Heart motion estimation method and device based on sparse key point driving and medium

Abstract

The invention relates to a heart motion estimation method and device based on sparse key point driving and a medium, and the method comprises the steps: respectively inputting a moving image and a fixed image into a key point detection network with time-space consistency, and carrying out the extraction to obtain a key point pair; wherein in the key point detection process, a local correlation graph is generated according to normalized cross-correlation of key points in a fixed image and key points in a moving image, and after the local correlation graph is converted into a probability distribution graph, the positions of the key points in the moving image are corrected; and constructing a sparse basic motion vector according to the relative motion between each pair of key points to obtain a sparse motion vector expressed by heart motion, and performing dense motion reconstruction on the moving image by using a deep neural network according to the sparse motion vector to generate a dense motion field. Compared with the prior art, the method has the advantages of being high in structure retentivity, high in individual generalization and high in heart motion detail modeling capacity.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, device and medium for estimating cardiac motion based on sparse keypoints. Background Technology

[0002] Cardiac motion analysis, as an important assessment tool, can reveal the mechanical properties of the heart by quantifying its elasticity and contractile function, thereby helping to deepen the understanding of the impact of cardiovascular diseases on myocardial motion function. Cardiac motion estimation and analysis can reveal the mechanical load and abnormal motion characteristics of the diseased area, thus improving the accuracy of identifying and assessing the degree of cardiac function impairment.

[0003] In 4D cardiac images, such as cine MRI, tagged MRI, and four-dimensional computed tomography (4D-CT), cardiac motion is typically estimated by extracting deformations between consecutive 3D image frames, usually through registration methods. Registration methods are mainly divided into two categories: pixel-value-based registration methods and image feature-based registration methods.

[0004] Pixel-based registration methods optimize the deformation field by minimizing the intensity similarity cost function between two images. However, this type of optimization is often an ill-posed problem, requiring regularization terms to constrain the deformation field, avoid generating unreasonable displacement fields, and reduce the risk of getting trapped in local minima during optimization. Even with regularization, most pixel-based registration methods rely on pixel-level assumptions, making them susceptible to fluctuations in imaging conditions and background noise. This leads to significant deviations in optimization results in local regions and a lack of global robustness to overall deformation. Therefore, these methods still face challenges in handling complex cardiac motion, especially under high-noise or complex background conditions, where optimization may be limited to local optima, making it difficult to stably capture realistic myocardial motion patterns.

[0005] Feature-based registration methods estimate a dense cardiac motion field by extracting and matching salient image features (such as corners, edges, or structure maps) between images. Unlike pixel-intensity-dependent registration methods, these methods utilize the spatial distribution and matching relationships of features to estimate the motion field, thus exhibiting strong robustness in situations with drastic changes in image intensity or the presence of noise. However, registration performance is highly dependent on the quality of features and the accuracy of matching; failure in feature extraction or matching directly affects the accuracy and stability of the entire motion field.

[0006] In summary, pixel-based registration methods are typically based on pixel-level assumptions and are easily affected by fluctuations in imaging conditions and background noise, leading to large deviations in optimization results in local areas and a lack of global robustness to overall deformation. Feature-based registration methods, on the other hand, exhibit strong robustness in situations with drastic changes in image intensity or the presence of noise, and perform well in areas with large-scale motion and clear structures. However, they still have limitations in adapting to individual differences, preserving cardiac topology, and extracting fine motion. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art by providing a cardiac motion estimation method, device and medium based on sparse keypoints. By using a keypoint detection network and loss constraint settings, the accuracy of keypoint pairing and the stability of cardiac motion modeling are improved.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] According to a first aspect of the present invention, a sparse keypoint-driven cardiac motion estimation method is provided, comprising:

[0010] The moving image and the fixed image are respectively input into the spatiotemporal consistency key point detection network to extract key point pairs. In the key point detection process, a local correlation map is generated based on the normalized cross-correlation between key points in the fixed image and key points in the moving image. After the local correlation map is converted into a probability distribution map, the position of key points in the moving image is corrected.

[0011] Based on the relative motion between each pair of key points, a sparse basic motion vector is constructed to obtain a sparse motion vector representing the heart motion. Based on the sparse motion vector, a deep neural network is used to reconstruct the motion of the moving image to generate a dense motion field.

[0012] Preferably, the spatiotemporal consistency key point detection network is a convolutional neural network based on a conjoined structure, including an upsampling path and a downsampling path; the downsampling path is provided with multiple convolutional modules, each convolutional module including a 3D convolutional layer, a 3D batch normalization layer and an activation layer; the features on the upsampling path are connected to the features of the corresponding spatial dimension on the downsampling path after the spatial dimension is repaired by 3D linear interpolation, so as to perform multi-scale fusion;

[0013] The upsampled features are further refined by 3D convolutional layers, 3D batch normalization layers, and activation layers. The output layer outputs multiple 3D heatmaps. Each 3D heatmap is subjected to softmax normalization to generate a corresponding confidence map. The key point location is obtained by calculating the weighted expectation of spatial coordinates.

[0014] Preferably, a local correlation map is generated, and the key point positions in the moving image are corrected after the local correlation map is converted into a probability distribution map. Specifically, this includes:

[0015] Calculate key points of a fixed image With candidate optimization key points Normalized cross-correlation between them generates a local correlation plot. in, Indicates the initial moving image key points Candidate optimization key points centered on Scope;

[0016] The softmax function is used to extract the local correlation map. Convert to probability distribution graph

[0017] Using probability distribution diagrams For the initial moving image key points 3D heat map After making corrections, the corrected 3D heat map is obtained.

[0018] The softmax function is used to refine the 3D heatmap. Normalization is performed to generate a confidence map. Based on the confidence map, the corrected key points of the moving image are obtained by calculating the expected value of spatial coordinates.

[0019] In this context, the subscripts M and F correspond to the moving image and the fixed image, respectively, and the superscript k represents the kth key point.

[0020] Preferably, the use of a probability distribution map For the initial moving image key points 3D heat map After making corrections, the corrected 3D heat map is obtained. The calculation expression is:

[0021]

[0022] Preferably, the loss function used for network training includes image reconstruction loss and keypoint reprojection loss, which are represented in a weighted sum form;

[0023] The keypoint reprojection loss includes local spatiotemporal consistency loss, which is characterized by the degree of matching between paired keypoints in a fixed image F and a moving image M.

[0024] Preferably, the keypoint reprojection loss further includes a distribution control loss, which is characterized by calculating the degree of overlap of the spatial influence range of paired keypoints in the fixed image F and the moving image M. The calculation expression is as follows:

[0025]

[0026] In the formula: K is the total number of keypoints; ReLU is the activation function; G(p) represents a Gaussian distribution centered at the keypoint location p, used to measure the spatial influence range of each keypoint in the image; σ is the variance; and u is the coordinate of each pixel in the 3D heatmap.

[0027] Preferably, the image reconstruction loss includes the registered image obtained by the moving image M through dense motion field distortion transform. The L1 norm difference between the image and the fixed image F, and the registered image The L2 norm difference between key points in the image and key points in a fixed image F.

[0028] Preferably, a deep neural network based on an encoder-decoder architecture is used to perform dense motion reconstruction on the moving image to generate a dense motion field.

[0029] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement any of the methods described above.

[0030] According to a third aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) This invention estimates cardiac motion based on a key point-driven approach. By constructing a sparse motion representation using representative key points, it enhances the ability to model the topology of different individual anatomical structures and the corresponding relationships in different images. This effectively avoids structural mismatch and topological destruction problems during the registration process. During key point detection, a local correlation map is generated based on the normalized cross-correlation between key points in the fixed image and key points in the moving image. After converting the local correlation map into a probability distribution map, the position of key points in the moving image is corrected, thereby avoiding sudden jumps or unreasonable movements of key point positions. This makes the adjustment process of key points smoother and more natural, and improves the ability to model significant coarse-grained motion changes in individualized hearts.

[0033] (2) This invention improves the stability of key point registration by introducing distributed control and local spatiotemporal consistency constraints. The key point detection network optimization process introduces distributed control and local spatiotemporal consistency constraints, which can significantly improve the pairing stability of key points in the cardiac cycle and maintain the consistency of the temporal topology, thereby improving the continuity and robustness of the entire motion estimation. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the cardiac motion estimation framework of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0036] Example

[0037] like Figure 1 As shown, this embodiment provides a sparse keypoint-driven cardiac motion estimation method, including:

[0038] S1. Input the moving image M and the fixed image F into the spatiotemporal consistency keypoint detection network respectively, and extract the corresponding keypoints. Among them, the former of two adjacent cardiac image frames is the fixed image F, and the latter is the moving image M.

[0039] In this embodiment, the spatiotemporal consistency keypoint detection network is a convolutional neural network based on a conjoined Siamese structure with a symmetrical architecture, including upsampling and downsampling paths. This structure effectively improves the robustness and spatial distribution quality of keypoint detection while maintaining spatial context continuity.

[0040] The downsampling path includes three convolutional modules, each consisting of a 3D convolutional layer (3×3×3 kernel size, stride 1), a 3D batch normalization layer (BatchNorm3D), and a ReLU activation layer. Downsampling is achieved through 3D average pooling (2×2×2 pooling kernel size, stride 2), with channel numbers of 32, 64, and 128 respectively. After downsampling, the features reach the minimum spatial resolution and maximum number of channels at the center of the network. Features on the upsampling path are then connected to the corresponding spatial dimensions of the features on the downsampling path via 3D linear interpolation to restore their spatial dimensions, enabling multi-scale fusion. The upsampling features are further refined by the 3D convolutional layer, the 3D batch normalization layer, and the activation layer, with channel numbers of 128, 64, and 32 respectively. The final output layer is a 1×1×1 3D convolutional layer, compressing the number of channels from 32 to K 3D heatmap channels (each keypoint corresponds to one 3D heatmap). In this embodiment, K is set to 35. The output 3D heatmap is the same size as the input image, i.e., 128×128×16. After Softmax normalization, a confidence map is generated from the 3D heatmap, and the key point locations are calculated using the weighted expectation of spatial coordinates.

[0041] For the kth key point The calculation expression is:

[0042]

[0043] In the formula: u=(X,Y,Z) are the coordinates of each pixel in the 3D heatmap, and Θ represents the search space of the coordinates; This represents the 3D heatmap corresponding to the k-th key point, with subscripts M and F representing the moving image and the fixed image, respectively; softmax is the normalized exponential function.

[0044] Finally, the key point sets corresponding to the moving image M and the fixed image F are obtained:

[0045]

[0046] To improve the spatial coverage and temporal consistency of keypoints in the heart region, two types of keypoint constraints are introduced:

[0047] (a) Distributed control constraints

[0048] To ensure keypoints are reasonably distributed in the image and to capture diverse and informative motion features, a distribution control constraint is introduced. This constraint penalizes keypoints that are too close together, preventing them from over-clustering in small areas.

[0049] Distributed control loss L dis Defined as:

[0050]

[0051] In the formula: the keypoint location p is represented by a Gaussian graph G(p) centered at p, σ is 0.005, u = (x, y, z) are the coordinates of each pixel in the heatmap, and the Gaussian distribution is used to measure the spatial influence range of each keypoint in the image. When the influence ranges of two keypoints overlap within their respective local regions, the sum of their Gaussian graphs may exceed 1. In this case, the distribution is penalized to prevent keypoints from clustering too densely. The ReLU function ensures that only significant overlaps are penalized, thus maintaining an appropriate distribution of keypoints in the image. In this way, keypoints can effectively cover the entire image, capturing rich motion information.

[0052] (b) Local spatiotemporal consistency constraints

[0053] To ensure the accuracy of keypoint pairing at different time points and maintain the spatial topological consistency of paired keypoints, a local spatiotemporal consistency constraint is introduced. This constraint guides the keypoint extraction and matching process, ensuring more accurate pairing of keypoints in the temporal dimension while maintaining a reasonable topological structure in space. The introduction of this constraint helps enhance the consistency of keypoints across different time frames, making the extracted motion information more reliable.

[0054] For key points in a fixed image F The initial keypoints are optimized by searching for better matching points within a local region of the moving image M. The search process is limited to initial key points. Within the centered search window, the window size is set to 1 / 8 of the feature map size. This strategy ensures that keypoint optimization is performed only within the relevant region around the initial keypoints, thereby reducing noise interference from regions far from the keypoints and lowering computational costs.

[0055] In this embodiment, during keypoint detection, a local correlation map is generated based on the normalized cross-correlation between keypoints in a fixed image and keypoints in a moving image. This local correlation map is then converted into a probability distribution map to correct the positions of keypoints in the moving image. This operation avoids sudden jumps or unreasonable movements in keypoint positions, making the adjustment process smoother and more natural. Specifically, it includes:

[0056] 1) Calculate key points of a fixed image With candidate optimization key points The normalized cross-correlation between them generates a local correlation plot.

[0057] Correlation is extracted based on local feature blocks around each keypoint, with the feature block size being 1 / 16 of the feature map size. The features are extracted based on three consecutive 3D convolutional layers, each with 16 channels. This method can more accurately match keypoint positions, ensuring consistency between the keypoints in the moving image and their corresponding keypoints in the fixed image, thereby improving overall alignment accuracy.

[0058] The expression for calculating the local correlation plot is:

[0059]

[0060] In the formula: Cov(*) represents the covariance between two feature blocks, and σ represents the standard deviation calculation; A window represents a local feature block centered at a keypoint p in an image. Indicates the initial key point Potential optimization key points centered on The range.

[0061] 2) To further highlight the possible locations of key optimization points and reduce noise interference, the softmax function is used to optimize the local correlation map. Convert to probability distribution graph The expression is:

[0062]

[0063] 3) Use probability distribution diagrams For initial key points 3D heat map After making corrections, the corrected 3D heat map is obtained.

[0064]

[0065] In the formula: K is the number of cardiac motion types.

[0066] This operation avoids sudden jumps or unreasonable movements in the position of key points, making the adjustment process of key points smoother and more natural.

[0067] 4) Based on equations (1) and (2), the corrected three-dimensional heat map is... After normalization using the softmax function, a confidence map is generated. The optimized keypoint locations are then calculated using the weighted expectation of spatial coordinates.

[0068] The local spatiotemporal consistency loss is used to measure the matching degree of paired keypoints in fixed and moving images, and is defined as:

[0069]

[0070] In the formula: G(p) represents a Gaussian distribution centered at the keypoint location p, which is used to measure the spatial influence range of each keypoint in the image, so that the influence of the keypoint location can be extended in space, thereby better capturing local consistency.

[0071] When the local spatiotemporal consistency loss L cor A lower value indicates a better match between paired keypoints in both stationary and moving images. This loss function penalizes inconsistencies between paired keypoints, ensuring spatial and temporal consistency and improving the model's accuracy in capturing motion information. This consistency loss is particularly important for dynamic image registration because it maintains the relative positional relationships of paired keypoints in both spatial and temporal dimensions.

[0072] S2. Construct sparse basic motion vectors based on the relative motion between each pair of key points to obtain sparse motion vectors representing cardiac motion. Based on the sparse motion vectors, use a deep neural network to reconstruct dense motion from the moving image to generate a dense motion field.

[0073] Specifically, a sparse basic motion vector is constructed based on the relative motion between each pair of key points to obtain a sparse motion vector representing cardiac motion. The sparse motion vector is then subjected to three-dimensional repetitive broadcasting to generate a dense basic motion field. Adopting dense basic sports fields Perform a spatial warp transformation on the moving image M to generate a coarsely registered image set. The coarsely registered image set The input is fed into a dense motion estimation network to obtain the weight mask M corresponding to each dense basic motion field. 1～K Each mask represents the spatial contribution intensity of the corresponding basic sports field, and the weighted combination yields a dense sports field.

[0074] In this embodiment, the dense motion estimation network consists of 4 encoders and 4 decoders. The number of encoders and decoders can be adjusted according to the task. For example, if the image size is large, the number of encoders and decoders can be increased.

[0075] Specifically, the encoder consists of 3D convolutional layers, 3D batch normalization layers (BatchNorm3D), ReLU activation layers, and 3D average pooling (AveragePool3D) downsampling layers, used to progressively extract spatial context information and compress feature dimensions. The decoder consists of 3D interpolation upsampling layers, 3D convolutional layers, 3D batch normalization layers (BatchNorm3D), and ReLU activation layers, used to progressively restore spatial resolution and fuse low-level features, achieving a fine estimation of the contribution weights of each sparse motion vector.

[0076] In this embodiment, the loss function L used for network training includes reconstruction loss L. Recon and keypoint reprojection loss L Keypoint Keypoint reprojection loss L KeyPoint Including distributed control loss L dis And loss of local spatiotemporal consistency.

[0077] Specifically, reconstruction loss L Recon Including the registered image obtained by dense motion field distortion transformation of the moving image M. The L1 norm difference between the image and the fixed image F, and the registered image Key points in a fixed image F The L2 norm difference between them; where the superscript k corresponds to the k-th keypoint, and the expression is:

[0078]

[0079] In the formula: This represents the spatial warp transformation operator, implemented using the `grid_sample` function in PyTorch. It performs 3D bilinear interpolation and is used to transform a moving image M and its key points. Based on the estimated motion field distortion, an image and key point locations aligned with the fixed image F are finally generated.

[0080] To further verify the effectiveness of the present invention, experiments were conducted on 4D-CT (3D CT with time dimension) and ACDC (dynamic contrast-enhanced imaging) datasets, respectively.

[0081] In this embodiment, the selected evaluation parameters include:

[0082] The DICE (Diabetic Intersection Index) measures the degree of overlap between two sets and is used to evaluate segmentation tasks. A value closer to 1 indicates better segmentation, while 0 indicates no overlap at all.

[0083] Hausdorff distance (HD): Evaluates the maximum mismatch between two shapes and is often used to assess the accuracy of segmentation boundaries. The smaller the value, the closer the two shapes are.

[0084] Mean Surface Memory (ASD): Evaluates the average surface error between two shapes and is used to assess the accuracy of organ or lesion boundaries in medical segmentation. The smaller the value, the closer the surfaces are.

[0085] As shown in Table 1, the distributed control loss L is also considered. dis and local spatiotemporal consistency loss L cor It achieved optimal results on both datasets.

[0086] Table 1

[0087]

[0088] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0089] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0090] The processing unit performs the various methods and processes described above. For example, in some embodiments, the methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the methods by any other suitable means (e.g., by means of firmware).

[0091] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0092] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0093] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0094] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cardiac motion estimation method based on sparse keypoints, characterized in that, include: The moving image and the fixed image are respectively input into the spatiotemporal consistency key point detection network to extract key point pairs. In the key point detection process, a local correlation map is generated based on the normalized cross-correlation between key points in the fixed image and key points in the moving image. After the local correlation map is converted into a probability distribution map, the position of key points in the moving image is corrected. Based on the relative motion between each pair of key points, a sparse basic motion vector is constructed to obtain a sparse motion vector representing the heart motion. Based on the sparse motion vector, a deep neural network is used to reconstruct the motion of the moving image to generate a dense motion field.

2. The cardiac motion estimation method based on sparse keypoints according to claim 1, characterized in that, The spatiotemporal consistency key point detection network is a convolutional neural network based on a conjoined structure, including an upsampling path and a downsampling path. The downsampling path is equipped with multiple convolutional modules, each of which includes a 3D convolutional layer, a 3D batch normalization layer, and an activation layer. The features on the upsampling path are connected to the features of the corresponding spatial dimension on the downsampling path after the spatial dimension is repaired by 3D linear interpolation, so as to perform multi-scale fusion. The upsampled features are further refined by 3D convolutional layers, 3D batch normalization layers, and activation layers. The output layer outputs multiple 3D heatmaps. Each 3D heatmap is subjected to softmax normalization to generate a corresponding confidence map. The key point location is obtained by calculating the weighted expectation of spatial coordinates.

3. The cardiac motion estimation method based on sparse keypoints according to claim 1, characterized in that, In the keypoint detection process, a local correlation map is generated based on the normalized cross-correlation between keypoints in the fixed image and keypoints in the moving image. After converting the local correlation map into a probability distribution map, the positions of keypoints in the moving image are corrected. Specifically, this includes: Calculate key points of a fixed image With candidate optimization key points Normalized cross-correlation between them generates a local correlation plot. in, Indicates the initial moving image key points Candidate optimization key points centered on Scope; The softmax function is used to extract the local correlation map. Convert to probability distribution graph Using probability distribution diagrams For the initial moving image key points 3D heat map After making corrections, the corrected 3D heat map is obtained. The softmax function is used to refine the 3D heatmap. Normalization is performed to generate a confidence map. Based on the confidence map, the corrected key points of the moving image are obtained by calculating the expected value of spatial coordinates. In this context, the subscripts M and F correspond to the moving image and the fixed image, respectively, and the superscript k represents the kth key point.

4. The cardiac motion estimation method based on sparse keypoints according to claim 3, characterized in that, The probability distribution map is used For the initial moving image key points 3D heat map After making corrections, the corrected 3D heat map is obtained. The calculation expression is:

5. The cardiac motion estimation method based on sparse keypoints according to claim 3, characterized in that, The loss functions used for network training include image reconstruction loss and keypoint reprojection loss, which are represented in a weighted sum form. The keypoint reprojection loss includes local spatiotemporal consistency loss, which is characterized by the degree of matching between paired keypoints in a fixed image F and a moving image M.

6. The cardiac motion estimation method based on sparse keypoints according to claim 5, characterized in that, The keypoint reprojection loss also includes distribution control loss, which is characterized by calculating the degree of overlap of paired keypoints in the spatial influence range of the fixed image F and the moving image M. The calculation expression is as follows: In the formula: K is the total number of keypoints; ReLU is the activation function; G(p) represents a Gaussian distribution centered at the keypoint location p, used to measure the spatial influence range of each keypoint in the image; σ is the variance; and u is the coordinate of each pixel in the 3D heatmap.

7. The cardiac motion estimation method based on sparse keypoints according to claim 5, characterized in that, The image reconstruction loss includes the registered image obtained from the moving image M after dense motion field distortion transformation. The L1 norm difference between the image and the fixed image F, and the registered image The L2 norm difference between key points in the image and key points in a fixed image F.

8. The cardiac motion estimation method based on sparse keypoints according to claim 1, characterized in that, A deep neural network based on an encoder-decoder architecture is used to perform dense motion reconstruction on the moving image, generating a dense motion field.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images