A Meta-learning-based Sparse-view 3D Echocardiography Rapid Imaging and Evaluation Method

Abstract

This invention discloses a rapid 3D cardiac ultrasound imaging and evaluation method based on meta-learning with a sparse perspective. The method includes the following steps: S1, acquiring a dataset of heart samples to be reconstructed; S2, constructing an imaging model, using the dataset, and training and initializing the imaging model based on a meta-learning algorithm to obtain initial model parameters; S3, the imaging model performs 3D reconstruction of the target heart in a new scene based on the initial model parameters; S4, evaluating cardiac function of the heart reconstructed in step S3. The innovation of this invention lies in introducing a meta-learning algorithm to learn common structural information among different hearts, providing prior knowledge, thereby achieving high-quality, high-performance rapid imaging with only sparse sample input. This method can be applied to existing 3D cardiac ultrasound imaging systems to improve imaging quality and speed.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional ultrasound imaging technology, and in particular to a method for rapid three-dimensional cardiac ultrasound imaging and cardiac function assessment based on meta-learning and sparse perspective. Background Technology

[0002] Echocardiography is a non-invasive, safe, simple, and practical diagnostic method for cardiovascular diseases, widely used in the diagnosis and treatment of heart conditions. It assesses the heart's structure and function by transmitting high-frequency sound waves to the heart using an ultrasound probe, generating echo images. Two-dimensional (2D) echocardiography is the most common technique, providing two-dimensional images that show the heart's structure and function; however, it only provides planar images and struggles to accurately display the heart's three-dimensional structure. To overcome this limitation, 3D echocardiography was developed. Compared to 2D ultrasound, 3D ultrasound can generate three-dimensional images containing more anatomical information, better showcasing the heart's structure and function. However, 3D echocardiography still faces some limitations in practical applications. First, the quality of 3D ultrasound imaging is affected by factors such as the heart's location, size, and motion; therefore, in some cases, the image quality of 3D ultrasound may be inferior to that of 2D ultrasound. Second, 3D ultrasound technology requires processing large amounts of data during imaging, necessitating longer acquisition times, higher computational power, and more expensive equipment, thus limiting its practical application. Therefore, achieving high-quality, low-time-consuming three-dimensional cardiac ultrasound imaging technology, combined with left and right ventricular image segmentation algorithms, to realize accurate evaluation of cardiac structure and function is of great significance in the diagnosis of heart diseases.

[0003] To achieve three-dimensional ultrasound imaging, methods are divided into two categories: imaging methods based on three-dimensional ultrasound probes and methods based on deep learning. The former acquires three-dimensional cardiac spatial resolution with poor resolution, requiring manual calibration by experienced sonographers for cardiac function assessment. The latter includes cardiac three-dimensional ultrasound imaging methods based on implicit neural representations, but these methods have longer reconstruction times and require a large number of multi-view two-dimensional ultrasound images as input; otherwise, it is difficult to reconstruct image information from unknown perspectives.

[0004] Recent research has shown that meta-learning algorithms, such as MAML and Reptile, can significantly improve the performance of neural networks with limited input. These algorithms have been used to learn the initial parameters of neural networks, providing prior information for subsequent network training. This allows the model to quickly adapt to new tasks that have never been seen before, even with limited training data.

[0005] Based on the aforementioned background conditions and technical problems, there is an urgent need for a sparse-view three-dimensional cardiac ultrasound reconstruction algorithm to at least solve the problems of large sample requirements and long time consumption in existing technologies. Summary of the Invention

[0006] In view of the above-mentioned prior art, the purpose of this invention is to provide a method for rapid imaging and evaluation of three-dimensional cardiac ultrasound based on meta-learning with sparse perspective, so as to achieve low-data, high-speed, and high-quality three-dimensional cardiac ultrasound imaging.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A rapid imaging and evaluation method for sparse-view 3D cardiac ultrasound based on meta-learning, the steps of which include:

[0009] S1, Obtain the dataset of the heart sample to be reconstructed;

[0010] S2, Construct an imaging model, and use the dataset to train and initialize the imaging model based on a meta-learning algorithm to obtain the initial parameters of the model;

[0011] S3, the imaging model performs three-dimensional reconstruction of the target heart in the new scene based on the initial parameters;

[0012] S4. Perform a cardiac function assessment on the heart reconstructed in step S3.

[0013] Further, in step S1, a dataset of heart samples to be reconstructed is obtained, specifically: for each sample, two-dimensional ultrasound images are sparsely acquired around the long axis of the heart in 360°, and then all sample hearts are collected as a set.

[0014] Furthermore, in step S1, the dataset is divided into meta-training data and test data according to a certain ratio.

[0015] Furthermore, in step S2, the Reptile meta-learning algorithm is used to randomly sample N samples from the meta-training data for inner loop training, with each sample looping M times to update the inner loop parameters; after N samplings, the outer loop parameters are updated to obtain the initial parameters of the imaging model.

[0016] Further, in step S3, the test data is input into the imaging model using the initial parameters to perform three-dimensional reconstruction; then, a set of two-dimensional slices are generated using the reconstructed three-dimensional heart, and a loss function is constructed with the input data. The backpropagation algorithm is used to update the imaging model parameters and optimize the reconstructed three-dimensional heart.

[0017] Furthermore, the ratio of the meta-training data to the test data is 5:1.

[0018] Furthermore, in step S2, the imaging model is an imaging model based on implicit neural expression, whose input is the three-dimensional coordinates of the heart space and whose output is an implicit function of the voxel value of the three-dimensional heart at the corresponding position.

[0019] Further, in step S4, a set of two-dimensional slices of the heart are generated using the implicit neural expression of the imaging model, and the left and right ventricular regions in the slices are segmented using a semantic segmentation algorithm, and the area of ​​the cavity in each slice is calculated; then, by accumulating the areas of the left and right ventricles in all slices, the volume of the left and right ventricles is obtained, and then the ejection fraction of the left and right ventricles is calculated, and then cardiac function is assessed.

[0020] The innovation and advantages of this invention are as follows:

[0021] (1) This invention innovatively introduces meta-learning algorithms into a cardiac three-dimensional ultrasound imaging system based on implicit neural representation, learns the common structural information between different hearts, provides prior knowledge, and thus achieves high-quality, high-performance, and fast imaging with only sparse sample input, which can solve the problems of large sample requirements and long time consumption in the prior art.

[0022] (2) This invention requires only a set of two-dimensional ultrasound images acquired from a sparse viewpoint to reconstruct a high-quality three-dimensional heart while preserving detailed information about the cavity. Compared to existing imaging methods, this invention can significantly shorten the two-dimensional ultrasound image acquisition time and the three-dimensional reconstruction time. This method is efficient and simple, and can be applied to existing three-dimensional cardiac ultrasound imaging systems to improve imaging quality and speed, leading to its widespread application and promotion in clinical diagnosis.

[0023] (3) The three-dimensional heart model reconstructed in this invention can provide a reliable basis for subsequent calculation of cardiac function indicators, thus playing a role in auxiliary diagnosis. Using image segmentation algorithms, important indicators such as left and right ventricular volumes and ejection fraction can be accurately calculated, thereby more accurately assessing the patient's cardiac function status. Furthermore, this method is highly repeatable and operable, providing important reference data for clinical medical research. Attached Figure Description

[0024] Figure 1 This is a block diagram illustrating the principle and structure of the method of the present invention;

[0025] Figure 2 This is a flowchart of sparse acquisition of two-dimensional ultrasound images in an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the meta-learning algorithm flow.

[0027] Figure 4 The flowchart of the fast three-dimensional cardiac ultrasound imaging method based on meta-learning in this embodiment of the invention is shown in (a), which represents the weight update process of the Reptile meta-learning algorithm; and (b) which represents the various components of the network model.

[0028] Figure 5 This is a flowchart of the cardiac function assessment method in an embodiment of the present invention;

[0029] Figure 6 This is a comparison chart showing the reconstruction speed of the present invention with that of existing implicit representation methods;

[0030] Figure 7 The image shows a comparison of a 3D heart reconstructed according to an embodiment of the present invention, a 3D heart acquired by a traditional 3D probe, and a 3D heart reconstructed based on existing implicit representation methods with dense viewpoint input. Detailed Implementation

[0031] Embodiments of the present invention will now be described in detail with reference to the accompanying drawings, examples of which are illustrated in the drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0032] Reference Figure 1 and 2 As shown, a 360° circumferential scan was performed at the apex of the heart using the most commonly used two-dimensional ultrasound probe in clinical practice. A small number of two-dimensional images were sparsely acquired at sampling intervals of 6°–10°, without the need for additional positioning sensors. The same acquisition mode was used on different patients to obtain a dataset of heart samples to be reconstructed.

[0033] Reference Figure 3 and Figure 4 As shown in (a), the initial imaging model is trained based on the meta-learning algorithm. The specific implementation method is as follows:

[0034] First, the sample dataset is divided into meta-training data and test data according to a certain ratio. After testing, this embodiment can achieve the best performance at a ratio of 5:1.

[0035] Furthermore, it is assumed that the three-dimensional volumetric intensity of different hearts with intrinsic structural similarities follows a specific distribution. Then, a meta-learning algorithm is used to find the initial network parameters θ. * The meta-training data is defined as the task set for 3D heart reconstruction. The loss function of a network is defined as This embodiment uses the Reptile algorithm to optimize the initial parameters θ of the model, and its optimization objective is:

[0036]

[0037] in, Indicates a given task The initial parameter θ is used to perform k iterations of gradient descent, which is also called the inner loop. The expected value of the loss function is represented by θ, which represents the optimized parameters. The goal of meta-learning is to enable a model trained with meta-learning initial parameters to achieve optimal accuracy and minimum loss on a new task. This allows for rapid convergence within the model with only sparse samples, demonstrating optimal performance. The step of optimizing the initial parameters of the model is called the outer loop: resampling the task. And update the initial parameters of the model, using the following formula:

[0038]

[0039] Where β is the update step size in the meta-learning process, and the parameter θ obtained after the outer loop is the initial parameter of the model optimized by meta-learning.

[0040] Reference Figure 4 As shown in (b), the target heart in the new scene is reconstructed in three dimensions using the initial parameters of the model learned through meta-learning.

[0041] Furthermore, the data in the test set represents the core of the target in the new scenario;

[0042] Furthermore, an initialized implicit neural network model is used to construct a three-dimensional representation of the heart structure, with the input being the three-dimensional coordinates of the heart space. The output is the voxel value of the three-dimensional heart at the corresponding location. implicit function F θ ,Right now:

[0043]

[0044] Furthermore, the physical imaging process of two-dimensional echocardiographic images is modeled as a process of acquiring slices from a three-dimensional simulated heart, and a loss function is constructed based on this physical process. To supervise the training of implicit neural representations:

[0045]

[0046] Among them, I s For sliced ​​images, I g The two-dimensional ultrasound images are the actual acquired images, and N is the number of two-dimensional images acquired.

[0047] Furthermore, the backpropagation algorithm is used to update the model parameters and optimize the reconstructed 3D heart.

[0048] Reference Figure 5 As shown, the ejection fraction was calculated on the reconstructed high-precision heart to assess cardiac function. The formula for calculating the ejection fraction is:

[0049]

[0050] Wherein EDV and ESV represent the end-diastolic and end-systolic volumes of the ventricles, respectively. This invention first generates a set of two-dimensional ultrasound slices of the heart using a constructed implicit neural representation of the heart. Then, a semantic segmentation algorithm is used to automatically segment the left and right ventricular regions, calculating the area of ​​the cavities in each slice. After segmentation, the areas of the left and right ventricles in all slices are summed to obtain the volumes of the left and right ventricles. Finally, the ejection fractions of the left and right ventricles are obtained based on the rate of change of their volumes during end-diastole and end-systole.

[0051] Reference Figure 6 , Figure 7 As shown, the three-dimensional heart reconstructed by the present invention under sparse perspective retains as much internal cavity information as the three-dimensional heart reconstructed under a large number of input perspectives by existing implicit neural representations, and has fewer artifacts and higher reconstruction speed, with an overall reconstruction speed improvement of about 75%.

[0052] Referring to Table 1, this invention compares with methods that initialize the implicit neural network using only random parameters without employing a meta-learning algorithm. The results show that the accuracy of left ventricular ejection fraction calculation using this invention is higher. Here, the mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the accuracy of ejection fraction estimation.

[0053] Table 1. Quantitative comparison of the accuracy of cardiac function assessment between the present invention and algorithms without meta-learning.

[0054]

[0055]

[0056] Because the cardiac structures of different individuals exhibit strong geometric correlations, they can be modeled as network parameters representing structural information. To utilize this prior information, this invention introduces a meta-learning algorithm into a cardiac 3D ultrasound imaging system based on implicit neural representations. This addresses the problems of large sample requirements and long processing times in existing technologies. Only a set of sparsely viewed 2D ultrasound images is needed to reconstruct a high-quality 3D heart, which means shorter acquisition time in clinical applications. This invention improves reconstruction quality, reduces computation time by 75%, and can accurately calculate cardiac function indicators on the reconstructed heart, thus providing diagnostic assistance.

Claims

1. A rapid imaging and evaluation method for sparse-view 3D cardiac ultrasound based on meta-learning, characterized in that, The steps of this method include: S1, obtaining a dataset of heart samples to be reconstructed from different patients, specifically: for each sample, the long axis of the heart... A small number of two-dimensional ultrasound images were sparsely acquired at sampling intervals of 6° to 10° without the need for additional positioning sensors. All sampled hearts were then collected as a set. The dataset was then divided into meta-training data and test data in a 5:1 ratio. S2, Construct an imaging model. Using the dataset, train and initialize the imaging model based on a meta-learning algorithm to obtain initial model parameters. Specifically, this includes: using the Reptile meta-learning algorithm, randomly sampling N samples from the meta-training data for inner loop training, with each sample looping M times, and updating the inner loop parameters; after N sampling cycles, updating the outer loop parameters to obtain the initial parameters of the imaging model; the imaging model is an imaging model based on implicit neural expression, whose input is the three-dimensional coordinates of the heart space, and whose output is an implicit function of the voxel values ​​of the three-dimensional heart at the corresponding position; wherein, the physical imaging process of the two-dimensional echocardiogram image is modeled as a process of simulating slices from the three-dimensional heart, and a loss function is constructed based on this physical process. To supervise the training of implicit neural representations: in, For sliced ​​images, These are actual two-dimensional ultrasound images; S3, the imaging model performs three-dimensional reconstruction of the target heart in the new scene according to the initial parameters; specifically: the test data is input into the imaging model using the initial parameters to perform three-dimensional reconstruction; then a set of two-dimensional slices are generated using the reconstructed three-dimensional heart, a loss function is constructed with the input data, and the backpropagation algorithm is used to update the imaging model parameters and optimize the reconstructed three-dimensional heart. S4. Perform cardiac function assessment on the heart reconstructed in step S3. Specifically, generate a set of two-dimensional slices of the heart using the implicit neural expression of the imaging model, and use a semantic segmentation algorithm to segment the left and right ventricular regions in the slices respectively, and calculate the area of ​​the cavity in each slice; then, by accumulating the areas of the left and right ventricles in all slices, obtain the volume of the left and right ventricles, and then calculate the ejection fraction of the left and right ventricles, and then perform cardiac function assessment. The formula for calculating the ejection fraction is: Where EF is ejection fraction, EDV is end-diastolic volume of the ventricle, and ESV is end-systolic volume of the ventricle.

Images