Non-invasive cardiac function evaluation system and method

Abstract

The invention relates to a non-invasive cardiac function evaluation system and method, and the system comprises an ultrasonic video collection module which is responsible for the collection of a cardiac image, and guarantees the collection of LVEDV and LVESV data. The preprocessing module locates a ventricle motion key time point, cuts a video and standardizes the image size. The feature extraction module extracts spatial and temporal features of the echocardiogram by using a C3D network. The multi-task learning network module optimizes heart function indexes such as EF, EDV and ESV through a shared feature extraction network and a task specific learning network, and controls the complexity of the model by applying L2 regularization. According to the non-invasive cardiac function evaluation system and method, space and time sequence characteristics of an echocardiography video are extracted through a C3D network with an attention mechanism; different weights are given to the sub-tasks, multi-task learning is carried out in a joint training mode, ejection fraction prediction is achieved, and doctors are assisted to carry out more accurate heart function evaluation.

Description

Technical Field

[0001] The present invention belongs to the field of medical technology, and in particular relates to a non-invasive cardiac function assessment system and method. Background Art

[0002] Traditional cardiac function assessment methods rely heavily on medical staff's professional interpretation of data from different sources such as electrocardiograms, echocardiograms, and hemodynamics. In addition, the cost of assessing cardiac function is high, time-consuming, harmful to the human body, and has low accuracy. In addition, traditional cardiac function assessment methods are easily limited by factors such as professional knowledge, high deployment costs, and complicated operations. Therefore, a more objective, efficient, and accurate cardiac function assessment method is urgently needed.

[0003] In recent years, with the rapid development of machine learning methods and deep learning technologies, cardiac function assessment methods based on deep learning and machine learning have been widely used in echocardiography assessment, such as the cardiac function assessment method and device with application number CN2014108539157, and the method and device for quantitatively assessing cardiac function based on electrocardiogram U wave and T wave with application number CN2020101504237. However, the current cardiac function assessment methods still have certain defects. For example, the effective information of the spatiotemporal feature map is easily lost in the feature extraction step, and there is a lack of mining the correlation between EF, EDV, and ESV.

[0004] The accuracy of cardiac function assessment is crucial for the prevention and control of cardiovascular diseases and the reduction of global mortality. In recent years, with the rapid development of machine learning and deep learning, the application of artificial intelligence technologies such as convolutional neural networks and multi-task learning has significantly improved the objectivity and accuracy of echocardiographic image assessment. However, traditional cardiac function assessment methods are easily limited by factors such as professional knowledge, high deployment costs, and complicated operations. In addition, cardiac ultrasound dynamic image processing methods are prone to losing effective information during feature extraction, and fail to fully explore the correlation between indicators such as ejection fraction (EF), end-diastolic volume (EDV), and end-systolic volume (ESV) in echocardiographic videos. Summary of the invention

[0005] In order to overcome the deficiencies and shortcomings of the prior art, the primary purpose of the present invention is to provide a non-invasive cardiac function assessment system and method, which extracts spatial and temporal features of ultrasonic cardiography videos through a C3D network with an attention mechanism; by assigning different weights to subtasks and using a joint training method for multi-task learning, ejection fraction prediction is achieved to assist doctors in performing more accurate cardiac function assessment.

[0006] The primary purpose of the present invention is achieved through the following technical solutions:

[0007] Non-invasive cardiac function assessment system, including:

[0008] Ultrasonic video acquisition module: used to acquire cardiac images through ultrasound. Each image is acquired at least twice, and at least three consecutive cardiac cycles are acquired each time. It is ensured that the acquired data includes the left ventricular end-diastolic volume LVEDV and the left ventricular end-systolic volume LVESV.

[0009] Ultrasound video preprocessing module: used to locate the time point T0 when ventricular relaxation begins, the time point T1 when the ventricular volume reaches the maximum and the ventricular contraction begins, and the time point T2 when the ventricular volume reaches the minimum and the ventricle begins to relax again, and to crop the video sequence according to the time points T1, T2, and T3, apply image segmentation technology to extract the heart structure, perform tracking processing, and use cubic interpolation method to standardize images of different sizes into ultrasound dynamic image sequences of uniform size;

[0010] Ultrasound dynamic image feature extraction module: It is used to extract spatiotemporal features of echocardiogram sequences, including heart movement, volume changes, and blood flow, by integrating the 3D convolutional neural network of the C3D structure with the spatial feature attention mechanism and the temporal feature attention mechanism, and to capture the temporal dynamic relationship and spatial features;

[0011] Multi-task learning network module: used to extract shared features in the time domain and spatial domain through the shared feature extraction network, design an independent learning network for each subtask through task segmentation, set task-specific weights, and optimize through a multi-task loss function. The loss function introduces L2 regularization to control the complexity of the model and calculates cardiac function indicators including ejection fraction EF, end-diastolic volume EDV, and end-systolic volume ESV.

[0012] Another technical problem to be solved by the present invention is to provide a non-invasive cardiac function assessment method, comprising the following steps:

[0013] Use ultrasound to collect cardiac images and obtain the left ventricular end-diastolic volume LVEDV and left ventricular end-systolic volume LVESV;

[0014] Locate the time point T0 when ventricular relaxation begins, the time point T1 when the ventricular volume reaches the maximum and the ventricle begins to contract, and the time point T2 when the ventricular volume reaches the minimum and the ventricle begins to relax again, crop the video sequence into diastolic and systolic cycles, exclude information irrelevant to the image content, separate the cardiac structure through image segmentation technology, and use cubic interpolation to standardize the image to a uniform size;

[0015] The spatial feature attention mechanism and the temporal feature attention mechanism are introduced into the 3D convolutional neural network. The 3D convolutional neural network with the C3D structure that integrates the spatial feature attention mechanism and the temporal feature attention mechanism is used to extract the spatiotemporal features of the standardized echocardiogram sequence to capture the changes in cardiac structure and function.

[0016] Based on the extracted features, a multi-task learning network is constructed. Task segmentation is performed through a shared feature extraction network. An independent learning network is designed for each subtask. Task-specific weights are set. The model is optimized using a multi-task loss function. L2 regularization is introduced to calculate the ejection fraction EF, end-diastolic volume EDV, and end-systolic volume ESV to provide a comprehensive cardiac function assessment.

[0017] Preferably, the method for obtaining the left ventricular end-diastolic volume LVEDV and the left ventricular end-systolic volume LVESV is:

[0018] Use ultrasound to collect cardiac images, collect each image at least twice, collect at least three consecutive cardiac cycles each time, and save the cardiac ultrasound dynamic image. During the image collection process, measure through the apical four-chamber heart section;

[0019] At the end of left ventricular diastole and end of left systole, the left ventricular border was outlined along the intimal line trajectory to obtain the left ventricular end-diastolic volume LVEDV and the left ventricular end-systolic volume LVESV.

[0020] Preferably, when using an ultrasound device to acquire cardiac images, for patients with arrhythmia, at least five consecutive cardiac cycles are acquired each time.

[0021] Preferably, the method of normalizing the image to a uniform size using cubic interpolation is:

[0022] Mark the time point T0 when ventricular relaxation begins, the time point T1 when the ventricular volume reaches its maximum and the ventricle begins to contract, and the time point T2 when the ventricular volume reaches its minimum and the ventricle begins to relax again, so as to define the diastolic cycle [T0, T1] and the systolic cycle [T1, T2];

[0023] The cropped image sequence was masked to exclude text, electrocardiogram, and spirometer information; the cardiac structure was separated from the background using image segmentation technology, and the position and shape changes of the cardiac structure at different time points were tracked using tracking technology;

[0024] The ultrasound images of different sizes and resolutions were normalized into ultrasound moving image sequences of uniform size W×H using cubic interpolation.

[0025] Preferably, the method of introducing the spatial feature attention mechanism in the 3D convolutional neural network is:

[0026] The spatial attention function f is calculated through the spatial convolution layer and the global average pooling operation s (X), generate weights for each pixel position, and then these weights A s It is applied to the original feature map X to achieve element-by-element weighted adjustment, namely:

[0027]

[0028] Among them, A s is the spatial attention weight, f s (X) is the output feature of the spatial convolutional layer, d s is the characteristic dimension of the space dimension.

[0029] Preferably, the method of introducing the temporal feature attention mechanism in the 3D convolutional neural network is:

[0030] By calculating the temporal attention weight A t Processing echocardiogram sequences, the temporal attention function f is calculated through temporal convolution layers and global average pooling t (X), generate the weights for each time step, and add these weights A t Application to echocardiographic sequence X t , weighted adjustment is performed on the features of each time step, namely:

[0031]

[0032] Among them, A t is the temporal attention weight, f t (X t ) is the output feature of the temporal convolutional layer, d t It is the characteristic dimension of the time dimension.

[0033] As a preferred method, the method of constructing a multi-task learning network is:

[0034] Calculate ejection fraction EF, end-diastolic volume EDV, and end-systolic volume ESV;

[0035] EF=((EDV-ESV) / EDV)×100%(3)

[0036] Different weights are assigned to each subtask and optimized through a multi-task loss function. The form of the loss function is as follows:

[0037] Loss=w1·Loss1+w2·Loss2+w3·Loss3+L2(4)

[0038] Among them, w1, w2 and w3 are the weight parameters of each task, and Loss1, Loss2 and Loss3 are the loss functions of the corresponding tasks;

[0039] The L2 regularization term is introduced into the loss function, and the loss function form after introduction is:

[0040]

[0041] Among them, λ is the regularization coefficient, θ i is the i-th parameter in the model, Represents the L2 norm sum of squares of the parameters.

[0042] Another technical problem to be solved by the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a non-invasive cardiac function assessment method as described above is implemented.

[0043] Another technical problem to be solved by the present invention is to provide a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the non-invasive cardiac function assessment method as described in any one of the above is implemented.

[0044] Compared with the prior art, the present invention has the following advantages:

[0045] The shared feature extraction network C3D with spatiotemporal attention mechanism is used to analyze the cardiac ultrasound animation video, accurately extracting the features of the heart's systolic and diastolic phases in the image sequence, as well as the texture, morphology and motion features of the ventricular wall; constructing a multi-task learning network and assigning different weights to the three subtasks to explore the relationship between ejection fraction (EF), end-diastolic volume (EDV) and end-systolic volume (ESV); through a joint training strategy, the shared feature extraction network and the multi-task learning network can work together to improve the model's prediction accuracy for ejection fraction and assist doctors in evaluating cardiac function. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The figure is a principle block diagram of a non-invasive cardiac function assessment method of the present invention. DETAILED DESCRIPTION

[0047] The specific implementation modes of the present invention are further described in detail to make the technical solutions of the present invention easier to understand and grasp.

[0048] Example

[0049] See also Figure 1 As shown in the figure, this method includes four main modules: ultrasound video acquisition, ultrasound video preprocessing, ultrasound dynamic image feature extraction, and multi-task learning network construction. The main process of this method is divided into the following steps:

[0050] Ultrasound video acquisition

[0051] The images are acquired by experienced ultrasound technicians using ultrasound equipment. Each image is acquired at least twice, with no less than three cardiac cycles recorded continuously each time. For patients with arrhythmias, at least five consecutive cardiac cycles are recorded each time, and the dynamic ultrasound images of the heart are saved. During the acquisition process, the technician measures through the apical four-chamber heart section to ensure that the endocardium is clearly visible and avoid the appearance of the left ventricular short-axis section. At the end of diastole and end of systole of the left ventricle, the endocardial line is traced to obtain the left ventricular end-diastolic volume (LVEDV) and left ventricular end-systolic volume (LVESV). This data acquisition method can provide a more accurate assessment of cardiac function and provide doctors with key clinical information. The saved dynamic ultrasound images of the heart can be used for subsequent analysis and research.

[0052] By providing detailed and consistent cardiac function data, the systolic and diastolic function of the heart can be assessed more accurately. Data from multiple acquisitions and multiple cardiac cycles increase the reliability of the results and reduce errors through precise measurements. These images and data not only help with real-time clinical judgment, but can also be used for subsequent diagnosis and research, helping doctors better track and manage patients' heart health.

[0053] Ultrasound video preprocessing

[0054] In echocardiography, the systolic and diastolic cycles reflect important aspects of cardiac function, especially changes in myocardial thickness and area. Therefore, accurate video cropping and selection of the cycle of interest are crucial for data accuracy.

[0055] Usually, this process is marked by experienced doctors: first, the doctor will locate the time point T0 when ventricular relaxation begins, then the doctor will find the time point T1 when the ventricular volume reaches the maximum and the ventricle begins to contract, and the time point T2 when the ventricular volume reaches the minimum and the ventricle begins to relax again. [T0, T1] is a diastolic cycle, and [T1, T2] is a systolic cycle.

[0056] Then, the present invention performs masking operations on the cropped image sequence to exclude information irrelevant to the image content, including text, electrocardiogram, and spirometer, etc. Considering that the contrast between the cardiac structure and the background noise is low and it is easily affected by noise, the present invention uses image segmentation technology to separate the cardiac structure from the background, and uses tracking technology to track the position and shape changes of the cardiac structure at different time points to improve the accuracy and stability of the data.

[0057] Finally, considering the differences in size and resolution between different ultrasound devices and operators, in order to ensure that the feature extraction network can process ultrasound images of various sizes, the present invention uses cubic interpolation to standardize images of different pixels into ultrasound dynamic image sequences of uniform size W×H.

[0058] This solution can effectively solve the size and resolution differences in ultrasound images. By using cubic interpolation to standardize the image to a uniform size W×H, it can ensure that the feature extraction network has a consistent input size when processing ultrasound images from different sources, thereby improving the generalization ability and stability of the model. In addition, this standardization method can also reduce the complexity of data processing caused by inconsistent image sizes and ensure the accuracy of data analysis and model training.

[0059] Spatiotemporal feature extraction

[0060] Echocardiography is a non-invasive medical imaging technology that is widely used to assess the heart health of patients. The dynamic image sequence generated by this technology records the movement of the heart, volume changes, and blood flow. However, these video data are usually large and complex, and precise analysis is required to extract valuable cardiac information, including the characteristics of cardiac systole and diastole, as well as the texture, shape, and motion characteristics of the ventricular wall. In order to accurately capture the spatiotemporal information of echocardiography videos and obtain accurate cardiac information, the present invention adopts a three-dimensional convolutional neural network with a C3D structure that integrates spatial feature attention and temporal feature attention mechanisms. This network architecture has been widely used in the field of video analysis, and its core feature is that its convolutional layer considers both temporal and spatial dimensions.

[0061] The C3D (Convolutional 3D) network is a deep learning model specially optimized for medical image video analysis, especially for processing dynamic medical images such as echocardiograms. The core innovation of this network is that its convolution operation can be performed simultaneously on the time axis and the space axis, so that the network can capture the time-varying cardiac structure and functional characteristics in the ultrasound image sequence. Specifically, the C3D network introduces three-dimensional convolution kernels in the convolution layer. These convolution kernels not only operate in the spatial dimension (i.e., the width and height of the image), but also slide and calculate simultaneously in the temporal dimension (i.e., the frame sequence). This design allows the C3D network to capture the temporal dynamic relationship between frames while extracting the spatial features of the image. In each 3D convolution layer, the convolution kernel operates on each frame of the video and between adjacent frames, so that the structural changes and motion patterns of the heart at different times can be extracted. This joint spatiotemporal feature extraction method, the stacked 3D convolution layers can gradually abstract and capture the complex spatiotemporal features in the video, such as the contraction and relaxation patterns of the heart, blood flow dynamics, etc. These features provide richer and more accurate information for the diagnosis of heart disease and the evaluation of cardiac function, making the C3D network of great application value in medical image analysis.

[0062] Furthermore, the C3D network usually fuses the pooling layer with the fully connected layer to further refine the features and perform the final classification or regression task. The pooling layer helps to reduce the data dimension while maintaining sensitivity to key information and extracting key spatiotemporal features, which helps to maintain sensitivity to important information; while the fully connected layer is responsible for mapping the learned high-level features to specific cardiac function parameters, such as ejection fraction (EF), end-diastolic volume (EDV), and end-systolic volume (ESV). The introduction of the C3D network has significantly improved the ability to extract key cardiac function information from echocardiography videos, making the automatic assessment of cardiac function faster and more accurate, and providing clinicians with a powerful auxiliary diagnostic tool. This automated analysis method can greatly reduce the workload of doctors and improve the consistency and reliability of diagnosis.

[0063] Different from the traditional C3D network, this paper introduces the spatial feature attention mechanism and the temporal feature attention mechanism to enhance the ability to capture spatiotemporal characteristics in echocardiography video analysis. The spatial attention mechanism enables the model to dynamically adjust the correlation between different pixel positions within each frame, thereby highlighting the key areas of the heart. This is achieved by calculating the spatial attention weight A s Specifically, we calculate the spatial attention function f through operations such as spatial convolutional layers and global average pooling. s (X), generate weights for each pixel position. Then, these weights A sIt is applied to the original feature map X to achieve element-by-element weighted adjustment, namely:

[0064]

[0065] Among them, A s is the spatial attention weight, f s (X s ) is the output feature of the spatial convolutional layer, d s is the characteristic dimension of the space dimension.

[0066] At the same time, the temporal attention mechanism is used to adjust the weights between different time steps to better capture the dynamic changes of cardiac motion. When processing echocardiogram sequences, we calculate the temporal attention weight A t Specifically, operations such as temporal convolutional layers and global average pooling are used to calculate the temporal attention function f t (X), generate the weights for each time step. Then, these weights A t Applied to echocardiographic sequence X t , weighted adjustment is performed on the features of each time step, namely:

[0067]

[0068] Among them, A t is the temporal attention weight, f t (X t ) is the output feature of the temporal convolutional layer, d t It is the characteristic dimension of the time dimension.

[0069] The introduction of the spatiotemporal attention mechanism provides a powerful tool for processing cardiac ultrasound animations. It allows the model to dynamically adjust the weights of different moments or image positions to better focus on key areas in the heart image. For example, it can automatically identify the motion trajectory of the heart chamber wall to evaluate heart function. In addition, spatiotemporal attention also helps to reduce redundant information and improve the utilization efficiency of data, thereby more accurately analyzing cardiac ultrasound animations. In addition, considering the long-term dependencies and information richness in cardiac ultrasound animations, the spatiotemporal attention mechanism can help the model better capture these characteristics, improve the accuracy and comprehensiveness of cardiac information, and deeply explore the spatiotemporal characteristics in ultrasound cardiography videos, providing more possibilities in the field of medical diagnosis.

[0070] Multi-task learning network construction

[0071] In clinical practice, EF is widely used to assess the health of the heart and the severity of heart disease. Different heart diseases can lead to different changes in EF values, and its accurate calculation is of great significance for timely detection of abnormal heart function, formulation of treatment plans and monitoring of disease progression. Cardiac function assessment is a key research direction in the cardiovascular field, which aims to objectively and quantitatively evaluate the heart's pumping function. Ejection fraction (EF) is one of the important parameters, which measures the proportion of blood pumped out during each heart contraction to the heart's filling blood. EF is calculated by the relationship between the left ventricular end-diastolic volume (EDV) and the end-systolic volume (ESV), usually presented as a percentage, and the calculation formula is:

[0072] EF = ((EDV - ESV) / ​​EDV) × 100% (3)

[0073] EF, ESV and EDV are commonly used indicators in cardiac function assessment, which can reflect the contraction and relaxation ability of the heart and are crucial for the diagnosis and treatment of cardiovascular diseases. Utilizing the rich correlation information between the three indicators, the present invention designs a multi-task learning strategy and loss function, considering the correlation between the EF, ESV and EDV calculation tasks, which can achieve a comprehensive assessment of cardiac function. Specifically, firstly, the shared features in the time domain and the spatial domain extracted by the shared feature extraction network are used to perform task segmentation, and a task-specific learning network is designed for each subtask, using an independent fully connected layer.

[0074] Next, we assign different weights to each subtask, set them according to the importance and difficulty of the task, and optimize them through a multi-task loss function:

[0075] Loss=w1·Loss1+w2·Loss2+w3·Loss3+L2 (4)

[0076] Among them, w1, w2 and w3 are the weight parameters of each task respectively, and Loss1, Loss2 and Loss3 are the loss functions of the corresponding tasks. The design of the loss function Loss not only considers the correlation between different tasks, but also focuses on the main task of EF prediction to improve the generalization ability of the model. In order to further control the complexity of the model and avoid overfitting, we introduced the L2 regularization term into the loss function. Specifically, L2 regularization limits the complexity of the model by adding the sum of the squares of the weight parameters to the loss function to penalize larger weights in the model. The complete loss function form is:

[0077]

[0078] Among them, λ is the regularization coefficient, θ iis the i-th parameter in the model, Represents the L2 norm square sum of the parameters. By introducing the L2 regularization term, the model can effectively control complexity during training, avoid overfitting, and improve performance on unseen data.

[0079] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the non-invasive cardiac function assessment method as described above when executing the program.

[0080] This embodiment also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the non-invasive cardiac function assessment method as described above is implemented.

[0081] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0082] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0083] Of course, the above are only typical examples of the present invention. In addition to these, the present invention may also have many other specific implementations. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the scope of protection required by the present invention. Of course, the above are only typical examples of the present invention. In addition to these, the present invention may also have many other specific implementations. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the scope of protection required by the present invention.

Claims

1. A non-invasive cardiac function assessment system, characterized in that: Included are: Ultrasonic video acquisition module: used to acquire cardiac images through ultrasound. Each image is acquired at least twice, and at least three consecutive cardiac cycles are acquired each time. It is ensured that the acquired data includes the left ventricular end-diastolic volume LVEDV and the left ventricular end-systolic volume LVESV. Ultrasound video preprocessing module: used to locate the time point T0 when ventricular relaxation begins, the time point T1 when the ventricular volume reaches the maximum and the ventricular contraction begins, and the time point T2 when the ventricular volume reaches the minimum and the ventricle begins to relax again, and to crop the video sequence according to the time points T1, T2, and T3, apply image segmentation technology to extract the heart structure, perform tracking processing, and use cubic interpolation method to standardize images of different sizes into ultrasound dynamic image sequences of uniform size; Ultrasound dynamic image feature extraction module: It is used to extract spatiotemporal features of echocardiogram sequences, including heart movement, volume changes, and blood flow, by integrating the 3D convolutional neural network of the C3D structure with the spatial feature attention mechanism and the temporal feature attention mechanism, and to capture the temporal dynamic relationship and spatial features; Multi-task learning network module: used to extract shared features in the time domain and spatial domain through the shared feature extraction network, design an independent learning network for each subtask through task segmentation, set task-specific weights, and optimize through a multi-task loss function. The loss function introduces L2 regularization to control the complexity of the model and calculates cardiac function indicators including ejection fraction EF, end-diastolic volume EDV, and end-systolic volume ESV.

2. A non-invasive cardiac function assessment method, characterized in that: The following steps are involved: Use ultrasound to collect cardiac images and obtain the left ventricular end-diastolic volume LVEDV and left ventricular end-systolic volume LVESV; Locate the time point T0 when ventricular relaxation begins, the time point T1 when the ventricular volume reaches the maximum and the ventricle begins to contract, and the time point T2 when the ventricular volume reaches the minimum and the ventricle begins to relax again, crop the video sequence into diastolic and systolic cycles, exclude information irrelevant to the image content, separate the cardiac structure through image segmentation technology, and use cubic interpolation to standardize the image to a uniform size; The spatial feature attention mechanism and the temporal feature attention mechanism are introduced into the 3D convolutional neural network. The 3D convolutional neural network with the C3D structure that integrates the spatial feature attention mechanism and the temporal feature attention mechanism is used to extract the spatiotemporal features of the standardized echocardiogram sequence to capture the changes in cardiac structure and function. Based on the extracted features, a multi-task learning network is constructed. Task segmentation is performed through a shared feature extraction network. An independent learning network is designed for each subtask. Task-specific weights are set. The model is optimized using a multi-task loss function. L2 regularization is introduced to calculate the ejection fraction EF, end-diastolic volume EDV, and end-systolic volume ESV to provide a comprehensive cardiac function assessment.

3. The non-invasive cardiac function assessment method according to claim 2, characterized in that: The method for obtaining the left ventricular end-diastolic volume LVEDV and the left ventricular end-systolic volume LVESV is as follows: Use ultrasound to collect cardiac images, collect each image at least twice, collect at least three consecutive cardiac cycles each time, and save the cardiac ultrasound dynamic image. During the image collection process, measure through the apical four-chamber heart section; At the end of left ventricular diastole and end of left systole, the left ventricular border was outlined along the intimal line trajectory to obtain the left ventricular end-diastolic volume LVEDV and the left ventricular end-systolic volume LVESV.

4. The non-invasive cardiac function assessment method according to claim 3, characterized in that: When using an ultrasound machine to acquire cardiac images, for patients with arrhythmias, at least five consecutive cardiac cycles should be acquired each time.

5. The non-invasive cardiac function assessment method according to claim 2, characterized in that: The method to normalize the image to a uniform size using cubic interpolation is: Mark the time point T0 when ventricular relaxation begins, the time point T1 when the ventricular volume reaches its maximum and the ventricle begins to contract, and the time point T2 when the ventricular volume reaches its minimum and the ventricle begins to relax again, so as to define the diastolic cycle [T0, T1] and the systolic cycle [T1, T2]; The cropped image sequence was masked to exclude text, electrocardiogram, and spirometer information; the cardiac structure was separated from the background using image segmentation technology, and the position and shape changes of the cardiac structure at different time points were tracked using tracking technology; The ultrasound images of different sizes and resolutions were normalized into ultrasound moving image sequences of uniform size W×H using cubic interpolation.

6. The non-invasive cardiac function assessment method according to claim 2, characterized in that: The method of introducing spatial feature attention mechanism in 3D convolutional neural network is: The spatial attention function f is calculated through the spatial convolution layer and the global average pooling operation s (X), generate weights for each pixel position, and then these weights A s It is applied to the original feature map X to achieve element-by-element weighted adjustment, namely: X s′ =A s ·X s , Among them, A s is the spatial attention weight, f s (X) is the output feature of the spatial convolutional layer, d s is the characteristic dimension of the space dimension.

7. The non-invasive cardiac function assessment method according to claim 6, characterized in that: The method of introducing the temporal feature attention mechanism in the 3D convolutional neural network is: By calculating the temporal attention weight A t Processing echocardiogram sequences, the temporal attention function f is calculated through temporal convolution layers and global average pooling t (X), generate the weights for each time step, and add these weights A t Application to echocardiographic sequence X t , weighted adjustment is performed on the features of each time step, namely: X t′ =A t ·X t , Among them, A t is the temporal attention weight, f t (X t ) is the output feature of the temporal convolutional layer, d t It is the characteristic dimension of the time dimension.

8. The non-invasive cardiac function assessment method according to claim 2, characterized in that: The method to construct a multi-task learning network is: Calculate ejection fraction EF, end-diastolic volume EDV, and end-systolic volume ESV; EF=((EDV-ESV) / EDV)×100% (3) Different weights are assigned to each subtask and optimized through a multi-task loss function. The form of the loss function is as follows: Loss=w1·Loss1+w2·Loss2+w3·Loss3+L2 (4) Among them, w1, w2 and w3 are the weight parameters of each task, and Loss1, Loss2 and Loss3 are the loss functions of the corresponding tasks; The L2 regularization term is introduced into the loss function, and the loss function form after introduction is: Among them, λ is the regularization coefficient, θ i is the i-th parameter in the model, Represents the L2 norm sum of squares of the parameters.

9. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the non-invasive cardiac function assessment method as claimed in any one of claims 2 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the non-invasive cardiac function assessment method as described in any one of claims 2-8 is implemented.

Images