Method and system for automated detection of ultrasound based on ultrasound images

Abstract

The application provides a method and system for automatic ultrasonic detection based on an ultrasonic image, wherein the method steps comprise the following steps: S100, defining an ultrasonic image quality parameter and a target standard image, constructing and training a reinforcement learning network model; S200, moving an ultrasonic probe to a target position, collecting an ultrasonic image containing a detection target, and recording a corresponding pose of the ultrasonic probe; S300, calculating the ultrasonic image quality parameter obtained in the step S200, and taking the corresponding pose as an input of the reinforcement learning network model to obtain an adjusted pose of the ultrasonic probe, so as to adaptively adjust the pose of the ultrasonic probe according to the ultrasonic image quality, and obtain an ideal ultrasonic detection image.

Description

Technical Field

[0001] This invention relates to automatic control technology, and more particularly to an automatic ultrasonic detection method and system for automatically controlling the probe of an ultrasonic robot based on ultrasonic image quality. Background Technology

[0002] B-mode ultrasound, commonly known as "B-scan," is widely used in medical examinations. During the examination, the doctor typically operates the probe, observing the ultrasound image displayed on the screen and adjusting the probe's direction, position, and angle based on their experience to control the probe to the target location and obtain a high-quality image of the target area.

[0003] Therefore, the control of the ultrasound probe is highly dependent on the experience of the operating physician. However, the demand for medical resources in today's society is constantly increasing, and training an experienced physician requires a certain amount of time and resources, so the shortage of ultrasound personnel cannot be filled in the short term.

[0004] To address this, existing technologies have proposed using multi-degree-of-freedom robots to mimic the techniques used by doctors during examinations, automatically moving ultrasound probes to perform human body detection. However, the problem with such existing technologies is that, due to the different clinical experiences of different doctors, the posture adjustments when operating the probe also vary from doctor to doctor. Therefore, there is a lack of quantitative analysis and standardization, which may result in unsatisfactory automatic detection results for different subjects.

[0005] On the other hand, if the automated testing method is planned solely based on the experience of certain physicians, such a highly experience-dependent approach may not be suitable and could reduce testing efficiency. Summary of the Invention

[0006] Therefore, the main objective of this invention is to provide a method and system for automatic ultrasound detection based on ultrasound images, so as to adaptively adjust the position and orientation of the ultrasound probe according to the quality of the ultrasound image in order to obtain an ideal ultrasound detection image.

[0007] To achieve the above objectives, according to a first aspect of the present invention, a method for automatic ultrasound detection based on ultrasound images is provided, comprising the steps of:

[0008] Step S100: Define ultrasound image quality parameters and target standard image, and construct and train a reinforcement learning network model;

[0009] Step S200: Move the ultrasonic probe to the target position, acquire an ultrasonic image containing the target, and record the corresponding pose of the ultrasonic probe.

[0010] Step S300: Calculate the ultrasound image quality parameters obtained in step S200, and use their corresponding poses as input to the reinforcement learning network model to obtain the adjustment pose of the ultrasound probe.

[0011] In a possible preferred embodiment, step S100, the step of defining the ultrasound image quality parameters and the target standard image, includes:

[0012] Step S110: Define the ultrasound image quality parameter Q based on the edge and spatial features of the ultrasound image.

[0013] ;

[0014] in and The grayscale difference between the inside and outside of the target tissue in the obtained ultrasound image and the preset target standard image are respectively, and S is the image integrity of the target tissue;

[0015] Step S111: Obtain several high-quality ultrasound images of the target tissue, calculate the mean values ​​of grayscale difference and image integrity as standard values ​​for calculation, and select the image closest to the mean value as the target standard image.

[0016] In a possible preferred embodiment, step S100, the step of constructing and training the reinforcement learning network model, includes:

[0017] Step S120: Construct an Actor network, taking the environmental state, including the ultrasound image quality parameter Q and the corresponding ultrasound probe pose, as input to define the corresponding pose transformation vector H for the output; Construct a Critic network, taking the pose transformation vector H and the environmental state as input to define the reward value r for evaluating the quality of the pose transformation vector H.

[0018] Step S121: Define the loss function:

[0019] ;

[0020] in, This represents the predicted ultrasound image quality Q value after the ultrasound probe pose change. The Q value represents the ultrasound image quality obtained after the actual execution of the pose transformation vector H. This is the root mean square error;

[0021] Step S122: Train the reinforcement learning network model: Perform gradient descent operation based on the loss function to update the network parameters, repeating the process several times, or based on the quality of the obtained ultrasound images. Training ends when the value exceeds the target standard image.

[0022] In a possible preferred embodiment, step S122 further includes the steps of: setting an experience replay pool, and the first few rounds are experience replay pool filling steps, without training; during the process of filling the experience replay pool, in the first few rounds, a teaching-based approach is adopted to collect sample data under the constraint of the robotic arm's movement space to fill the experience replay pool.

[0023] In a possible preferred embodiment, the step of filling the experience playback pool includes:

[0024] Step S130: Based on the ultrasound probe pose corresponding to the target standard image as the center point, establish a planar dot matrix with a preset distance as the unit, record the pose when moving the ultrasound probe to each point as a round sample, and fill the sample into the experience playback pool.

[0025] Step S131 Record the ultrasound image quality Q value under each grid, select the grids with the highest scores, and rotate the ultrasound probe with the pose of each grid as the starting point. Record the ultrasound image quality Q value sample every 60° clockwise rotation, for a total of five rotations, so as to fill the obtained samples into the experience playback pool.

[0026] In a possible preferred embodiment, the step of filling the experience playback pool includes:

[0027] Step S132: Set the filtering parameters Determine the sample reward value after the first few rounds, and when it reaches the average reward value in the experience replay pool... When the sample size is increased by a factor of 1, it is considered a valid sample and added to the experience replay pool.

[0028] In a possible preferred embodiment, the step of filling the experience playback pool includes:

[0029] Step S133 Define the attenuation factor Where t is the number of training iterations, and the filtering parameters are defined. ; Determine the sample reward value after the first few rounds, and when it reaches the average reward value in the experience replay pool. When the sample size is increased by a factor of 1, it is considered a valid sample and added to the experience replay pool.

[0030] To achieve the above objectives, according to a second aspect of the present invention, a system for automatic ultrasonic detection based on ultrasonic images is also provided, comprising:

[0031] The storage unit is used to store a program including the method steps of automatic ultrasound detection based on ultrasound images as described above, so that the control unit and the processing unit can retrieve and execute it in a timely manner.

[0032] The control unit is used to control the robotic arm to move the ultrasonic probe carrying the ultrasonic unit to the target position, acquire the ultrasonic image containing the target to be detected, and record the corresponding pose of the ultrasonic probe.

[0033] The processing unit is used to calculate the quality parameters of the acquired ultrasound image and use the corresponding pose as input to the reinforcement learning network model to obtain the adjustment pose of the ultrasound probe.

[0034] The method and system for automatic ultrasonic detection based on ultrasonic images provided by this invention realize servo feedback control of the robotic arm posture based on ultrasonic images, thus solving the problem of adaptive control of the ultrasonic robot for the probe posture.

[0035] This allows the solution of the present invention to no longer rely on the physician's experience and techniques, but instead uses the quality of the obtained ultrasound image as an objective indicator, thereby making the final detection effect tend to obtain an ideal ultrasound detection image, thus improving the actual detection efficiency and practical value of this type of automatic detection solution. Attached Figure Description

[0036] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0037] Figure 1 This is a schematic diagram of the steps of the method in the first embodiment of the present invention;

[0038] Figure 2 This is a logical schematic diagram of the method in the first embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the reinforcement learning network Actor-Critic model of the method in the first embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of a planar dot matrix of samples screened in the experience playback pool according to the first embodiment of the present invention.

[0041] Figure 5 This is a schematic diagram of the rotational training sampling of the experience replay pool screening samples in the first embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram of the system structure according to the second embodiment of the present invention. Detailed Implementation

[0043] To enable those skilled in the art to better understand the technical solutions of the present invention, the specific technical solutions of the present invention will be clearly and completely described below in conjunction with embodiments, so as to help those skilled in the art further understand the present invention. Obviously, the embodiments described in this application are merely some embodiments of the present invention, and not all embodiments. It should be noted that, for those skilled in the art, the embodiments and features in the embodiments of this application can be combined with each other without departing from the concept of the present invention and without conflict. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the disclosure and protection scope of the present invention.

[0044] Furthermore, the terms "first," "second," "S1," "S2," etc., used in the specification, claims, and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those described herein. At the same time, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. Unless otherwise expressly specified and limited, the terms "set," "arranged," "installed," "connected," and "linked" should be interpreted broadly, for example, as a fixed connection, a detachable connection, or an integral connection; a mechanical connection or an electrical connection; a direct connection or an indirect connection through an intermediate medium; or a connection within two elements. Those skilled in the art can understand the specific meaning of the above terms in this case based on the specific circumstances and in conjunction with existing technology.

[0045] like Figures 1 to 5 As shown, in order to achieve adaptive adjustment of the ultrasound probe's pose based on the ultrasound image quality to obtain an ideal ultrasound detection image, this invention provides a method for automatic ultrasound detection based on ultrasound images, the steps of which include:

[0046] Step S100: Define ultrasound image quality parameters and target standard image, and construct and train a reinforcement learning network model.

[0047] Specifically, addressing the issue that image evaluation in the medical field primarily relies on subjective assessment by physicians and lacks quantitative description, this invention defines an image quality parameter. This parameter is defined based on the edge features (echo intensity at tissue edges and internal grayscale information) and spatial features (relative positional relationships of tissues) of ultrasound images. ,in and Let S be the grayscale difference between the inside and outside of the target tissue in the acquired image and the standard image, respectively, and S be the image integrity of the target tissue. Acquire 4-5 standard images, calculate the average of the grayscale difference and image integrity as standard values, and select the image closest to the average as the final standard image. For the final selected standard image, let its image quality be... .

[0048] Using objective indicators to quantitatively analyze ultrasound images effectively reduces the reliance on subjective judgment by doctors, avoiding the influence of factors such as display equipment, the observer's visual condition, and emotional state. Image quality assessment based on raw images is computationally simple and fast.

[0049] Therefore, step S100, which defines the ultrasound image quality parameters and the target standard image, includes:

[0050] Step S110: Define the ultrasound image quality parameter Q based on the edge and spatial features of the ultrasound image.

[0051] ;

[0052] in and The values ​​represent the grayscale difference between the inside and outside of the target tissue in the obtained ultrasound image and the preset target standard image, respectively, and S represents the image integrity of the target tissue.

[0053] Step S111: Obtain several high-quality ultrasound images of the target tissue, calculate the mean values ​​of grayscale difference and image integrity as standard values ​​for calculation, and select the image closest to the mean value as the target standard image.

[0054] Furthermore, in this embodiment, the step of constructing and training the reinforcement learning network model includes:

[0055] Step S120: Construct an Actor network, taking the environmental state, including the ultrasound image quality parameter Q and the corresponding ultrasound probe pose, as input to define the corresponding pose transformation vector H for the output; construct a Critic network, taking the pose transformation vector H and the environmental state as input to define the reward value r for evaluating the quality of the pose transformation vector H.

[0056] In network model training:

[0057] Forward propagation: Set the current end effector pose of the robotic arm to... The ultrasound image obtained in this state is input into the Actor network to obtain the pose transformation vector of the robotic arm's end effector. Where x, y, and z represent the translations of the robotic arm's end effector along the X, Y, and Z axes of the base coordinate system, respectively, and rx, ry, and rz are the angles between the tool coordinate system and the base coordinate system along the X, Y, and Z axes, respectively. Based on the decisions made by the policy network, the pose of the robotic arm's end effector is transformed to a new state. Obtain again Ultrasound images under certain conditions. The pose transformation vector H... The ultrasound image under the given conditions is input into the Critic network to obtain a reward value. ,in , The robotic arm end effector is located at... , The quality parameter Q of the ultrasound image in the state.

[0058] Furthermore, it should be noted that in this invention, the ultrasonic probe pose and the robotic arm end effector pose can be understood as the same pose.

[0059] Backpropagation: Step S121 Define the Huber loss function:

[0060] ;

[0061] in, This represents the predicted ultrasound image quality Q value after the ultrasound probe pose change. The Q value represents the ultrasound image quality obtained after the actual execution of the pose transformation vector H. This is the root mean square error.

[0062] Step S122: Train the reinforcement learning network model: Perform gradient descent operation based on the loss function to update the network parameters, repeating the process several times, or based on the quality of the obtained ultrasound images. Training ends when the value exceeds the target standard image.

[0063] For example, for existing standard images, their image quality... Set as a threshold, when Exceed Training ends when the obtained image meets the requirements for ultrasound image quality, or when the required number of training sessions is met.

[0064] Furthermore, step S122 also includes the steps of: setting up an experience replay pool, and the first few rounds are for filling the experience replay pool without training; during the process of filling the experience replay pool, in the first few rounds, a teaching-based approach is used to collect sample data under the constraint of the robotic arm's movement space to fill the experience replay pool.

[0065] The steps for filling the experience replay pool include:

[0066] Step S130: Based on the ultrasound probe pose corresponding to the target standard image as the center point, establish a planar dot matrix with a preset distance as the unit, record the pose when moving the ultrasound probe to each point as a round sample, and fill the sample into the experience playback pool.

[0067] Step S131 Record the ultrasound image quality Q value under each grid, select the grids with the highest scores, and rotate the ultrasound probe with the pose of each grid as the starting point. Record the ultrasound image quality Q value sample every 60° clockwise rotation, for a total of five rotations, so as to fill the obtained samples into the experience playback pool.

[0068] Specifically, since the robotic arm explores a large space in reinforcement learning, resulting in high computational complexity, this invention simplifies the pose control of the robotic arm. At the same time, during the filling of the experience replay pool, a teaching-based approach is used to constrain the movement space of the robotic arm for the first 50 rounds.

[0069] For example, in this training process, the experience replay pool size is designed to be 5000. The first 50 rounds are for filling the experience replay pool, and no training is performed. Regarding the pose state of the robotic arm, the more joints there are, the more complex the possible state space obtained by the permutation and combination of the rotation transformations of each joint becomes. For a commonly used six-degree-of-freedom robotic arm, even if the joint angles are discretized at 10° intervals, each joint has 36 possible states. As the degrees of freedom increase, the final number of possible state spaces is in the tens of millions.

[0070] For robotic arms, it's not enough to just consider joint rotation; the length of each segment must also be calculated to prevent collisions, which introduces a huge computational burden. Since reinforcement learning-based probe control methods are used for fine-tuning, in coarse-tuning, the robotic arm controls the probe's movement to center the target in the image. The moved ultrasound probe is now near the actual detection location. Subsequent fine-tuning aims to obtain a clearer ultrasound image; the range of posture changes is small, and the robotic arm doesn't need excessive exploration space.

[0071] Considering practical needs, we chose to separate the rotation of the robotic arm's end joint from other joints, treating the remaining joints as a whole and only calculating the changes in three-dimensional coordinates. This greatly reduced the amount of state space and eliminated redundant pose states, resulting in reduced computation and improved training efficiency.

[0072] To further improve learning efficiency while avoiding increased computational complexity due to increased training iterations and a larger experience replay pool, the replay strategy is improved. When filling the experience replay pool, spatial exploration is performed based on the robotic arm pose corresponding to existing standard images. Through teaching, the robotic arm is controlled to reach the standard pose, and coordinate and rotation angle changes are made based on the secondary pose to obtain higher-quality samples.

[0073] like Figure 4 As shown, assuming that during the teaching-based filling of the experience playback pool, 25 target locations (5x5) are generated around the current position in 1cm increments on the two-dimensional xy-plane. The robotic arm is moved to each location, and each move to a location is considered a round sample, which is then added to the experience playback pool. The ultrasound image quality obtained at each location is recorded. The five positions with the highest scores were selected, and the end was rotated 60° clockwise from the current pose as one sample, for a total of five rotations. Each of the five selected positions underwent one rotation in the first 25 rounds, forming the last 25 sample data based on the teaching method. This reinforcement learning based on teaching can significantly reduce the exploration space and effectively improve the training rate.

[0074] like Figure 4 As shown, The point represents the xy-plane position of the robotic arm's end effector at the moment the coarse adjustment is completed, immediately after the network training begins filling the experience backdoor pool. Centered on the x and y axes, generate in 1cm increments along the positive and negative directions. to There are 25 coordinate points in total. The robotic arm moves to these points sequentially, and performs one round of training at each point, generating data samples that are added to the experience replay pool.

[0075] After the first 25 rounds, the ultrasound image quality corresponding to each location was obtained. ,like Figure 4 As shown, the one with the highest Q value is , , , , Five points. Move the end effector of the robotic arm to... At this position, the end effector is at 0°, which is the pose of the robotic arm during the second round of training. For example... Figure 5 As shown, the robotic arm's end effector is rotated 60° clockwise, and this is recorded as the 26th training iteration, added to the experience replay pool. The end effector continues to rotate clockwise, with each 60° rotation constituting one training iteration, and the sample is added to the experience replay pool. After five rotations, training iterations 26 to 30 are completed, including iteration 2. The robotic arm's end effector... The point generates data in 6 rounds and completes one rotation.

[0076] The same steps were performed at the remaining four points, with an additional 5 rounds of samples recorded at each location. Finally, the selected five points underwent training for 25 rounds, from the 26th to the 50th, through end-rotation, and were then filled into the experience replay pool.

[0077] After the first 50 rounds of teaching-based exploration, subsequent samples need to be screened before being added to the experience replay pool. Therefore, the steps for filling the experience replay pool include:

[0078] Step S132: Set the filtering parameters Determine the sample reward value after the first few rounds, and when it reaches the average reward value in the experience replay pool... When the sample size is increased by a factor of 1, it is considered a valid sample and added to the experience replay pool.

[0079] Therefore, the reward value of subsequent samples needs to reach the average reward value in the experience replay pool. Only samples that are multiples of the original sample size are considered valid and added to the experience replay pool. Samples selected through this process are of higher quality, resulting in better training stability.

[0080] However, as the number of training iterations increases and the sample quality gradually improves, more and more samples will fail to meet the standards and be rejected, resulting in an increase in the number of samples that cannot enter the replay pool. Therefore, the steps for filling the experience replay pool also include:

[0081] Step S133 Define the attenuation factor Where t is the number of training iterations, and the filtering parameters are defined. ; Determine the sample reward value after the first few rounds, and when it reaches the average reward value in the experience replay pool. When the sample size is increased by a factor of 1, it is considered a valid sample and added to the experience replay pool.

[0082] As the number of training iterations increases, the selection parameter gradually decreases until it decays to 1.

[0083] Step S200: Move the ultrasonic probe to the target position, acquire an ultrasonic image containing the target, and record the corresponding pose of the ultrasonic probe.

[0084] Specifically, the location of the patient and lesion is first roughly determined by a vision camera. Combined with the robotic arm posture and pressure sensor data corresponding to the standard image, the control unit controls the robotic arm to move to the corresponding position and manipulates the ultrasound probe to acquire images.

[0085] The generated ultrasound image is transmitted to a processing unit such as a computer. The image comparison and judgment section analyzes the image and performs a rough adjustment of the probe's pose. The analysis process is as follows: First, the acquired image is filtered to remove noise. Then, edge detection is performed based on the image's grayscale values. A known standard image is used for identification. If the target can be identified, the direction and distance the ultrasound probe needs to move to the target detection position can be calculated by comparing the translational and rotational relationship between the target and the standard image. Subsequently, this information is sent to the position correction parameter section, which, combined with the current end-effector pose information of the robotic arm, calculates and converts it into the target pose that the robotic arm's end-effector needs to move to.

[0086] The target pose is then sent to the robotic arm, which moves the ultrasonic probe to the target position. The above steps are repeated for adjustment until the ultrasonic probe 5 moves to the corresponding target position. If the obtained image cannot be recognized, the robotic arm is controlled to move randomly in all directions to re-image until the target is recognized in the image, and then position correction and other operations are performed.

[0087] When the actual position of the target being detected differs from the predicted position in the ultrasound image by less than a set threshold, the probe pose can be finely adjusted through a trained reinforcement learning model to ensure high-quality output images.

[0088] Step S300: Calculate the ultrasound image quality parameters obtained in step S200, and use their corresponding poses as input to the reinforcement learning network model to obtain the adjustment pose of the ultrasound probe.

[0089] Therefore, this invention first performs a coarse adjustment to center the target in the ultrasound image. Based on a pre-obtained standard image and the corresponding robotic arm pose, the robotic arm is controlled to reach the appropriate position. Due to individual differences in the target object, this pose will have a certain error compared to the actual target pose. By matching and comparing the actual acquired image with the standard image, the probe is controlled to move, ensuring the target is centered in the ultrasound image, thus completing the coarse adjustment.

[0090] Since coarse adjustment only roughly adjusts the probe position based on whether the detection object is identified in the image and whether it is in the center of the image, it will have some errors and cannot guarantee that the ultrasound image quality meets the needs of subsequent diagnosis by doctors. Therefore, reinforcement learning is used to fine-tune the probe pose. The optimal solution to the problem can be obtained through reinforcement learning.

[0091] The Actor-Critic algorithm is used for model training. In reinforcement learning, the agent is defined as the ultrasound probe, the state as its current pose, the interactive environment as the quality of the ultrasound image (including occlusion and sharpness), the action as the probe's next position, and the reward as the difference in image quality parameters between the current and previous states. When the ultrasound probe is in a certain state, it analyzes the current ultrasound image and makes decisions, including rotating, tilting, and moving the probe. After making a decision, the environment information is updated, and a reward is given based on this decision. Based on feedback and environmental analysis results, the ultrasound probe continues to make decisions, forming a loop to continuously learn how to control its pose and thus acquire high-quality ultrasound images that meet the requirements.

[0092] By setting thresholds based on standard images, the image quality is guaranteed to meet requirements, while reducing the time and resources consumed in calculating the optimal solution, thus lowering computational complexity and improving control speed.

[0093] On the other hand, this invention addresses the problem of excessive exploration space and high computational complexity in reinforcement learning for robotic arms by simplifying the pose control of the robotic arm. At the same time, during the process of filling the experience replay pool, a teaching-based approach is used to constrain the movement space of the robotic arm for the first 50 rounds.

[0094] Since the robotic arm has already controlled the ultrasonic probe to be near the target during the coarse pose adjustment process, the purpose of the subsequent fine adjustment is to compensate for the error caused by the coarse adjustment and to control the probe to obtain a clear image. Therefore, there is no need to generate excessive pose adjustment. So, in the fine adjustment process based on reinforcement learning, the control of the robotic arm is simplified. The multiple joint states of the robotic arm are divided into the rotation of the end effector and the overall three-dimensional coordinate change of the remaining segments, making the state space explored by the robotic arm more effective.

[0095] During the teaching-based filling of the experience playback pool, to compensate for errors caused by coarse adjustments, 25 target positions (5x5 grids) are generated around the current position in 1cm increments on the two-dimensional xy-plane. The robotic arm is moved to each position, and each move to a position is considered a round sample, which is then added to the experience playback pool. The ultrasound image quality obtained at each position is recorded. Select the five positions with the highest scores, and record each 60° clockwise rotation of the end from the current posture as a sample, for a total of five rotations.

[0096] The five positions selected in the first 25 rounds are each rotated once, forming the last 25 sample data based on the teaching method. This reinforcement learning based on teaching can greatly reduce the exploration space and effectively improve the training rate.

[0097] Corresponding to the above method embodiments, such as Figure 6 As shown, in a second aspect, the present invention also provides a system for automatic ultrasound detection based on ultrasound images, comprising:

[0098] The storage unit is used to store a program including the method steps of automatic ultrasound detection based on ultrasound images as described above, so that the control unit and the processing unit can retrieve and execute it in a timely manner.

[0099] The control unit is used to control the robotic arm to move the ultrasonic probe carrying the ultrasonic unit to the target position, acquire the ultrasonic image containing the target to be detected, and record the corresponding pose of the ultrasonic probe.

[0100] The processing unit is used to calculate the quality parameters of the acquired ultrasound image and use the corresponding pose as input to the reinforcement learning network model to obtain the adjustment pose of the ultrasound probe.

[0101] In summary, the method and system for automatic ultrasonic detection based on ultrasonic images provided by this invention realize servo feedback control of the robotic arm's pose based on ultrasonic images, thus solving the problem of adaptive control of the probe pose by the ultrasonic robot.

[0102] This allows the solution of the present invention to no longer rely on the physician's experience and techniques, but instead uses the quality of the obtained ultrasound image as an objective indicator, thereby making the final detection effect tend to obtain an ideal ultrasound detection image, thus improving the actual detection efficiency and practical value of this type of automatic detection solution.

[0103] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The present invention is limited only by the claims and their full scope and equivalents. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

[0104] Those skilled in the art will understand that, besides implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0105] Furthermore, all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0106] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A method for ultrasound automated detection based on ultrasound images, characterized by the steps of Comprise: Step S100 defines the ultrasound image quality parameters and target standard image, constructs and trains the reinforcement learning network model, the steps include: step S110 defines the ultrasound image quality parameters Q according to the edge features and spatial features of the ultrasound image: ； wherein and respectively are the internal and external gray scale differences of the target tissue in the obtained ultrasound image and the preset target standard image, and S is the image integrity of the target tissue display. Step S111 obtains several ultrasound images of high-quality target tissues, calculates the mean value of the gray difference and image integrity as the standard value for calculation, and selects the image closest to the mean value as the target standard image; Step S120 constructs an Actor network, takes the environment state including the ultrasound image quality parameters Q and the corresponding ultrasound probe pose as input, and defines the output corresponding pose transformation vector H; a Critic network is constructed, and the pose transformation vector H and the environment state are taken as input, and the reward value r for evaluating the quality of the pose transformation vector H is defined; Step S121 defines the loss function: ； wherein, is the predicted ultrasound image quality Q value after the ultrasound probe pose transformation, is the ultrasound image quality Q value obtained after the real execution of the pose transformation vector H, is the root mean square error; Step S122 trains the reinforcement learning network model: gradient descent operation is performed according to the loss function, the network parameters are updated, and the loop is performed for a plurality of times or the obtained ultrasound image quality value exceeds the target standard image, the training is ended; Step S200 moves the ultrasound probe to the target position, collects the ultrasound image containing the detection target, and records the corresponding pose of the ultrasound probe; Step S300 calculates the ultrasound image quality parameters obtained in step S200, and takes the corresponding pose as the input of the reinforcement learning network model to obtain the adjustment pose of the ultrasound probe.

2. The method of claim 1, wherein, In step S122, the step of setting an experience replay pool is also included, and the initial several rounds are experience replay pool filling steps, and no training is performed; in the process of filling the experience replay pool, the sample data is collected under the constraint of the movement space of the mechanical arm in the initial several rounds based on teaching to fill the experience replay pool.

3. The method of claim 2, wherein, The step of filling the experience replay pool includes: Step S130 takes the ultrasound probe pose corresponding to the target standard image as the center point, establishes a plane point array with a preset distance as the unit, records the pose of the moving ultrasound probe to each point as a round sample, and fills the sample into the experience replay pool; Step S131 records the ultrasound image quality Q value under each grid respectively, selects several grids with the highest scores, respectively rotates the ultrasound probe with the pose corresponding to each grid as the starting point, records the ultrasound image quality Q value sample once every 60 degrees of clockwise rotation, and rotates five times to fill the obtained sample into the experience replay pool.

4. The method of claim 3, wherein, The step of filling the experience replay pool includes: Step S132 sets the screening parameter The sample reward value after the first several rounds is determined, and when it reaches the average reward value in the experience replay pool times, it is considered an effective sample and filled into the experience replay pool.

5. The method of claim 3, wherein, The step of filling the experience replay pool includes: Step S133 defines the attenuation factor where t is the number of training, defines the screening parameter ; determine the sample reward value after the first several rounds, when it reaches the average reward value in the experience replay pool times, as an effective sample filled in the experience replay pool.

6. System for ultrasound automatic detection based on ultrasound images, characterized in that Comprise: A storage unit for storing programs including the steps of the method for automatic ultrasound detection based on ultrasound images according to any one of claims 1 to 5, for the control unit, the processing unit, to execute in time; A control unit for controlling the mechanical arm carrying the ultrasound probe of the ultrasound unit to move to the target position, collect the ultrasound image containing the detection target, and record the corresponding pose of the ultrasound probe; A processing unit for calculating the obtained ultrasound image quality parameters, and taking the corresponding pose as the input of the reinforcement learning network model to obtain the adjustment pose of the ultrasound probe.

Images