Ultrasound scan trajectory planning method and system based on variational autoencoder

Abstract

The present application relates to the technical field of ultrasonic scanning path planning, and discloses an ultrasonic scanning trajectory planning method and system based on a variational autoencoder. The ultrasonic scanning trajectory planning method is applied to an ultrasonic scanning robot and specifically comprises the following steps: S101: receiving an ultrasonic image acquisition instruction sent by a user terminal, wherein the ultrasonic image acquisition instruction comprises an activation command of a carried ultrasonic probe, target region data collected by an ultrasonic image, and feature information of the ultrasonic probe when collecting the image; S102: acquiring ultrasonic wave image data collected by the target region, and sequentially placing the ultrasonic wave image data in an image preprocessing model to sequentially denoise the ultrasonic wave image data. The present application trains and optimizes an ultrasonic diagnosis model through few-sample learning, plans a trajectory through high-precision control and intelligent decision of an ultrasonic scanning robot, and automatically completes ultrasonic scanning operation, thereby improving the accuracy and stability of a diagnosis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ultrasonic scanning path planning, in particular to an ultrasonic scanning trajectory planning method and system based on a variational autoencoder. BACKGROUND

[0002] Ultrasound scanning is a non-invasive and radiation-free medical imaging technology widely used in diagnosis and treatment, especially in gynecology and obstetrics, cardiology, tumor detection, etc. It uses high-frequency sound waves to image the internal structure of the human body, and has significant advantages in real-time dynamic monitoring and image acquisition.

[0003] Currently, most ultrasound systems rely on manual adjustment of the scanning trajectory of the probe or use simple mechanical aids, which have limitations in accuracy, response speed and adaptability in complex environments, and cannot meet the needs of high precision and high efficiency. SUMMARY

[0004] The purpose of the present application is to provide an ultrasonic scanning trajectory planning method and system based on a variational autoencoder, which automatically completes ultrasonic scanning operations through high-precision control and intelligent decision planning of the scanning trajectory of an ultrasonic scanning robot, improves the accuracy and stability of the diagnostic results, and reduces the dependence on operators, aiming to solve the problems in the prior art.

[0005] The present application is implemented as follows: an ultrasonic scanning trajectory planning method based on a variational autoencoder is applied to an ultrasonic scanning robot, and specifically includes the following steps:

[0006] S101: receiving an ultrasonic image acquisition instruction sent by a user terminal, the ultrasonic image acquisition instruction including an activation command for the ultrasonic probe carried, target region data acquired by the ultrasonic image, and feature information of the ultrasonic probe when acquiring the image;

[0007] S102: obtaining ultrasonic wave image data acquired by the target region, sequentially placing the ultrasonic wave image data in an image preprocessing model to sequentially denoise the ultrasonic wave image data, and obtaining balanced ultrasonic wave image data after denoising processing;

[0008] S103: placing the obtained balanced ultrasonic wave image data in a multi-item variational autoencoder for image enhancement, obtaining an enhanced image of the multi-item variational autoencoder, and performing initial global feature fusion on the enhanced image of the multi-item variational autoencoder through a fully connected neural network to obtain a multi-item initial global feature matrix;

[0009] S104: adding a preset latent variable to each of the multiple variational autoencoders, and concatenating the multiple initial global feature matrices extracted by the fully connected neural network with the latent variables in the multiple variational autoencoders in the first dimension to obtain multiple first feature matrices;

[0010] S105: Input the multiple first feature matrices into the fully connected neural network respectively to obtain the output multiple second feature matrices, split the multiple second feature matrices separately, retain only the multiple latent feature matrices at the end, and then splice the multiple latent feature matrices in the second dimension to obtain the trajectory vector, and input the trajectory vector into the fully connected neural network to obtain the final output scanning trajectory matrix.

[0011] Furthermore, in S101, receiving an ultrasound image acquisition instruction sent by a user terminal includes:

[0012] receiving a connection request sent by a user terminal through a preset network, wherein the connection request is used to request to establish a connection with the ultrasonic scanning robot;

[0013] Detecting whether the current account of the user terminal is the target account;

[0014] If the current account of the user terminal is the target account, a connection is made with the user terminal according to the connection request. After the connection is completed, an ultrasound image acquisition instruction sent by the user terminal is received.

[0015] Furthermore, the preset network includes one or a combination of 3G network, 4G network, 5G network, and WIFI network.

[0016] Furthermore, in S102, ultrasonic image data collected in the target area is obtained, including:

[0017] The operator manually controls the ultrasonic scanning robot to move the ultrasonic probe to the starting point Q0 of the area to be scanned;

[0018] The ultrasound image is acquired with the starting point Q0 as the starting point, and the image I(0) at the initial moment is acquired and then the key medical features are highlighted through denoising and image enhancement.

[0019] Furthermore, in S103, the multiple variational autoencoders are at least two, namely a first variational autoencoder and a second variational autoencoder.

[0020] Furthermore, the obtained balanced ultrasonic image data is placed in a multinomial variational autoencoder for image enhancement to obtain an enhanced image of the multinomial variational autoencoder, including:

[0021] Input X1 and X2 in the first and second variational autoencoders respectively, wherein the X1 and X2 are the same data-enhanced enhanced image;

[0022] The input X1 in the first variational autoencoder is a 64x64x1 grayscale image, which is deformed into a 4096x1 bar vector through a flattening operation. The bar vector is subjected to an initial global feature fusion through a fully connected neural network with a network parameter size of 1x512 to obtain an initial global feature A1 with a size of 4096x512. The X2 is also subjected to the same processing in the second variational autoencoder to obtain an initial global feature A2 with a size of 4096x512.

[0023] Further, in S104, the multiple initial global feature matrices extracted through the fully connected neural network are spliced with the multiple latent variables in the first dimension in the variational autoencoder, including:

[0024] A preset latent variable z is added in the first and second variational autoencoders respectively, wherein the z is a 1x512 zero vector.

[0025] The initial global feature A1 and the initial global feature A2 extracted through the fully connected neural network are spliced with the respective latent variables z in the first dimension to obtain a first feature matrix B1 in the first variational autoencoder and a first feature matrix B2 in the second variational autoencoder.

[0026] Further, in S105, the multiple first feature matrices are input into the fully connected neural network to obtain multiple second feature matrices, and the multiple second feature matrices are respectively split to retain only the multiple latent feature matrices at the end, including:

[0027] The first feature matrix B1 and the first feature matrix B2 are input into a fully connected neural network with a size of 512x32 to obtain a second feature matrix C1 and a second feature matrix C2, respectively. The size of the second feature matrix C1 and the second feature matrix C2 is 4097x32. The second feature matrix C1 and the second feature matrix C2 are respectively split to obtain a latent feature matrix Z1 and a latent feature matrix Z2 at the end.

[0028] Further, the multiple latent feature matrices are spliced in the second dimension to obtain a trajectory vector, and the trajectory vector is input into the fully connected neural network to obtain a final output scan trajectory matrix, including:

[0029] The hidden feature matrix Z1 and the hidden feature matrix Z2 are spliced in the second dimension to obtain a vector V with a size of 1*64, the vector V is input into a fully connected neural network with a size of 64*6, and a final output scanning trajectory matrix Q1 = [Sx1, Sy1, Sz1, Rx1, Ry1, Rz1] is obtained, wherein Sx1, Sy1 and Sz1 represent respective coordinates in three spatial coordinate axis directions at t = 1, and Rx1, Ry1 and Rz1 represent end posture of the mechanical arm in three posture coordinate axis directions at t = 1.

[0030] After the ultrasonic scanning robot moves to the trajectory coordinates Q1, the step is repeated to plan the trajectory of the entire scanning task until the ultrasonic scanning robot is manually stopped, that is, the ultrasonic scanning trajectory is obtained.

[0031] Compared with the prior art, the ultrasonic scanning trajectory planning method and system based on the variational autoencoder provided by the application have the following beneficial effects:

[0032] 1. By training the scanning samples of the ultrasonic scanning robot, the problem of large image data error of mechanical ultrasonic scanning is solved. Considering that the quality of the ultrasonic image is highly dependent on the experience of the operator and there is significant individual difference, it is necessary to achieve good generalization ability under limited training data and to process the difference in geometry and image features of different patients. Therefore, the mutual information learning method based on the variational autoencoder is adopted, the ultrasonic diagnosis model is trained and optimized through few-sample learning, the ultrasonic scanning operation is automatically completed through high-precision control and intelligent decision planning trajectory of the ultrasonic scanning robot, and the accuracy and stability of the diagnosis result are improved, and the dependence on the operator is reduced.

[0033] 2. The ultrasonic scanning robot acquires the ultrasonic image in real time, and inputs the ultrasonic image at each moment into the first variational autoencoder and the second variational autoencoder to extract the hidden feature 1 and the hidden feature 2. After mutual information is disentangled in the hidden feature 1 and the hidden feature 2, the scanning trajectory is generated by inputting into the fully connected network. The acquired scanning trajectory is followed to perform ultrasonic scanning during subsequent ultrasonic scanning, which liberates personnel, and the collected ultrasonic image is of high accuracy.

[0034] The ultrasonic scanning trajectory planning system based on the variational autoencoder is used to execute the ultrasonic scanning trajectory planning method described above, and the ultrasonic scanning trajectory planning system comprises:

[0035] The acquisition module is configured to receive an ultrasonic image acquisition instruction sent by a user terminal.

[0036] The acquisition module is configured to receive an ultrasonic image acquisition instruction sent by a user terminal.

[0037] The pre-processing module is used for sequentially placing the ultrasonic image data into an image pre-processing model to sequentially denoise the ultrasonic image data, and obtaining balanced ultrasonic image data after denoising processing.

[0038] The feature fusion module is used for obtaining enhanced images of the multiple variational autoencoders, and the enhanced images of the multiple variational autoencoders are subjected to initial global feature fusion through the fully connected neural network to obtain multiple initial global feature matrices.

[0039] The matrix generation module is used for splicing the multiple initial global feature matrices and the hidden variables in the first dimension in the multiple variational autoencoders, and separately splitting the multiple second feature matrices, and only retaining the multiple hidden feature matrices at the end.

[0040] The generation module is used for splicing the multiple hidden feature matrices in the second dimension to obtain a trajectory vector, inputting the trajectory vector into the fully connected neural network to obtain a final output scanning trajectory matrix, repeating the steps to plan the trajectory of the entire scanning task until the ultrasonic scanning robot is manually stopped, that is, the ultrasonic scanning trajectory is obtained. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A flowchart of the image data pre-processing process in the ultrasonic scanning trajectory planning method based on the variational autoencoder proposed in the application;

[0042] Figure 2 A flowchart of the ultrasonic scanning robot intelligently generating a planning trajectory in the ultrasonic scanning trajectory planning method based on the variational autoencoder proposed in the application;

[0043] Figure 3 A coordinate graph of the ultrasonic scanning trajectory generated in the ultrasonic scanning trajectory planning method based on the variational autoencoder proposed in the application;

[0044] Figure 4 A structural diagram of the ultrasonic scanning trajectory planning system based on the variational autoencoder proposed in the application. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the application clearer, the application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application, and are not used to limit the application.

[0046] The implementation of the application is described in detail below in combination with specific examples.

[0047] The same or similar reference numerals in the drawings of the embodiments correspond to the same or similar components; in the description of the present application, it is understood that if the orientations or positional relationships indicated by the terms "upper", "lower", "left", "right" and the like are based on the orientations or positional relationships shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for exemplary illustration, and cannot be understood as a limitation on the present application, for those skilled in the art, the specific meanings of the above terms can be understood according to the specific circumstances.

[0048] Referring to Figures 1-2 The ultrasonic scan trajectory planning method based on the variational autoencoder is applied to an ultrasonic scanning robot, and specifically includes the following steps:

[0049] S101: receiving an ultrasonic image acquisition instruction sent by a user terminal, the ultrasonic image acquisition instruction including an activation command of a mounted ultrasonic probe, target region data collected by an ultrasonic image, and feature information of the ultrasonic probe when collecting the image;

[0050] The ultrasonic image acquisition instruction sent by the user terminal includes:

[0051] The connection request is used to request to establish a connection with the ultrasonic scanning robot;

[0052] Detecting whether the current account of the user terminal is a target account;

[0053] If the current account of the user terminal is the target account, connect with the user terminal according to the connection request, and after the connection is completed, receive the ultrasonic image acquisition instruction sent by the user terminal;

[0054] S102: obtaining ultrasonic wave image data collected by a target region, sequentially placing the ultrasonic wave image data in an image preprocessing model to sequentially denoise the ultrasonic wave image data, and obtaining balanced ultrasonic wave image data after denoising processing;

[0055] The ultrasonic wave image data collected by the target region includes:

[0056] An operator manually controls the ultrasonic scanning robot to move the ultrasonic probe to the starting point Q0 of the region to be scanned;

[0057] The ultrasound image acquisition is performed with the starting point Q0 as the starting point, the image I(0) at the initial time obtained is highlighted the key medical features through denoising and image enhancement, and then the ultrasound scanning robot moves to the track coordinate Q1, the track planning of the entire scanning task is repeated, and the ultrasound scanning robot is manually stopped, that is, the ultrasound scanning track is obtained;

[0058] S103: The obtained balanced ultrasound image data is placed in the multiple variational autoencoder for image enhancement, and the enhanced image of the multiple variational autoencoder is obtained, and the enhanced image of the multiple variational autoencoder is subjected to initial global feature fusion through the fully connected neural network to obtain a plurality of initial global feature matrices;

[0059] S104: A preset hidden variable is added to the multiple variational autoencoders respectively, the plurality of initial global feature matrices extracted through the fully connected neural network are spliced with the hidden variables in the first dimension in the multiple variational autoencoders to obtain a plurality of first feature matrices;

[0060] S105: The plurality of first feature matrices are respectively input into the fully connected neural network to obtain a plurality of second feature matrices output, the plurality of second feature matrices are respectively split, only the plurality of hidden feature matrices at the end are retained, and the plurality of hidden feature matrices are spliced in the second dimension to obtain a track vector, and the track vector is input into the fully connected neural network to obtain a final output scanning track matrix.

[0061] In this embodiment, the preset network includes one or a combination of a 3G network, a 4G network, a 5G network, and a WIFI network.

[0062] In this embodiment, in the first step, the multiple variational autoencoders are at least two, which are a first variational autoencoder and a second variational autoencoder, and X1 and X2 are respectively input into the first variational autoencoder and the second variational autoencoder, X1 and X2 are the same enhanced image subjected to data enhancement; X1 input into the first variational autoencoder is a gray image of 64x64x1, X1 is deformed into a bar vector of 4096x1 through a flattening operation, the bar vector is subjected to initial global feature fusion through a fully connected neural network with a network parameter size of 1x512 to obtain an initial global feature A1 with a size of 4096x512, and X2 is also subjected to the same processing in the second variational autoencoder to obtain an initial global feature A2 with a size of 4096x512;

[0063] In the second step, a preset hidden variable z is added to the first variational autoencoder and the second variational autoencoder respectively, and z is a zero vector of 1x512;

[0064] The initial global features A1 and A2 extracted by the fully connected neural network are concatenated with their respective latent variables z in the first dimension, and the first feature matrix B1 is obtained in the first variational autoencoder, and the first feature matrix B2 is obtained in the second variational autoencoder;

[0065] The third step is to input the first feature matrix B1 and the first feature matrix B2 into a fully connected neural network with a size of 512×32 to obtain the output second feature matrix C1 and the second feature matrix C2 respectively. The size of the second feature matrix C1 and the second feature matrix C2 are both 4097×32. Then, the second feature matrix C1 and the second feature matrix C2 are split separately to obtain the latent feature matrix Z1 and the latent feature matrix Z2 at the end;

[0066] The fourth step is to concatenate the latent feature matrix Z1 and the latent feature matrix Z2 in the second dimension to obtain a vector V of size 1×64, and input the vector V into a fully connected neural network of size 64×6 to obtain the final output scanning trajectory matrix Q1 = [Sx1, Sy1, Sz1, Rx1, Ry1, Rz1], where Sx1, Sy1, Sz1 represent the coordinates of the three spatial coordinate axes at time t = 1, and Rx1, Ry1, Rz1 represent the postures of the end of the manipulator in the three posture coordinate axes at time t = 1; after the ultrasonic scanning robot moves to the trajectory coordinate Q1, the steps are repeated to plan the trajectory of the entire scanning task until the ultrasonic scanning robot is manually stopped, and the ultrasonic scanning trajectory is obtained. Figure 3 As shown, it is a trajectory curve diagram of three-dimensional coordinates, which is the connection curve from Q0 to Qn, and each coordinate point is determined by the global feature fusion judgment calculated by the first variational autoencoder and the second variational autoencoder. When performing subsequent ultrasound scanning, the ultrasound scanning is performed in accordance with the obtained scanning trajectory, which frees up personnel and at the same time the collected ultrasound images are more accurate.

[0067] Reference Figure 4 As shown in the figure, the ultrasound scanning trajectory planning system based on variational autoencoder is used to execute the above-mentioned ultrasound scanning trajectory planning method. By training the scanning samples of the ultrasound scanning robot, the problem of large image data errors in mechanical ultrasound scanning is solved. Considering that the quality of ultrasound images is highly dependent on the operator's experience and there are significant individual differences, it is necessary to achieve good generalization ability with limited training data and to deal with the differences in geometric and image features of different patients. Therefore, the mutual information learning method based on variational autoencoder is adopted to train and optimize the ultrasound diagnosis model through few-sample learning. Through the high-precision control and intelligent decision-making trajectory planning of the ultrasound scanning robot, the ultrasound scanning operation is automatically completed, the accuracy and stability of the diagnostic results are improved, and the dependence on the operator is reduced. The ultrasound scanning trajectory planning system includes:

[0068] The acquisition module is configured to receive an ultrasonic image acquisition instruction sent by a user terminal; the acquisition module is configured to acquire ultrasonic image data acquired by a target region; the preprocessing module is configured to sequentially place the ultrasonic image data in an image preprocessing model to sequentially perform denoising on the ultrasonic image data, and obtain balanced ultrasonic image data after denoising processing; the feature fusion module is configured to acquire enhanced images of multiple variational autoencoders, and the enhanced images of the multiple variational autoencoders are subjected to initial global feature fusion through a fully connected neural network to obtain multiple initial global feature matrices; the matrix generation module is configured to splice the multiple initial global feature matrices and hidden variables in the multiple variational autoencoders in a first dimension, and separately split the multiple second feature matrices, and only retain the multiple hidden feature matrices at the end; the generation module is configured to splice the multiple hidden feature matrices in a second dimension to obtain a trajectory vector, input the trajectory vector into a fully connected neural network to obtain a final output scan trajectory matrix, repeat the steps to plan a trajectory for an entire scan task, and stop the ultrasonic scanning robot manually until an ultrasonic scan trajectory is obtained; the ultrasonic image is acquired in real time through the ultrasonic scanning robot, and each moment of the ultrasonic image is input into the first variational autoencoder and the second variational autoencoder to extract hidden feature 1 and hidden feature 2; after mutual information disentanglement of the hidden feature 1 and the hidden feature 2, the ultrasonic image is input into a fully connected network to generate a scan trajectory; and the ultrasonic scanning is performed according to the obtained scan trajectory in subsequent ultrasonic scanning, which releases personnel and improves the accuracy of the acquired ultrasonic image.

[0069] In the embodiment, the entire operation process can be controlled by a computer to realize automatic operation control, and in each operation link, a sensor can be set to perform signal feedback to realize sequential performance of the steps. These are all conventional knowledge of automatic control, and will not be described in detail in the embodiment.

[0070] The above only describes preferred embodiments of the present application and should not be used to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. Ultrasonic scanning trajectory planning method based on variational autoencoder, characterized in that: Applied to ultrasonic scanning robots, specifically including the following steps: S101: Receive an ultrasound image acquisition instruction sent by a user terminal, where the ultrasound image acquisition instruction includes a command to activate an onboard ultrasound probe, target area data acquired by the ultrasound image, and feature information of the ultrasound probe when acquiring the image; S102: Acquire ultrasonic image data collected in the target area, place the ultrasonic image data in an image preprocessing model in sequence to perform denoising on the ultrasonic image data in sequence, and obtain balanced ultrasonic image data after the denoising process; S103: placing the acquired balanced ultrasonic image data in a multinomial variational autoencoder for image enhancement to obtain an enhanced image of the multinomial variational autoencoder, performing initial global feature fusion on the enhanced image of the multinomial variational autoencoder through a fully connected neural network to obtain multiple initial global feature matrices; S104: adding a preset latent variable to each of the multiple variational autoencoders, and concatenating the multiple initial global feature matrices extracted by the fully connected neural network with the latent variables in the multiple variational autoencoders in the first dimension to obtain multiple first feature matrices; S105: Input the multiple first feature matrices into the fully connected neural network respectively to obtain the output multiple second feature matrices, split the multiple second feature matrices separately, retain only the multiple latent feature matrices at the end, and then splice the multiple latent feature matrices in the second dimension to obtain the trajectory vector, and input the trajectory vector into the fully connected neural network to obtain the final output scanning trajectory matrix.

2. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 1, characterized in that: In S101, receiving an ultrasound image acquisition instruction sent by a user terminal includes: receiving a connection request sent by a user terminal through a preset network, wherein the connection request is used to request to establish a connection with the ultrasonic scanning robot; Detecting whether the current account of the user terminal is the target account; If the current account of the user terminal is the target account, a connection is made with the user terminal according to the connection request. After the connection is completed, an ultrasound image acquisition instruction sent by the user terminal is received.

3. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 2, characterized in that: The preset network includes one of 3G network, 4G network, 5G network, WIFI network or a combination thereof.

4. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 3, characterized in that: In S102, ultrasonic image data collected in the target area is obtained, including: The operator manually controls the ultrasonic scanning robot to move the ultrasonic probe to the starting point Q0 of the area to be scanned; The ultrasound image is acquired with the starting point Q0 as the starting point, and the image I(0) at the initial moment is acquired and then the key medical features are highlighted through denoising and image enhancement.

5. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 4, characterized in that: In S103, the multiple variational autoencoders are at least two, namely a first variational autoencoder and a second variational autoencoder.

6. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 5, characterized in that: The obtained balanced ultrasonic image data is placed in a multinomial variational autoencoder for image enhancement to obtain an enhanced image of the multinomial variational autoencoder, including: Input X1 and X2 into the first variational autoencoder and the second variational autoencoder respectively, where X1 and X2 are the same enhanced image after data augmentation; In the first variational autoencoder, the input X1 is a 64×64×1 grayscale image. X1 is transformed into a 4096×1 strip vector through a flattening operation. The strip vector undergoes initial global feature fusion through a fully connected neural network with a network parameter size of 1×512 to obtain an initial global feature A1 of size 4096×512. X2 is also processed in the second variational autoencoder in the same way to obtain an initial global feature A2 of size 4096×512.

7. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 6, characterized in that: In S104, multiple initial global feature matrices extracted by the fully connected neural network are concatenated with multiple latent variables in the variational autoencoder in the first dimension, including: Add a preset latent variable z to the first variational autoencoder and the second variational autoencoder respectively, where z is a 1×512 zero vector; The initial global features A1 and A2 extracted by the fully connected neural network are concatenated with their respective latent variables z in the first dimension, and the first feature matrix B1 is obtained in the first variational autoencoder, and the first feature matrix B2 is obtained in the second variational autoencoder.

8. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 7, characterized in that: In S105, the multiple first feature matrices are respectively input into the fully connected neural network to obtain multiple second feature matrices as output, and the multiple second feature matrices are respectively split, and only the multiple latent feature matrices at the end are retained, including: The first feature matrix B1 and the first feature matrix B2 are respectively input into a fully connected neural network of size 512×32 to obtain the output second feature matrix C1 and the second feature matrix C2. The size of the second feature matrix C1 and the second feature matrix C2 are both 4097×32. Then the second feature matrix C1 and the second feature matrix C2 are respectively split to obtain the latent feature matrix Z1 and the latent feature matrix Z2 at the end.

9. The ultrasonic scanning trajectory planning method based on variational autoencoder according to claim 8, characterized in that: Then, multiple latent feature matrices are concatenated in the second dimension to obtain the trajectory vector, which is then input into the fully connected neural network to obtain the final output scanning trajectory matrix, including: The latent feature matrix Z1 and the latent feature matrix Z2 are concatenated in the second dimension to obtain a vector V of size 1×64. The vector V is input into a fully connected neural network of size 64×6 to obtain the final output scanning trajectory matrix Q1 = [Sx1, Sy1, Sz1, Rx1, Ry1, Rz1], where Sx1, Sy1, and Sz1 represent the coordinates of the three spatial coordinate axes at time t = 1, and Rx1, Ry1, and Rz1 represent the postures of the end of the robot in the three posture coordinate axes at time t = 1. After the ultrasonic scanning robot moves to the trajectory coordinate Q1, the steps are repeated to perform trajectory planning for the entire scanning task until the ultrasonic scanning robot is manually stopped, and the ultrasonic scanning trajectory is obtained.

10. Ultrasonic scanning trajectory planning system based on variational autoencoder, characterized in that: Used to execute the ultrasound scanning trajectory planning method according to any one of claims 1 to 9, the ultrasound scanning trajectory planning system comprises: An acquisition module, configured to receive an ultrasound image acquisition instruction sent by a user terminal; An acquisition module, used to acquire ultrasonic image data collected from a target area; A preprocessing module is used to sequentially place the ultrasonic image data into an image preprocessing model to perform denoising on the ultrasonic image data, and obtain balanced ultrasonic image data after denoising; The feature fusion module is used to obtain the enhanced images of multiple variational autoencoders. The enhanced images of multiple variational autoencoders are subjected to initial global feature fusion through a fully connected neural network to obtain multiple initial global feature matrices; The matrix generation module is used to concatenate multiple initial global feature matrices and multiple latent variables in the variational autoencoder in the first dimension, and to split the multiple second feature matrices separately, retaining only the multiple latent feature matrices at the end; The generation module is used to concatenate multiple latent feature matrices in the second dimension to obtain a trajectory vector, input the trajectory vector into a fully connected neural network to obtain the final output scanning trajectory matrix, and repeat the steps to plan the trajectory of the entire scanning task until the ultrasonic scanning robot is manually stopped, thus obtaining the ultrasonic scanning trajectory.

Images