Segmental wall motion analysis method and system based on left ventricular opacification

Abstract

This invention belongs to the field of artificial intelligence cardiac motion analysis technology, and relates to a method and system for segmental ventricular wall motion analysis based on left ventricular opacity. The method includes: segmenting the left ventricular contour using a segmentation network in four-chamber, two-chamber, and three-chamber sections at the apex of an LVO video sequence; dividing the left ventricular wall into several segments and extracting node features to obtain a sample set; constructing three identical spatiotemporal graph convolutional networks; and inputting the node features into classifiers of the three spatiotemporal graph convolutional networks to detect the segmental ventricular wall motion states related to three coronary artery branches. This invention utilizes a segmentation network to segment the left ventricular contour, dividing the left ventricular wall into several segments, greatly reducing the sample size for model training; it uses GCN to fuse and model multiple sections, combining three-dimensional myocardial topology, and considering the correspondence between each coronary artery branch and myocardial segment, making the classification of ventricular wall motion more detailed.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence cardiac motion analysis technology, specifically, it relates to a method and system for segmental ventricular wall motion analysis based on left ventricular turbidity. Background Technology

[0002] Coronary artery disease (CAD) refers to heart disease caused by myocardial ischemia and hypoxia due to coronary atherosclerosis, seriously affecting human health. The branches of the coronary arteries supply blood to different areas of the myocardium, ensuring its motor function. If some branches become narrowed or blocked, the myocardial areas they control will experience ischemia and even reduced mobility, a condition known as segmental wall motion abnormalities. In clinical practice, assessing left ventricular motion function, especially segmental wall motion, is becoming increasingly important in determining the severity and prognosis of CAD.

[0003] Echocardiography, with its high spatiotemporal resolution, is an ideal non-invasive method for assessing ventricular wall motion. Left ventricular opacification (LVO) using intravenous contrast agents provides even clearer visualization of cardiac structures, further improving the accuracy of wall motion analysis. However, traditional wall motion analysis relies on manual observation and measurement of endocardial displacement and myocardial thickening, depending on the sonographer's subjective experience, resulting in low reproducibility.

[0004] Most current deep learning methods for automatically analyzing myocardial function are based on quantitative measurements, typically measuring ejection fraction and strain based on automatic left ventricular segmentation. Existing methods use convolutional neural networks to directly classify segmental wall motion qualitatively, but do not perform left ventricular segmentation. Automatically measured myocardial function indicators can only reflect wall motion to a certain extent and are used for auxiliary diagnosis, and are highly dependent on the accuracy of left ventricular segmentation. Due to issues such as poor quality and strong interference in ultrasound images, existing deep learning image segmentation methods show unstable performance in echocardiography, especially contrast echocardiography. Measurement methods that do not rely on left ventricular segmentation suffer from a large amount of redundant information outside the left ventricular region in ultrasound images. Without extracting the region of interest (i.e., left ventricular image segmentation), more data is needed to accurately train the network; however, large-scale and diverse samples are difficult to obtain clinically. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for segmental wall motion analysis based on left ventricular turbidity.

[0006] In a first aspect, the present invention provides a method for segmental wall motion analysis based on left ventricular turbidity, comprising:

[0007] Dilated convolution is introduced into residual convolution, and a segmentation network is built based on U-Net and Transformer;

[0008] The left ventricular contour was segmented using a segmentation network in the four-chamber, two-chamber, and three-chamber views of the apex of LVO video sequences.

[0009] Based on the segmented left ventricular contour, the left ventricular wall is divided into several segments. The center of each segment is taken as a node, and the two-dimensional coordinates are extracted as node features to obtain a sample set.

[0010] Three identical spatiotemporal graph convolutional networks are constructed. The spatiotemporal graph convolutional network consists of alternating multi-faceted spatial convolutions and multi-scale temporal convolutions, and is finally classified by a classifier. The multi-faceted spatial convolution includes several multi-head self-attention graph convolutional modules.

[0011] The node features of each node in the sample set are input into the classifiers of three spatiotemporal graph convolutional networks to detect the segmental wall motion state related to the left anterior descending branch, the left circumflex branch and the right coronary artery, respectively, and the trained spatiotemporal graph convolutional networks are obtained.

[0012] The segmental wall motion state detected was analyzed.

[0013] Secondly, the present invention provides a left ventricular segmental wall motion analysis system, comprising a first construction unit, a segmentation unit, a sample extraction unit, a second construction unit, a training unit, and a detection unit;

[0014] The first building unit is used to introduce dilated convolution into residual convolution and build a segmentation network based on U-Net and Transformer;

[0015] Segmentation unit, used to segment the left ventricular contour in the four-chamber, two-chamber and three-chamber sections of the apex of LVO video sequences using a segmentation network;

[0016] The sample extraction unit is used to divide the left ventricular wall into several segments, take the center of each segment as a node, extract two-dimensional coordinates as node features, and obtain a sample set.

[0017] The second building unit is used to construct three identical spatiotemporal graph convolutional networks. The spatiotemporal graph convolutional network consists of alternating multi-faceted spatial convolutions and multi-scale temporal convolutions, and is finally classified by a classifier. The multi-faceted spatial convolution includes several multi-head self-attention graph convolutional modules.

[0018] The training unit is used to input the node features of each node in the sample set into the classifiers of three spatiotemporal graph convolutional networks, respectively, to detect the segmental wall motion state related to the left anterior descending branch, the left circumflex branch and the right coronary artery, and to obtain the trained spatiotemporal graph convolutional network.

[0019] Based on the above technical solution, the present invention can be further improved as follows.

[0020] Furthermore, the segmentation network comprises a residual dilated convolution, a Transformer encoder, a Transformer decoder, a 2x2 max pooling layer, a 3x3 convolutional layer, an activation layer, and a 1x1 convolutional layer connected in sequence.

[0021] Furthermore, let Q be the query value, K be the key value, and d be the intermediate dimension. Let be the total number of extended heads in the multi-head self-attention map convolutional module, i be the number of extended heads in the multi-head self-attention map convolutional module, and A be the myocardial topology of the multi-head self-attention map convolutional module. Let be a hyperparameter, and let be the attention matrix of the i-th extended head. , For the query value of the i-th extended head number, Let be the key value of the i-th extended header, and T be the number of frames in a video sequence. Let Q and K be the activation functions. In the multi-head self-attention map convolution module, Q and K are obtained through convolution and average pooling over T.

[0022] .

[0023] Furthermore, let the node feature input of each node in the sample set be... The Value of the i-th extended head obtained through convolution is , The total number of extended heads in the multi-head self-attention map convolutional module. The first self-attention graph convolutional module of the multi-head system The transpose of the attention matrices of each extended head, and the output of the multi-head self-attention graph convolutional module is... ,but:

[0024] .

[0025] Furthermore, the first layer of multi-head self-attention map convolution module in the multi-faceted spatial convolution uses the two-dimensional left ventricular structural topology of each facet, and the second layer of multi-head self-attention map convolution module uses the three-dimensional global topology of the left ventricle; the correspondence between the three-dimensional global topology of the left ventricle and the 18 segments of the left ventricular wall under the short axis section is established.

[0026] Furthermore, the multi-scale temporal convolution includes four channels and a ReLU activation layer; the first channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 1, the second channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 2, the third channel includes a 1x1 convolution and a 4x1 average pooling layer, and the fourth channel includes a 1x1 convolution; the outputs of the four channels are respectively connected to the ReLU activation layer.

[0027] The beneficial effects of this invention are as follows: After refining the left ventricular wall into multiple segments, this invention uses spatiotemporal graph convolution for modeling and classification. Unlike methods that classify directly in a single cross-section without image segmentation, this invention uses a segmentation network to segment the left ventricular contour, dividing the left ventricular wall into several segments. It first extracts and refines the region of interest, greatly reducing the sample size for model training. By using GCN to fuse and model multiple cross-sections, it fully considers the correspondence between each coronary artery branch and myocardial segments, making the classification of ventricular wall motion more detailed. This facilitates targeted adjustments to the myocardial topology and improves the sensitivity of apical segment detection. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the segmental wall motion analysis method based on left ventricular turbidity provided in Embodiment 1 of the present invention;

[0029] Figure 2 This is a flowchart illustrating the segmental wall motion analysis method based on left ventricular turbidity according to the present invention.

[0030] Figure 3 This is a schematic diagram of the multi-faceted spatial convolution and MSAGC module;

[0031] Figure 4 This is a schematic diagram showing the distribution of the 18 segments of the left ventricular wall and its global topology.

[0032] Figure 5 This is a schematic diagram illustrating the principle of multi-scale temporal convolution.

[0033] Figure 6 This is a schematic diagram of the left ventricular segmental wall motion analysis system provided in Embodiment 2 of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0035] Example 1

[0036] As an example, see the attached document. Figure 1 As shown, to solve the above-mentioned technical problems, this embodiment provides a method for segmental wall motion analysis based on left ventricular turbidity, including:

[0037] Dilated convolution is introduced into residual convolution, and a segmentation network is built based on U-Net and Transformer;

[0038] The left ventricular contour was segmented using a segmentation network in the apical four-chamber (A4C), two-chamber (A2C), and three-chamber (A3C) views of the LVO video sequence.

[0039] Based on the segmented left ventricular contour, the left ventricular wall is divided into several segments. The center of each segment is taken as a node, and the two-dimensional coordinates are extracted as node features to obtain a sample set.

[0040] Three identical spatiotemporal graph convolutional networks are constructed. The spatiotemporal graph convolutional network consists of alternating multi-faceted spatial convolutions and multi-scale temporal convolutions, and is finally classified by a classifier. The multi-faceted spatial convolution includes several multi-head self-attention graph convolutional modules.

[0041] The node features of each node in the sample set are input into the classifiers of three spatiotemporal graph convolutional networks to detect the segmental wall motion state related to the left anterior descending branch, the left circumflex branch and the right coronary artery, respectively, and the trained spatiotemporal graph convolutional networks are obtained.

[0042] The segmental wall motion state detected was analyzed.

[0043] In practical applications, as shown in the appendix Figure 2As shown, the left ventricular contour is segmented in the four-chamber, two-chamber, and three-chamber views at the apex of the LVO video sequence. The left ventricular wall is divided into 18 segments (6 segments per view), and the center of each segment is taken as a node. These nodes are input into three Spatial Temporal Graph Convolution Network (ST-GCN) classifiers to detect segmental wall motion abnormalities related to the left anterior descending artery (LAD), left circumflex artery (LCX), and right coronary artery (RCA), respectively. Segmental wall motion states include normal and abnormal states; a decrease in the amplitude of segmental wall motion is considered an abnormal state. An improved image segmentation network based on U-Net and Transformer is used to segment the left ventricular contour in low-quality and highly interfering LVO angiography echocardiography.

[0044] This invention introduces dilated convolution into residual convolution, which improves the ability to extract edge features from complex textures in images. It uses an improved multi-head self-attention (MSA) mechanism to reduce time complexity and change the position encoding method, allowing the Transformer to be applied to shallower layers of the U-Net network, further improving the global consistency of segmentation.

[0045] The left ventricular wall was divided into 18 segments (6 segments per segment) across three sections. The center of each segment was taken as a node, and its two-dimensional coordinates were obtained as node features. Let T be the number of frames in a video sequence, N be the number of nodes in each frame (18), and C be the number of node feature channels (2), to obtain the sample. .

[0046] The three ST-GCNs have identical structures but are trained using different case samples. The training set for the RCA classifier includes standard samples and abnormal samples from the RCA segment; the training set for the LAD classifier includes standard samples and abnormal samples from the LAD segment; and the training set for the LCX classifier includes standard samples and abnormal samples from the LCX segment. Reduced amplitude of myocardial segment motion is considered an abnormal sample. Each ST-GCN classifier performs a binary classification function, diagnosing the normality of the three coronary artery branches separately. For example, each ST-GCN consists of alternating 10 multi-plane spatial convolutions and 10 multi-scale temporal convolutions, and is finally classified using a Softmax classifier.

[0047] Multi-faceted spatial convolution: as shown in the appendix Figure 3 As shown, Figure 3 In diagram (a), the Multi-Head Self-Attention Graph Convolution (MSAGC) module is shown. Q represents the query value, K represents the key value, and V represents the value of the extended head. The values ​​of Q, K, and V represent the query, key, and value of the Transform model, respectively. The query value of the first extended header. X:[T,N,h] represents the key value of the first extended head. Figure (b) shows a multi-faceted spatial convolution, which consists of several MSAGCs. Each MSAGC uses a different topology. X:[T,18,h] represents the input node features. X:[T,6,h] is split from X:[T,18,h]. X:[T,18,C'] represents the output node features.

[0048] Optionally, the segmentation network includes a residual dilated convolution, a Transformer encoder, a Transformer decoder, a 2x2 max pooling layer, a 3x3 convolutional layer, an activation layer, and a 1x1 convolutional layer connected in sequence.

[0049] Optionally, let Q be the query value, K be the key value, and d be the intermediate dimension. Let be the total number of extended heads in the multi-head self-attention map convolutional module, i be the number of extended heads in the multi-head self-attention map convolutional module, and A be the myocardial topology of the multi-head self-attention map convolutional module. Let be a hyperparameter, and let be the attention matrix of the i-th extended head. , For the query value of the i-th extended head number, Let be the key value of the i-th extended header, and T be the number of frames in a video sequence. Let Q and K be the activation functions. In the multi-head self-attention map convolution module, Q and K are obtained through convolution and average pooling over T.

[0050] .

[0051] Optionally, let the node feature input of each node in the sample set be... The Value of the i-th extended head obtained through convolution is , The total number of extended heads in the multi-head self-attention map convolutional module. The first self-attention graph convolutional module of the multi-head system The transpose of the attention matrices of each extended head, and the output of the multi-head self-attention graph convolutional module is... ,but:

[0052] .

[0053] Optionally, the first layer of multi-head self-attention map convolution module in the multi-faceted spatial convolution uses the two-dimensional left ventricular structural topology of each facet, and the second layer of multi-head self-attention map convolution module uses the three-dimensional global topology of the left ventricle; establish the correspondence between the three-dimensional global topology of the left ventricle and the 18 segments of the left ventricular wall under the short axis section.

[0054] As attached Figure 3 As shown in (b), the first MSAGC module in the multi-plane spatial convolution uses the two-dimensional left ventricular structural topology of each plane, while the second MSAGC module uses the three-dimensional global topology of the left ventricle. The correspondence between the structure and the 18 segments of the left ventricular wall in the short axis section is shown in the appendix. Figure 4 As shown. (Attached) Figure 4 (a) shows a schematic diagram of the 18 segments of the left ventricle. Figure 4 (b) shows the global topology of the 18 segments. The 16 nodes in the apical region form a complete graph because the apical region is mainly dominated by the LAD, and the segments are highly interconnected. In addition, motion abnormalities in the apical segment are smaller than in other parts. Using a complete graph can enhance information propagation between nodes and increase the model's sensitivity to motion abnormalities in the apical segment.

[0055] Optionally, the multi-scale temporal convolution includes four channels and a ReLU activation layer; the first channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 1, the second channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 2, the third channel includes a 1x1 convolution and a 4x1 average pooling layer, and the fourth channel includes a 1x1 convolution; the outputs of the four channels are respectively connected to the ReLU activation layer.

[0056] As attached Figure 5 As shown, multi-scale temporal convolution has four channels, extracting features at different scales along the time dimension T. The maximum receptive field is 9, close to half a heartbeat cycle. The large receptive field and average pooling operation effectively reduce the error caused by inaccurate segmentation of a single frame during the image segmentation stage.

[0057] This invention refines the left ventricular wall into 18 segments and then uses spatiotemporal graph convolution for modeling and classification. Unlike methods that classify directly on a single cross-section without image segmentation, this invention uses a segmentation network to segment the left ventricular contour, dividing the left ventricular wall into several segments. It first extracts and refines the region of interest, greatly reducing the sample size for model training. By using GCN to fuse and model multiple cross-sections, it fully considers the correspondence between each coronary artery branch and myocardial segments, making the classification of ventricular wall motion more detailed. This facilitates targeted adjustments to the myocardial topology and improves the sensitivity of apical segment detection.

[0058] Example 2

[0059] Based on the same principle as the method shown in Embodiment 1 of the present invention, as illustrated in the appendix. Figure 6 As shown, an embodiment of the present invention also provides a left ventricular segmental wall motion analysis system, including a first construction unit, a segmentation unit, a sample extraction unit, a second construction unit, a training unit, and a detection unit;

[0060] The first building unit is used to introduce dilated convolution into residual convolution and build a segmentation network based on U-Net and Transformer;

[0061] Segmentation unit, used to segment the left ventricular contour in the four-chamber, two-chamber and three-chamber sections of the apex of LVO video sequences using a segmentation network;

[0062] The sample extraction unit is used to divide the left ventricular wall into several segments, take the center of each segment as a node, extract two-dimensional coordinates as node features, and obtain a sample set.

[0063] The second building unit is used to construct three identical spatiotemporal graph convolutional networks. The spatiotemporal graph convolutional network consists of alternating multi-faceted spatial convolutions and multi-scale temporal convolutions, and is finally classified by a classifier. The multi-faceted spatial convolution includes several multi-head self-attention graph convolutional modules.

[0064] The training unit is used to input the node features of each node in the sample set into the classifiers of three spatiotemporal graph convolutional networks, respectively, to detect the segmental wall motion state related to the left anterior descending branch, the left circumflex branch and the right coronary artery, and to obtain the trained spatiotemporal graph convolutional network.

[0065] Optionally, the segmentation network includes a residual dilated convolution, a Transformer encoder, a Transformer decoder, a 2x2 max pooling layer, a 3x3 convolutional layer, an activation layer, and a 1x1 convolutional layer connected in sequence.

[0066] Optionally, let Q be the query value, K be the key value, and d be the intermediate dimension. Let be the total number of extended heads in the multi-head self-attention map convolutional module, i be the number of extended heads in the multi-head self-attention map convolutional module, and A be the myocardial topology of the multi-head self-attention map convolutional module. Let be a hyperparameter, and let be the attention matrix of the i-th extended head. , For the query value of the i-th extended head number, Let be the key value of the i-th extended header, and T be the number of frames in a video sequence. Let Q and K be the activation functions. In the multi-head self-attention map convolution module, Q and K are obtained through convolution and average pooling over T.

[0067] .

[0068] Optionally, let the node feature input of each node in the sample set be... The Value of the i-th extended head obtained through convolution is , The total number of extended heads in the multi-head self-attention map convolutional module. The first self-attention graph convolutional module of the multi-head system The transpose of the attention matrices of each extended head, and the output of the multi-head self-attention graph convolutional module is... ,but:

[0069] .

[0070] Optionally, the first layer of multi-head self-attention map convolution module in the multi-faceted spatial convolution uses the two-dimensional left ventricular structural topology of each facet, and the second layer of multi-head self-attention map convolution module uses the three-dimensional global topology of the left ventricle; establish the correspondence between the three-dimensional global topology of the left ventricle and the 18 segments of the left ventricular wall under the short axis section.

[0071] Optionally, the multi-scale temporal convolution includes four channels and a ReLU activation layer; the first channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 1, the second channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 2, the third channel includes a 1x1 convolution and a 4x1 average pooling layer, and the fourth channel includes a 1x1 convolution; the outputs of the four channels are respectively connected to the ReLU activation layer.

[0072] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for segmental wall motion analysis based on left ventricular opacity, characterized in that, include: Dilated convolution is introduced into residual convolution, and a segmentation network is built based on U-Net and Transformer; The segmentation network consists of a residual dilated convolution, a Transformer encoder, a Transformer decoder, a 2x2 max pooling layer, a 3x3 convolutional layer, an activation layer, and a 1x1 convolutional layer connected in sequence. The left ventricular contour was segmented using a segmentation network in the four-chamber, two-chamber, and three-chamber views of the apex of LVO video sequences. Based on the segmented left ventricular contour, the left ventricular wall is divided into several segments. The center of each segment is taken as a node, and the two-dimensional coordinates are extracted as node features to obtain a sample set. Three identical spatiotemporal graph convolutional networks are constructed. The spatiotemporal graph convolutional network consists of alternating multi-faceted spatial convolutions and multi-scale temporal convolutions, and is finally classified by a classifier. The multi-faceted spatial convolution includes several multi-head self-attention graph convolutional modules. The first layer of multi-head self-attention map convolution module in the multi-faceted spatial convolution uses the two-dimensional left ventricular structural topology of each facet, and the second layer of multi-head self-attention map convolution module uses the three-dimensional global topology of the left ventricle; establish the correspondence between the three-dimensional global topology of the left ventricle and the 18 segments of the left ventricular wall under the short axis section; The multi-scale temporal convolution consists of four channels and a ReLU activation layer; the first channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 1, the second channel includes a 1x1 convolution and a 5x1 convolution with a dilation rate of 2, the third channel includes a 1x1 convolution and a 4x1 average pooling layer, and the fourth channel includes a 1x1 convolution; the outputs of the four channels are connected to the ReLU activation layer respectively. The node features of each node in the sample set are input into the classifiers of three spatiotemporal graph convolutional networks to detect the segmental wall motion state related to the left anterior descending branch, the left circumflex branch and the right coronary artery, respectively, and the trained spatiotemporal graph convolutional networks are obtained. The segmental wall motion state detected was analyzed.

2. The method for segmental wall motion analysis based on left ventricular turbidity according to claim 1, characterized in that, Let Q be the query value, K be the key value, and d be the intermediate dimension. Let be the total number of extended heads in the multi-head self-attention map convolutional module, i be the number of extended heads in the multi-head self-attention map convolutional module, and A be the myocardial topology of the multi-head self-attention map convolutional module. Let be a hyperparameter, and let be the attention matrix of the i-th extended head. , For the query value of the i-th extended head number, Let be the key value of the i-th extended header, and T be the number of frames in a video sequence. Let Q and K be the activation function obtained through convolution and average pooling over T. 。 3. The method for segmental wall motion analysis based on left ventricular turbidity according to claim 1, characterized in that, Let the node features of each node in the sample set be input as follows: The Value of the i-th extended head obtained through convolution is , The total number of extended heads in the multi-head self-attention map convolutional module. The first self-attention graph convolutional module of the multi-head system The transpose of the attention matrices of each extended head, and the output of the multi-head self-attention graph convolutional module is... ,but: 。 4. A left ventricular segmental wall motion analysis system, wherein the left ventricular segmental wall motion analysis system is used to perform the segmental wall motion analysis method according to any one of claims 1-3, characterized in that, It includes a first building unit, a segmentation unit, a sample extraction unit, a second building unit, a training unit, and a detection unit; The first building unit is used to introduce dilated convolution into residual convolution and build a segmentation network based on U-Net and Transformer; Segmentation unit, used to segment the left ventricular contour in the four-chamber, two-chamber and three-chamber sections of the apex of LVO video sequences using a segmentation network; The sample extraction unit is used to divide the left ventricular wall into several segments, take the center of each segment as a node, extract two-dimensional coordinates as node features, and obtain a sample set. The second building unit is used to construct three identical spatiotemporal graph convolutional networks. The spatiotemporal graph convolutional network consists of alternating multi-faceted spatial convolutions and multi-scale temporal convolutions, and is finally classified by a classifier. The multi-faceted spatial convolution includes several multi-head self-attention graph convolutional modules. The training unit is used to input the node features of each node in the sample set into the classifiers of three spatiotemporal graph convolutional networks, respectively, to detect the segmental wall motion state related to the left anterior descending branch, the left circumflex branch and the right coronary artery, and to obtain the trained spatiotemporal graph convolutional network. The detection unit is used to analyze the detected segmental wall motion state.

Images