Virtual operation evaluation model training method, evaluation system and evaluation method

Abstract

The invention provides a virtual operation evaluation model training method. The method comprises the steps that S1, multiple pieces of virtual operation kinematics data and operation level labels corresponding to the virtual operation kinematics data are acquired; s2, each piece of virtual operation kinematics data is preprocessed, and preprocessing comprises the steps that a feature curve of partial features in each piece of virtual operation kinematics data is obtained; based on a plurality of characteristic curves obtained by each piece of virtual operation kinematics data, determining area characteristics corresponding to part of time points on each characteristic curve according to an area formed by a plurality of coordinates on each characteristic curve; adding area features corresponding to a part of time points on a plurality of feature curves obtained by each piece of virtual operation kinematics data to own virtual operation kinematics data to obtain each piece of preprocessed virtual operation kinematics data; and S3, training the virtual operation evaluation model according to the preprocessed multiple pieces of virtual operation kinematics data until the model converges.

Description

Technical Field

[0001] The present invention relates to the field of medical training, in particular to a virtual surgery evaluation technology in the field of medical training, and more particularly to a virtual surgery evaluation model training method, an evaluation system and an evaluation method. Background Art

[0002] In a virtual surgical environment, doctors can practice various surgical operations on a safe and controllable platform, and conduct in-depth analysis of the training process and operation performance after surgical training. With feedback from the evaluation method, doctors can self-evaluate their skill level and make necessary adjustments in repeated training to continuously improve their own operation capabilities. The development of virtual surgical evaluation has gone through two significant stages: initially, it was a traditional evaluation method that relied on postoperative indicators, and then gradually evolved into an intelligent evaluation method with deep learning as the core.

[0003] Traditional evaluation methods mainly rely on quantitative analysis of postoperative indicators, including operation time, path length, number of collisions, and path ratio (ratio of actual path to standard path). These data are collected by sensors and detectors, which can reflect the proficiency of the surgical operator to a certain extent. For example, studies have shown that skilled surgical operators usually show shorter path lengths, faster operation speeds, and lower error rates. In addition, in virtual shoulder arthroscopy, injury assessment is also used as a direct indicator to measure the fine movements of the surgical operator's hands and tissue processing capabilities, among which collision frequency and collision depth become the core parameters for evaluating surgeons with different skill levels. At the same time, force and torque data are also included in the evaluation system to analyze the dynamic characteristics of surgical tools. Although traditional evaluation methods objectively evaluate the skills of surgeons from multiple perspectives, their limitations are also significant: on the one hand, the evaluation dimensions based on paths and error operations are relatively single, and it is difficult to fully reflect the complexity of surgical operations; on the other hand, traditional evaluation methods lack fine-grained analysis capabilities and it is difficult to deeply portray those subtle but crucial operational details during the operation.

[0004] In recent years, evaluation methods based on machine learning have gradually emerged, promoting the development of virtual surgery evaluation towards intelligence. Evaluation methods based on machine learning improve the efficiency and accuracy of virtual surgery evaluation by automatically extracting high-dimensional features from kinematic data. For example, convolutional neural networks (CNNs) are used to perform high-precision skill classification of three different types of surgeries with an accuracy rate of over 91%. At the same time, in order to better cope with the diversity of data, researchers have also designed full convolutional networks to cluster data using different channels. However, these full convolutional networks are limited by the receptive field, resulting in incomplete feature extraction. To this end, some researchers have introduced dilated convolutions to expand the receptive field, thereby effectively capturing the representative features of kinematic multi-time series data (MTS). In addition, studies based on wavelet decomposition and ensemble learning have shown that combining different levels of decomposition with classifiers can achieve the best classification effect. However, although evaluation methods based on machine learning perform well in overall pattern recognition and feature processing, this method mainly focuses on the input and output of global data, pays insufficient attention to the detailed changes during the operation, and is difficult to fully reflect the details of the operation of the operator in a small time interval.

[0005] In summary, although the existing virtual surgery evaluation methods based on machine learning can evaluate the operator's surgical level by analyzing virtual surgery kinematic data, they still have two shortcomings in practical applications. On the one hand, the ability to capture local details is insufficient: the existing virtual surgery evaluation methods based on machine learning mainly rely on global feature extraction strategies, such as statistical analysis of the overall motion trajectory or operation time, but fail to effectively capture the fine-grained detail features during the surgical operation. This limitation causes the model to only reflect the operator's macroscopic behavior pattern (such as overall fluency), but cannot identify the accuracy of key actions (such as instrument contact angle, tissue cutting depth, etc., the accuracy of these key actions is an important indicator that effectively reflects the level of surgery). Since surgical evaluation is highly sensitive to operation details, the lack of fine-grained detail feature extraction capabilities will significantly reduce the accuracy of the evaluation results; on the other hand, the data fusion capability is limited: virtual surgery kinematic data is essentially multi-source heterogeneous information, which contains not only spatial features such as three-dimensional spatial coordinates and instrument posture matrix, but also dynamic information such as motion acceleration and operation timing, and implicitly contains the interactive features of visual feedback and mechanical feedback. However, the existing virtual surgery evaluation methods based on machine learning have weak fusion processing capabilities for these heterogeneous data and cannot make full use of these heterogeneous data, resulting in incomplete and inaccurate model evaluation results.

[0006] It should be noted that: This background art is only used to introduce relevant information of the present invention to facilitate understanding of the technical solution of the present invention, but it does not necessarily mean that the relevant information is prior art. Without evidence indicating that the relevant information has been made public before the filing date of the present invention, the relevant information should not be regarded as prior art. Summary of the Invention

[0007] Therefore, an object of the present invention is to overcome the defects of the above-mentioned prior art and provide a virtual surgery evaluation model training method, a virtual surgery evaluation system, a virtual surgery evaluation method, and a virtual surgery analysis method.

[0008] The object of the present invention is achieved by the following technical solutions.

[0009] According to a first aspect of the present invention, a virtual surgery evaluation model training method, the method comprising: Step S1, obtaining a plurality of virtual surgery kinematic data and their respective corresponding surgery level labels; wherein, the virtual surgery kinematic data all include a plurality of consecutive time points, each time point includes a plurality of features, and each feature corresponds to a feature value; the surgery level labels include expert, proficient, and novice; Step S2, preprocessing each virtual surgery kinematic data, wherein the preprocessing includes: obtaining a feature curve of some features in each virtual surgery kinematic data, wherein the feature curve of each feature is obtained by connecting the coordinates mapped from the feature values of the feature at different time points to a two-dimensional coordinate system or a three-dimensional coordinate system, and one coordinate corresponds to one time point; based on the plurality of feature curves obtained from each virtual surgery kinematic data, determining the area features corresponding to some time points on each feature curve by the area formed by a plurality of coordinates on each feature curve, obtaining an area feature sequence corresponding to each feature curve; wherein, the area feature sequence reflects the change trend of the feature value; adding the area features corresponding to some time points on the plurality of feature curves obtained from each virtual surgery kinematic data to the virtual surgery kinematic data itself, obtaining each preprocessed virtual surgery kinematic data; Step S3, training the virtual surgery evaluation model with the preprocessed plurality of virtual surgery kinematic data until the model converges.

[0010] In some embodiments of the present invention, in step S2, each virtual surgical kinematics data is preprocessed in the following manner: one or more features to be processed are determined, and a feature curve of each feature to be processed in each virtual surgical kinematics data is determined in the following manner: based on an existing mapping method, the feature value corresponding to each time point in the current feature to be processed is mapped to a two-dimensional coordinate system or a three-dimensional coordinate system to determine the coordinates of the feature value corresponding to each time point in the current feature to be processed in the coordinate system, and the coordinates of the feature value corresponding to each time point in the current feature to be processed in the coordinate system are connected to obtain the feature curve of the current feature to be processed; the area features corresponding to some time points on the feature curve of each feature to be processed are calculated in the following manner: starting from the first coordinate on the feature curve of the current feature to be processed, the area formed by a plurality of consecutive preset coordinates is calculated in sequence with a step size of 1, and the area calculated in each step is used as the area feature of the time point corresponding to the second coordinate in the step, and the area calculated is The area features of the time points corresponding to the coordinates constitute the area feature sequence of the current feature to be processed; and the area feature sequence of each feature to be processed obtained from each virtual surgical kinematics data is added to the own virtual surgical kinematics data in the following manner: if the corresponding feature value of the current feature to be processed is linear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, all areas in the area sequence features are added as new features to the corresponding time points in the own virtual surgical kinematics data, and all features of the time points without corresponding areas in the own virtual surgical kinematics data are deleted; if the corresponding feature value of the current feature to be processed is nonlinear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, the feature value of the time point corresponding to the current feature to be processed in the own virtual surgical kinematics data is replaced with the corresponding area in the area sequence features, and all features of the time points without corresponding areas in the own virtual surgical kinematics data are deleted.

[0011] In some embodiments of the present invention, the preset multiple coordinates are 3 coordinates.

[0012] Preferably, the virtual surgery evaluation model is a random forest.

[0013] Preferably, the time interval between adjacent time points in the virtual surgery kinematics data is 0.2 seconds.

[0014] According to a second aspect of the present invention, a virtual surgery evaluation system is provided, the system comprising: a data acquisition module, used to acquire virtual surgery kinematic data to be evaluated; a data processing module, used to preprocess the virtual surgery kinematic data to be evaluated; a virtual surgery evaluation model trained by the method described in the first aspect of the present invention, used to evaluate the preprocessed virtual surgery kinematic data to be evaluated, so as to obtain the surgical level corresponding to the virtual surgery kinematic data to be evaluated.

[0015] According to a third aspect of the present invention, a virtual surgery evaluation method is provided, the method comprising: using the system according to the second aspect of the present invention to evaluate virtual surgery kinematic data.

[0016] According to a fourth aspect of the present invention, a virtual surgery analysis method is provided, the method comprising: step T1, obtaining virtual surgery kinematic data to be analyzed and its corresponding expert virtual surgery kinematic data; wherein, the virtual surgery kinematic data to be analyzed and the expert virtual surgery kinematic data have three-dimensional coordinate data at each time point, and the corresponding surgical trajectories are determined based on the three-dimensional coordinate data at all time points of the virtual surgery kinematic data to be analyzed and the expert virtual surgery kinematic data; step T2, calculating the global similarity between the surgical trajectory corresponding to the virtual surgery kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgery kinematic data; wherein the global similarity is calculated as follows: Obtain the path vector of the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed, and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data; based on the path vector of the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed, and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data, use cosine similarity to calculate the global similarity between the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data; step T3, calculate the local similarity between the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data, wherein the local similarity is calculated in the following manner: using the method described in the first embodiment of the present invention The system described in the second aspect evaluates the virtual surgical kinematic data to be analyzed to obtain the importance score of each feature at each time point in the virtual surgical kinematic data to be analyzed, and adds up all the feature importance scores at each time point to obtain the contribution value of each time point; wherein the contribution value of each time point represents the degree of influence of all the features at that time point on the evaluation result; based on the contribution value of each time point in the virtual surgical kinematic data to be analyzed, the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed is segmented according to a preset segmentation method to obtain multiple local surgical trajectories corresponding to the virtual surgical kinematic data to be analyzed; wherein the preset segmentation method is: adding up the contribution values ​​of multiple consecutive time points, and The surgical trajectory corresponding to multiple consecutive time points whose sum of contribution values ​​satisfies greater than or equal to a preset threshold is regarded as a key local surgical trajectory; and the surgical trajectory corresponding to multiple consecutive time points whose sum of contribution values ​​does not satisfy greater than or equal to the preset threshold is regarded as a non-key local surgical trajectory; the path vector of each local surgical trajectory of the virtual surgical kinematic data to be analyzed is obtained; based on the path vector of each local surgical trajectory of the virtual surgical kinematic data to be analyzed and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data, the cosine similarity is used to calculate the local similarity between each local surgical trajectory of the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data.

[0017] Compared with the prior art, the advantages of the present invention are: (1) a feature extraction operation based on area segmentation is set up to refine the different features in the virtual surgery kinematics data, so as to accurately capture the local operation features, and then refine the analysis of the dynamic changes during the operation, which solves the problem of over-reliance on global features and neglect of key details in traditional evaluation methods, and helps to improve the evaluation accuracy of the virtual surgery evaluation model; (2) a local data fusion operation is set up, which can effectively improve the consistency of heterogeneous data in the time dimension without changing the overall data distribution of the virtual surgery kinematics data, thereby providing a more complete and accurate feature expression for model training, which helps to improve the evaluation accuracy of the virtual surgery evaluation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0019] Figure 1 A flowchart of a virtual surgery assessment model training method according to an embodiment of the present invention;

[0020] Figure 2 It is a schematic diagram of an example of area feature extraction of a characteristic curve according to an embodiment of the present invention;

[0021] Figure 3 It is a schematic diagram of an example of area feature extraction of uniform circular motion according to an embodiment of the present invention;

[0022] Figure 4 is a schematic diagram of an example of local data fusion operation according to an embodiment of the present invention;

[0023] Figure 5 FIG. 4 is a schematic diagram of a virtual surgery evaluation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0025] As mentioned in the background technology section, although the existing virtual surgery evaluation method based on machine learning can evaluate the operator's surgical level by analyzing the virtual surgery kinematic data, it still has two shortcomings in practical applications. On the one hand, the ability to capture local details is insufficient: the existing virtual surgery evaluation methods based on machine learning mainly rely on global feature extraction strategies, such as statistical analysis of the overall motion trajectory or operation time, but fail to effectively capture the fine-grained detail features during the surgical operation. This limitation causes the model to only reflect the operator's macroscopic behavior pattern (such as the overall smoothness of the surgical trajectory), but cannot identify the accuracy of key actions (such as instrument contact angle, tissue cutting depth, etc., the accuracy of these key actions is an important indicator that effectively reflects the surgical level). Since surgical evaluation is highly sensitive to operation details, the lack of fine-grained detail feature extraction capabilities will significantly reduce the accuracy of the evaluation results; on the other hand, the data fusion capability is limited: virtual surgery kinematic data is essentially multi-source heterogeneous information, which contains not only spatial features such as three-dimensional spatial coordinates and instrument posture matrix, but also dynamic information such as motion acceleration and operation timing, and implicitly contains the interactive features of visual feedback and mechanical feedback. However, the existing virtual surgery evaluation methods based on machine learning have weak fusion processing capabilities for these heterogeneous data and cannot make full use of these heterogeneous data, resulting in incomplete and inaccurate model evaluation results.

[0026] To address the problem of insufficient local detail capture, the inventors proposed that different feature data in the virtual surgery kinematics data can be refined to capture fine-grained detail features during the surgical operation. To address the problem of limited data fusion capabilities, the inventors proposed that the captured fine-grained detail features can be aligned based on time points to achieve heterogeneous data fusion.

[0027] Based on the above analysis, the inventors proposed a model training method, in which the virtual surgery kinematics data is processed by designing fine-grained feature extraction and local data fusion mechanisms, and the virtual surgery evaluation model is trained with the processed virtual surgery kinematics data to improve the evaluation accuracy and generalization ability of the virtual surgery evaluation model. Among them, fine-grained feature extraction refers to mapping the characteristic values ​​of some features in the virtual surgery kinematics data at different time points to the coordinate system to obtain the characteristic curves of some features, and determining the area features corresponding to some time points on the characteristic curve by calculating the area constructed by multiple consecutive coordinate points on the characteristic curve, and obtaining the area feature sequence corresponding to each characteristic curve, so as to capture the changes in the operation details during the operation through the change trend of the characteristic values ​​reflected by the area feature sequence. The local data fusion mechanism refers to adding the area feature sequences corresponding to all characteristic curves to the virtual surgery kinematics data to enrich the feature expression in the virtual surgery kinematics data, thereby improving the evaluation accuracy of the virtual surgery evaluation model.

[0028] In summary, if Figure 1 As shown, the present invention provides a virtual surgery evaluation model training method, the method comprising: step S1, obtaining multiple virtual surgery kinematic data and their corresponding surgical level labels; wherein the virtual surgery kinematic data each comprises multiple continuous time points, each time point comprises multiple features, and each feature corresponds to a feature value; the surgical level labels comprise expert, skilled and novice; step S2, preprocessing each virtual surgery kinematic data, wherein the preprocessing comprises: obtaining characteristic curves of some features in each virtual surgery kinematic data, wherein the characteristic curve of each feature is mapped from the characteristic value of the feature at different time points to a two-dimensional coordinate system or a three-dimensional coordinate system The coordinates are connected, and one coordinate corresponds to one time point; based on the multiple characteristic curves obtained from each virtual surgical kinematics data, the area characteristics corresponding to some time points on each characteristic curve are determined by the area composed of multiple coordinates on each characteristic curve, and the area characteristic sequence corresponding to each characteristic curve is obtained; wherein the area characteristic sequence reflects the changing trend of the characteristic value; the area characteristics corresponding to some time points on the multiple characteristic curves obtained from each virtual surgical kinematics data are added to the virtual surgical kinematics data itself, and each virtual surgical kinematics data after preprocessing is obtained; step S3, training the virtual surgery evaluation model with the multiple preprocessed virtual surgical kinematics data until the model converges.

[0029] In order to better understand the present invention, each step is described in detail below in conjunction with specific embodiments.

[0030] 1. Step S1

[0031] In step S1, a plurality of virtual surgical kinematic data and their corresponding surgical level labels are obtained; wherein the virtual surgical kinematic data each includes a plurality of continuous time points, each time point includes a plurality of features, and each feature corresponds to a feature value; the surgical level labels include expert, skilled, and novice. The features included at each time point in the virtual surgical kinematic data include kinematic features such as three-dimensional coordinates, rotation matrix, acceleration, angular velocity, Euler angle, and posture angle. Specifically, the virtual surgical kinematic data can be expressed as ,in, represents the kinematic eigenvector of the first time point; represents the kinematic eigenvector of the second time point; represents the kinematic eigenvector at the Tth time point; It represents the surgical level label corresponding to the virtual surgical kinematic data, which can be expert, skilled or novice; the kinematic feature vector at each time point contains features of multiple dimensions, such as three-dimensional coordinates, rotation matrix, acceleration, angular velocity, Euler angle, posture angle and other kinematic features.

[0032] Step S2

[0033] In step S2, each piece of virtual surgical kinematics data is preprocessed, wherein the preprocessing includes: obtaining characteristic curves of some features in each piece of virtual surgical kinematics data, wherein the characteristic curve of each feature is obtained by mapping the characteristic values ​​of the feature at different time points to the coordinates of a two-dimensional coordinate system or a three-dimensional coordinate system, and one coordinate corresponds to one time point; based on multiple characteristic curves obtained from each piece of virtual surgical kinematics data, the area characteristics corresponding to some time points on each characteristic curve are determined by the area composed of multiple coordinates on each characteristic curve, and an area characteristic sequence corresponding to each characteristic curve is obtained; wherein the area characteristic sequence reflects the changing trend of the characteristic value; the area characteristics corresponding to some time points on the multiple characteristic curves obtained from each piece of virtual surgical kinematics data are added to the virtual surgical kinematics data itself, and each piece of virtual surgical kinematics data after preprocessing is obtained.

[0034] In one embodiment of the present invention, in the step S2, each virtual surgical kinematics data is preprocessed in the following manner: one or more features to be processed are determined, and a feature curve of each feature to be processed in each virtual surgical kinematics data is determined in the following manner: based on an existing mapping method, the feature value corresponding to each time point in the current feature to be processed is mapped to a two-dimensional coordinate system or a three-dimensional coordinate system to determine the coordinates of the feature value corresponding to each time point in the current feature to be processed in the coordinate system, and the coordinates of the feature value corresponding to each time point in the current feature to be processed in the coordinate system are connected to obtain the feature curve of the current feature to be processed; the area features corresponding to some time points on the feature curve of each feature to be processed are calculated in the following manner: starting from the first coordinate on the feature curve of the current feature to be processed, the area formed by a plurality of consecutive preset coordinates is calculated in sequence with a step size of 1, and the area calculated in each step is used as the area feature of the time point corresponding to the second coordinate in the step, and the area calculated is The area features of the time points corresponding to the coordinates constitute the area feature sequence of the current feature to be processed; and the area feature sequence of each feature to be processed obtained from each virtual surgical kinematics data is added to the own virtual surgical kinematics data in the following manner: if the corresponding feature value of the current feature to be processed is linear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, all areas in the area sequence features are added as new features to the corresponding time points in the own virtual surgical kinematics data, and all features of the time points without corresponding areas in the own virtual surgical kinematics data are deleted; if the corresponding feature value of the current feature to be processed is nonlinear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, the feature value of the time point corresponding to the current feature to be processed in the own virtual surgical kinematics data is replaced with the corresponding area in the area sequence features, and all features of the time points without corresponding areas in the own virtual surgical kinematics data are deleted.

[0035] In one embodiment of the present invention, the preset multiple coordinates are 3 coordinates.

[0036] Based on the above embodiments, it can be known that in step S2, the preprocessing includes a feature extraction operation based on area segmentation and a local data fusion operation. The following describes these two operations in detail.

[0037] Among them, the feature extraction operation based on area segmentation includes: first, based on the existing mapping method, the feature value corresponding to each time point of the current feature to be processed in the virtual surgical kinematics data is mapped to a two-dimensional coordinate system or a three-dimensional coordinate system to determine the coordinates of the feature value corresponding to each time point in the current feature to be processed in the coordinate system, and the coordinates in the coordinate system are connected to obtain the feature curve of the current feature to be processed; then starting from the first coordinate on the feature curve, the area formed by three consecutive coordinates is calculated in sequence with a step size of 1, and the area calculated in each step is used as the area feature of the time point corresponding to the second coordinate in the step, and the area features of all the coordinates corresponding to the time points calculated constitute the area feature sequence of the current feature to be processed.

[0038] In order to better understand the feature extraction operation based on area segmentation, such as Figure 2 As shown, assuming there are two characteristic curves and ,in, and The points on the , , , and , , , . Starting from the first time step, the curve is divided into triangular areas at every three points, thus the curve can be obtained The triangle , and curve of , ;Will , and , Arranged in chronological order, the characteristic curve can be obtained and Through these area characteristic sequences, it can be seen that the characteristic curve The area of ​​the triangle in and Less than the characteristic curve The area of ​​the triangle in , , because the characteristic curve With characteristic curve Compared with the smoother, this means that the change of the eigenvalue corresponding to the characteristic curve B is smaller. Among them, the triangle area formula is: , and Represents the vectors of two adjacent sides in a triangle. It should be noted that for ease of understanding, Figure 2 The characteristic curves shown have omitted coordinate axes, and the coordinate axes are also omitted in other drawings related to characteristic curves.

[0039] It should be noted that the purpose of setting up a feature extraction operation based on area segmentation is to refine the different features in the virtual surgery kinematics data in order to accurately capture the local operation features, and then refine the analysis of dynamic changes during the operation. For example, the changing trend of the motion speed, the smoothness of the surgical trajectory, and the slight changes in the rotation angle during the operation. Specifically, for the changing trend of the motion speed, when the values ​​in the area feature sequence of the characteristic curve of the three-dimensional coordinate feature are continuously greater than 0 and the change value remains within 5%, it can be inferred that the motion pattern corresponding to the virtual surgery kinematics data may be as follows Figure 3 The uniform circular motion shown; when the value in the area feature sequence of the feature curve of the three-dimensional coordinate feature is close to zero, it can be inferred that the motion mode corresponding to the virtual surgery kinematics data may be uniform motion or static; when the value in the area feature sequence of the feature curve of the three-dimensional coordinate feature shows a significant data difference (greater than 5%), it can be inferred that the motion mode corresponding to the virtual surgery kinematics data has changed, such as changing from uniform motion to accelerated motion. For the smoothness of the surgical trajectory, if the area feature in the area feature sequence of the feature curve of the three-dimensional coordinate feature changes little, it means that the surgical trajectory is smooth; if the area feature in the area feature sequence of the feature curve of the three-dimensional coordinate feature changes greatly, it means that the surgical trajectory is steep. For the rotation angle, if the area feature in the area feature sequence of the feature curve of the rotation angle changes little, it means that the surgical operation is stable and there is no large shaking or adjustment; if the area feature in the area feature sequence of the feature curve of the rotation angle changes greatly, it means that the surgical operation is unstable and there is a large shaking or adjustment.

[0040] The local data fusion operation includes: if the feature value corresponding to the current feature to be processed is linear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, all areas in the area sequence feature are added as new features to the corresponding time points in the own virtual surgical kinematics data, and all features of the time points in the own virtual surgical kinematics data that do not have corresponding areas are deleted; if the feature value corresponding to the current feature to be processed is nonlinear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, the feature value of the time point corresponding to the current feature to be processed in the own virtual surgical kinematics data is replaced with the corresponding area in the area sequence feature, and all features of the time points in the own virtual surgical kinematics data that do not have corresponding areas are deleted. It should be noted that the local data fusion operation synchronously processes different types of data at a unified time point without changing the data distribution, so that heterogeneous data can be consistent in the time dimension, thereby achieving deep data fusion.

[0041] In order to better understand the preprocessing process, virtual surgery kinematics data containing only three-dimensional coordinate features and rotation matrix features is taken as an example to illustrate how to preprocess virtual surgery kinematics data.

[0042] The first step is to determine the characteristic curves corresponding to the three-dimensional coordinate features and the rotation matrix features. For the three-dimensional coordinate features, directly map the three-dimensional coordinate values ​​corresponding to each time point to the three-dimensional coordinate system, and connect the three-dimensional coordinate values ​​corresponding to each time point to obtain Figure 4 The characteristic curve corresponding to the three-dimensional coordinate feature shown , where the characteristic curve Including time points For the rotation matrix feature, the rotation matrix at each time point is Since the rotation matrix cannot be directly mapped to the coordinate system, the rotation angle corresponding to the rotation matrix at each time point must be calculated using the existing calculation method. Then, rotate the As the y-axis coordinate, with the rotation angle The corresponding time point is used as the x-axis coordinate to determine the coordinate of the feature matrix of each time point in the coordinate system, and the coordinates of the feature matrix of each time point in the coordinate system are connected to obtain the following Figure 4 The characteristic curve corresponding to the rotation matrix characteristics shown , where the characteristic curve Including time points . Among them, the rotation matrix The corresponding rotation angle .

[0043] The second step is to select the characteristic curves and The first coordinate on (time point ), the area formed by three consecutive coordinates is calculated in sequence with a step length of 1, and the area calculated in each step is used as the area feature of the time point corresponding to the second coordinate in the step. The area features of all the coordinates corresponding to the time points calculated constitute the area feature sequence of the current feature to be processed. Among them, the characteristic curve The corresponding area characteristic sequence is { }, For time point The area characteristics of Corresponding to the time point The area of ​​the composition; For time point The area characteristics of Corresponding to the time point The area of ​​the composition; For time point The area characteristics of Corresponding to time point The area of ​​the composition; For time point The area characteristics of Corresponding to the time point The area of ​​the composition; For time point The area characteristics of Corresponding to time point The area formed by the characteristic curve The corresponding area characteristic sequence is { }, For time point The area characteristics of Corresponding to the time point The area of ​​the composition; For time point The area characteristics of Corresponding to time point The area of ​​the composition; For time point The area characteristics of Corresponding to the time point The area of ​​the composition; For time point The area characteristics of Corresponding to the time point The area of ​​the composition; For time point The area characteristics of Corresponding to time point The area formed.

[0044] Step 3: Since the three-dimensional coordinate value corresponding to the three-dimensional coordinate feature is linear data, the characteristic curve The corresponding area characteristic sequence { All areas in the virtual surgery kinematics data are added as new features to the corresponding time points in the virtual surgery kinematics data, and all features of the time points without corresponding areas in the virtual surgery kinematics data are deleted; the rotation matrix corresponding to the rotation matrix feature is nonlinear data, so the characteristic curve is used The corresponding area characteristic sequence { }Replace the rotation matrix of the corresponding time node in the virtual surgery kinematics data, and delete the three-dimensional coordinate value and rotation matrix of the time point without area corresponding to the virtual surgery kinematics data; obtain Figure 4 It should be noted that, since the three-dimensional coordinate values ​​and rotation matrices of the time points without area corresponding to the virtual surgery kinematics data are deleted (the first and last time points are deleted), the time points in the preprocessed virtual surgery kinematics data are Time points in the virtual surgery kinematic data before preprocessing It is not a one-to-one relationship.

[0045] Step S3

[0046] In step S3, the virtual surgery evaluation model is trained with the preprocessed multiple virtual surgery kinematics data until the model converges.

[0047] In one embodiment of the present invention, the virtual surgery evaluation model is a random forest.

[0048] In one embodiment of the present invention, the time interval between adjacent time points in the virtual surgery kinematics data is 0.2 seconds.

[0049] Based on the above embodiments, it can be known that during the training process, the time interval between adjacent time points in the virtual surgery kinematics data is 0.2 seconds, that is, the sampling rate is 0.2 seconds; the virtual surgery evaluation model is a random forest, wherein the sampling rate and the random forest are determined through experiments.

[0050] The reason why we need to conduct experiments to determine the virtual surgery evaluation model and sampling rate is that too low a sampling rate may not be able to fully capture the details, while too high a sampling rate may result in too much data, affecting the efficiency of preprocessing and model training; at the same time, the evaluation performance of different neural network models is also different. Therefore, it is necessary to conduct experiments to determine the optimal sampling rate and the optimal virtual surgery evaluation model. The following is a detailed introduction to the experimental process.

[0051] In order to ensure the universality of the optimal sampling rate, a variety of virtual surgery kinematic data that meet different sampling rates are used to train different neural network models in the experiment, and the optimal sampling rate and the optimal virtual surgery evaluation model are determined by comparing the evaluation accuracy of different neural network models.

[0052] Before conducting the experiment, the sampling rate, training data, and neural network model were determined. The sampling rate was set to 2SPS, 2.5SPS, 5SPS, 10SPS, 20SPS, and 30SPS. SPS (Samples Per Second) means the number of samples per second. For example, 2SPS means two samples per second. The training data is the time series data of virtual counterclockwise capsulorhexis. ), Dataset of suturing, )、Dataset of needle passing, ) and Dataset of knottying, ). The neural network models included support vector machine (SVM), random forest (RF) and feedforward neural network (FNN); the kernel function in the support vector machine (SVM) was set as linear kernel, and the One-vs-Rest method was used as the decision function; 100 random forest trees were constructed in the random forest (RF), the weight of each surgical level category was 1, and the random state was set to 0; 128 hidden layers were set in the feedforward neural network (FNN), and the cross entropy loss function and Adam optimizer with a learning rate of 0.001 were used; ReLU was used as the activation function, and the epoch size was 200.

[0053] Based on the set sampling rate, training data and neural network model, a triple cross-validation was performed to analyze the performance of each neural network model in evaluating the virtual surgical kinematics data with different sampling rates, and the experimental results were obtained as shown in Table 1. The data in Table 1 represent the accuracy of the neural network model in evaluating (classifying) the virtual surgical kinematics data corresponding to the surgical level of expert, skilled or novice.

[0054] As shown in Table 1, the accuracy of random forest is better than that of support vector machine and feedforward neural network at different sampling rates of different data. Among them, random forest has the highest average accuracy (0.944) at 5 SPS (sampling 5 times per second, sampling once every 0.2 seconds). The accuracy on the data is 0.915. The accuracy on the data is 0.949. The accuracy on the data is 0.967. The accuracy on the test is 0.944. Random forest performs well not only at a low sampling rate of 5 SPS, but also at a high sampling rate of 30 SPS, which shows that the integration characteristics of random forest effectively reduce the risk of overfitting and cope well with the data complexity and potential noise brought by higher sampling rates. Although the accuracy of random forest decreases slightly at a sampling rate of 20 SPS, overall, random forest still performs better than support vector machine and feedforward neural network, so random forest is selected as the virtual surgery evaluation model. At the same time, based on the experimental results, the sampling rate is set to 0.2 (5SPS) seconds. At this sampling rate, the virtual surgery evaluation model can not only capture the key features in the virtual surgery trajectory data, but also avoid the data redundancy or noise that may be introduced by high sampling rates, and achieve the best model training effect.

[0055] Table 1

[0056]

[0057] Furthermore, in order to compare the performance difference between the virtual surgery assessment model trained by the training method proposed in the present invention and the existing virtual surgery assessment method, a comparative experiment was conducted, and the performance difference between different methods was analyzed by accuracy and macro average recall. In the comparative experiment, the macro average recall was calculated in the following way:

[0058]

[0059]

[0060] in, represents the macro average recall, Indicates category Recall rate of (expert, skilled, novice), Indicates category The number of samples correctly predicted as positive, Representation Analogy The number of samples that are incorrectly predicted as negative.

[0061] In the comparative experiment, a method based on postoperative indicators and three methods based on machine learning were selected to compare the performance with the virtual surgery evaluation model trained by the training method proposed in the present invention. Among them, the three methods based on machine learning were an evaluation method using WPD for skill learning, an evaluation method based on DCNN, and an evaluation method based on TCN. Specifically, the method based on postoperative indicators: time consumption and collision are used to evaluate whether the surgical level corresponding to the virtual surgical kinematic data is an expert, skilled, or a novice. The evaluation method using WPD for skill learning: the skill characteristics of the surgical operator are extracted by using wavelet packet decomposition (WPD) technology to accurately evaluate whether the surgical level corresponding to the virtual surgical kinematic data is an expert, skilled, or a novice. The evaluation method based on DCNN evaluates whether the surgical level corresponding to the virtual surgical kinematic data is an expert, skilled, or a novice through the trained DCNN model. The evaluation method based on TCN evaluates whether the surgical level corresponding to the virtual surgical kinematic data is an expert, skilled, or a novice through the trained TCN model.

[0062] In the comparative experiment, the virtual surgery evaluation model trained by the training method proposed in the present invention and the existing four evaluation methods are used to evaluate the surgical level corresponding to the timing data, suturing data, needle threading data and knotting data of the counterclockwise virtual capsulorhexis. Among them, the evaluation method based on postoperative indicators uses the experimental features related to collision and time consumption (Collision & Time) as the input of the random forest, and the network parameter setting of the random forest is consistent with the parameter setting of the trained random forest of the present invention. The evaluation method for skill learning using WPD selects the support vector machine (SVM) and sets the decomposition level to 4 to determine the best combination for distinguishing a specific surgical action. The evaluation method based on DCNN sets the DCNN model to consist of four convolutional layers and three dropout layers. The evaluation method based on TCN sets the TCN model to consist of five one-dimensional convolutional layers, batch normalization, activation function, maximum pooling and dropout layers. At the same time, for the DCNN model and the TCN model, the ratios of training, testing and validation sets are 0.8:0.1:0.1 respectively.

[0063] Based on the above settings, comparative experiments were performed to obtain the experimental results shown in Table 2. It can be seen from Table 2 that the accuracy and macro-recall rate of the virtual surgery evaluation model trained by the training method of the present invention are better than the existing virtual surgery evaluation methods. Among them, although the method based on postoperative indicators is effective in specific scenarios, it may not be able to fully capture the complex patterns and changes in the virtual surgery kinematic data due to the limitations of its feature representation. Compared with other methods, this limitation leads to lower accuracy and recall. The evaluation method using WPD for skill learning is good at capturing multi-resolution information and frequency components of data by decomposing enhanced feature extraction. However, if this method is not properly adjusted, WPD may introduce noise or irrelevant information, affecting the overall effect of the evaluation, and the performance of this method is affected by the selection of decomposition parameters and the nature of the data, so the parameters need to be carefully selected to obtain the best results. The evaluation method based on DCNN uses dilated convolution to capture long-distance dependencies in the data, enabling it to capture complex patterns. However, the performance of this evaluation method depends largely on the network architecture and hyperparameters, and requires careful optimization to obtain the best results in data sets with different characteristics. TCN-based evaluation methods use temporal convolution to model sequential dependencies and are therefore suitable for time series data analysis. However, the TCN model focuses on temporal patterns, which may limit its ability to exploit non-temporal features or complex interactions within the data, resulting in poor evaluation performance.

[0064] Table 2

[0065]

[0066] The virtual surgery evaluation model trained based on the above-mentioned embodiment can be used to evaluate the virtual surgery kinematics data. Figure 5 As shown, the present invention proposes a virtual surgery evaluation system, which includes: a data acquisition module, used to acquire virtual surgery kinematic data to be evaluated; a data processing module, used to preprocess the virtual surgery kinematic data to be evaluated; a virtual surgery evaluation model trained according to the method described in the aforementioned embodiment of claim 1, used to evaluate the preprocessed virtual surgery kinematic data to be evaluated, so as to obtain the surgical level corresponding to the virtual surgery kinematic data to be evaluated.

[0067] Based on the virtual surgery evaluation system proposed in the above embodiment, the present invention further proposes a virtual surgery evaluation method, which includes: using the system described in the above embodiment to evaluate virtual surgery kinematics data.

[0068] Furthermore, in order to better analyze the virtual surgery kinematics data, the present invention also proposes a virtual surgery analysis method, which includes step T1, step T2 and step T3.

[0069] Among them, in the step T1, the virtual surgical kinematic data to be analyzed and its corresponding expert virtual surgical kinematic data are obtained; there are three-dimensional coordinate data at each time point in the virtual surgical kinematic data to be analyzed and the expert virtual surgical kinematic data, and the corresponding surgical trajectories are determined based on the three-dimensional coordinate data at all time points of the virtual surgical kinematic data to be analyzed and the expert virtual surgical kinematic data.

[0070] In step T2, the global similarity between the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data is calculated; wherein, the global similarity is calculated in the following manner: the path vector of the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data are obtained; based on the path vector of the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data, the cosine similarity is used to calculate the global similarity between the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data. For example, assuming that the path vectors of surgical trajectory A and surgical trajectory B are respectively and , then the global similarity between surgical trajectory A and surgical trajectory B is:

[0071] In step T3, the local similarity between the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data is calculated, wherein the local similarity is calculated as follows: the virtual surgical kinematic data to be analyzed is evaluated by using the virtual surgical evaluation system described in the aforementioned embodiment to obtain the importance score of each feature at each time point in the virtual surgical kinematic data to be analyzed, and all the feature importance scores at each time point are added together to obtain the contribution value of each time point; wherein the contribution value of each time point represents the influence of all features at that time point on the evaluation result; based on the contribution value of each time point in the virtual surgical kinematic data to be analyzed, the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed is segmented according to a preset segmentation method to obtain the virtual surgical kinematic data to be analyzed. Corresponding multiple local surgical trajectories; wherein the preset segmentation method is: adding the contribution values ​​of multiple consecutive time points, and taking the surgical trajectories corresponding to the multiple consecutive time points whose contribution values ​​are greater than or equal to the preset threshold as the key local surgical trajectories; and taking the surgical trajectories corresponding to the multiple consecutive time points whose contribution values ​​are not greater than or equal to the preset threshold as the non-key local surgical trajectories; obtaining the path vector of each local surgical trajectory of the virtual surgical kinematic data to be analyzed; based on the path vector of each local surgical trajectory of the virtual surgical kinematic data to be analyzed and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data, the cosine similarity is used to calculate the local similarity between each local surgical trajectory of the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data. wherein the preset threshold can be set to 0.8. It should be noted that since the virtual surgery evaluation system will pre-process the virtual surgery kinematic data to be analyzed, and all features at the first time point and the last time point in the virtual surgery kinematic data to be analyzed will be deleted during the pre-processing, when calculating the contribution value of each time point, the contribution value of the first time point and the last time point in the virtual surgery kinematic data to be analyzed can be set to 0.

[0072] The similarity between the surgical trajectory corresponding to the virtual surgical kinematic data and the surgical trajectory corresponding to the expert virtual surgical kinematic data can be comprehensively evaluated by calculating the global similarity and local similarity. Among them, the global similarity measures the consistency of the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data at the global level. The global similarity can quickly determine whether the two surgical trajectories are roughly similar or meet the standards. The local similarity focuses on analyzing the local differences between the local surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data, and deeply analyzes the inconsistencies in the details during the surgical operation, which helps to accurately discover problems during the surgical operation.

[0073] The beneficial effects of the present invention are: (1) a feature extraction operation based on area segmentation is set up to refine the different features in the virtual surgery kinematics data, so as to accurately capture the local operation features, and then refine the analysis of the dynamic changes during the operation, which solves the problem of over-reliance on global features and neglect of key details in traditional evaluation methods, and helps to improve the evaluation accuracy of the virtual surgery evaluation model; (2) a local data fusion operation is set up, which can effectively improve the consistency of heterogeneous data in the time dimension without changing the overall data distribution of the virtual surgery kinematics data, thereby providing a more complete and accurate feature expression for model training, which helps to improve the evaluation accuracy of the virtual surgery evaluation model.

[0074] It should be noted that although the above describes the various steps in a specific order, it does not mean that the various steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order as long as the required functions can be achieved.

[0075] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0076] A computer-readable storage medium may be a tangible device that holds and stores instructions used by an instruction execution device. Computer-readable storage media may include, for example, but are not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a protruding structure in a groove on which instructions are stored, and any suitable combination thereof.

[0077] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A virtual surgery evaluation model training method, characterized in that: The method comprises: Step S1, obtaining a plurality of virtual surgery kinematics data and their corresponding surgery level labels; wherein the virtual surgery kinematics data each includes a plurality of continuous time points, each time point includes a plurality of features, and each feature corresponds to a feature value; the surgery level labels include expert, skilled, and novice; Step S2: preprocessing each piece of virtual surgery kinematics data, wherein the preprocessing includes: Acquire characteristic curves of some features in each virtual surgery kinematics data, wherein the characteristic curve of each feature is obtained by mapping characteristic values ​​of the feature at different time points to coordinates of a two-dimensional coordinate system or a three-dimensional coordinate system, and one coordinate corresponds to one time point; Based on multiple characteristic curves obtained from each virtual surgery kinematics data, the area characteristics corresponding to some time points on each characteristic curve are determined by the area composed of multiple coordinates on each characteristic curve, and the area characteristic sequence corresponding to each characteristic curve is obtained; wherein the area characteristic sequence reflects the change trend of the characteristic value; Adding area features corresponding to some time points on multiple characteristic curves obtained from each virtual surgery kinematics data to the virtual surgery kinematics data itself to obtain each virtual surgery kinematics data after preprocessing; Step S3: training the virtual surgery evaluation model with the preprocessed multiple virtual surgery kinematics data until the model converges.

2. The method according to claim 1, characterized in that In step S2, each piece of virtual surgery kinematics data is preprocessed in the following manner: One or more features to be processed are determined, and a feature curve of each feature to be processed in each virtual surgery kinematics data is determined as follows: Based on the existing mapping method, the feature value corresponding to each time point in the current feature to be processed is mapped to a two-dimensional coordinate system or a three-dimensional coordinate system to determine the coordinates of the feature value corresponding to each time point in the current feature to be processed in the coordinate system, and the coordinates of the feature value corresponding to each time point in the current feature to be processed in the coordinate system are connected to obtain the feature curve of the current feature to be processed; The area features corresponding to some time points on the feature curve of each feature to be processed are calculated as follows: Starting from the first coordinate on the characteristic curve of the current feature to be processed, the area formed by the continuous preset multiple coordinates is calculated in sequence with a step length of 1, and the area calculated in each step is used as the area feature of the time point corresponding to the second coordinate in the step. The area features of all the calculated coordinates corresponding to the time points constitute the area feature sequence of the current feature to be processed; And the area feature sequence of each feature to be processed obtained from each virtual surgery kinematics data is added to the virtual surgery kinematics data itself in the following manner: If the feature value corresponding to the current feature to be processed is linear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, all areas in the area sequence feature are added as new features to the corresponding time points in the own virtual surgery kinematics data, and all features of the time points without area corresponding to the own virtual surgery kinematics data are deleted; If the eigenvalue corresponding to the current feature to be processed is nonlinear data, then based on the time points corresponding to each area in the area feature sequence of the current feature to be processed, the eigenvalue of the time point corresponding to the current feature to be processed in the own virtual surgical kinematics data is replaced with the corresponding area in the area sequence feature, and all features of the time points without corresponding areas in the own virtual surgical kinematics data are deleted.

3. The method according to claim 2, characterized in that The preset multiple coordinates are 3 coordinates.

4. The method according to claim 3, characterized in that The virtual surgery evaluation model is a random forest.

5. The method according to claim 4, characterized in that The time interval between adjacent time points in the virtual surgery kinematics data is 0.2 seconds.

6. A virtual surgery evaluation system, characterized in that: The system comprises: A data acquisition module, used for acquiring the kinematic data of the virtual surgery to be evaluated; A data processing module, used for preprocessing the kinematic data of the virtual surgery to be evaluated; The virtual surgery evaluation model trained by the method according to any one of claims 1 to 5 is used to evaluate the preprocessed virtual surgery kinematic data to be evaluated, so as to obtain the surgical level corresponding to the virtual surgery kinematic data to be evaluated.

7. A virtual surgery evaluation method, characterized in that: The method comprises: The system of claim 6 is used to evaluate virtual surgery kinematic data.

8. A virtual surgery analysis method, characterized in that: The method comprises: Step T1, obtaining the virtual surgery kinematics data to be analyzed and its corresponding expert virtual surgery kinematics data; wherein the virtual surgery kinematics data to be analyzed and the expert virtual surgery kinematics data have three-dimensional coordinate data at each time point, and the corresponding surgical trajectories are determined based on the three-dimensional coordinate data at all time points of the virtual surgery kinematics data to be analyzed and the expert virtual surgery kinematics data; Step T2, calculating the global similarity between the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data; wherein the global similarity is calculated in the following manner: Obtaining a path vector of a surgical trajectory corresponding to the virtual surgical kinematics data to be analyzed, and a path vector of a surgical trajectory corresponding to the expert virtual surgical kinematics data; Based on the path vector of the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data, the cosine similarity is used to calculate the global similarity between the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data; Step T3: Calculate the local similarity between the surgical trajectory corresponding to the virtual surgical kinematics data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematics data, wherein the local similarity is calculated in the following manner: The system as claimed in claim 6 is used to evaluate the virtual surgical kinematics data to be analyzed, so as to obtain the importance score of each feature at each time point in the virtual surgical kinematics data to be analyzed, and the importance scores of all features at each time point are added together to obtain the contribution value of each time point; wherein the contribution value of each time point represents the degree of influence of all features at that time point on the evaluation result; Based on the contribution value of each time point in the virtual surgical kinematic data to be analyzed, the surgical trajectory corresponding to the virtual surgical kinematic data to be analyzed is segmented according to a preset segmentation method to obtain multiple local surgical trajectories corresponding to the virtual surgical kinematic data to be analyzed; wherein the preset segmentation method is: adding the contribution values ​​of multiple consecutive time points, and taking the surgical trajectories corresponding to the multiple consecutive time points whose contribution values ​​are added to satisfy a value greater than or equal to a preset threshold as the key local surgical trajectory; and taking the surgical trajectories corresponding to the multiple consecutive time points whose contribution values ​​are added to satisfy a value greater than or equal to a preset threshold as the non-key local surgical trajectory; Obtaining a path vector of each local surgical trajectory of the virtual surgical kinematic data to be analyzed; Based on the path vector of each local surgical trajectory of the virtual surgical kinematic data to be analyzed and the path vector of the surgical trajectory corresponding to the expert virtual surgical kinematic data, the cosine similarity is used to calculate the local similarity between each local surgical trajectory of the virtual surgical kinematic data to be analyzed and the surgical trajectory corresponding to the expert virtual surgical kinematic data.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of any method described in claims 1-5, 7-8.

10. An electronic device, characterized in that: include: one or more processors, and A memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method of any one of claims 1-5, 7-8 by executing the executable instructions.

Images