Orthopedic surgery imitation learning system oriented to sparse data and small samples

Abstract

According to the sparse data and small sample-oriented orthopedic surgery imitation learning system provided by the invention, imitation learning is combined with data interpolation and data prediction, complete and effective demonstration data can be generated according to small samples and sparse demonstration data, and actions of a robot are regenerated through an imitation learning algorithm; operation precision of the robot is improved, and burden of doctors is relieved; therefore, the problem that the simulation learning precision is reduced due to the fact that data loss inevitably exists in the collection process of demonstration data can be solved, and the defect that the operation time is greatly prolonged due to the fact that a large number of demonstration data samples are needed for learning in a traditional simulation learning algorithm is overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of imitation learning, and in particular relates to an orthopedic surgery imitation learning system for sparse data and small samples. Background Art

[0002] Robots can achieve precise positioning and rapid response, making them widely used in the medical field. In orthopedic surgery, there are some relatively simple and repetitive operations, which consume the doctor's physical strength. Imitation learning can greatly reduce the burden on doctors. In order for robots to imitate doctors in performing surgical operations, a large number of data samples need to be provided for the robot to learn. However, before the operation, doctors cannot be given enough time to demonstrate the operation, and the data samples of the doctor's operation demonstration may be lost. Based on this, a method that allows doctors to provide small samples of demonstration data and can still be used for robot imitation learning when the data is sparse has emerged.

[0003] In the prior art, in order to enable a robot to perform a certain operation, the commonly used method is to use teaching programming or offline programming, etc., but this method has poor environmental adaptability and cannot be used in non-structural surgical environments. Imitation learning is a method that can quickly learn operating skills from the teaching data provided by the instructor. Commonly used imitation learning algorithms include dynamic motion primitives (DMP), Gaussian mixture models (GMM), kernel motion primitives (KMP), probabilistic motion primitives (ProMP), etc. These algorithms are usually used in industry. The algorithm for imitation learning during surgery is Surgical Robot Transformer, but this algorithm is only applicable to laparoscopic surgery and requires a large number of learning samples. Therefore, this method is not applicable to imitation learning in orthopedic surgery. Summary of the invention

[0004] To solve the above problems, the present invention provides an orthopedic surgery imitation learning system for sparse data and small samples, which can improve the operation accuracy of the robot and reduce the burden on doctors.

[0005] An orthopedic surgery imitation learning system for sparse data and small samples, including a data acquisition module, an imitation learning module, a motion vector acquisition module, a node mapping module, and a motion control module;

[0006] The data acquisition module is used to construct a small sample data set based on the actual trajectory of the surgical instrument held by the doctor within a set time period, wherein each sample consists of the trajectory position information x0, speed information x1 and x2 of the surgical instrument at each moment. and the actual operating force F0 between the surgical instrument and the bone tissue;

[0007] The imitation learning module is used to perform trajectory position information x0 and speed information x1 of the surgical instrument in the small sample data set according to the Gaussian mixture regression method. and the actual operating force F0 to obtain the probability distribution of the operating force F used to characterize the mapping relationship between the robot motion trajectory and the operating force;

[0008] The motion vector acquisition module is used to obtain the starting point P of the robot motion trajectory represented by the operating force F according to the probability distribution of the operating force F. s Point to the next end point P e Motion vector

[0009] The node mapping module is used to determine the initial position of the robot specified by the doctor for the bone tissue to be removed. and the motion vector at the next moment Get the next end point of the actual resection trajectory of the bone tissue to be resected

[0010] The motion control module is used to adjust the initial position and the next moment end Control the robot's movements so that it can remove the bone tissue according to the set actual removal trajectory.

[0011] Furthermore, an orthopedic surgery imitation learning system for sparse data and small samples also includes a data processing module;

[0012] The data processing module is used to use a data interpolation algorithm to fill in missing data in a small sample data set, then use a time series prediction algorithm to extend the time period covered by the small sample data set after the missing data filling is completed, and finally use a time alignment algorithm to time align the small sample data set after the time extension is completed, so as to obtain a time-continuous small sample data set for the imitation learning module to perform imitation learning.

[0013] Furthermore, the next moment end point of the actual resection trajectory of the bone tissue to be resected The method to obtain is:

[0014]

[0015] Among them, R is the set scaling factor.

[0016] Furthermore, the data acquisition module acquires the trajectory position information x and speed information of the surgical instrument. And the method of the operating force F between the surgical instrument and the bone tissue is:

[0017] Optical navigation markers and force sensors are mounted on surgical instruments, and data collection is accomplished by direct demonstration by the surgeon holding the surgical instruments.

[0018] Furthermore, the data acquisition module acquires the trajectory position information x and speed information of the surgical instrument. And the method of the operating force F between the surgical instrument and the bone tissue is:

[0019] The surgical instruments are installed on the robotic arm, and the doctor drags the robotic arm or remotely operates it for demonstration, thereby completing data collection.

[0020] Furthermore, the probability distribution of the operating force F is obtained as follows:

[0021] According to the actual operating force F0, the small sample data set is divided into K segments of sample data, and each segment of sample data corresponds to a Gaussian distribution;

[0022] The occurrence probability p(F0) of each actual operation force F0 in the small sample data set is constructed by the Gaussian mixture model as follows:

[0023]

[0024] in, is the prior probability of the kth Gaussian distribution of the small sample data set, and μ k is the mean value of the kth Gaussian distribution of the small sample data set; Σ k is the covariance matrix of the kth Gaussian distribution of the small sample data set; is the probability density function of the kth Gaussian distribution of the small sample data set;

[0025] Obtain the unknown parameters in the probability of occurrence p(F0) through the expectation maximization algorithm μ k ,Σ k ;

[0026] Gaussian mixture regression μ k ,Σ k The determined occurrence probability p(F0) is regressed to obtain the probability distribution p(F|x) of any operating force F:

[0027]

[0028] Among them, p(k|x) is the prior probability of the kth Gaussian distribution with any robot motion trajectory position information x as the input variable, is the probability density function of the kth Gaussian distribution with any robot motion trajectory position information x as input variable, is the mean of the kth Gaussian distribution with any robot motion trajectory position information x as input variable, is the covariance matrix of the kth Gaussian distribution with any robot motion trajectory position information x as input variable.

[0029] Beneficial effects:

[0030] The present invention provides an orthopedic surgery imitation learning system for sparse data and small samples, which combines imitation learning with data interpolation and data prediction, and can generate complete and effective demonstration data based on small samples and sparse demonstration data, and regenerate the robot's movements through an imitation learning algorithm, thereby improving the robot's operating accuracy and reducing the burden on doctors; therefore, the present invention can solve the problem of reduced imitation learning accuracy due to inevitable data loss during the collection of demonstration data, and overcome the defect that traditional imitation learning algorithms require a large number of demonstration data samples for learning, which greatly prolongs the operation time. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A principle block diagram of an orthopedic surgery imitation learning system for sparse data and small samples provided by the present invention;

[0032] Figure 2 A flow chart of an orthopedic surgery imitation learning method for sparse data and small samples provided by the present invention;

[0033] Figure 3 A new task effect diagram generated by the imitation learning provided by the present invention. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0035] Traditional imitation learning algorithms require a large number of demonstration data samples for learning, which will greatly prolong the operation time. At the same time, due to the inevitable data loss during the collection of demonstration data, the accuracy of imitation learning may be reduced. Therefore, the present invention combines imitation learning with data interpolation and data prediction, generates complete and effective demonstration data from small samples and sparse demonstration data, and regenerates the actions through the imitation learning algorithm.

[0036] like Figure 1 As shown, an orthopedic surgery imitation learning system for sparse data and small samples includes a data acquisition module, a data processing module, an imitation learning module, a motion vector acquisition module, a node mapping module, and a motion control module;

[0037] The data acquisition module is used to construct a small sample data set based on the actual trajectory of the surgical instrument held by the doctor within a set time period, wherein each sample consists of the trajectory position information x0, speed information x1 and x2 of the surgical instrument at each moment. And the actual operating force F0 between the surgical instrument and the bone tissue.

[0038] For example, the doctor demonstrates the operation with the corresponding surgical equipment for 2-5 minutes before the operation, and records the position information x0 and speed information of the doctor's handheld instrument operation trajectory. Together with the operation force information F0, a small sample data set is formed. To record the operation trajectory and operation force, the optical navigation marker and force sensor can be installed on the surgical instrument, and the doctor can hold the surgical instrument to directly collect demonstration data; or the surgical instrument can be installed on a robotic arm, and the doctor can drag the robotic arm or remotely operate to collect demonstration data.

[0039] The data processing module is used to use a data interpolation algorithm to fill in missing data in a small sample data set, then use a time series prediction algorithm to extend the time period covered by the small sample data set after the missing data filling is completed, and finally use a time alignment algorithm to time align the small sample data set after the time extension is completed, so as to obtain a time-continuous small sample data set for the imitation learning module to perform imitation learning.

[0040] It should be noted that due to the disconnection of sensor communication or the obstruction of optical navigation markers during the doctor's demonstration, the demonstration data at a certain moment may be lost, so the data interpolation algorithm is used to fill the missing data. Since the doctor's demonstration time is short, only 2-5 minutes, and the actual working time may be much longer than the demonstration time, the time series prediction algorithm is used to predict the demonstration data of a small sample to generate the data length required for the actual working time. During the data collection process, the doctor cannot guarantee the consistency of the start time and time length in the collection of demonstration data from different groups, so the time alignment algorithm is used to find the best match between different groups of data to achieve time alignment. It is worth noting that there are many types of algorithms for data interpolation algorithms, time series prediction algorithms, and time alignment algorithms that can achieve the above functions, including but not limited to:

[0041] 1. Data interpolation algorithms: principal component analysis interpolation, Bayesian interpolation, random forest interpolation, K nearest neighbor interpolation, mean / median / mode interpolation, etc.

[0042] 2. Time series prediction algorithms: Long Short-Term Memory Network (LSTM), TimesNet, Transformer, etc.

[0043] 3. Time alignment algorithms: dynamic time warping (DTW), cross correlation, phase synchronization, feature alignment, etc.

[0044] The imitation learning module is used to perform trajectory position information x0 and speed information x1 of the surgical instrument in the small sample data set according to the Gaussian mixture regression method. The mapping relationship between the robot motion trajectory and the actual operating force F0 is imitated and learned to obtain the probability distribution of the operating force F used to characterize the mapping relationship between the robot motion trajectory and the operating force.

[0045] The motion vector acquisition module is used to obtain the starting point P of the robot motion trajectory represented by the operating force F according to the probability distribution of the operating force F. s Point to the next end point P e Motion vector

[0046] The node mapping module is used to determine the initial position of the robot specified by the doctor for the bone tissue to be removed. and the motion vector at the next moment Get the next end point of the actual resection trajectory of the bone tissue to be resected Among them, R is the set scaling factor.

[0047] The motion control module is used to adjust the initial position and the next moment end Control the robot's movements so that it can remove the bone tissue according to the set actual removal trajectory.

[0048] It can be seen that the present invention aims to map the robot motion trajectory and the initial point of the operating force of the new bone tissue resection task to the target surgical position through node mapping, and move the robot to the target surgical position. In order to adapt to the unstructured environment, it is necessary to ensure that the surgical position arbitrarily selected by the doctor can be used as the starting point of the new task trajectory and force data. Therefore, the present invention proposes a node mapping method that can map the data generated by imitation learning to the real surgical environment. The doctor then selects a suitable position as the initial position of the robot operation In the actual operation process, the trajectory information or force information has a certain method ratio, so the design scale factor is K, then the trajectory position and operation force information end point of the actual operation are

[0049] It should be noted that the mapping of arbitrary trajectory position and operation force information can also be achieved through this method, but the premise is that the relative relationship of any point in the task trajectory position and force information generated by imitation learning to the starting point / end point needs to be determined. The surgical robot begins to perform the surgical operation according to the task trajectory and force obtained by imitation learning. After mapping the task trajectory and force information of imitation learning to the surgical environment, the robot begins to execute according to the established trajectory and force. The relevant control algorithm can be MPC, LQR, etc.

[0050] Furthermore, the motion vector acquisition module re-uses the end point P at the previous moment e As a new starting point s , calculate the motion vector at the next moment The node mapping module then calculates the motion vector at the next moment. Get the end point of the next moment And so on, until the bone tissue is completely removed.

[0051] The following is a detailed description of the method for the imitation learning module to obtain the probability distribution of the operating force F:

[0052] According to the actual operating force F0, the small sample data set is divided into K segments of sample data, and each segment of sample data corresponds to a Gaussian distribution;

[0053] The occurrence probability p(F0) of each actual operation force F0 in the small sample data set is constructed by Gaussian mixture model (GMM) as follows:

[0054]

[0055] in, is the prior probability of the kth Gaussian distribution of the small sample data set, and μ k is the mean value of the kth Gaussian distribution of the small sample data set; Σ k is the covariance matrix of the kth Gaussian distribution of the small sample data set; is the probability density function of the kth Gaussian distribution of a small sample data set, and the probability density function can be calculated by the distance from the data point to the mean value:

[0056]

[0057] Obtain the unknown parameters in the probability of occurrence p(F0) through the expectation maximization algorithm (EM) μ k ,Σ k ;

[0058] Gaussian mixture regression μ k,Σ k The determined occurrence probability p(F0) is regressed to obtain the probability distribution p(F|x) of any operating force F:

[0059]

[0060] Among them, p(k|x) is the prior probability of the kth Gaussian distribution with any robot motion trajectory position information x as the input variable, is the probability density function of the kth Gaussian distribution with any robot motion trajectory position information x as input variable, is the mean of the kth Gaussian distribution with any robot motion trajectory position information x as input variable, is the covariance matrix of the kth Gaussian distribution with any robot motion trajectory position information x as input variable.

[0061] That is to say, in order to associate the trajectory position information with the operating force information, the present invention uses the trajectory position information as an input variable and the operating force information as an output variable, thereby calculating the probability distribution of the operating force information through the Gaussian mixture model, and finally calculating the conditional expectation of the operating force when the trajectory position information is given:

[0062]

[0063] The mean and covariance matrix of the Gaussian components are iteratively solved using the EM algorithm. The method of calculating the conditional expectation of the output through the input variables is called Gaussian mixture regression. This method ultimately obtains the new task trajectory and force data based on imitation learning. The effect is as follows Figure 2 shown.

[0064] Figure 3 Generate a new task effect diagram for imitation learning, where each ellipse is a Gaussian mixture model, and the smooth curve in the middle is the task data generated by imitation learning that contains trajectory position and operation force information.

[0065] In summary, the present invention provides an orthopedic surgery imitation learning system for sparse data and small samples, including a data acquisition module, an imitation learning module, a motion vector acquisition module, a node mapping module, and a motion control module;

[0066] Among them, the data acquisition module is used to collect the doctor's demonstration data before the operation; the data processing module is used to perform time alignment, data interpolation and prediction on the demonstration data to generate complete and effective demonstration data; the motion vector acquisition module is used to obtain the motion direction; the node mapping module generates new task data through a Gaussian mixture regression algorithm and maps it to the corresponding surgical environment; the motion control module is used to execute the task data generated by imitation learning in the surgical environment; the present invention can generate complete and effective demonstration data based on small samples and sparse demonstration data, and regenerate the robot's movements through an imitation learning algorithm, thereby improving the robot's operating accuracy and reducing the burden on doctors.

[0067] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may certainly make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. An orthopedic surgery imitation learning system for sparse data and small samples, characterized by: It includes data acquisition module, imitation learning module, motion vector acquisition module, node mapping module and motion control module; The data acquisition module is used to construct a small sample data set based on the actual trajectory of the surgical instrument held by the doctor within a set time period, wherein each sample consists of the trajectory position information x0, speed information x1 and x2 of the surgical instrument at each moment. and the actual operating force F0 between the surgical instrument and the bone tissue; The imitation learning module is used to perform trajectory position information x0 and speed information x1 of the surgical instrument in the small sample data set according to the Gaussian mixture regression method. and the actual operating force F0 to obtain the probability distribution of the operating force F used to characterize the mapping relationship between the robot motion trajectory and the operating force; The motion vector acquisition module is used to obtain the starting point P of the robot motion trajectory represented by the operating force F according to the probability distribution of the operating force F. s Point to the next end point P e Motion vector The node mapping module is used to determine the initial position of the robot specified by the doctor for the bone tissue to be removed. and the motion vector at the next moment Get the next end point of the actual resection trajectory of the bone tissue to be resected The motion control module is used to adjust the initial position and the next moment end Control the robot's movements so that it can remove the bone tissue according to the set actual removal trajectory.

2. The orthopedic surgery imitation learning system for sparse data and small samples as claimed in claim 1, characterized in that: Also includes a data processing module; The data processing module is used to use a data interpolation algorithm to fill in missing data in a small sample data set, then use a time series prediction algorithm to extend the time period covered by the small sample data set after the missing data filling is completed, and finally use a time alignment algorithm to time align the small sample data set after the time extension is completed, so as to obtain a time-continuous small sample data set for the imitation learning module to perform imitation learning.

3. The orthopedic surgery imitation learning system for sparse data and small samples as claimed in claim 1, characterized in that: The next end point of the actual resection trajectory of the bone tissue to be resected The method to obtain is: Among them, R is the set proportional factor.

4. The orthopedic surgery imitation learning system for sparse data and small samples as claimed in claim 1, characterized in that: The data acquisition module acquires the trajectory position information x and speed information of the surgical instrument. And the method of the operating force F between the surgical instrument and the bone tissue is: Optical navigation markers and force sensors are mounted on surgical instruments, and data collection is accomplished by direct demonstration by the surgeon holding the surgical instruments.

5. The orthopedic surgery imitation learning system for sparse data and small samples as claimed in claim 1, characterized in that: The data acquisition module acquires the trajectory position information x and speed information of the surgical instrument. And the method of the operating force F between the surgical instrument and the bone tissue is: The surgical instruments are installed on the robotic arm, and the doctor drags the robotic arm or remotely operates it for demonstration, thereby completing data collection.

6. The orthopedic surgery imitation learning system for sparse data and small samples as claimed in claim 1, characterized in that: The probability distribution of the operating force F is obtained as follows: According to the actual operating force F0, the small sample data set is divided into K segments of sample data, and each segment of sample data corresponds to a Gaussian distribution; The occurrence probability p(F0) of each actual operation force F0 in the small sample data set is constructed by the Gaussian mixture model as follows: in, is the prior probability of the kth Gaussian distribution of the small sample data set, and μ k is the mean value of the kth Gaussian distribution of the small sample data set; Σ k is the covariance matrix of the kth Gaussian distribution of the small sample data set; is the probability density function of the kth Gaussian distribution of the small sample data set; Obtain the unknown parameters in the probability of occurrence p(F0) through the expectation maximization algorithm μ x ,Σ k ; Gaussian mixture regression μ k ,Σ k The determined occurrence probability p(F0) is regressed to obtain the probability distribution p(F|x) of any operating force F: Among them, p(k|x) is the prior probability of the kth Gaussian distribution with any robot motion trajectory position information x as input variable, is the probability density function of the kth Gaussian distribution with any robot motion trajectory position information x as input variable, is the mean of the kth Gaussian distribution with any robot motion trajectory position information x as input variable, is the covariance matrix of the kth Gaussian distribution with any robot motion trajectory position information x as input variable.

Images