Ultrasonic cardiogram myocardial motion estimation method and system based on structure perception diffusion model

Abstract

The invention discloses an echocardiogram myocardial motion estimation method and system based on a structure perception diffusion model, and relates to the technical field of ultrasonic video data processing. The invention aims to solve the problem of myocardial motion estimation in an ultrasonic video. According to the technical key points, the method comprises the following steps: (1) preprocessing collected echocardiogram video data; (2) establishing an echocardiogram myocardial motion estimation model; (3) constructing a diffusion model motion estimation module of structure perception; (4) constructing a spatial adaptive up-sampling module; (5) training the network by using the existing data to obtain a network model; and (6) estimating the myocardial movement of the echocardiogram video by using the trained model. According to the method, the cardiac structure information and the multi-scale correlation contained in the echocardiogram can be fully utilized to automatically estimate the cardiac muscle movement of the echocardiogram, and the accuracy and the robustness of the algorithm are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of ultrasound video data processing technology, specifically to a method and system for estimating myocardial motion in echocardiography based on a structure-aware diffusion model. Background Technology

[0002] Ultrasound imaging[1] is a non-invasive medical imaging technique that uses the physical properties of ultrasound waves and the reflection characteristics of ultrasound waves by different tissues to obtain information on the internal structure and function of human tissues. Compared with techniques such as computed tomography and magnetic resonance imaging, ultrasound imaging is fast and has high temporal resolution. It can observe the structure, function and motion state of organs and tissues in real time and is widely used in the diagnosis of heart disease, with broad application prospects. However, traditional methods of myocardial motion assessment often rely on subjective visual interpretation, resulting in large differences between observers. These limitations highlight the urgent need for automatic motion estimation in order to more accurately and objectively assess myocardial dynamics[2].

[0003] To this end, researchers have proposed many algorithms for automatic estimation of myocardial motion in echocardiography. Traditional intensity matching-based algorithms usually have high computational complexity, are sensitive to speckle noise, and have low estimation accuracy [3]. Existing deep learning-based algorithms are mostly based on the PWC-Net framework to process echocardiographic data [4], which do not utilize rich scale information and perform poorly in myocardial motion estimation tasks. In recent years, some researchers have begun to focus on how to utilize the multi-scale correlation of the heart to improve the accuracy of myocardial motion estimation in echocardiography.

[0004] In addition, most methods use point tracking or U-Net [5], rarely considering the inherent characteristics of echocardiographic data. Therefore, it is necessary to design a motion estimation method that utilizes the multi-scale correlation of echocardiography to improve the tracking accuracy and robustness of myocardial tissue. Summary of the Invention

[0005] The technical problem to be solved by this invention is:

[0006] The purpose of this invention is to provide a method and system for estimating myocardial motion in echocardiography based on a structure-aware diffusion model, in order to solve the problem of automatic estimation of myocardial motion in ultrasound videos.

[0007] The technical solution adopted by the present invention to solve the above problems is as follows:

[0008] (1) Preprocessing of the acquired echocardiogram video data;

[0009] (2) Establish an echocardiographic model for estimating myocardial motion;

[0010] (3) Construct a structure-aware diffusion model motion estimation module;

[0011] (4) Construct a spatial adaptive upsampling module;

[0012] (5) Train the network using the existing data to obtain the network model;

[0013] (6) Use the trained model to estimate myocardial motion in echocardiogram videos.

[0014] Furthermore, in step (1) above, echocardiogram data is acquired using an ultrasound imaging device, and the data is preprocessed, including unifying the data size, randomly augmenting the data, normalizing and standardizing the data, and sampling the data.

[0015] Specifically, the Cubic interpolation method was used to unify the echocardiogram frames and labels to 224×224. Random horizontal flipping, scaling, rotation, contrast enhancement, and gamma enhancement were applied to the data to avoid overfitting. The data was normalized to [0, 1] using a maximum-minimum method, as shown in the following equation.

[0016]

[0017] Where x is the original data, x max and x min These are the maximum and minimum values ​​of the data, respectively.

[0018] z-score standardization transforms the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula.

[0019]

[0020] Where μ is the mean of the data and σ is the standard deviation of the data.

[0021] Each echocardiogram video sequence is divided into frame pairs between two adjacent frames for model training.

[0022] Furthermore, the motion estimation model construction in step (2) above can be carried out according to the following steps:

[0023] Given two echocardiographic video frames of size H×W, where H and W represent the spatial resolution of each frame, a shared feature encoder is first used to extract multi-scale basic features. and Where l∈{1, 2, ... L} represents the scale level. Furthermore, a content encoder is applied to I1 to extract contextual features C. l In order to model the pixel-level correspondence between frames, calculations are performed. and The dot product between them constructs a four-dimensional cost function. And pixel-level motion affinity is represented as:

[0024]

[0025] Where H l and W l Represents the resolution of the feature. Indices (g, h) and (i, j) respectively represent... and The spatial location within the vector. k corresponds to the feature channel dimension.

[0026] To integrate myocardial segmentation structure priors into the motion estimation process, a lightweight segmentation head is used to extract contextual features C. l Generate a rough myocardial mask. As a priori information about cardiac structure, it guides the adaptive enhancement of myocardial features, improving feature representation within myocardial regions and enabling the model to focus on anatomically relevant areas. Enhanced features and cost function This was then used as a conditional input to the diffusion-based motion estimation module (CMD). This module employs a diffusion-based iterative denoising framework to estimate a coarse myocardial motion field through iterative refinement. Assign a displacement vector <Δu, Δv> to each myocardial pixel.

[0027] To improve the accuracy of motion estimation, a Spatial Adaptive Upsampling Module (SGMU) was also constructed. This module utilizes a content-aware adaptive strategy to optimize the output motion stream of the motion estimation module. Upsampling for high-resolution motion flow It makes full use of cross-scale feature similarity.

[0028] Furthermore, the structure-aware diffusion model motion estimation module constructed in step (3) above mainly utilizes cardiac structure priors to enhance the motion estimation process based on the diffusion model. First, in the context feature C l Using a lightweight mask head, generate a probability map of myocardial segmentation. This probability map serves as a structural prior, guiding the network to focus on the myocardial region while suppressing irrelevant background information. To embed this prior into the contextual feature space, spatial adaptive modulation is applied to obtain... The formula is:

[0029]

[0030] in This is instance normalization. The modulation parameters α and β are applied by two independent convolutional layers. The generated spatial variation map enables the network to adaptively enhance features related to myocardial structure. These features are then modulated using ReLU activation and convolutional layers. Refine it, and then combine it with the input feature C l Perform residual connections to produce enhanced output

[0031] Subsequently, the myocardial motion flow is estimated through a conditional diffusion-based denoising process. Specifically, from the initial noise flow input... Initially, a series of noise reduction steps were used to iteratively optimize the process, ultimately resulting in a clean sports field. Each noise reduction step will transfer the flow from Convert to Its guiding conditional input includes enhanced contextual features. and cost function This process can be described as follows:

[0032]

[0033] Where θ is a learnable parameter for the denoising step.

[0034] Following the RAFT decoding process, the current stream is used at the beginning of each iteration. Perform a relevant pyramid lookup, and then use a motion encoder. Motion feature extraction is performed, represented as To enable the model to precisely adjust the feature distribution at each diffusion time step, a temporal embedding control block is introduced, with the following formula:

[0035]

[0036] Where τ c and τ h Represents the scale and translation vector obtained through the time embedding function. This is the key to aligning the diffusion iteration with its corresponding time step.

[0037] Operator This represents the attention-motion aggregation module, used to fuse contextual features. and motion characteristics M l This generates aggregated motion-sensing features. Module The module consists of 3×3 convolutions, GELU activation, and group normalization. It consists of 3×3 convolutions, GELU activation, and group normalization. φ is a learnable parameter. The final features... It integrates time-aware modulation and spatial motion aggregation, and is then passed to the GRU and the flow head module for final flow prediction.

[0038] During training, a forward diffusion process is employed, based on the real flow. Generate noise stream This process follows a Markov chain, gradually adding Gaussian noise to the clean stream, defined as:

[0039]

[0040] in, It is the noise flow at the diffusion step t on scale l. These are control parameters, calculated using the following formula:

[0041]

[0042] β s It is the noise variance at step s.

[0043] To ensure numerical stability, the noise flow Normalize and rescale to the range [-b, b]. During inference, apply a reverse diffusion process stepwise to... Denoising is performed, and the clean flow is estimated. This generation process simulates a non-Markovian forward process with parameter σ, as shown in the following formula:

[0044]

[0045] in Represents standard Gaussian noise, function For use based on Approximate prediction

[0046] Approximate noise at time step t The calculation is as follows:

[0047]

[0048] Furthermore, traditional upsampling techniques, such as bilinear interpolation or convex upsampling, often lose fine-grained spatial details due to their content-independent nature. To overcome this limitation, a Spatial Adaptive Upsampling Module (SGMU) is used. This module performs spatial adaptive upsampling of the motion field by incorporating cross-scale feature similarity. By leveraging high-resolution features to guide the refinement of coarse optical flow predictions, SGMU can effectively recover motion details and improve the accuracy of motion estimation.

[0049] In step (4), a spatial adaptive upsampling module is constructed, given a coarse motion field estimated at level l. and the corresponding contextual features {C} from two consecutive scales lC l-1 First, establish the projection between the low-resolution position (h, w) at scale l-1 and its corresponding high-resolution coordinates (h′, w′) = (h·s, w·s) at scale l, where s is the resolution scaling factor.

[0050] To guide motion upsampling with content awareness, a similarity map is calculated using the dot product between cross-scale features, and then refined using a learnable transformation function Φ(·), as shown in the following formula:

[0051]

[0052] Where S l-1 This indicates guidance on cross-scale similarity. It is a two-dimensional neighborhood of size Λ×Λ centered at (h, w), located on the low-resolution optical flow. The function Φ(·) is implemented as a lightweight subnetwork consisting of convolutional layers, ReLU, and batch normalization, allowing for spatial adaptive adjustment of the similarity map.

[0053] Similarity guidance S l-1 Used for modulating low-resolution motion prediction This generates upsampling optical flow. The formula is as follows:

[0054]

[0055] To further improve motion fidelity, a reverse attention mechanism for residual correction is proposed to refine the upsampled optical flow. Specifically, the upsampled optical flow is first used... basic features Towards Directional deformation, to obtain Then, calculate and The cosine similarity between them is calculated, and the inverse operation is performed to generate a back attention map A. r This is used to highlight areas with high motion uncertainty or mismatch.

[0056] The reverse attention diagram A r Contextual features and motion characteristics M l-1 It is connected to the FlowCorrector module, which consists of multiple residual layers for estimating residual optical flow correction. Then, by adding this residual to the upsampled optical flow, the output corrected motion field is obtained.

[0057]

[0058] This attention-guided improvement enables the network to dynamically focus on difficult regions, thereby enhancing the fidelity of the upsampled motion field. By combining context-aware similarity modulation with inverse attention correction, this method achieves robust and accurate myocardial motion estimation in echocardiographic imaging.

[0059] The constructed network can be trained using a preprocessed dataset to obtain optimal network parameters for myocardial motion estimation of subsequent echocardiographic data; to ensure the segmentation performance of the network, the dataset used contains 6,165 sets of echocardiographic data.

[0060] Furthermore, in step (6) above, the collected data is input into the trained network to obtain the corresponding myocardial motion estimation flow field. This process can be represented as:

[0061] y = f net (x) (17)

[0062] Among them, f net The constructed echocardiographic myocardial motion estimation network is represented by x and y, which represent the input echocardiographic data and the output myocardial motion estimation results, respectively.

[0063] The present invention has the following beneficial technical effects:

[0064] This invention provides a method for estimating myocardial motion in echocardiography based on a structure-aware diffusion model, effectively solving the problem of myocardial motion estimation in ultrasound videos. The invention includes the following steps: preprocessing the acquired echocardiographic video data; establishing an echocardiographic myocardial motion estimation model; constructing a structure-aware diffusion model motion estimation module; constructing a spatial adaptive upsampling module; training the network using existing data to obtain the network model; and using the trained model to estimate myocardial motion in the echocardiographic video. This invention can fully utilize the cardiac structural information and multi-scale correlations contained in echocardiograms to automatically estimate myocardial motion, effectively improving the accuracy and robustness of the algorithm and reducing the problem of incorrect motion vector estimation.

[0065] The present invention provides a method for estimating myocardial motion in echocardiography based on a structure-aware diffusion model, which can automatically estimate the myocardial motion field from echocardiographic video data. The algorithm proposed in this invention has high computational efficiency and can be easily accelerated using GPUs, making it a highly efficient algorithm. Experiments show that, compared with existing methods based on PWC-Net, this invention can significantly improve the accuracy of myocardial estimation, achieving the goal of automatic myocardial estimation from echocardiographic videos. Attached Figure Description

[0066] Figure 1 This is a flowchart of the method of the present invention.

[0067] Figure 2 This is a schematic diagram of the overall structure of the network proposed in this invention (a schematic diagram of the overall structure of the echocardiographic myocardial motion estimation method based on the structure-aware diffusion model proposed in this invention);

[0068] Figure 3 This is a schematic diagram of the structure perception module;

[0069] Figure 4 This is a schematic diagram of the spatial adaptive upsampling module structure;

[0070] Figure 5 This is a comparison chart of motion estimation results between the method of this invention and other currently advanced methods. Detailed Implementation

[0071] The following is in conjunction with the appendix Figure 1-5 The invention will be described in detail with examples.

[0072] See the flowchart of the method of this invention. Figure 1 The specific implementation steps are as follows:

[0073] (1) Preprocessing of the acquired echocardiographic data;

[0074] (2) Establish an echocardiographic model for estimating myocardial motion;

[0075] (3) Construct a structure-aware diffusion model motion estimation module;

[0076] (4) Construct a spatial adaptive upsampling module;

[0077] (5) Train the network using the existing data to obtain the network model;

[0078] (6) Use the trained model to estimate myocardial motion in echocardiogram videos.

[0079] The above step (1) is performed as follows:

[0080] Echocardiographic data were acquired using ultrasound imaging equipment. Preprocessing of the data included standardizing data size, random augmentation, normalization, and sampling. Specifically, Cubic interpolation was used to standardize the image size and labels of each echocardiographic frame to 224×224. Random horizontal flipping, scaling, rotation, contrast enhancement, and gamma enhancement were applied to the data to avoid overfitting. The data was normalized to [0, 1] using a maximum-minimum normalization method, as shown in the following equation.

[0081]

[0082] Where x is the original data, xmax and x min These are the maximum and minimum values ​​of the data, respectively.

[0083] z-score standardization transforms the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula.

[0084]

[0085] Where μ is the mean of the data and σ is the standard deviation of the data.

[0086] Each echocardiogram video sequence is divided into frame pairs between two adjacent frames for model training.

[0087] Step (2) above shall be performed as follows:

[0088] Given two echocardiographic video frames of size H×W, where H and W represent the spatial resolution of each frame, a shared feature encoder is first used to extract multi-scale basic features. and Where l∈{1, 2, ... L} represents the scale level. Furthermore, a content encoder is applied to I1 to extract contextual features C. l In order to model the pixel-level correspondence between frames, calculations are performed. and The dot product between them constructs a four-dimensional cost function. And pixel-level motion affinity is represented as:

[0089]

[0090] Where H l and W l Represents the resolution of the feature. Indices (g, h) and (i, j) respectively represent... and The spatial location within the vector. k corresponds to the feature channel dimension.

[0091] To integrate myocardial segmentation structure priors into the motion estimation process, a lightweight segmentation head is used to extract contextual features C. l Generate a rough myocardial mask. As a priori information about cardiac structure, it guides the adaptive enhancement of myocardial features, improving feature representation within myocardial regions and enabling the model to focus on anatomically relevant areas. Enhanced features and cost function This was then used as a conditional input to the diffusion-based motion estimation module (CMD). This module employs a diffusion-based iterative denoising framework to estimate a coarse myocardial motion field through iterative refinement. Assign a displacement vector <Δu, Δv> to each myocardial pixel.

[0092] To improve the accuracy of motion estimation, a Spatial Adaptive Upsampling Module (SGMU) was also constructed. This module utilizes a content-aware adaptive strategy to optimize the output motion stream of the motion estimation module. Upsampling for high-resolution motion flow It makes full use of cross-scale feature similarity.

[0093] The above step (3) shall be performed as follows:

[0094] In the CMD module, firstly, in the context feature C l Using a lightweight mask head, generate a probability map of myocardial segmentation. This probability map serves as a structural prior, guiding the network to focus on the myocardial region while suppressing irrelevant background information. To embed this prior into the contextual feature space, spatial adaptive modulation is applied to obtain... The formula is:

[0095]

[0096] in This is instance normalization. The modulation parameters α and β are applied by two independent convolutional layers. The generated spatial variation map enables the network to adaptively enhance features related to myocardial structure. These features are then modulated using ReLU activation and convolutional layers. Refine it, and then combine it with the input feature C l Perform residual connections to produce enhanced output

[0097] Subsequently, the myocardial motion flow is estimated through a conditional diffusion-based denoising process. Specifically, from the initial noise flow input... Initially, a series of noise reduction steps were used to iteratively optimize the process, ultimately resulting in a clean sports field. Each noise reduction step will transfer the flow from Convert to Its guiding conditional input includes enhanced contextual features. and cost function This process can be described as follows:

[0098]

[0099] Where θ is a learnable parameter for the denoising step.

[0100] Following the RAFT decoding process, the current stream is used at the beginning of each iteration. Perform a relevant pyramid lookup, and then use a motion encoder. Motion feature extraction is performed, represented as To enable the model to precisely adjust the feature distribution at each diffusion time step, a temporal embedding control block is introduced, with the following formula:

[0101]

[0102] Where τ c and τ h Represents the scale and translation vector obtained through the time embedding function. This is crucial for aligning the diffusion iteration with its corresponding time step. Operator This represents the attention-motion aggregation module, used to fuse contextual features. and motion characteristics M l This generates aggregated motion-sensing features. Module It consists of 3×3 convolutions, GELU activation, and group normalization. Module It consists of 3×3 convolutions, GELU activation, and group normalization. φ is a learnable parameter. The final features... It integrates time-aware modulation and spatial motion aggregation, and is then passed to the GRU and the flow head module for final flow prediction.

[0103] During training, a forward diffusion process is employed, based on the real flow. Generate noise stream This process follows a Markov chain, gradually adding Gaussian noise to the clean stream, defined as:

[0104]

[0105] in, It is the noise flow at the diffusion step t on scale l. These are control parameters, calculated using the following formula:

[0106]

[0107] β s It is the noise variance at step s.

[0108] To ensure numerical stability, the noise flow Normalize and rescale to the range [-b, b]. During inference, apply a reverse diffusion process stepwise to... Denoising is performed, and the clean flow is estimated. This generation process simulates a non-Markovian forward process with parameter σ, as shown in the following formula:

[0109]

[0110] in Represents standard Gaussian noise, function For use based on Approximate prediction

[0111] Approximate noise at time step t The calculation is as follows:

[0112]

[0113] The above step (4) shall be performed as follows:

[0114] Traditional upsampling techniques, such as bilinear interpolation or convex upsampling, often lose fine-grained spatial details due to their content-independent nature. To overcome this limitation, a Spatial Adaptive Upsampling (SGMU) module is used. This module performs spatial adaptive upsampling of the motion field by incorporating cross-scale feature similarity. By leveraging high-resolution features to guide the refinement of coarse optical flow predictions, SGMU can effectively recover motion details and improve the accuracy of motion estimation.

[0115] Given a rough motion field estimated at level l and the corresponding contextual features {C} from two consecutive scales l C l-1 First, establish the projection between the low-resolution position (h, w) at scale l-1 and its corresponding high-resolution coordinates (h′, w′) = (h·s, w·s) at scale l, where s is the resolution scaling factor.

[0116] To guide motion upsampling with content awareness, a similarity map is calculated using the dot product between cross-scale features, and then refined using a learnable transformation function Φ(·), as shown in the following formula:

[0117]

[0118] Where S l-1 This indicates guidance on cross-scale similarity. It is a two-dimensional neighborhood of size Λ×Λ centered at (h,w) on the low-resolution optical flow. The function Φ(·) is implemented as a lightweight subnetwork consisting of convolutional layers, ReLU, and batch normalization, allowing for spatial adaptive adjustment of the similarity map.

[0119] Similarity guidance S l-1 Used for modulating low-resolution motion prediction This generates upsampling optical flow. The formula is as follows:

[0120]

[0121] To further improve motion fidelity, a reverse attention mechanism for residual correction is proposed to refine the upsampled optical flow. Specifically, the upsampled optical flow is first used... basic features Towards Directional deformation, to obtain Then, calculate and The cosine similarity between them is calculated, and the inverse operation is performed to generate a back attention map A. r This is used to highlight areas with high motion uncertainty or mismatch.

[0122] The reverse attention diagram A r Contextual features and motion characteristics M l-1 It is connected to the FlowCorrector module, which consists of multiple residual layers for estimating residual optical flow correction. This residual is then added to the upsampled optical flow to obtain the output corrected motion field.

[0123]

[0124] This attention-guided improvement enables the network to dynamically focus on difficult regions, thereby enhancing the fidelity of the upsampled motion field. By combining context-aware similarity modulation with inverse attention correction, this method achieves robust and accurate myocardial motion estimation in echocardiographic imaging.

[0125] Step (5) above shall be performed as follows:

[0126] The constructed network can be trained using a preprocessed dataset to obtain optimal network parameters for myocardial motion estimation of subsequent echocardiographic data; to ensure the segmentation performance of the network, the dataset used contains 6,165 sets of echocardiographic data.

[0127] The above step (6) shall be performed as follows:

[0128] By inputting the collected data into the trained network, the corresponding estimated flow field of myocardial motion can be obtained. This process can be represented as:

[0129] y = f net (x) (17)

[0130] Among them, f net The constructed echocardiographic myocardial motion estimation network is represented by x and y, which represent the input echocardiographic data and the output myocardial motion estimation results, respectively.

[0131] To quantitatively evaluate the performance of the method proposed in this invention, the mean endpoint error (EPE) was used as a quantitative evaluation index to assess the effectiveness of echocardiographic video myocardial motion estimation. EPE is defined as follows:

[0132]

[0133] Where N is the total number of pixels in the image, f i Let f represent the predicted optical flow vector of the i-th pixel. i gt Let |||||2 represent the true optical flow vector of the i-th pixel. ||·||2 represents the L2 norm (i.e., Euclidean distance).

[0134] Figure 5 This diagram presents a comparison of motion estimation results between the method of this invention and other state-of-the-art methods. From left to right, it shows echocardiographic video frames, myocardial motion field labels, and myocardial motion estimation results from RAFT, FlowFormer, FlowDiffuser, EchoPWC, and the method proposed in this invention. Figure 5 As can be seen, the method proposed in this invention for echocardiographic motion estimation yields the most accurate myocardial tissue motion estimation results, closest to the given labels. Table 1 shows the quantitative comparison results of different methods for myocardial motion estimation on the test dataset.

[0135]

[0136] Table 1 lists the quantitative comparison results of the motion estimation method of this invention with other state-of-the-art methods, using the EPE index. The quantitative comparison results show that, compared with the existing state-of-the-art methods, this invention can significantly improve the accuracy of echocardiographic motion estimation.

[0137] The following is a list of references cited in this invention:

[0138] [1]Antico, Maria, et al. "Ultrasound guidance in minimally invasiverobotic procedures." Medical image analysis 54(2019):149-167.

[0139] [2]Leclerc,Sarah,et al."Deeplearning for segmentation using anopenlarge-scale dataset in 2D echocardiography."IEEE transactions on medicalimaging 38.9(2019):2198-2210.

[0140] [3]Heyde B,Barbosa D,Claus P,et al.Three-dimensional cardiac motionestimation based on non-rigid image registration using a novel transformationmodel adapted to the heart[C] / / International Workshop on Statistical Atlasesand Computational Models of the Heart.Berlin,Heidelberg:Springer BerlinHeidelberg,2012:142-150.

[0141] [4] A,Salte I M,Smistad E,et al.Myocardial function imaging inechocardiography using deep learning[J].ieee transactions on medical imaging,2021,40(5):1340-1351.

[0142] [5]Ta K,Ahn S S,Thorn S L,et al.Multi-task learning for motionanalysis and segmentation in 3D echocardiography[J].IEEE Transactions onMedical Imaging,2024,43(5):2010-2020.

Claims

1. A method for estimating myocardial motion in echocardiography based on a structure-aware diffusion model, characterized in that, The method, which utilizes prior knowledge of myocardial structure from echocardiography and adaptive upsampling based on a diffusion model to improve the accuracy and robustness of myocardial motion estimation, includes the following steps: (1) Preprocessing of the acquired echocardiogram video data; (2) Establish an echocardiographic model for estimating myocardial motion; (3) Construct a structure-aware diffusion model motion estimation module; (4) Construct a spatial adaptive upsampling module; (5) Train the network using the existing data to obtain the network model; (6) Use the trained model to estimate myocardial motion in echocardiogram videos.

2. The method as described in claim 1, characterized in that, In step (1), echocardiographic data is acquired using ultrasound imaging equipment. The data is preprocessed, including standardizing data size, random augmentation, normalization, and sampling. The specific process is as follows: The Cubic interpolation method is used to unify the echocardiogram frames and optical flow field labels to 224×224. Random horizontal flipping, scaling, rotation, contrast adjustment, and gamma enhancement are applied to the data to avoid overfitting. The data is normalized to [0, 1] using the maximum-minimum ratio, as shown in the following formula. Where x is the original data, x max and x min These are the maximum and minimum values ​​of the data, respectively. z-score normalization converts the image to a standard normal distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula. Where μ is the mean of the data and σ is the standard deviation of the data; In each sequence, a total of 10 uniformly distributed frames were sampled, starting with a relaxation frame and ending with a contraction frame.

3. The method as described in claim 1 or 2, characterized in that, Step (2) is performed as follows: Given two echocardiographic video frames of size H×W, where H and W represent the spatial resolution of each frame, a shared feature encoder is first used to extract multi-scale basic features. and Where l∈{1,2,…L} represents the scale level; Extracting contextual features C from the I1 application content encoder l Model the pixel-level correspondence between frames and calculate and The dot product between them is used to construct a four-dimensional cost function. And pixel-level motion affinity is represented as: Among them, H l and W l The resolution of the feature is represented by indices (g,h) and (i,j), respectively. and In the spatial location, k corresponds to the feature channel dimension; The myocardial segmentation structure prior is integrated into the motion estimation process, utilizing a lightweight segmentation head and inputting contextual features C. l Generate a rough myocardial mask. As a priori information about cardiac structure, it guides the adaptive enhancement of myocardial features, improving feature representation within myocardial regions and enabling the model to focus on anatomically relevant areas, thereby enhancing the features. and cost function Subsequently, this serves as a conditional input to the diffusion-based motion estimation module CMD, which employs a diffusion-based iterative denoising framework to estimate a coarse myocardial motion field through iterative refinement. Assign a displacement vector <△u,△v> to each myocardial pixel; A spatial adaptive upsampling module (SGMU) is constructed to improve the accuracy of motion estimation. Based on a content-aware adaptive strategy, the output motion stream of the motion estimation module is... Upsampling for high-resolution motion flow Make full use of cross-scale feature similarity.

4. The method as described in claim 3, characterized in that, Step (3) is performed as follows: In the CMD module, firstly, in the context feature C l Using a lightweight mask head, generate a probability map of myocardial segmentation. This probabilistic map serves as a structural prior, guiding the network to focus on the myocardial region while suppressing irrelevant background information. To embed this prior into the contextual feature space, spatial adaptive modulation is applied to obtain... The formula is: in It is instance normalization, and the modulation parameters α and β are applied by two independent convolutional layers. The generated spatial variation map enables the network to adaptively enhance features related to myocardial structure; Then, the modulation features are processed through ReLU activation and convolutional layers. Refine it, and then combine it with the input feature C l Perform residual connections to produce enhanced output Subsequently, the myocardial motion flow is estimated through a conditional diffusion-based denoising process, specifically: from the initial noise flow input... Initially, a series of noise reduction steps were used to iteratively optimize the process, ultimately resulting in a clean sports field. Each noise reduction step will redirect the flow from F t l Convert to Its guiding conditional input includes enhanced contextual features. and cost function The process is described as follows: Where θ is a learnable parameter for the denoising step; Following the RAFT decoding process, the current stream F is used at the beginning of each iteration. t l Perform a relevant pyramid lookup, and then use a motion encoder. Motion feature extraction is performed, represented as To enable the model to precisely adjust the feature distribution at each diffusion time step, a temporal embedding control block is introduced, with the following formula: Where τ c and τ h Represents the scale and translation vector obtained through the time embedding function. It is crucial to align the diffusion iteration with its corresponding time step; operator This represents the attention-motion aggregation module, used to fuse contextual features. and motion characteristics M l This generates aggregated motion-sensing features. Module It consists of 3×3 convolution, GELU activation, and group normalization; module It consists of 3×3 convolutions, GELU activation, and group normalization; φ is a learnable parameter; the final feature It integrates time-aware modulation and spatial motion aggregation, and is then passed to the GRU and the flow head module for final flow prediction; During training, a forward diffusion process is employed, based on the real flow. Generate noise stream This process follows a Markov chain, gradually adding Gaussian noise to the clean stream, defined as: Among them, F t l It is the noise flow at the diffusion step t on scale l. These are control parameters, calculated as follows: β s It is the noise variance at step s; To ensure numerical stability, the noise flow F t l Normalize and rescale to the range [-b, b]. During inference, apply the inverse diffusion process stepwise to F. t l Denoising is performed, and the clean flow is estimated. This generation process simulates a non-Markovian forward process with parameter σ, as shown in the following formula: in Represents standard Gaussian noise, function Used based on F l t Approximate prediction Approximate noise at time step t The calculation is as follows:

5. The method as described in claim 4, characterized in that, Step (4) is performed as follows: The Spatial Adaptive Upsampling Module (SGMU) is used to overcome the loss of fine-grained spatial details. This module performs spatial adaptive upsampling of the motion field by combining cross-scale feature similarity. By utilizing high-resolution features to guide the refinement of coarse optical flow prediction, SGMU is used to recover motion details and improve the accuracy of motion estimation. Given a rough motion field estimated at level l and the corresponding contextual features {C} from two consecutive scales l C l-1 First, establish the projection between the low-resolution position (h,w) at scale l-1 and its corresponding high-resolution coordinates (h′,w′) = (h·s,w·s) at scale l, where s is the resolution scaling factor; Motion upsampling is guided by content awareness, and a similarity map is calculated using the dot product between cross-scale features. This map is then refined using a learnable transformation function Φ(·), as shown in the following formula: Where S l-1 Indicates cross-scale similarity guidance; It is a two-dimensional neighborhood of size Λ×Λ centered at (h,w) on the low-resolution optical flow; the function Φ(·) is implemented as a lightweight subnetwork consisting of convolutional layers, ReLU and batch normalization, which allows for spatial adaptive adjustment of the similarity map; Similarity guidance S l-1 Used for modulating low-resolution motion prediction This generates upsampling optical flow. The formula is as follows: A reverse attention mechanism for residual correction is proposed to refine the upsampled optical flow to improve motion fidelity. Specifically, the upsampled optical flow is first used... basic features Towards Directional deformation, to obtain Then, calculate and The cosine similarity between them is calculated, and the inverse operation is performed to generate a back attention map A. r This is used to highlight areas with high motion uncertainty or mismatch; The reverse attention diagram A r Contextual features and motion characteristics M l-1 It is connected to the FlowCorrector module, which consists of multiple residual layers for estimating residual optical flow correction; then, by adding this residual to the upsampled optical flow, the output corrected motion field is obtained. Improvements in attention guidance enable the network to dynamically focus on difficult regions, enhancing the fidelity of the upsampled motion field; and robust and accurate myocardial motion estimation is achieved in echocardiographic imaging by combining context-aware similarity modulation with inverse attention correction.

6. The method as described in claim 5, characterized in that, Step (5) is performed as follows: The constructed network was trained using a preprocessed dataset to obtain optimal network parameters for subsequent myocardial motion estimation from echocardiographic data. To ensure the network's segmentation performance, the dataset used contained 6,165 sets of echocardiographic data.

7. The method as described in claims 2 and 6, characterized in that, Step (6) is performed as follows: By inputting the collected data into the trained network, the corresponding estimated flow field of myocardial motion can be obtained. This process can be represented as follows: y=f net (x) (17) Among them, f net The constructed echocardiographic myocardial motion estimation network is represented by x and y, which represent the input echocardiographic data and the output myocardial motion estimation results, respectively.

8. A myocardial motion estimation system based on a structure-aware diffusion model in echocardiography, characterized in that: The system has a program module corresponding to the steps of any one of claims 1-7 above, and executes the steps in the above-described method for echocardiographic myocardial motion estimation based on the structure-aware diffusion model when it is run.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the echocardiographic myocardial motion estimation method based on a structure-aware diffusion model according to any one of claims 1-7.

Images