A myocardial motion tracking method based on myocardial geometric prior and image feedback correction

Abstract

The application discloses a myocardial motion tracking method based on myocardial geometry prior and image feedback correction. The myocardial region geometry prior model is constructed to express the myocardial tissue as a two-dimensional triangular mesh structure with spatial topological relationship, and the spatial constraint relationship in the myocardial tissue is described by using the connection relationship between the mesh nodes. The local image features in the reference frame image and the target frame image are combined to predict the mesh node motion, and the motion information is interpolated according to the spatial relationship in the mesh unit to generate an initial dense motion field covering the myocardial region, so that the spatial continuity and anatomical structure rationality of the motion estimation result are improved. The motion correction mechanism based on image feedback is introduced, the initial dense motion field is used for motion compensation of the target frame image, the image difference information between the reference image and the motion compensated image is calculated, and the image feedback residual reflecting the current motion estimation error is obtained.

Description

Technical Field

[0001] This invention relates to a method for tracking myocardial motion based on myocardial geometric priors and image feedback correction, belonging to the field of computer artificial intelligence. Background Technology

[0002] Myocardial motion reflects the dynamic changes of cardiac tissue during contraction and relaxation, serving as a crucial basis for evaluating cardiac function and aiding in the diagnosis of cardiovascular diseases. Compared to traditional evaluation methods based on overall cardiac function parameters (such as ejection fraction (EF) and stroke volume (SV), local myocardial motion information can more sensitively reflect myocardial dysfunction. Therefore, accurately, rapidly, and non-invasively acquiring myocardial motion information has become an important research direction in the field of medical image analysis.

[0003] Currently, myocardial motion tracking methods mainly include image feature matching-based methods, deformation constraint-based methods, and deep learning-based methods. Two-dimensional speckle tracking technology detects speckle texture changes in echocardiograms to establish correspondences between local tissue regions in adjacent image frames, thereby obtaining myocardial motion information. Deformation constraint-based methods typically optimize the displacement between consecutive image frames by constructing image similarity constraints, motion smoothing constraints, or predefined deformation models. With the development of deep learning technology, neural network-based motion estimation methods are increasingly being applied to myocardial motion analysis, enabling the prediction of dense myocardial motion fields by learning the feature change relationships in image sequences.

[0004] However, most of the aforementioned methods rely on image texture information, local feature matching, or pixel-level correspondences for motion estimation, lacking explicit constraints on the geometric structure and topological relationships of myocardial tissue. In practical applications, medical images often exhibit low signal-to-noise ratios, weakly textured regions, and complex non-rigid deformations, easily leading to inaccurate local motion estimation, insufficient spatial continuity of the motion field, and abnormal deformations in some areas that do not conform to the structural patterns of myocardial tissue. While some deformation-based methods introduce prior constraints to improve the stability of motion estimation, their priors typically rely on manually set parameterized models, making it difficult to adaptively express complex local motion changes based on individual differences in myocardial morphology.

[0005] Because myocardial tissue has a well-defined spatial structure and topological connections, its motion changes exhibit a certain degree of continuity and regional consistency. Therefore, how to utilize prior information about myocardial geometry to reduce the uncertainty of motion estimation, and how to combine image information to provide feedback correction for motion results, thereby achieving a balance between overall anatomical consistency and the ability to express local motion details, remains a problem that needs further resolution in the field of medical image analysis.

[0006] The existing technology has the following technical problems:

[0007] 1) Existing technologies, in the process of tracking myocardial motion in continuous medical imaging sequences, cannot simultaneously guarantee the consistency of the overall anatomical structure, the ability to express local motion details, and the accuracy under complex motion conditions in motion estimation results, which easily leads to motion estimation bias.

[0008] 2) Existing technologies lack effective constraints on myocardial geometry and spatial topology. Under conditions of weak texture, low signal-to-noise ratio, and complex non-rigid deformation, local motion estimation bias and abnormal deformation that do not conform to the motion law of myocardial tissue are easily generated.

[0009] 3) Existing technologies are insufficient in modeling the spatial relationships within the myocardial region, making it difficult to fully utilize motion correlation information between different locations. When there are differences in motion in local areas, motion estimation is prone to discontinuity.

[0010] 4) Existing technologies lack a feedback correction mechanism based on the difference between motion compensation results and target image information. When there are local errors in the initial motion estimation, it is difficult to use image feedback information to dynamically correct the motion estimation results. Summary of the Invention

[0011] This invention addresses the problems of insufficient structural consistency in motion estimation results, limited ability to express local motion information, and difficulty in further correcting motion errors in existing myocardial motion tracking methods during continuous medical image sequence analysis. It provides a myocardial motion tracking method based on myocardial geometric priors and image feedback correction. This method utilizes the inherent geometric morphology and spatial topological features of the myocardial region to construct structured motion constraints, and combines this with motion-compensated image feedback information to dynamically correct the initial motion estimation results. This achieves effective fusion of structural prior information and dynamic image information during myocardial motion estimation, improving the accuracy, continuity, and stability of myocardial motion tracking results.

[0012] Specifically, this method incorporates prior geometric information of the myocardium into the motion estimation process, integrating the spatial structural features of the myocardial region to make the motion results more consistent with the anatomical characteristics of myocardial tissue. Simultaneously, it utilizes image feedback information generated from the motion estimation results to correct local errors in the initial motion results, forming a feedback optimization mechanism based on image information in the motion estimation process. This results in a more accurate and dense motion field that conforms to the laws of myocardial motion. This invention is achieved through the following technical solutions:

[0013] This invention discloses a myocardial motion tracking method based on myocardial geometric prior and image feedback correction, comprising:

[0014] S1: Acquire a medical image sequence of continuous cardiac cycles, select a reference frame image and a target frame image from the medical image sequence, and obtain the myocardial region segmentation mask of the reference frame image and the target frame image;

[0015] S2: Construct a two-dimensional myocardial geometric mesh model based on the myocardial region segmentation mask of the reference frame image. The two-dimensional myocardial geometric mesh model includes mesh nodes, mesh edges, and mesh cells.

[0016] S3: Based on the spatial position information of the grid nodes in the two-dimensional myocardial geometric grid model, extract the corresponding local image features from the reference frame image and the target frame image, combine them with the spatial position information of the grid nodes, input them into the displacement prediction network, and predict the two-dimensional motion displacement of each grid node.

[0017] S4: Based on the two-dimensional motion displacement of each grid node and the grid elements in the two-dimensional myocardial geometric grid model, the two-dimensional motion displacement of each grid node is mapped to the myocardial region to generate an initial dense motion field.

[0018] S5: Use the initial dense motion field to perform motion compensation on the reference frame image to obtain a motion-compensated image, and obtain the image feedback residual based on the difference between the target frame image and the motion-compensated image;

[0019] S6: Based on the initial dense motion field and image feedback residuals, obtain the local residual correction amount through the residual correction model;

[0020] S7: Correct the initial dense motion field based on the local residual correction amount to generate the final dense motion field of the myocardium.

[0021] As a further improvement, in step S2 of this invention, constructing a two-dimensional myocardial geometric mesh model based on the myocardial region segmentation mask of the reference frame image includes:

[0022] Extract the intima and adventitia boundary contours from the myocardial region segmentation mask of the reference frame image, sample the intima and adventitia boundary contours to generate boundary grid nodes, and sample the internal region of the myocardial region segmentation mask to generate internal grid nodes.

[0023] Constrained triangulation is performed on the boundary and internal mesh nodes to generate a two-dimensional myocardial geometric mesh model; the two-dimensional myocardial geometric mesh model is represented as follows: ,in Represents a set of grid nodes. Represents the set of grid edges. This represents a set of grid cells.

[0024] As a further improvement, in step S3 of this invention, predicting the two-dimensional motion displacement of each grid node includes:

[0025] Based on the spatial location of each grid node, local image regions centered on each grid node are extracted in the reference frame image and the target frame image, and convolutional neural networks are used to extract local image features.

[0026] Local image features are fused with the spatial location information of grid nodes to obtain the node features of each grid node;

[0027] The node features are input into the graph neural network, and the features are propagated according to the topological relationship corresponding to the grid edges to obtain the updated node features.

[0028] The updated node features are input into the displacement prediction network to obtain the two-dimensional motion displacement corresponding to each grid node.

[0029] As a further improvement, in step S4 of the present invention, generating the initial dense motion field includes:

[0030] Determine the grid cell containing the pixel to be interpolated within the corresponding region of the myocardial region segmentation mask;

[0031] Calculate the centroid coordinates of the pixel to be interpolated relative to the three grid nodes in the grid cell;

[0032] Based on the centroid coordinates, the two-dimensional motion displacements corresponding to the three grid nodes in the grid cell are weighted and summed to obtain the motion displacement at the pixel to be interpolated. The initial dense motion field is formed by the motion displacements at each pixel to be interpolated within the corresponding region of the myocardial region segmentation mask.

[0033] As a further improvement, in step S6 of this invention, the residual correction model includes a lightweight convolutional feature extraction module and an implicit neural network, and its construction and prediction process includes:

[0034] The coordinates of each pixel in the initial dense motion field are encoded to obtain positional features. The positional features, local image features, motion displacement at the corresponding position in the initial dense motion field, and image feedback residuals are fused to obtain the input features of the implicit neural network.

[0035] The input features are fed into an implicit neural network composed of multilayer perceptrons to predict the local residual correction amount corresponding to each pixel position.

[0036] As a further improvement, the S3 displacement prediction network and the S6 motion residual correction model described in this invention are trained jointly:

[0037] Total loss function during joint training Represented as:

[0038] ;

[0039] in, This represents the mesh motion consistency loss used in the displacement prediction network output. This represents the residual correction loss used in the output of the residual correction model. This indicates the loss of motion consistency in the final dense motion field of the myocardium. , as well as These represent the weighting coefficients of the corresponding loss terms.

[0040] As a further improvement, the mesh motion consistency loss described in this invention is used to constrain the displacement prediction network, and it is expressed as follows:

[0041] ;

[0042] in, For the set of grid edges, and Adjacent grid nodes and The corresponding two-dimensional motion displacement;

[0043] The residual correction loss is used to constrain the residual correction model, and it is expressed as follows:

[0044] ;

[0045] in, This is the set of pixel locations in the myocardial region. For pixels The local residual correction amount at the location.

[0046] As a further improvement, the motion consistency loss described in this invention includes image similarity loss, motion field smoothing constraint loss, and mesh deformation constraint loss:

[0047] Image similarity loss is expressed as:

[0048] ;

[0049] in, For the target frame image, For reference frame image, For the final dense movement field of the myocardium, Indicated according to the sports field Motion compensation for the image. This represents the normalized cross-correlation coefficient.

[0050] The smoothing constraint loss of the motion field is expressed as:

[0051] ;

[0052] in, pixel position The spatial gradient of the final dense motion field of the myocardium.

[0053] As a further improvement, the mesh deformation constraint loss described in this invention is expressed as:

[0054] Mesh deformation constraint loss avoids mesh cell flipping based on the area of ​​the triangular mesh cells after deformation:

[0055] ;

[0056] in, The grid cells represent the two-dimensional myocardial mesh model. This represents the directed area of ​​the triangular mesh cell after deformation.

[0057] The beneficial effects of this invention are as follows:

[0058] (1) By introducing myocardial geometric prior information, the myocardial region of the reference frame is constructed as a two-dimensional mesh model containing mesh nodes, mesh edges, and mesh cells. The spatial topological connection relationship between mesh nodes is used to explicitly describe the geometric structural constraints within the myocardial tissue. Compared with the problem of abnormal motion easily generated by existing technologies, this invention can use the myocardial anatomical structure prior to constrain the overall motion trend, thereby improving the spatial continuity and structural rationality of the motion field. Experimental results show that by using only the myocardial geometric mesh model to generate the initial dense motion field, the myocardial region overlap similarity index (Dice) is increased from 56.12% in the unregistered state to 75.60%, an increase of 19.48 percentage points, and the 95% Hausdorff distance (HD95) is reduced from 6.68 mm to 4.20 mm, a decrease of 2.48 mm, indicating that the geometric prior constraints can effectively improve the spatial alignment effect of the myocardial region.

[0059] (2) By utilizing the node connection relationships in the two-dimensional myocardial geometric mesh model, the spatial location features of the mesh nodes, local image features, and mesh topology information are fused, and feature propagation is performed based on the topological relationships corresponding to the mesh edges, enabling the motion information between different spatial locations to be interconnected. Compared with the existing technology, which lacks modeling of the internal spatial relationships of the myocardium and is prone to discontinuities in local regional motion, this invention can utilize the topological correlation within the myocardial tissue to constrain node motion changes, making the predicted motion results more consistent with the continuous deformation law and reducing the possibility of local abnormal motion. When tested under different motion span conditions, when the motion offset increased to 14, the unregistered result Dice decreased to 70.29%, while it remained at 81.13% after adopting the method of this invention, indicating that the constraint mechanism can improve the tracking stability under complex motion conditions.

[0060] (3) By using the discrete motion displacements predicted from the grid nodes and combining them with the grid cell structure, a spatial interpolation method based on the centroid coordinates is adopted to continuously map the motion information of the discrete nodes to the entire myocardial region, generating an initial dense motion field covering the myocardial region. This method can maintain the continuous representation of the myocardial motion field while avoiding the problem of local motion discontinuity that is prone to occur in traditional motion estimation methods. Based on the geometric prior motion estimation, the Dice of the myocardial region reaches 75.60%, and the normalized cross-correlation coefficient (NCC) increases from 0.37 in the unregistered state to 0.46, verifying that generating a dense motion field based on the grid structure can effectively improve the consistency of motion estimation.

[0061] (4) By introducing a motion correction mechanism based on image feedback residuals, the initial dense motion field is used to perform motion compensation on the target frame image, and the image feedback residual is obtained based on the difference between the reference frame image and the motion-compensated image. Then, the local residual correction amount is predicted based on the feedback residual, and the initial motion field is dynamically optimized. Compared with the problem of lacking feedback correction based on the relationship between motion results and image changes in the prior art, the present invention can further compensate for local small deformations that are difficult to express by geometric models and improve the ability to express local motion details. Experimental results show that after adding image feedback residual correction, the final Dice reaches 82.20%, which is 26.08 percentage points higher than the unregistered result, a relative improvement of 46.67%; NCC reaches 0.62, which is 67.57% higher than the unregistered result; the structural similarity index (SSIM) reaches 0.81, and the peak signal-to-noise ratio (PSNR) reaches 34.33dB, indicating that the image feedback correction mechanism can further reduce local motion errors and improve motion estimation accuracy and image consistency.

[0062] (5) By fusing large-scale motion information obtained based on myocardial geometry priors with small-scale residual motion information obtained based on image feedback, a motion tracking framework combining structural constraints and image-driven approaches is formed, which improves the adaptability of complex motion regions while ensuring the consistency of the overall anatomical structure. Compared with existing medical image registration and myocardial motion estimation methods, the method of this invention achieves an average Dice of 82.20%, which is higher than VoxelMorph's 80.27%, DeepTag's 81.15%, and DiffuseMorph's 80.56%. At the same time, HD95 is reduced to 3.73 mm and PSNR reaches 34.33 dB, indicating that the method of this invention has better overall performance in terms of myocardial motion tracking accuracy and image registration consistency.

[0063] (6) By adopting a motion tracking method based on the fusion of myocardial geometric structure information and medical imaging information, this method can be applied to different types of cardiac medical imaging sequences, including echocardiogram sequences and cardiac magnetic resonance cine images, providing more accurate and reliable motion information for myocardial function evaluation, detection of local motion abnormalities and auxiliary analysis of cardiovascular diseases, thus improving the practical application value of the method. Attached Figure Description

[0064] Figure 1 This is the overall flowchart of the method of the present invention. Detailed Implementation

[0065] Exemplary embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0066] The purpose of this invention is to address the problems of existing myocardial motion tracking methods, such as the lack of myocardial structural constraints, inaccurate motion estimation in locally complex deformation regions, and the susceptibility of motion results to image quality. This invention proposes a myocardial motion tracking method based on myocardial geometric priors and image feedback correction, which can achieve accurate estimation of continuous motion changes in the myocardial region and improve the stability and reliability of myocardial motion tracking under complex motion conditions.

[0067] The specific implementation method of the present invention is as follows:

[0068] S1: Acquire continuous cardiac cycle medical image sequences and obtain reference frame images, target frame images, and myocardial region segmentation masks.

[0069] Specifically, a continuous cardiac cycle medical image sequence is acquired, a reference frame image and a target frame image are selected from the medical image sequence, and the myocardial region segmentation mask of the reference frame image and the target frame image is obtained; in this embodiment, the myocardial region segmentation mask is obtained by manual annotation by a clinical expert.

[0070] Furthermore, bilinear interpolation scaling is applied to both the reference frame image and the target frame image, and the pixel grayscale values ​​are normalized. Further, the spatial geometric location of the myocardial tissue is calculated based on the myocardial mask of the reference frame, the minimum bounding rectangle of the myocardial mask is extracted, and the range is expanded outward from the minimum bounding rectangle. The outer boundary of the image is used to crop a region of interest (ROI) image block containing the myocardial region and adjacent tissue regions. In this embodiment, the size of the cropped ROI image block is [size missing]. Simultaneously, the corresponding local myocardial region segmentation mask is obtained.

[0071] S2: Construct a two-dimensional myocardial geometric mesh model based on the myocardial region segmentation mask of the reference frame image.

[0072] Based on the preprocessed reference frame's myocardial region segmentation mask, boundary extraction is performed on the myocardial region to obtain its spatial contour information. Within the region of interest in the myocardial region, a two-dimensional triangular mesh model covering the entire myocardial tissue is constructed based on the spatial distribution of the myocardial region.

[0073] Specifically, the myocardial tissue extent is determined based on a myocardial region segmentation mask, and multiple discrete spatial sampling points are generated within the myocardial region as grid nodes, while preserving the spatial constraint information of the myocardial inner and outer boundary regions. Node connections are established based on the spatial adjacency relationships between grid nodes to form grid edges, and further, triangular grid cells enclosed by adjacent grid nodes are constructed. The two-dimensional myocardial geometric grid model is represented as follows:

[0074] ;

[0075] in, This represents a set of grid nodes used to represent discrete locations within the myocardial tissue space. This represents the total number of grid vertices; Represents the set of grid edges, used to represent adjacent vertices. and The topological connections between them; This represents a set of mesh cells used to represent a local triangular region consisting of three adjacent vertices.

[0076] Based on the spatial extent of the myocardial ROI and the image resolution, the number of grid vertices was determined. The spatial distribution is adaptively adjusted to allow the generated triangular mesh to adapt to myocardial regions of different sizes and shapes. In this embodiment, the number of mesh nodes is set to 300-1000 according to the size of the myocardial region. Approximately 500 mesh nodes are preferentially used to discretize the myocardial tissue, and corresponding mesh edges and triangular mesh units are generated through triangulation to finally form a two-dimensional myocardial geometric mesh model.

[0077] S3: Extract local image features based on the spatial location information of grid nodes in the two-dimensional myocardial geometric grid model, and predict the two-dimensional motion displacement of each grid node.

[0078] Specifically, based on the two-dimensional myocardial geometric mesh model obtained in S2, the spatial location information of each mesh node is acquired. Taking each mesh node as the center, corresponding local image patches are extracted from the reference frame ROI image and the target frame ROI image respectively, forming image patch pairs corresponding to the nodes.

[0079] In this embodiment, a local image region of size 33×33 is extracted centered on a grid node, and the local image patch is input into a convolutional neural network to obtain the local image feature vector of the corresponding node. The convolutional neural network (CNN) comprises three stages for feature extraction. The first stage uses two 3×3 convolutional layers to extract features from local image patches. The outputs of each convolutional layer are activated by ReLU and then downsampled using max pooling to obtain a 16-dimensional feature representation. The second stage further uses two 3×3 convolutional layers for deeper feature extraction, combining ReLU activation and max pooling to output a 32-dimensional feature representation. The third stage uses a single 3×3 convolutional layer for feature fusion. After ReLU activation, a fixed-length image feature vector is obtained through global average pooling. Finally, a fully connected layer maps the feature vector to obtain a 128-dimensional local image feature. .

[0080] Furthermore, local image features are fused with the node spatial features in a two-dimensional myocardial geometric mesh model to construct mesh node features. ,in This represents the spatial coordinates of the mesh nodes in the two-dimensional myocardial geometric mesh model.

[0081] Furthermore, all grid node features are input into a graph neural network (GNN), and feature propagation and updates are performed based on the node adjacency relationships represented by the grid edges in the two-dimensional myocardial geometric grid model. The GNN aggregates information from adjacent grid nodes, ensuring that the current node features simultaneously include its own local image information and the spatial motion constraints of surrounding nodes. In this embodiment, the GNN employs a Graph Sampling and Aggregation Network (GraphSAGE) structure, comprising two graph feature propagation layers. Each layer updates the node representation through adjacent node feature aggregation. After propagation by the GNN, updated node features incorporating myocardial spatial topology information are obtained.

[0082] Furthermore, the updated node features are input into the displacement prediction network, which outputs the two-dimensional motion displacement corresponding to each grid node. In this embodiment, the displacement prediction network adopts a multilayer perceptron structure, including two fully connected layers. The first fully connected layer maps the input node features to a 64-dimensional hidden space and performs feature transformation through the ReLU activation function. The second fully connected layer maps the hidden features to a two-dimensional output space to obtain the two-dimensional motion displacement of the corresponding grid node.

[0083] S4: Generate an initial dense motion field based on grid interpolation.

[0084] Specifically, based on the two-dimensional motion displacement of each grid node obtained in S3, and combined with the two-dimensional myocardial geometric grid model constructed in S2, the motion information on the discrete grid nodes is mapped to the entire myocardial region to generate an initial dense motion field.

[0085] Furthermore, for each triangular mesh unit in the two-dimensional myocardial geometric mesh model, the motion displacement corresponding to each pixel position inside the triangular mesh unit is calculated by spatial interpolation based on the two-dimensional motion displacement corresponding to its three vertices.

[0086] For any pixel located within the triangular mesh cell, the corresponding motion displacement is calculated using barycentric coordinate interpolation based on its spatial position within the triangular cell. This method propagates the discrete motion displacements of all mesh nodes to the corresponding triangular mesh cells, achieving a transformation from node-level motion estimation to pixel-level motion estimation. Furthermore, the motion displacements corresponding to the pixel positions within each triangular mesh cell are combined to obtain a two-dimensional initial dense motion field covering the entire myocardial region.

[0087] S5: Generate a motion-compensated image based on the initial dense motion field and obtain the image feedback residual.

[0088] Specifically, based on the initial dense motion field obtained in S4, a spatial transformation is performed on the reference frame image to generate a motion-compensated image. In this embodiment, a differentiable spatial transformer is used as the image deformation module. The initial dense motion field is used as the spatial sampling displacement field, and the reference frame image is warped and reconstructed using bilinear interpolation to obtain a motion-compensated image corresponding to the spatial position of the target frame image. Furthermore, the motion-compensated image With target frame image Perform pixel-level difference calculations to obtain the image feedback residual:

[0089]

[0090] in, Represents the pixel coordinates of the image. This indicates image inconsistencies that still exist after compensation by the initial dense motion field.

[0091] S6: Obtain the local residual correction amount based on the initial dense motion field and the image feedback residual.

[0092] Specifically, the input information for the residual correction model is constructed based on the initial dense motion field obtained in S4 and the image feedback residual obtained in S5. The residual correction model includes a lightweight convolutional feature extraction module and an implicit neural network.

[0093] First, the reference frame image, target frame image, motion-compensated image, and image feedback residual are combined to form multi-channel input data, which is then input into a lightweight convolutional feature extraction module to extract local image change features. In this embodiment, the lightweight convolutional feature extraction module adopts a three-level convolutional structure. The first and second stages each use two 3×3 convolutional layers for feature extraction, and the third stage uses a single 3×3 convolutional layer for feature fusion. After each convolutional layer, a ReLU activation function is used to complete the nonlinear mapping, and finally, 64-dimensional local image features are output.

[0094] Furthermore, within the myocardial region, positional encoding is performed based on the spatial coordinates of each pixel to obtain corresponding positional features. These positional features, local image features, motion displacement information at corresponding positions in the initial dense motion field, and image feedback residual information are then fused to form the input features of the implicit neural network.

[0095] Furthermore, the fused input features are input into an implicit neural network to predict the local residual correction amount corresponding to each pixel position. In this embodiment, the implicit neural network adopts a multilayer perceptron structure, including 3 hidden layers with a hidden layer dimension of 128. To avoid local abnormal corrections affecting the overall motion trend, the magnitude of the output local residual correction amount is constrained and limited to a preset range. In this embodiment, the maximum correction magnitude is set to 5 pixels.

[0096] S7: Correct the initial dense motion field based on the local residual correction amount to generate the final dense motion field of the myocardium.

[0097] Specifically, based on the local residual correction amount obtained in S6, the initial dense motion field obtained in S4 is corrected pixel by pixel, and the local residual correction information is superimposed on the initial motion estimation result to obtain the final myocardial dense motion field.

[0098] The model training process is as follows:

[0099] During model training, the displacement prediction network and motion residual correction model are jointly optimized so that the two networks can jointly learn the mapping relationship between myocardial structural constraints and image feedback information.

[0100] During joint training, the total loss function consists of grid motion consistency loss, residual correction loss, and final motion field consistency loss:

[0101] ;

[0102] in, This represents the mesh motion consistency loss used in the displacement prediction network output. This represents the residual correction loss used in the output of the residual correction model. This indicates the loss of motion consistency in the final dense motion field of the myocardium. , as well as These represent the weighting coefficients of the corresponding loss terms.

[0103] The grid motion consistency loss is used to constrain the continuity of motion changes between adjacent myocardial grid nodes, ensuring that adjacent nodes maintain a motion relationship that conforms to the continuous deformation law of myocardial tissue. It is expressed as:

[0104] ;

[0105] in, For the set of grid edges, and Adjacent grid nodes and The corresponding two-dimensional motion displacement;

[0106] The residual correction loss is used to constrain the local residual correction amount of the implicit neural network output, avoiding excessive residual correction that could lead to abnormal changes in the motion field. It is expressed as:

[0107] ;

[0108] in, This is the set of pixel locations in the myocardial region. For pixels The local residual correction amount at the location.

[0109] Motion consistency loss is used to constrain the image consistency and motion field spatial continuity between the final dense motion field of the myocardium and the target frame image, including image similarity loss, motion field smoothing constraint loss and mesh deformation constraint loss.

[0110] Image similarity loss is used to measure the degree of similarity between the motion-compensated image and the target frame image, and is expressed as:

[0111] ;

[0112] in, For the target frame image, For reference frame image, For the final dense movement field of the myocardium, Indicated according to the sports field Motion compensation for the image. This represents the normalized cross-correlation coefficient.

[0113] Motion field smoothing constraint loss is used to limit the motion changes of adjacent pixels and ensure the continuity of the motion field, and is expressed as:

[0114] ;

[0115] in, pixel position The spatial gradient of the final dense motion field of the myocardium.

[0116] The mesh deformation constraint loss is used to prevent triangular mesh cells from flipping and to maintain the stability of the myocardial geometry. It is expressed as follows:

[0117] ;

[0118] in, The grid cells represent the two-dimensional myocardial mesh model. This represents the directed area of ​​the triangular mesh cell after deformation.

[0119] Performance evaluation and experimental results:

[0120] To verify the effectiveness of the method of this invention in myocardial motion tracking tasks, experiments were conducted using the publicly available ACDC cardiac MRI dataset. The experimental data was divided into training, validation, and test sets according to the principle of patient independence. The training set contained 105 patients (9413 image samples), the validation set contained 22 patients (2209 image samples), and the test set contained 23 patients (2156 image samples). During the experiment, the network structure and parameter configuration described in the above embodiments were used for model training. The model was jointly trained end-to-end using the Adam optimizer and trained for 1500 epochs on an NVIDIA RTX 4090 GPU.

[0121] The accuracy of motion estimation and the effect of image registration are comprehensively evaluated by using the overlap similarity index (Dice), 95% Hausdorff distance (HD95), normalized cross-correlation coefficient (NCC), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR).

[0122] Based on the comprehensive evaluation results of the entire sample, the method of this invention demonstrates good effectiveness in two-dimensional myocardial motion estimation and image registration tasks. As shown in Table 1, in the performance evaluation of stepwise registration across the entire sequence, without motion compensation, the baseline results are: Dice 56.12%, HD95 6.68 mm, NCC 0.37, SSIM 0.58, and PSNR 24.86 dB.

[0123] After introducing the initial dense motion field based on myocardial geometry mesh constraints generated in step S4, the myocardial region achieved preliminary spatial alignment, with Dice increasing to 75.60%, an improvement of 19.48 percentage points compared to the unregistered result, HD95 decreasing to 4.20 mm, and NCC increasing to 0.46. This indicates that motion estimation based on myocardial geometry priors can effectively maintain the overall motion trend of the myocardial region and improve the accuracy of structural matching.

[0124] Furthermore, after image feedback residual correction in step S6 and motion field fusion optimization in step S7, the myocardial region overlap index Dice after final motion field registration reached 82.20%, which is 26.08 percentage points higher than the unregistered result. HD95 was further reduced to 3.73 mm, NCC was increased to 0.62, SSIM reached 0.81, and PSNR reached 34.33 dB.

[0125] The experimental results above show that the method of the present invention provides structural prior constraints through a two-dimensional myocardial geometric mesh model, so that the motion estimation process conforms to the spatial continuity of myocardial tissue; at the same time, the motion error in the local complex deformation area is further compensated by the image feedback residual correction mechanism, so that the final motion field takes into account both the overall structural consistency and the local image consistency, thereby improving the accuracy and stability of myocardial motion tracking.

[0126] Table 1. Performance Improvement Analysis of Step-by-Step Registration in Each Example

[0127] ;

[0128] To verify the contribution of each key step in the method of this invention to the performance of myocardial motion estimation, and to analyze the stability of motion tracking under different motion span conditions, the unregistered results, the initial dense motion field results generated based on step S4, and the final motion estimation results after residual correction and motion field fusion combined with steps S6 and S7 are compared and analyzed.

[0129] As shown in Table 2, under the condition of a small motion span (Offset01), the degree of inter-frame deformation is limited, and the Dice is 86.61% without registration. After generating the initial dense motion field in step S4, the contour registration and image similarity are improved, the Dice increases to 88.75%, and the HD95 decreases to 1.36. Based on this, after introducing image feedback residual correction in step S6 and motion field fusion in step S7, the Dice reaches 91.90%, the HD95 decreases to 1.29mm, the PSNR increases from 28.69dB to 29.53dB, and the SSIM reaches 0.89.

[0130] As the motion span increases, the inter-frame misalignment intensifies significantly. Under Offset05 conditions, the unregistered Dice drops to 72.48%, and HD95 reaches 4.10mm. After obtaining the initial dense motion field through S4, Dice increases to 84.08%, and HD95 decreases to 3.72mm. After residual correction and motion field fusion, the final Dice increases to 85.41%, an improvement of 12.93 percentage points compared to the uncompensated result, and HD95 further decreases to 3.04mm. This indicates that motion estimation based on myocardial geometric priors can effectively suppress structural shifts caused by large-scale motion, and residual fusion can further improve contour alignment and grayscale matching.

[0131] Furthermore, under a large motion span of Offset14, the unregistered Dice decreased to 70.29%, while HD95 increased to 4.81 mm. After obtaining the initial dense motion field via S4, Dice reached 78.77%, and HD95 decreased to 3.77 mm. After residual correction and motion field fusion, Dice further improved to 81.13%, HD95 decreased to 3.53 mm, and PSNR reached 22.98 dB. Under large deformation conditions, the residual correction module significantly improved NCC, SSIM, and other indicators, indicating that the image feedback residual correction mechanism can effectively compensate for local errors in the initial motion estimation and improve motion tracking stability.

[0132] The above results show that the method of the present invention provides structural prior constraints through a two-dimensional myocardial geometric mesh model, enabling the initial motion estimation to maintain the overall motion trend of myocardial tissue. Furthermore, by combining image feedback residual correction and motion field fusion mechanism, it can adaptively correct local deformation errors under different motion span conditions, and simultaneously improve contour registration accuracy and image grayscale similarity, thereby improving the accuracy and robustness of myocardial motion tracking.

[0133] Table 2 Comparison of motion estimation and registration performance at different stages under different motion spans

[0134]

[0135] Compared with existing mainstream methods for myocardial motion estimation and medical image registration, the method of this invention exhibits better overall performance. As shown in Table 3, under the same test conditions, the average Dice of the VoxelMorph (VM) medical image registration method based on convolutional neural networks is 80.27%, and the average HD95 is 4.62 mm; the average Dice of the DeepTag myocardial motion estimation method based on deep learning is 81.15%, and the PSNR is 33.68 dB; the average Dice of the DiffuseMorph registration method based on diffusion models is 80.56%, and the average HD95 is 4.38 mm. Compared with the above methods, the method of this invention improves the average Dice to 82.20%, reduces HD95 to 3.73 mm, and achieves a PSNR of 34.33 dB, indicating that the method of this invention can effectively improve the performance of myocardial motion estimation and medical image registration.

[0136] Table 3. Performance Comparison of Different Myocardial Motion Estimation and Medical Image Registration Methods

[0137]

[0138] The above experiments demonstrate that the method proposed in this invention can effectively combine prior knowledge of myocardial geometry with image feedback correction mechanisms, while maintaining the spatial continuity of myocardial motion and improving the estimation accuracy of complex motion regions, thus enhancing the adaptability to complex deformation regions during myocardial motion tracking.

[0139] The above is not intended to limit the present invention. It should be noted that, for those skilled in the art, various changes, modifications, additions or substitutions can be made without departing from the essential scope of the present invention, and these improvements and refinements should also be considered within the scope of protection of the present invention.

Claims

1. A method for tracking myocardial motion based on myocardial geometric prior and image feedback correction, characterized in that, include: S1: Acquire a medical image sequence of continuous cardiac cycles, select a reference frame image and a target frame image from the medical image sequence, and obtain the myocardial region segmentation mask of the reference frame image and the target frame image; S2: Construct a two-dimensional myocardial geometric mesh model based on the myocardial region segmentation mask of the reference frame image. The two-dimensional myocardial geometric mesh model includes mesh nodes, mesh edges, and mesh cells. S3: Based on the spatial position information of the grid nodes in the two-dimensional myocardial geometric grid model, extract the corresponding local image features from the reference frame image and the target frame image, combine them with the spatial position information of the grid nodes, input them into the displacement prediction network, and predict the two-dimensional motion displacement of each grid node. S4: Based on the two-dimensional motion displacement of each grid node and the grid elements in the two-dimensional myocardial geometric grid model, the two-dimensional motion displacement of each grid node is mapped to the myocardial region to generate an initial dense motion field. S5: Use the initial dense motion field to perform motion compensation on the reference frame image to obtain a motion-compensated image, and obtain the image feedback residual based on the difference between the target frame image and the motion-compensated image; S6: Based on the initial dense motion field and image feedback residuals, obtain the local residual correction amount through the residual correction model; S7: Correct the initial dense motion field based on the local residual correction amount to generate the final dense motion field of the myocardium.

2. The myocardial motion tracking method according to claim 1, characterized in that, In step S2, constructing a two-dimensional myocardial geometric mesh model based on the myocardial region segmentation mask of the reference frame image includes: Extract the intima boundary contour and the adventitia boundary contour from the myocardial region segmentation mask of the reference frame image, sample the intima boundary contour and the adventitia boundary contour to generate boundary mesh nodes, and sample the internal region of the myocardial region segmentation mask to generate internal mesh nodes. Constrained triangulation is performed on boundary and internal mesh nodes to generate a two-dimensional myocardial geometric mesh model; the two-dimensional myocardial geometric mesh model is represented as follows: ,in Represents a set of grid nodes. Represents the set of grid edges. This represents a set of grid cells.

3. The myocardial motion tracking method according to claim 2, characterized in that, In step S3, the prediction of the two-dimensional motion displacement of each grid node includes: Based on the spatial location of each grid node, local image regions centered on each grid node are extracted in the reference frame image and the target frame image, and convolutional neural networks are used to extract local image features. Local image features are fused with the spatial location information of grid nodes to obtain the node features of each grid node; The node features are input into the graph neural network, and the features are propagated according to the topological relationship corresponding to the grid edges to obtain the updated node features. The updated node features are input into the displacement prediction network to obtain the two-dimensional motion displacement corresponding to each grid node.

4. The myocardial motion tracking method according to claim 3, characterized in that, In step S4, generating the initial dense motion field includes: Determine the grid cell containing the pixel to be interpolated within the corresponding region of the myocardial region segmentation mask; Calculate the centroid coordinates of the pixel to be interpolated relative to the three grid nodes in the grid cell; Based on the centroid coordinates, the two-dimensional motion displacements corresponding to the three grid nodes in the grid cell are weighted and summed to obtain the motion displacement at the pixel to be interpolated. The initial dense motion field is formed by the motion displacements at each pixel to be interpolated within the corresponding region of the myocardial region segmentation mask.

5. The myocardial motion tracking method according to claim 4, characterized in that, In step S6, the residual correction model includes a lightweight convolutional feature extraction module and an implicit neural network, and its construction and prediction process includes: The coordinates of each pixel in the initial dense motion field are encoded to obtain positional features. The positional features, local image features, motion displacement at the corresponding position in the initial dense motion field, and image feedback residuals are fused to obtain the input features of the implicit neural network. The input features are fed into an implicit neural network composed of a multilayer perceptron to predict the local residual correction amount corresponding to each pixel position.

6. The myocardial motion tracking method according to claim 5, characterized in that, The displacement prediction network in S3 and the residual correction model in S6 are jointly trained. Total loss function during joint training Represented as: ； in, This represents the mesh motion consistency loss used in the displacement prediction network output. This represents the residual correction loss used in the output of the residual correction model. This indicates the loss of motion consistency in the final dense motion field of the myocardium. , as well as These represent the weighting coefficients of the corresponding loss terms.

7. The myocardial motion tracking method according to claim 6, characterized in that: The mesh motion consistency loss is used to constrain the displacement prediction network, and it is expressed as follows: ； in, For the set of grid edges, and Adjacent grid nodes and The corresponding two-dimensional motion displacement; The residual correction loss is used to constrain the residual correction model, and it is expressed as follows: ； in, This is the set of pixel locations in the myocardial region. For pixels The local residual correction amount at the location.

8. The myocardial motion tracking method according to claim 6, characterized in that, The motion consistency loss includes image similarity loss, motion field smoothing constraint loss, and mesh deformation constraint loss: The image similarity loss is expressed as: ； in, For the target frame image, For reference frame image, For the final dense movement field of the myocardium, Indicated according to the sports field Motion compensation for the image. Represents the normalized cross-correlation coefficient; The motion field smoothing constraint loss is expressed as: ； in, pixel position The spatial gradient of the final dense motion field of the myocardium.

9. The myocardial motion tracking method according to claim 8, characterized in that, The mesh deformation constraint loss is expressed as: The mesh deformation constraint loss avoids mesh cell flipping based on the area of ​​the deformed triangular mesh cells: ； in, The grid cells represent the two-dimensional myocardial mesh model. This represents the directed area of ​​the triangular mesh cell after deformation.

Images