Myocardial tracking method and system in ultrasound myocardial perfusion video

Abstract

The invention discloses a myocardial tracking method and system in an ultrasonic myocardial perfusion video, and belongs to the technical field of medical image processing. Performing image preprocessing on the ultrasonic myocardial perfusion video frame by frame to obtain a preprocessed single-frame ultrasonic image; performing multi-scale feature extraction on the preprocessed single-frame ultrasonic image through a feature extraction module constructed on the basis of a Swin Transform model; inputting the multi-scale features output by the feature extraction module into a time sequence modeling and dynamic tracking module constructed based on a Swin Transform model according to a time sequence, and capturing a motion mode of the myocardium along with time; and performing subsequent processing and result optimization on the myocardial position estimation of each frame through a post-processing optimization module, and outputting a final tracking result. According to the myocardial tracking method and system in the ultrasonic myocardial perfusion video, high-precision and high-robustness myocardial tracking is realized through multi-scale feature extraction, time sequence information fusion and post-processing optimization.

Description

Technical Field

[0001] The present application belongs to the field of medical image processing technology, and specifically relates to a method and system for myocardial tracking in ultrasonic myocardial perfusion videos. Background Art

[0002] With the advancement of medical imaging technology and computer vision, ultrasound myocardial perfusion videos have become widely used in cardiovascular disease diagnosis. However, common ultrasound views such as the A2C apical two-chamber heart, the A3C apical three-chamber heart, and the A4C apical four-chamber heart suffer from high image noise, complex myocardial motion, and varying viewing angles, posing significant challenges to myocardial tracking.

[0003] The existing technology has the following defects and deficiencies:

[0004] 1. Insufficient local feature recognition: Existing convolutional neural network-based or manual feature extraction methods are unable to simultaneously take into account local details and global information, resulting in low myocardial tracking accuracy under complex motion conditions.

[0005] 2. Limited ability to capture cross-scale information: Existing methods are prone to feature loss or incomplete expression when processing myocardial features at different scales and views.

[0006] 3. Unsatisfactory noise resistance: Ultrasound images themselves contain a lot of noise and artifacts. Existing algorithms are not robust enough to noise, which may affect the final tracking effect. Summary of the Invention

[0007] In order to solve the technical defects and problems existing in the above-mentioned existing technologies, the present application provides a method and system for myocardial tracking in ultrasonic myocardial perfusion videos. Through local window self-attention and inter-layer feature fusion, it can simultaneously obtain the detailed texture and global semantic information of the myocardium, and realize the complete capture of multi-scale and multi-view myocardial features.

[0008] The technical solution is as follows:

[0009] In one aspect, a method for myocardial tracking in ultrasonic myocardial perfusion video is provided, comprising:

[0010] Perform image preprocessing on ultrasonic myocardial perfusion video frame by frame to obtain preprocessed single-frame ultrasonic images;

[0011] The feature extraction module based on the Swin Transformer model is used to extract multi-scale features from the pre-processed single-frame ultrasound image, and a hierarchical feature map containing multi-scale features is output.

[0012] The multi-scale features output by the feature extraction module are sequentially input into the time series modeling and dynamic tracking module based on the Swin Transformer model. This module captures the myocardial motion pattern over time, generates smooth myocardial motion and contrast agent perfusion trajectories, and outputs an estimated myocardial position for each frame. Through local window self-attention and inter-layer feature fusion, it simultaneously captures myocardial texture details and global semantic information, achieving complete capture of multi-scale, multi-view myocardial features.

[0013] The post-processing optimization module performs subsequent processing and result optimization on the myocardial position estimation of each frame and outputs the final tracking result.

[0014] Furthermore, the ultrasonic myocardial perfusion video is subjected to image preprocessing frame by frame, including:

[0015] Adaptive median filtering is used to suppress speckle noise in ultrasound images while preserving myocardial edge details;

[0016] Adaptive histogram equalization is used to enhance the contrast between the myocardium and the background;

[0017] Normalize the processed image.

[0018] Furthermore, the feature extraction module constructed based on the Swin Transformer model performs multi-scale feature extraction on the pre-processed single-frame ultrasound image, including:

[0019] A single-frame ultrasound image is divided into several local windows of fixed size, and local features are extracted in each window through a multi-head self-attention mechanism:

[0020] The pixels in the window are linearly mapped to generate query vectors, key vectors, and value vectors. After multi-head attention calculation, the local feature representation of the window is obtained.

[0021] Inter-layer feature fusion is performed through the multi-scale attention module. Each layer merges 2×2 adjacent window features, doubles the number of channels, and outputs a hierarchical feature map containing multi-scale features.

[0022] Furthermore, the Swin Transformer model adopts a hierarchical structure and gradually merges adjacent patches through the Patch Merging module. The Patch Merging module merges adjacent patches and 2 adjacent frames in the time dimension into a new patch, while doubling the number of channels.

[0023] Furthermore, the Swin Transformer model adopts a shift window mechanism to achieve information interaction between windows through cyclic shift operations;

[0024] On even layers, move the window horizontally and vertically. pixels, synchronous movement in the time dimension frame,

[0025]

[0026] Among them, X window is the original window feature, X shifted is the shifted window feature, ΔM is the number of pixels the window moves, ΔT is the number of frames the window moves, and Shift represents the shift operation.

[0027] Furthermore, the feature extraction module adopts relative position encoding to retain spatial structure information.

[0028] Furthermore, the timing modeling and dynamic tracking module integrates the Transformer timing encoder architecture.

[0029] Furthermore, the multi-scale features output by the feature extraction module are input into the time series modeling and dynamic tracking module in chronological order to capture the myocardial motion pattern over time, including:

[0030] Adopting an adaptive error correction strategy: when a significant difference is detected between the current frame prediction and the actual observation, the current position is corrected by comparing the feature matching results of the previous and next frames;

[0031] Output accurate myocardial position estimation for each frame.

[0032] Furthermore, the post-processing optimization module performs subsequent processing on the myocardial position estimation of each frame, including:

[0033] Gaussian filtering is used to remove high-frequency jitter in the trajectory;

[0034] Morphological operations are used to modify the myocardial edge: appropriate structural elements are used to fill in the breaks in the edge or to delete isolated noise points;

[0035] Combining the tracking history trajectory and feature similarity, the confidence of each frame estimation is evaluated: if the myocardial edge or motion in a certain frame is incoherent, the confidence will be significantly reduced, and the low-confidence results can be re-estimated or eliminated accordingly.

[0036] In another aspect, a myocardial tracking system in ultrasonic myocardial perfusion video is provided, which adopts the above-mentioned myocardial tracking method in ultrasonic myocardial perfusion video, comprising:

[0037] An image preprocessing module is used to perform image preprocessing on the ultrasonic myocardial perfusion video frame by frame to obtain a preprocessed single-frame ultrasonic image;

[0038] A feature extraction module is used to extract multi-scale features from the preprocessed single-frame ultrasound image and output a hierarchical feature map containing multi-scale features;

[0039] The temporal modeling and dynamic tracking module is used to input the multi-scale features output by the feature extraction module into the temporal coding network in chronological order, capture the myocardial motion pattern over time, generate smooth myocardial motion and contrast agent perfusion trajectories, and output the myocardial position estimation for each frame;

[0040] The post-processing optimization module performs subsequent processing and result optimization on the myocardial position estimation of each frame and outputs the final tracking result.

[0041] The technical solution includes at least the following technical effects:

[0042] (1) High-precision tracking: A time series modeling and dynamic tracking module was designed. A time series transformer was used to model continuous frame features and a smoothing algorithm was used to predict motion trajectories. In particular, an adaptive error correction mechanism was introduced to automatically correct tracking deviations through inter-frame feature matching, achieving continuous and robust real-time myocardial tracking.

[0043] (2) Multi-scale feature extraction: Through a hierarchical structure and local window self-attention mechanism, myocardial features at different scales can be captured, which is applicable to various ultrasound views.

[0044] (3) Excellent robustness: Image pre-processing and post-processing modules work together to reduce the impact of noise and ensure that myocardial motion can still be accurately tracked in low-quality images.

[0045] (4) Real-time: The time series modeling module supports rapid processing of continuous frame data to meet the needs of clinical real-time monitoring and auxiliary diagnosis.

[0046] (5) Adaptive error correction: Error correction is introduced so that the model can automatically correct itself when local errors occur, further improving tracking accuracy.

[0047] (6) Completely identify and extract myocardial perfusion information throughout the entire cardiac cycle.

[0048] It should be understood that the foregoing general description and the following detailed description are merely illustrative and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0050] Figure 1 A flowchart of a method for myocardial tracking in ultrasonic myocardial perfusion video provided in a preferred embodiment of the present application;

[0051] Figure 2 A preferred embodiment of the present application is a structural diagram of a feature extraction module based on Swin Transformer;

[0052] Figure 3 A structural diagram of a myocardial tracking system in ultrasonic myocardial perfusion video provided in a preferred embodiment of the present application;

[0053] Figure 4 This is a comparison chart of the visualization results of A2C, A3C, and A4C. The green lines are marked by experienced doctors, and the red lines are the model tracking results.

[0054] Figure 5 The following is a comparison chart of the visualization results of A2C, A3C, and A4C. DETAILED DESCRIPTION

[0055] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0056] Related terms explanation:

[0057] 1. Swin Transformer: Window sliding self-attention model.

[0058] 2. Transformer: Self-attention model.

[0059] 2. Query vector: query vector.

[0060] 3. Key vector: key vector.

[0061] 4. Value vector: value vector.

[0062] 5. Patch merging module: image block merging module.

[0063] 6. Patch: image block.

[0064] like Figure 1-Figure 3As shown, this application provides a method and system for myocardial tracking in ultrasonic myocardial perfusion videos. The feature extraction method based on the Swin Transformer model aims to extract features of medical images from different scales and ranges, and is effectively used for myocardial tracking in ultrasonic myocardial perfusion videos (including A2C apical two-chamber heart, A3C apical three-chamber heart, A4C apical four-chamber heart, etc.). Through local window self-attention and inter-layer feature fusion, it is possible to simultaneously obtain the detailed texture and global semantic information of the myocardium, achieving complete capture of multi-scale and multi-view myocardial features.

[0065] Example 1, as Figure 1 and Figure 2 As shown, a method for myocardial tracking in ultrasonic myocardial perfusion video is provided, and the specific steps are as follows:

[0066] Step S1: performing image preprocessing on the ultrasonic myocardial perfusion video frame by frame to obtain a preprocessed single-frame ultrasonic image;

[0067] The input ultrasonic myocardial perfusion video is preprocessed frame by frame to highlight the myocardial structural features and suppress noise interference.

[0068] First, adaptive median filtering is used to denoise the speckle noise unique to ultrasound images: by calculating the median value in the local neighborhood of each pixel and replacing the original value, the noise is effectively filtered out while retaining the myocardial edge details.

[0069] Adaptive histogram equalization is then used to enhance the contrast between the myocardium and the background: this method divides the image into small areas and performs independent histogram equalization, highlighting the myocardial structure while maintaining local contrast and avoiding excessive amplification of noise.

[0070] Finally, the processed image is normalized to ensure the consistency of subsequent network input and accelerate convergence.

[0071] Step S2: Multi-scale feature extraction is performed on the preprocessed single-frame ultrasound image through a feature extraction module constructed based on the Swin Transformer model (window sliding self-attention model), and a hierarchical feature map containing multi-scale features is output.

[0072] The feature extraction module uses the independently designed multi-layer Swin Transformer model (window sliding self-attention model) to perform multi-scale feature extraction on the preprocessed single-frame image.

[0073] The input image is first divided into several fixed-size local windows, such as 8×8 or 16×16 pixels. Within each window, fine-grained features are extracted using a multi-head self-attention mechanism. Pixels within the window are linearly mapped to generate a query vector, a key vector, and a value vector. After multi-head attention calculations, the local feature representation of the window is obtained, which is then fused into the final feature output through a feedforward network. After parallel calculations across all windows, the first-layer local feature map is generated. These features are then fed into a hierarchical, multi-layer Transformer structure. As the layers deepen, the feature maps are progressively downsampled and the number of channels is increased, enabling cross-scale fusion.

[0074] The feature extraction module also incorporates relative position encoding, which takes into account the relative position offset between pixels within the window when calculating self-attention. This preserves spatial structural information and ensures accurate identification of myocardial regions under different ultrasound views and myocardial motion conditions. The output of the entire module is a hierarchical feature map containing multi-scale features. Each layer encodes myocardial structural information at a different downsampling ratio, providing rich input for the temporal modeling module. The specific steps are as follows:

[0075] Ultrasound myocardial perfusion video Input the model designed in this application, where H is the height of the input video, W is the width of the input video, C is the number of channels of the input video, and T is the number of time frames, which is set to 16 to represent the continuous frame sequence within the cardiac cycle. In the patch division stage, each time frame is independently divided into non-overlapping patches with a patch size of (P×P). Through the convolution operation with a 3D convolution kernel size of (P×P×1), the spatiotemporal patch is mapped to a low-dimensional space to obtain the spatiotemporal embedding vector Right now:

[0076]

[0077] in, It represents a 3D convolution using a (P×P) convolution kernel in the spatial dimension and a single frame in the temporal dimension, with the number of output channels being C0.

[0078] In this embodiment, the Swin Transformer model adopts a shift window (SW-MSA) mechanism to solve the problem of lack of cross-window connections. The shift window mechanism realizes information exchange between windows through cyclic shift operations.

[0079] On even layers, move the window horizontally and vertically. pixels, synchronous movement in the time dimension frame,

[0080]

[0081] Among them, X window is the original window feature, X shifted is the shifted window feature, ΔM is the number of pixels the window moves, ΔT is the number of frames the window moves, and Shift represents the shift operation.

[0082] When calculating attention, it is necessary to fill the boundary pixels to keep the window size consistent, so as to achieve cross-window information exchange. For a window feature, when calculating self-attention, it is first projected into the Query (Q), Key (K), and Value (V) matrices, namely:

[0083] Q=W q X shifted

[0084] K=W k X shifted

[0085] V=W v X shifted

[0086] Among them, W q 、W k 、W v is the projection matrix.

[0087] The calculation of self-attention is:

[0088]

[0089] Among them, d k are the dimensions of Q and K, and PosEmb is the spatiotemporal position encoding, which includes the spatial 2D position and time frame index.

[0090] The Swin Transformer model (windowed sliding self-attention model) uses a hierarchical structure and gradually merges adjacent patches through the PatchMerging module. Taking layers l to l+1 as an example, the Patch Merging module merges 2×2 adjacent patches and 2 adjacent frames in the temporal dimension into a new patch, doubling the number of channels.

[0091] Assume that the feature map of layer l After Patch Merging, the l+1th layer feature map The calculation process is: X l+1 =PatchMerging(X l ).

[0092] Each stage consists of multiple consecutive Swin Transformer blocks. As the network layer deepens, the receptive field gradually expands, and features of different scales can be extracted.

[0093] The feature extraction module uses the attention mechanism, specifically:

[0094] The spatial attention (SA) module obtains spatial attention weights through adaptive maximum pooling and adaptive average pooling across channels. Perform adaptive maximum pooling MaxPool(X) and adaptive average pooling AvgPool(X) respectively to obtain two feature maps M and A. Concatenate them in the channel dimension, that is: Then through the convolution layer Conv 3×3,1 (Concat(M,A)) gets the spatial attention weight

[0095] The temporal attention (TA) module draws on the “squeeze-excitation” mechanism of channel attention and designs a temporal attention branch to transform the feature map Compressed into a time vector, Through two layers of fully connected layers (reduction ratio r) and Sigmoid activation function, the temporal attention weight is obtained:

[0096] TemporalAttention=Sigmoid(F2(ReLU(F1(t)))),

[0097] Among them, Sigmoid and Relu are both activation functions, H is the height of the input video, W is the width of the input video, T is the number of time frames, (F1:C→C / r), (F2:C / r→C) means that the number of channels in the F1 layer model changes from C to C / r, and the number of channels in the F2 layer model changes from C / r to C.

[0098] The feature extraction module performs inter-layer feature fusion through a multi-scale attention module. Each layer merges 2×2 adjacent window features, doubles the number of channels, and outputs a hierarchical feature map containing multi-scale features.

[0099] The fusion strategy uses spatiotemporal attention product weighting:

[0100]

[0101] Among them, X i Represents the spatial attention mechanism of the feature map at layer i (Spatial Attention), the temporal attention mechanism of the feature map at layer i (Temporal Attention), the feature map at layer i (X i), ⊙ represents the multiplication of the corresponding dimensions. The spatiotemporal three-dimensional attention weights are used to guide the network to focus on the spatiotemporally consistent regions of interest (such as myocardial motion trajectories), thereby enhancing the discrimination of dynamic semantic features.

[0102] Compared with traditional convolutional networks or single pyramid feature extraction methods, the feature extraction module constructed based on the SwinTransformer model (window sliding self-attention model) adopted in this application has significant advantages in feature expression. The traditional CNN convolutional neural network has a limited receptive field and can only capture local information, while the Transformer modeling based on self-attention has long-range dependencies and is suitable for capturing the global morphology of the myocardial area; at the same time, this module uses a local window mechanism to reduce computational complexity and focus on local details, which is more efficient than global self-attention. The hierarchical structure realizes cross-scale information transmission, enabling the network to extract low-level edges and textures, and encode high-level semantic context, solving the problem of feature loss in multiple views and multiple scales. In addition, the introduction of specially designed position encoding enhances the model's perception of spatial structure, and has better perspective adaptability than existing feature extraction methods without position information. In summary, the feature extraction module of this technical solution can more completely obtain subtle changes in multiple views of the myocardium and improve subsequent tracking accuracy.

[0103] Step S3: The multi-scale features output by the feature extraction module are sequentially input into the temporal modeling and dynamic tracking module based on the SwinTransformer model (window sliding self-attention model). This module captures the myocardial motion pattern over time, generates smooth myocardial motion and contrast agent perfusion trajectories, and outputs a myocardial position estimate for each frame. Through local window self-attention and inter-layer feature fusion, both detailed myocardial texture and global semantic information are captured simultaneously, achieving complete capture of multi-scale, multi-view myocardial features.

[0104] The temporal modeling and dynamic tracking module constructs temporal features for continuous frame sequences and enables real-time tracking. First, the multi-scale features output by the feature extraction module are sequentially input into the temporal encoding network. The temporal encoding network uses a multi-layered Swin Transformer model to capture myocardial motion patterns over time. The temporal encoding network uses a self-attention layer to correlate features across frames in the temporal dimension, extracting patterns in myocardial motion and outputting a temporal contextual representation of each frame.

[0105] The time series modeling and dynamic tracking module employs an adaptive error correction strategy to address positioning deviations caused by occlusion or noise. When a significant difference is detected between the current frame prediction and the previous prediction, the current position is corrected by comparing the feature matching results of the previous and subsequent frames to prevent error accumulation. Ultimately, the system outputs a precise myocardial position estimate for each frame, enabling continuous, real-time myocardial tracking.

[0106] Step S4: The post-processing optimization module performs subsequent processing and optimization on the myocardial position estimation of each frame, outputting the final tracking result. The dynamic tracking result is carefully optimized to improve the consistency and accuracy of the tracking curve.

[0107] First, a Gaussian filter is used to remove high-frequency jitter from the trajectory. Morphological operations (such as dilation, erosion, and opening and closing operations) are then applied to correct the myocardial edges. Appropriate structural elements are used to fill in any gaps in the edges or remove isolated noise points, ensuring a closed and smooth myocardial contour. Finally, a confidence assessment is performed on each frame's estimate, combining historical tracking trajectories and feature similarity. If the myocardial edge or motion in a frame is inconsistent, the confidence level will be significantly reduced, and low-confidence results can be re-estimated or eliminated accordingly. These subsequent processing steps ensure that the output tracking curve is anatomically coherent and smooth, while also filtering out abnormal results, improving the stability and reliability of the final myocardial motion curve.

[0108] Example 2, as Figure 3 As shown, a myocardial tracking system in ultrasonic myocardial perfusion video is provided, which adopts the myocardial tracking method in ultrasonic myocardial perfusion video in embodiment 1, including:

[0109] An image preprocessing module is used to perform image preprocessing on the ultrasonic myocardial perfusion video frame by frame to obtain a preprocessed single-frame ultrasonic image;

[0110] A feature extraction module is used to extract multi-scale features from the preprocessed single-frame ultrasound image and output a hierarchical feature map containing multi-scale features;

[0111] The temporal modeling and dynamic tracking module is used to input the multi-scale features output by the feature extraction module into the temporal coding network in chronological order, capture the myocardial motion pattern over time, generate smooth myocardial motion and contrast agent perfusion trajectories, and output the myocardial position estimation for each frame;

[0112] The post-processing optimization module performs subsequent processing and result optimization on the myocardial position estimation of each frame and outputs the final tracking result.

[0113] The performance of four different models on four validation sets (external validation set 1, external validation set 2, external validation set 3, and test set) is evaluated using IoU as the evaluation metric, as shown in Table 1.

[0114] Table 1

[0115]

[0116]

[0117] In comparative tests against several models, as shown in Table 1, the Swin Transformer model using this embodiment demonstrates significant advantages in myocardial tracking accuracy, as measured by the Intersection over Union (IoU) metric. This model achieved IoU values ​​of 0.7500, 0.7625, 0.8032, and 0.8229 on the test set and external validation sets 1-3, respectively. These performances are significantly higher than those achieved by traditional convolutional-based Unet and Unet++. In particular, on external validation set 3, the model achieved a 1.7% higher IoU than Unet++ and a 24.4% higher IoU than Unet. Furthermore, these performances also surpassed those achieved by the Transformer-based VIT and the popular large-scale segmentation model SAM2. Furthermore, ablation experiments revealed that removing the multi-scale attention module reduced the average IoU by 5-7%, demonstrating the critical role of the multi-scale feature fusion strategy in extracting myocardial features. Furthermore, the model's high IoU values ​​on the external validation set demonstrate its superior generalization across diverse ultrasound views and noise distributions, addressing the inadequate cross-scale information capture inherent in existing techniques.

[0118] Judging from the visualization results of A2C, A3C, and A4C, the tracking effect of this model is better than that of existing models in terms of detail fit and noise resistance. Figure 4 The middle red line represents the model's predicted myocardial contour, while the green line represents the standard used by experienced physicians. The model's red tracking line closely matches the green gold standard, especially at subtle edges like the apex. Unet++, on the other hand, exhibits jagged edges, and Unet+VIT misidentifies invasion into non-myocardial areas in areas with dense noise. Figure 5 Although SAM2 can track the position of the myocardium relatively well, it is not effective in tracking the contour of the myocardial edge.

[0119] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0120] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0121] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for myocardial tracking in ultrasonic myocardial perfusion video, characterized in that: include: Perform image preprocessing on ultrasonic myocardial perfusion video frame by frame to obtain preprocessed single-frame ultrasonic images; The feature extraction module based on the Swin Transformer model is used to extract multi-scale features from the pre-processed single-frame ultrasound image, and a hierarchical feature map containing multi-scale features is output. The multi-scale features output by the feature extraction module are input into the time series modeling and dynamic tracking module based on the Swin Transformer model in chronological order. This module captures the myocardial motion pattern over time, generates smooth myocardial motion and contrast agent perfusion trajectories, and outputs the myocardial position estimation for each frame. The post-processing optimization module performs subsequent processing and result optimization on the myocardial position estimation of each frame and outputs the final tracking result.

2. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 1, characterized in that: The ultrasonic myocardial perfusion video is subjected to image preprocessing frame by frame, including: Adaptive median filtering is used to suppress speckle noise in ultrasound images while preserving myocardial edge details; Adaptive histogram equalization is used to enhance the contrast between the myocardium and the background; Normalize the processed image.

3. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 1, characterized in that: The feature extraction module constructed based on the Swin Transformer model performs multi-scale feature extraction on the pre-processed single-frame ultrasound image, including: A single-frame ultrasound image is divided into several local windows of fixed size, and local features are extracted in each window through a multi-head self-attention mechanism: The pixels in the window are linearly mapped to generate query vectors, key vectors, and value vectors. After multi-head attention calculation, the local feature representation of the window is obtained. Inter-layer feature fusion is performed through the multi-scale attention module. Each layer merges 2×2 adjacent window features, doubles the number of channels, and outputs a hierarchical feature map containing multi-scale features.

4. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 3, characterized in that: The SwinTransformer model adopts a hierarchical structure and gradually merges adjacent patches through the Patch Merging module. The PatchMerging module merges adjacent patches and 2 adjacent frames in the time dimension into a new patch, while doubling the number of channels.

5. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 3, characterized in that: The SwinTransformer model adopts a shift window mechanism to achieve information interaction between windows through cyclic shift operations; On even layers, move the window horizontally and vertically. pixels, synchronous movement in the time dimension frame, Among them, X window is the original window feature, X shifted is the shifted window feature, ΔM is the number of pixels the window moves, ΔT is the number of frames the window moves, and Shift represents the shift operation.

6. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 3, characterized in that: The feature extraction module adopts relative position encoding to retain spatial structure information.

7. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 1, characterized in that: The timing modeling and dynamic tracking module integrates the Transformer timing encoder architecture.

8. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 7, characterized in that: The multi-scale features output by the feature extraction module are input into the time series modeling and dynamic tracking module in chronological order to capture the myocardial motion pattern over time, including: Adopting an adaptive error correction strategy: when a significant difference is detected between the current frame prediction and the actual observation, the current position is corrected by comparing the feature matching results of the previous and next frames; Output accurate myocardial position estimation for each frame.

9. The method for myocardial tracking in ultrasonic myocardial perfusion video according to claim 1, characterized in that: The post-processing optimization module performs subsequent processing on the myocardial position estimation of each frame, including: Gaussian filtering is used to remove high-frequency jitter in the trajectory; Morphological operations are used to modify the myocardial edge, filling the breaks in the edge or deleting isolated noise points through appropriate structural elements; Combining the tracking history trajectory and feature similarity, the confidence of each frame estimation is evaluated: if the myocardial edge or motion in a certain frame is incoherent, the confidence will be significantly reduced, and the low-confidence results can be re-estimated or eliminated accordingly.

10. A myocardial tracking system in ultrasonic myocardial perfusion video, using the myocardial tracking method in ultrasonic myocardial perfusion video according to any one of claims 1 to 9, characterized in that: include: An image preprocessing module is used to perform image preprocessing on the ultrasonic myocardial perfusion video frame by frame to obtain a preprocessed single-frame ultrasonic image; A feature extraction module is used to extract multi-scale features from the preprocessed single-frame ultrasound image and output a hierarchical feature map containing multi-scale features; The temporal modeling and dynamic tracking module is used to input the multi-scale features output by the feature extraction module into the temporal coding network in chronological order, capture the myocardial motion pattern over time, generate smooth myocardial motion and contrast agent perfusion trajectories, and output the myocardial position estimation for each frame; The post-processing optimization module performs subsequent processing and result optimization on the myocardial position estimation of each frame and outputs the final tracking result.

Images