise and compensate for the artifacts caused by cardiac motion.

[0045] Step 2: Myocardial region segmentation: Segment the myocardial region of the preprocessed time-series images. Use an improved U-Net network combined with an attention mechanism, with the contrast characteristics of the left ventricular cavity as the boundary constraint, to automatically extract the inner and outer myocardial contours and divide the myocardial segments. At the same time, introduce a dynamic region growing algorithm to adaptively segment the regions with abnormal contrast agent filling, and construct a myocardial perfusion partition model. This step can accurately divide the myocardial region, providing an accurate regional basis for subsequent extraction of dynamic myocardial perfusion characteristics.

[0046] Step 3: Extract dynamic myocardial perfusion characteristics and screen key time-series characteristics: Based on the segmentation results, use a three-dimensional convolutional neural network to construct a spatio-temporal feature encoder to extract dynamic characteristics such as the contrast agent filling rate, peak intensity decay curve, and spatial heterogeneity parameters of each myocardial segment. Combine the reinforcement learning method, and optimize the feature selection process through the Q-learning strategy to screen the key time-series characteristics related to myocardial viability. These characteristics can reflect the dynamic changes of myocardial perfusion and are of great significance for evaluating myocardial viability.

[0047] Step 4: Construct a multimodal fusion evaluation model: Construct a multimodal fusion evaluation model to perform cross-modal alignment of dynamic characteristics and clinical parameters. Use a graph neural network to establish an association map between myocardial perfusion and viability, and fuse spatio-temporal characteristics and topological relationships through node embedding technology to generate a probability distribution map of myocardial viability. This model comprehensively considers various information and can more accurately evaluate myocardial viability.

[0048] Step 5: Quantitative evaluation and generation of a visualization report: Based on the probability distribution map, perform quantitative evaluation, design an adaptive threshold segmentation algorithm, and dynamically divide the ischemic core area and the marginal area in combination with fuzzy logic reasoning. Use the Bayesian optimization method to iteratively adjust the evaluation parameters, output the myocardial perfusion defect index and viability score, and generate a visualization three-dimensional evaluation report. This makes the evaluation results more intuitive and accurate, facilitating doctors' diagnosis and treatment decisions.

[0049] The following further details the specific implementation manners of the present invention with 5 embodiments:

[0050] Example 1:

[0051] After obtaining the dynamic imaging data of cardiac ultrasound contrast, the preprocessing of the sequence images is started. For the problems of noise interference and motion artifacts in the original images, an image denoising and motion compensation algorithm based on generative adversarial network is adopted.

[0052] The image denoising algorithm of the generative adversarial network adopts a dual-path discriminator structure. Among them, the first discriminator focuses on spatial resolution restoration. By learning the spatial features of the image, the denoised image is clearer in details and can restore the real myocardial tissue structure as much as possible. The second discriminator optimizes the temporal continuity based on the cycle consistency loss function, and the cycle consistency loss function can be expressed as:

[0053]

[0054] where T represents the total number of frames in the time series, is the original image of the t-th frame, F is the generator used to convert the original image into a denoised image, G is the inverse generator used to restore the denoised image back to the original image, and ‖·‖1 represents the L1 norm. This function ensures that the denoised image remains coherent in the time series and avoids frame skipping or discontinuity.

[0055] The motion compensation algorithm combines optical flow field estimation and deformation field interpolation to achieve non-rigid registration of local myocardial motion. Optical flow field estimation is used to calculate the motion vectors of myocardial tissue between adjacent frames. Assuming that I(x, y, t) represents the image pixel value at the coordinate (x, y) at time t, according to the basic assumption of optical flow, there is:

[0056]

[0057] where u and v are the components of the optical flow in the x and y directions respectively. By solving this equation, the optical flow field can be obtained. Then, a deformation field is generated based on the optical flow field, and non-rigid registration of the image is performed using deformation field interpolation to align the myocardial tissue of different frames in space and reduce the influence of motion artifacts.

[0058] The training of the dual-path discriminator adopts a progressive learning strategy. In the first stage, only the spatial resolution is optimized. By minimizing the loss function of the first discriminator, the generated denoised image is closer to the real image in spatial details. In the second stage, the temporal consistency loss is jointly optimized, and the cycle consistency loss function is incorporated into the optimization objective to make the denoised image more coherent in the time series. In the final stage, adversarial perturbations are introduced to enhance the robustness of the model. During the training process, perturbations such as random noise are added to make the model adapt to different noise environments and improve the stability of the denoising effect.

[0059] Example 2:

[0060] When performing myocardial region segmentation on the preprocessed temporal images, the improved U-Net network introduces a multi-scale dilated convolution module in the encoder part. The multi-scale dilated convolution module can capture multi-resolution features of the myocardial boundary. The convolution kernel of dilated convolution introduces holes on the basis of the ordinary convolution kernel. By adjusting the dilation rate, the image features can be felt at different scales. Assuming that the size of the convolution kernel of dilated convolution is k and the dilation rate is d, for the input feature map x, the calculation formula of dilated convolution is:

[0061]

[0062] where y[i] is the output of dilated convolution, w[j] is the convolution kernel weight, and i and j are indices. By parallel computing with convolution kernels of different dilation rates, features at different scales can be obtained, so as to capture the myocardial boundary information more comprehensively.

[0063] The attention mechanism adopts a channel-spatial dual attention module to dynamically weight the contrast intensity information of different regions. The channel attention module calculates the importance weight of each channel through global average pooling and fully connected layers. Assuming that the input feature map is F ∈ R C×H×W , the channel attention weight M c ∈ R C×1×1 is calculated as follows:

[0064] M c = σ(FC2(ReLU(FC1(GAP(F)))))

[0065] where σ is the Sigmoid function, FC1 and FC2 are fully connected layers, and GAP is the global average pooling operation. The spatial attention module obtains the spatial attention weight M s ∈ R 1×H×W :

[0066] M s = σ(Conv([MaxPool(F); AvgPool(F)]))

[0067] The final attention weight represents element-wise multiplication. The original feature map is weighted by this weight to highlight the contrast intensity information of important regions.

[0068] The seed points of the dynamic region growing algorithm are selected based on the contrast agent first-pass time map. The contrast agent first-pass time map reflects the time when the contrast agent first reaches each position in the myocardial tissue. Seed points are selected in the regions with the earliest contrast agent first-pass time. These regions are usually the regions where the contrast agent is filled first, which is beneficial to accurately segment the abnormal contrast agent filling regions. The growth criterion combines local intensity similarity and the contrast agent diffusion kinetic model, and optimizes the smoothness of the segmentation boundary through the Markov random field. The regional growth criterion is expressed as:

[0069]

[0070] where \(G(p)\) is the growth probability of pixel point \(p\), \(I(p)\) is the intensity value of pixel point \(p\), \(I(s)\) is the average intensity of the seed point \(s\), \(\sigma\) is the standard deviation of the intensity distribution, \(D(p)\) is the contrast agent diffusion coefficient, is the time gradient vector, and \(\alpha,\beta\) are the spatial and temporal weight coefficients respectively. This criterion comprehensively considers the intensity similarity between the pixel point and the seed point and the diffusion of the contrast agent, making the segmentation result more accurate and smooth.

[0071] Example 3:

[0072] When extracting the dynamic features of myocardial perfusion based on the segmentation result, a spatio-temporal feature encoder is constructed using a three-dimensional convolutional neural network. The three-dimensional convolutional kernel of the spatio-temporal feature encoder adopts an asymmetric structure, and a long short-term memory unit (LSTM) is used in the time dimension to capture the dynamic evolution law of the contrast agent filling. Assume that the input three-dimensional feature map is \(X\in R C×T×H×W \), where \(C\) is the number of channels, \(T\) is the time dimension, and \(H\) and \(W\) are the spatial dimensions. The calculation formula of the three-dimensional convolution is:

[0073]

[0074] where \(Y\) is the output of the three-dimensional convolution, \(w\) is the weight of the three-dimensional convolutional kernel, \(M,N,L\) are the sizes of the convolutional kernel in the three dimensions respectively, and \(i,j,k\) are the indices of the output feature map. The LSTM unit can remember information for a long time and effectively capture the dynamic changes during the contrast agent filling process, such as the filling rate of the contrast agent, the attenuation of the peak intensity, etc.

[0075] The feature selection process is optimized through the Q-learning strategy to screen out the key temporal features related to myocardial viability. The Q-learning strategy designs a reward function to preferentially retain the feature parameters with a correlation higher than 0.8 with the clinical viability gold standard. The design of the reward function includes a feature stability index and a redundancy penalty term. The stability index calculates the reciprocal of the variance of the feature parameters through Bootstrap resampling, and the redundancy penalty term quantifies the correlation between features based on the mutual information theory. The reward function is expressed as:

[0076] R = γ·(1 / Var(X)) + η·(1 - MI(X,Y))

[0077] Among them, R is the final reward value, Var(X) is the variance of the feature parameter, and the smaller the variance, the more stable the feature; MI(X,Y) is the mutual information between features, and the larger the mutual information, the stronger the correlation between features; γ is the stability weight coefficient, which is used to adjust the importance of the stability index in the reward function; η is the redundancy penalty coefficient, which is used to adjust the importance of the redundancy penalty term in the reward function. During the Q-learning process, the intelligent agent selects features according to the current state, obtains rewards according to the reward function, and continuously adjusts the strategy to select the key temporal features that are most conducive to evaluating myocardial viability.

[0078] Example 4:

[0079] When constructing the multi-modal fusion evaluation model, a graph neural network is used to establish the association map between myocardial perfusion and viability. The nodes of the graph neural network represent myocardial segments, and the edge weights are calculated by the hemodynamic similarity between segments. Suppose the hemodynamic parameters of myocardial segments i and j are H i and H j , and the calculation of the edge weight w ij can adopt the following formula:

[0080]

[0081] Among them, ‖·‖ 2 represents the square of the Euclidean distance, and σ is a hyperparameter used to control the attenuation speed of the edge weight. In this way, the edge weights between myocardial segments with similar hemodynamics are larger, and vice versa.

[0082] Topological relation fusion adopts the graph attention mechanism to dynamically adjust the contribution degree of different modal features in node embedding. The graph attention mechanism weights the features of different nodes by calculating the attention coefficients between nodes. Suppose the feature of node i is f i , and the calculation of the attention coefficient α ij is as follows:

[0083]

[0084] Among them, W is the learnable weight matrix, [f i ‖f j represents concatenating the features of nodes i and j together, LeakyReLU is the ReLU activation function with leakage, and N i is the set of neighbor nodes of node i. Through the graph attention mechanism, the model can dynamically adjust its contribution in node embedding according to the features of different nodes, better fuse spatio-temporal features and topological relations, and generate a more accurate myocardial viability probability distribution map.

[0085] Example 5:

[0086] When performing quantitative evaluation based on the probability distribution map, the adaptive threshold segmentation algorithm adopts the level set method combined with the region competition model. The level set method transforms the curve evolution problem into a high-dimensional function solving problem by representing the segmentation curve as the zero level set of a high-dimensional function. Assume that the segmentation curve C can be represented as the zero level set of the level set function That is, The evolution equation of the level set function can be expressed as:

[0087]

[0088] where is the Dirac function, which is used to control the behavior of the level set function near the zero level set; div is the divergence operator; λ1 and λ2 are weight coefficients; I1 and I2 are the average gray values of the image inside and outside the segmentation curve respectively. By continuously iteratively solving this equation, the level set function can gradually converge to the target segmentation curve.

[0089] The region competition model makes the segmentation result more accurate by introducing a region competition term. The rule base of fuzzy logic reasoning is driven by expert knowledge, including the contrast agent retention time, the intensity gradient change rate, and the spatial continuity constraint. For example, when the contrast agent retention time is long and the intensity gradient change rate is small, combined with the spatial continuity constraint, it can be judged that this region is more likely to be the ischemic core area.

[0090] The Bayesian optimization method is used to iteratively adjust the evaluation parameters. The evaluation parameters include the ischemic area area weight, the survival probability threshold, and the spatial continuity penalty coefficient. The objective function is expressed as:

[0091]

[0092] where θ is the parameter combination to be optimized, including the ischemic area area weight w1, the parameter w2 related to the survival probability threshold, and the spatial continuity penalty coefficient w3; A is the ischemic area area, which reflects the range of myocardial ischemia; P is the survival probability, which is used to evaluate the viability of the myocardium; P0 is the probability normalization constant, which is used to normalize the survival probability; C is the spatial continuity measurement value, which is used to measure the spatial continuity of the segmentation result. The Bayesian optimization method constructs a surrogate model, combines the Gaussian process and the deep neural network, uses variational inference to accelerate the hyperparameter search, continuously adjusts the evaluation parameters, makes the objective function reach the optimal, and thus outputs a more accurate myocardial perfusion defect index and viability score, and generates a visual three-dimensional evaluation report, providing an intuitive and accurate diagnostic basis for doctors.

[0093] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0094] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac ultrasound contrast, characterized in that, It includes the following steps: Step 1: Obtain dynamic cine contrast echocardiography (DCE) image data, including myocardial perfusion sequence images with multiple angles and multiple time phases; further preprocess the sequence images; Step 2: Segment the myocardial region in the preprocessed temporal images. An improved U-Net network combined with an attention mechanism is used. With the contrast characteristics of the left ventricular cavity as the boundary constraint, automatically extract the endocardial and epicardial contours of the myocardium and divide the myocardial segments; at the same time, introduce a dynamic region growing algorithm to adaptively segment the regions with abnormal contrast agent filling, and construct a myocardial perfusion partition model; Step 3: Extract the dynamic myocardial perfusion features based on the segmentation results. Use a three-dimensional convolutional neural network to construct a spatio-temporal feature encoder to extract the contrast agent filling rate, peak intensity decay curve and spatial heterogeneity parameters of each myocardial segment; combined with a reinforcement learning method, optimize the feature selection process through the Q-learning strategy, and screen the key temporal features related to myocardial viability; Step 4: Construct a multimodal fusion evaluation model to perform cross-modal alignment of the dynamic features and clinical parameters; use a graph neural network to establish an association map between myocardial perfusion and viability, and fuse the spatio-temporal features and topological relationships through node embedding technology to generate a myocardial viability probability distribution map; Step 5: Perform quantitative evaluation based on the probability distribution map. Design an adaptive threshold segmentation algorithm, and dynamically divide the ischemic core area and the marginal area in combination with fuzzy logic reasoning; use the Bayesian optimization method to iteratively adjust the evaluation parameters, output the myocardial perfusion defect index and viability score, and generate a visualized three-dimensional evaluation report.

2. The method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac contrast echocardiography according to claim 1, wherein In the above Step 1, the method for preprocessing the sequence images includes: For the problems of noise interference and motion artifacts in the original images, use an image denoising and motion compensation algorithm based on a generative adversarial network (GAN), combined with the time series optical flow method to extract the local motion trajectory of the myocardium, and generate high signal-to-noise ratio myocardial contrast temporal images; The image denoising algorithm of the GAN adopts a dual-path discriminator structure, where the first discriminator focuses on spatial resolution recovery, and the second discriminator optimizes the temporal continuity based on the cycle consistency loss function; the motion compensation algorithm combines optical flow field estimation and deformation field interpolation to achieve non-rigid registration of local myocardial motion.

3. The myocardial perfusion and viability assessment method based on intelligent analysis of cardiac contrast echocardiography according to claim 1, wherein In the above Step 2, the improved U-Net network introduces a multi-scale dilated convolution module in the encoder part to capture multi-resolution features of the myocardial boundary; The attention mechanism adopts a channel-spatial dual attention module to dynamically weight the contrast intensity information of different regions.

4. The method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac contrast echocardiography according to claim 1, wherein In the above Step 3, the three-dimensional convolutional kernel of the spatio-temporal feature encoder adopts an asymmetric structure, and a long short-term memory unit is used in the time dimension to capture the dynamic evolution law of contrast agent filling; the Q-learning strategy designs a reward function to preferentially retain the feature parameters with a correlation higher than 0.8 with the clinical viability gold standard.

5. The method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac ultrasound contrast as claimed in claim 1, wherein, In the above Step 4, the nodes of the graph neural network represent myocardial segments, and the edge weights are calculated by the hemodynamic similarity between segments; the topological relationship fusion adopts a graph attention mechanism to dynamically adjust the contribution of different modal features in node embedding.

6. The myocardial perfusion and viability assessment method based on intelligent analysis of cardiac contrast echocardiography according to claim 1, wherein In step 5, the adaptive threshold segmentation algorithm adopts the level set method combined with the region competition model. The rule base of fuzzy logic inference is driven by expert knowledge, including the contrast agent retention time, the intensity gradient change rate, and the spatial continuity constraint.

7. The myocardial perfusion and viability assessment method based on intelligent analysis of cardiac contrast echocardiography according to claim 2, characterized in that, The training of the dual-path discriminator adopts a progressive learning strategy. In the first stage, only the spatial resolution is optimized. In the second stage, the temporal consistency loss is jointly optimized. In the final stage, adversarial perturbations are introduced to enhance the robustness of the model.

8. The method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac ultrasound contrast according to claim 3, wherein The seed point selection of the dynamic region growing algorithm is based on the contrast agent first-pass time map. The growth criterion combines local intensity similarity and the contrast agent diffusion kinetics model. The smoothness of the segmentation boundary is optimized through the Markov random field. The region growth criterion is expressed as: Among them, G(p) is the growth probability of pixel point p, I(p) is the intensity value of pixel point p, I(s) is the mean intensity of seed point s, σ is the standard deviation of intensity distribution, D(p) is the contrast agent diffusion coefficient, is the time gradient vector, and α and β are the spatial and temporal weight coefficients respectively.

9. The method for evaluating myocardial perfusion and viability based on intelligent analysis of cardiac contrast echocardiography according to claim 4, characterized in that The design of the reward function includes a feature stability index and a redundancy penalty term. The stability index calculates the reciprocal of the variance of the feature parameters through Bootstrap resampling. The redundancy penalty term quantifies the correlation between features based on the mutual information theory. The reward function is expressed as: R = γ·(1 / Var(X)) + η·(1 - MI(X,Y)) Where, R is the final reward value, Var(X) is the variance of the feature parameters, MI(X,Y) is the mutual information between features, γ is the stability weight coefficient, and η is the redundancy penalty coefficient.

10. The myocardial perfusion and viability assessment method based on intelligent analysis of cardiac contrast echocardiography according to claim 6, characterized in that, The surrogate model of the Bayesian optimization method adopts a deep kernel learning framework, which combines Gaussian processes and deep neural networks, and accelerates the hyperparameter search through variational inference. The evaluation parameters include the ischemic area area weight, the survival probability threshold, and the spatial continuity penalty coefficient. The objective function is expressed as: Where, θ is the parameter combination to be optimized, A is the ischemic area area, P is the survival probability, P0 is the probability normalization constant, C is the spatial continuity metric value, and w1, w2, w3 are the weight coefficients.

Images