Method for identifying progress of spinal canal puncture needle head and puncture system

Abstract

The invention discloses a spinal canal puncture needle progress recognition method and a puncture system.The method comprises the steps that puncture real-time data are obtained firstly, standardized puncture real-time data are obtained through filtering and normalization processing, a time sequence is generated through time step division, then a pre-trained puncture progress recognition model is called to recognize the time sequence, and a puncture progress recognition result is obtained; and outputting a corresponding puncture state recognition result. Through standardization processing and time sequence conversion of puncture real-time data (such as force signals), automatic identification of a tissue layer where a needle head is located is achieved in combination with a pre-training model, dependence on doctor experience and subjective judgment is avoided, and complications caused by misjudgment are effectively reduced. Meanwhile, the puncture progress can be dynamically tracked through cooperation of real-time data processing and model recognition, and the accuracy and safety of puncture operation are further guaranteed.

Description

Technical Field

[0001] The embodiments of this application relate to, but are not limited to, the field of surgical puncture, and particularly to a method and puncture system for recognizing the progression of a spinal canal puncture needle. Background Technology

[0002] Spinal puncture is widely used in clinical diagnosis and treatment, but its traditional operation relies heavily on the doctor's manual experience. In practice, due to differences in patient anatomy (such as obesity, spinal deformities, etc.), doctors cannot accurately determine the tissue layer (such as skin, ligamentum flavum, target cavity, etc.) where the puncture needle is located by touch alone, which can easily lead to complications such as puncture failure, cerebrospinal fluid leakage, and nerve damage. To improve accuracy, existing technologies have introduced image-guided equipment such as ultrasound and X-rays, but these devices are expensive, complex to operate, and have limited clinical applicability. Furthermore, doctors still need to subjectively judge the imaging results, making it impossible to achieve real-time, automated identification of needle progress during the puncture process. This makes it difficult to completely solve the problems of over-reliance on experience and insufficient safety in traditional operations. Summary of the Invention

[0003] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0004] This application provides a method and system for identifying the progression of a spinal canal puncture needle, which can reduce the over-reliance on doctors' experience and subjective judgment, effectively reduce the risk of complications caused by misjudgment, and thus ensure the accuracy and safety of the puncture operation.

[0005] In a first aspect, embodiments of this application provide a method for identifying the progression of a spinal canal puncture needle, comprising: acquiring real-time puncture data; filtering and normalizing the real-time puncture data to obtain standardized real-time puncture data; dividing the standardized real-time puncture data into time steps to obtain a time series; calling a pre-trained puncture progression identification model, performing identification processing based on the time series, and obtaining a corresponding puncture status identification result.

[0006] In conjunction with the first aspect, in one embodiment of this application, the method further includes: acquiring puncture data samples; preprocessing the puncture data samples to obtain multiple standard time series samples, the standard time series samples including multiple time-step puncture data samples and corresponding label samples, the label samples being used to indicate the puncture degree category corresponding to the time-step puncture data samples; acquiring a target model parameter combination; constructing an initial puncture progress recognition model based on the target model parameter combination; and training the initial puncture progress recognition model based on the multiple standard time series samples to obtain the puncture progress recognition model.

[0007] In conjunction with the first aspect, in one embodiment of this application, the preprocessing of the puncture data sample to obtain multiple standard time series samples includes: filtering and labeling the puncture data sample to obtain puncture-labeled data samples; dividing the puncture-labeled data samples into multiple time series samples, and normalizing each time series sample to obtain multiple standard time series samples.

[0008] In conjunction with the first aspect, in one embodiment of this application, obtaining the target model parameter combination includes: generating an initial population containing several sets of hyperparameter combinations based on a preset hyperparameter search space and reference points, wherein the parameters of each individual in the initial population satisfy the constraint range of the hyperparameter search space; determining the fitness value of each individual based on the reference points; selecting parent individuals from the initial population according to the fitness values ​​of each individual, and obtaining candidate individuals by performing crossover and mutation operations on the parent individuals based on the hyperparameter search space; determining the candidate fitness value of the candidate individuals based on the reference points; updating the initial population according to the candidate fitness values ​​and the fitness values ​​of each individual in the parent individuals to obtain an updated population; using the updated population as the initial population for a new round of iteration for iterative calculation until the number of iterations reaches a preset number of generations; and selecting a target individual from the population as the target model parameter combination.

[0009] In conjunction with the first aspect, in one embodiment of this application, determining the fitness value of each individual based on the reference point includes: acquiring the performance data of each individual; calculating the comprehensive difference between the performance data of each individual and the reference point using a multi-objective decomposition method based on the reference point; and using the comprehensive difference as the fitness value of the corresponding individual.

[0010] In conjunction with the first aspect, in one embodiment of this application, training the initial puncture progress recognition model based on the plurality of standard time series samples to obtain the puncture progress recognition model includes: dividing the plurality of standard time series samples into a training set and a validation set; inputting the standard time series samples in the training set into the initial puncture progress recognition model, calculating the predicted labels of the model for the standard time series samples, calculating the error between the predicted labels and the corresponding label samples of the standard time series samples using a cross-entropy loss function, calculating the gradients of the network parameters of each layer of the model based on the error and a preset loss function, and updating the network parameters based on the gradients to obtain the optimized initial puncture progress recognition model; evaluating the classification accuracy of the optimized initial puncture progress recognition model using the validation set, stopping iteration when the number of iterations reaches a preset number of training rounds, or when the classification accuracy does not exceed the historical benchmark value for a consecutive preset number of rounds; and using the model at the point where iteration stops as the completed puncture progress recognition model.

[0011] Secondly, this application provides a puncture system applied to the aforementioned spinal canal puncture needle progression recognition method. The system includes a data processing module, a data monitoring module, a puncture robot, and a wearable positioning mechanism. The wearable positioning mechanism is worn on the user's waist. The puncture robot is equipped with an end effector for puncturing the user's waist and acquiring real-time puncture data. The data processing module filters and normalizes the real-time puncture data to obtain standardized real-time puncture data and divides the standardized real-time puncture data into time steps to obtain a time series. The data monitoring module is equipped with a puncture progression recognition model for performing recognition processing based on the time series to obtain the corresponding puncture status recognition result.

[0012] In conjunction with the second aspect, in one embodiment of this application, the end-effector is equipped with a force sensor, which communicates with the data processing module. The force sensor is used to collect real-time puncture data during the puncture process and send the real-time puncture data to the data processing module.

[0013] In conjunction with the second aspect, in one embodiment of this application, the puncture robot further includes: a trolley housing equipped with a control device, the control device being electrically connected to the data monitoring module; a first robotic arm, disposed on the trolley housing and electrically connected to the control device; an end effector puncture actuator disposed on the first robotic arm, the first robotic arm being used to adjust the position of the end effector puncture actuator under the control of the control device; a second robotic arm disposed on the trolley housing; and a display device disposed on the second robotic arm, the second robotic arm being used to adjust the position of the display device, the display device being used to display the puncture status identification result.

[0014] In conjunction with the second aspect, in one embodiment of this application, the wearable positioning mechanism includes a lumbar support unit, a needle insertion limiting cylinder, a ratchet-like telescopic connecting device, a telescopic arm, a locking contact unit, and a waist belt mounting component; the needle insertion limiting cylinder is engaged with the locking grooves of the lumbar support unit via locking points at both ends; the waist belt mounting component is symmetrically arranged on both sides of the lumbar support unit; the telescopic arm is connected to the lumbar support unit via the ratchet-like telescopic connecting device, the ratchet-like telescopic connecting device including a ratchet-like device switch for adjusting the position of the telescopic arm; the locking contact unit is located at the tail of the telescopic arm for achieving spinal positioning.

[0015] This application provides a method for identifying the progress of a spinal canal puncture needle. First, real-time puncture data is acquired, filtered, and normalized to obtain standardized real-time puncture data. Then, a time series is generated by dividing the data into time steps. Subsequently, a pre-trained puncture progress identification model is called to identify the time series and output the corresponding puncture status identification result. This application, through the standardization and time series transformation of real-time puncture data (such as force signals), combined with a pre-trained model, achieves automatic identification of the tissue layer where the needle is located, eliminating reliance on physician experience and subjective judgment, and effectively reducing complications caused by misjudgment. Simultaneously, the synergy between real-time data processing and model identification allows for dynamic tracking of puncture progress, further ensuring the accuracy and safety of the puncture operation. Attached Figure Description

[0016] Figure 1 This is a flowchart of a spinal canal puncture needle progression identification method provided in one embodiment of this application;

[0017] Figure 2 This is a flowchart of the training process of a puncture progress recognition model provided in one embodiment of this application;

[0018] Figure 3 This is provided in one embodiment of the present application. Figure 2 The detailed flowchart of step 230;

[0019] Figure 4 This is provided in one embodiment of the present application. Figure 3 The detailed flowchart of step 340;

[0020] Figure 5 This is a flowchart of the overall algorithm for needle progression recognition provided in one embodiment of this application;

[0021] Figure 6 This is a structural diagram of a puncture system module provided in one embodiment of this application;

[0022] Figure 7 This is a structural diagram of a puncture robot provided in one embodiment of this application;

[0023] Figure 8 This is a structural diagram of an end-effector puncture actuator provided in one embodiment of this application;

[0024] Figure 9 This is a structural diagram of a wearable positioning mechanism provided in one embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] It should be noted that although the flowchart shows a logical order, in some cases, the steps shown or described may be performed in a different order than that shown in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the structures, proportions, sizes, etc., depicted in the drawings are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and purposes achieved by this application, should still fall within the scope of the technical content disclosed in this application. Similarly, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are only for clarity of description and are not used to limit the scope of implementation of this application. Changes or adjustments in their relative relationships, without substantially altering the technical content, should also be considered within the scope of implementation of this application.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0028] Spinal puncture is widely used in clinical diagnosis and treatment, but its traditional operation relies heavily on the doctor's manual experience. In practice, due to differences in patients' anatomical structures (such as obesity, spinal deformities, etc.), doctors find it difficult to accurately determine the tissue layer (such as skin, ligamentum flavum, target cavity, etc.) where the puncture needle is located by touch alone, which can easily lead to complications such as puncture failure, cerebrospinal fluid leakage, and nerve damage.

[0029] To improve accuracy, existing technologies have introduced image-guided devices such as ultrasound and X-rays. However, these devices are expensive, complex to operate, and have limited clinical applicability. Furthermore, doctors still need to make subjective judgments on the image results, which cannot achieve real-time and automated identification of needle progress during puncture. This makes it difficult to completely solve the problems of over-reliance on experience and insufficient safety in traditional operations.

[0030] In view of this, this application provides a method and system for identifying the progress of a spinal puncture needle. The method first acquires real-time puncture data, then filters and normalizes it to obtain standardized real-time puncture data. Next, it generates a time series by dividing the data into time steps. Subsequently, a pre-trained puncture progress identification model is called to identify the time series and output the corresponding puncture status identification result. This application, through the standardization and time series transformation of real-time puncture data (such as force signals), combined with a pre-trained model, achieves automatic identification of the tissue layer where the needle is located, eliminating reliance on physician experience and subjective judgment, and effectively reducing complications caused by misjudgment. Simultaneously, the synergy between real-time data processing and model identification allows for dynamic tracking of puncture progress, further ensuring the accuracy and safety of the puncture operation.

[0031] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0032] Reference Figure 1 , Figure 1 This is a flowchart of the spinal canal puncture needle progression identification method provided in the embodiments of this application. The process may specifically include, but is not limited to, steps 110 to 140.

[0033] Step 110: Obtain real-time puncture data;

[0034] Step 120: Filter and normalize the real-time puncture data to obtain standardized real-time puncture data;

[0035] Step 130: Divide the standardized real-time puncture data into time steps to obtain a time series;

[0036] Step 140: Call the pre-trained puncture progress recognition model, perform recognition processing based on the time series, and obtain the corresponding puncture status recognition result.

[0037] Steps 110 to 140 will be described in detail below.

[0038] In a feasible embodiment, in step 110, the real-time puncture data refers to the data collected in real time during the puncture operation (such as clinical surgery or simulation experiment), including but not limited to dynamic data captured in real time by a force sensor (such as an ATI six-axis force sensor) installed at the end of the robotic arm at a fixed frequency (which needs to be consistent with the acquisition frequency during model training, for example, once every 0.08 seconds), the core of which includes the resistance (i.e., puncture force) experienced by the puncture needle in the Z-axis direction.

[0039] In a feasible embodiment, in step 120, noise is introduced into the real-time puncture data during acquisition due to factors such as robotic arm vibration and sensor signal fluctuations. Furthermore, the dimensions and numerical ranges of data such as force and needle insertion depth differ. The pre-trained puncture progress recognition model has strict requirements for adapting to the noise level and numerical distribution of the input data (it must be consistent with the sample characteristics from the training phase). Therefore, it is necessary to filter and normalize the real-time puncture data to obtain standardized real-time puncture data (referring to real-time puncture data that eliminates noise interference, unifies the numerical range, and is fully compatible with the model training input format). The specific operation is as follows: First, an exponentially weighted average filtering algorithm consistent with the model training phase is used, setting the same smoothing factor (e.g., 0.1), to process the real-time acquired puncture force, needle insertion speed, and other data step-by-step. Then, the global extrema statistically obtained during the model training phase (i.e., the minimum and maximum values ​​of all samples in the training dataset) are reused to normalize the filtered real-time data, thereby obtaining standardized real-time puncture data.

[0040] In a feasible embodiment, in step 130, time step division refers to cutting continuous standardized real-time data into fixed-length segments using a sliding window technique. The number of time steps in each segment must be consistent with the sample length required by the model (e.g., 20 time steps, corresponding to a 1.6-second puncture process), so that the resulting time series is an input format that the model can directly accept.

[0041] Understandably, a puncture progress recognition model refers to an artificial intelligence model that learns from data during the puncture process and can automatically determine the current stage of the puncture operation (such as subcutaneous tissue penetration, muscle layer advancement, and intervertebral space arrival). This model is used to assist doctors in monitoring the puncture progress in real time and improve the accuracy of the operation.

[0042] In one feasible embodiment, such as Figure 2As shown, the training process of the pre-trained puncture progress recognition model may include, but is not limited to, steps 210 to 250.

[0043] Step 210: Obtain puncture data samples;

[0044] Step 220: Preprocess the puncture data samples to obtain multiple standard time series samples. The standard time series samples include multiple time step puncture data samples and corresponding label samples. The label samples are used to indicate the puncture degree category corresponding to the time step puncture data samples.

[0045] Step 230: Obtain the target model parameter combination;

[0046] Step 240: Construct an initial puncture progress recognition model based on the target model parameter combination;

[0047] Step 250: Train the initial puncture progress recognition model based on multiple standard time series samples to obtain the puncture progress recognition model.

[0048] In a feasible embodiment, in step 210, the puncture data sample refers to various types of data collected during the puncture operation (such as clinical surgery or simulation experiment), including but not limited to puncture force data at the end of the robotic arm (acquired through a force sensor, such as an ATI force sensor), motion data such as the insertion depth, angle, and speed of the puncture needle, and the timestamp corresponding to the recorded data. The puncture data sample can be a one-dimensional force data sample, specifically referring to single-dimensional force information (such as Z-axis force data collected by a force sensor), for example, axial force data of the puncture needle collected every 0.08 seconds. This type of data can reflect the resistance changes when the needle comes into contact with different tissues during the puncture process (such as the sudden change in force value when piercing the skin or ligamentum flavum).

[0049] In a feasible embodiment, step 220 aims to transform the raw puncture data into a standardized format suitable for model learning and simultaneously annotate key information. This process is specifically divided into two consecutive processing stages: first, filtering and labeling the puncture data samples to obtain puncture-labeled data samples; then, dividing the puncture-labeled data samples into multiple time-series samples and normalizing each time-series sample to obtain multiple standard time-series samples. Specifically, the puncture data samples are first filtered using an exponentially weighted average filtering method (e.g., setting the smoothing factor to 0.1) to smooth out noise in the original data (such as instantaneous fluctuations in force caused by robotic arm vibration) while preserving the true force change trend during the puncture process (such as the abrupt force change characteristics when puncturing tissue). After filtering, medical experts can combine the imaging records during the puncture process (such as ultrasound images) to label the puncture data at each time point (i.e., each time step) in the filtered data with the corresponding puncture degree category. The puncture severity category here is defined based on the tissue layer where the puncture needle is located. For example, it can be divided into category 0 (before skin penetration), category 1 (after skin penetration to pre-entry into the ligamentum flavum), category 2 (entry into the ligamentum flavum to penetration), and category 3 (penetration into the ligamentum flavum to reach the target cavity). After the above two steps, the resulting dataset is the puncture-labeled data sample, which contains both denoised puncture data and information associated with the corresponding puncture stage. After obtaining the puncture-labeled data sample, a sliding window technique can be used to segment the continuous puncture-labeled data into multiple independent subsamples. For example, each subsample can contain force data and corresponding labels for 20 consecutive time steps. Here, the time step puncture data sample refers to the puncture data at each moment in the time series (such as the force value, needle depth, etc. at a certain moment). In the above 20 consecutive time-moment data, the data at each moment belongs to a time step sample. This division method ensures that each subsample can completely reflect the characteristics of a continuous puncture process. Subsequently, the various time series samples obtained from the division were normalized using the MinMax normalization method, mapping the data to the [0,1] interval (the calculation formula is: normalized value = (original value - minimum value) / (maximum value - minimum value)). This eliminates the dimensional differences between different types of data (such as force value and needle insertion depth), preventing the model from ignoring key features due to differences in the numerical magnitude of the data. After the above two stages of processing, the final multiple standard time series samples are structured data sets arranged in chronological order. Each sample contains multiple consecutive time-step puncture data samples and corresponding label samples (i.e., markers used to indicate the puncture degree category of the corresponding time-step puncture data, such as numbers 0-3).

[0050] In a feasible embodiment, in step 230, the target model parameter combination refers to the set of CNN hyperparameters optimized by the MOEA / D multi-objective evolutionary algorithm. This process simulates natural evolution and simultaneously optimizes multiple objectives such as model performance (e.g., puncture stage recognition accuracy) and efficiency (e.g., inference speed) within a preset hyperparameter search space. Ultimately, it selects the optimal CNN hyperparameters that balance recognition accuracy and computational efficiency, providing the core configuration for constructing a puncture progress recognition model.

[0051] In one feasible embodiment, such as Figure 3 As shown, the execution process of obtaining the target model parameter combination in step 230 may include, but is not limited to, steps 310 to 360.

[0052] Step 310: Based on the preset hyperparameter search space and reference point, generate an initial population containing several sets of hyperparameter combinations. The parameters of each individual in the initial population satisfy the constraints of the hyperparameter search space.

[0053] Step 320: Using the reference point as a baseline, determine the fitness value for each individual;

[0054] Step 330: Based on the fitness value of each individual, select parent individuals from the initial population, and based on the hyperparameter search space, obtain candidate individuals by performing crossover and mutation operations on the parent individuals;

[0055] Step 340: Using the reference point as a benchmark, determine the candidate fitness value of the candidate individuals;

[0056] Step 350: Update the initial population based on the candidate fitness values ​​and the fitness values ​​of each individual in the parent generation to obtain the updated population;

[0057] Step 360: Use the updated population as the initial population for the new round of iterations and perform iterative calculations until the number of iterations reaches the preset number of generations. Select the target individual from the population as the target model parameter combination.

[0058] In a feasible embodiment, step 310 aims to define the selectable range of hyperparameters, generate an initial hyperparameter combination (population), and clarify the optimization objective. It can be understood that the hyperparameter search space (Θ) refers to the range of values ​​for CNN hyperparameters, defining the parameter boundaries for optimization. The population refers to a set of multiple hyperparameter combinations, each group called an individual (e.g., a specific combination of the number of convolutional filters or fully connected units). Reference point (z) * ) refers to the ideal optimization target (such as 95% accuracy and 50ms inference time) as the evaluation benchmark.

[0059] The hyperparameter search space is defined as shown in Equation (1), where conv_filters refers to the number of filters in the convolutional layer, affecting the ability to extract puncture force features (16 → basic features, 256 → complex features). fc_units refers to the number of neurons in the fully connected layer, determining the complexity of puncture feature fusion (32~128 cover different needs). batch_size refers to the training batch size, affecting training efficiency (32 / 64 / 128 are commonly used values).

[0060] Θ={conv_filters∈[16,256],

[0061] fc_units∈[32,128],

[0062] batch_size∈{32,64,128}}(1)

[0063] Multiple sets of hyperparameters are randomly generated within Θ to form the initial population.

[0064] In one feasible embodiment, step 320 aims to quantify the merits of a combination of hyperparameters using a formula, while balancing classification accuracy and inference speed. Fitness measures how close an individual is to a reference point; generally, a smaller value is better. The fitness value g for each individual is calculated... t In this case, the Tchebycheff function (a method that aggregates multi-objective biases into a single value) can be used. As shown in equation (2),

[0065]

[0066] Where x represents the current hyperparameter combination (such as the specific number of filters, batch size); λ = [λ1, λ2] represents the target weights (e.g., λ1 = 0.7 emphasizes accuracy, λ2 = 0.3 takes speed into account). f1(x) represents the model's classification accuracy for puncture data, and f2(x) represents the model's inference time. The operational logic is: take the weighted maximum value of the deviations between the two targets and the reference point; the smaller the value, the closer it is to the ideal state.

[0067] In one feasible embodiment, step 330 aims to generate a better combination through crossover (fusing superior parameters) and mutation (exploring new parameters). Crossover refers to exchanging hyperparameter fragments of individuals (e.g., exchanging filter numbers with neighboring individuals) to fuse superior genes. Mutation refers to mechanically adjusting hyperparameters to explore new regions and avoid local optima. Neighborhood crossover x′ k The definition is shown in equation (3):

[0068]

[0069] Where, x k Represents the parent individual, θ j / θk η represents the hyperparameters of the same type in the neighborhood / current individual (such as fc_units for both), η represents the learning rate (controlling the magnitude of parameter changes to ensure it is within Θ), and pc represents the crossover probability.

[0070] The definition of Gaussian mutation is shown in equation (4):

[0071] θ i ′=θ i +N(0,σ) 2 ), σ=0.1×θ i (4)

[0072] Where, θ i N(0, σ) represents a hyperparameter of an individual (such as batch_size). 2 ) represents Gaussian noise (simulated random variation), σ = 0.1 × θ i This indicates that the noise standard deviation is proportional to the hyperparameter (e.g., when batch_size = 32, σ = 3.2, ensuring reasonable variation).

[0073] In a feasible embodiment, in step 340, the classification accuracy and reasoning speed of the candidate individual are evaluated using the formula from step 320, and its fitness g can be obtained. t (x').

[0074] In one feasible embodiment, step 350 aims to retain only the better individuals, driving the population toward higher fitness. The elite retention strategy used is shown in equation (5):

[0075] If g t (x') <g t (x) → Replace x with x' (5)

[0076] Replace the parent only when the candidate individual has better fitness (closer to the reference point) to ensure improved population quality.

[0077] In a feasible embodiment, step 360 aims to select the most balanced combination of hyperparameters from the Pareto front after iterative evolution. The Pareto front (P) refers to the optimal boundary of a multi-objective system, i.e., the combination on the front, where it is impossible to optimize two objectives simultaneously (e.g., improving accuracy inevitably reduces speed). The compromise solution refers to the most balanced combination selected from the front by weights (balancing accuracy and speed).

[0078] The Pareto front analysis is shown in equation (6):

[0079]

[0080] That is, no other combination can be superior to θ in both accuracy and speed.

[0081] The compromise solution is chosen as shown in equation (7):

[0082]

[0083] in, This represents the maximum accuracy in the frontier. This represents the maximum velocity in the frontier; the weighting coefficients w1 = 0.7, w2 = 0.3.

[0084] After iterating to the preset algebra, select the weighted sum with the smallest θ. opt The parameters are used as the target model parameter combination (optimal CNN hyperparameters).

[0085] In one feasible embodiment, such as Figure 4 As shown, the execution process of step 340 may include, but is not limited to, steps 410 to 430.

[0086] Step 410: Obtain performance data for each individual;

[0087] Step 420: Using the reference point as a benchmark, calculate the overall difference between the performance data of each individual and the reference point using a multi-objective decomposition method;

[0088] Step 430: Use the overall gap as the fitness value of the corresponding individual.

[0089] In a feasible embodiment, step 410 mainly involves collecting the actual performance data of the models corresponding to candidate individuals (i.e., hyperparameter combinations generated through crossover and mutation) in the puncture task. Here, candidate individuals refer to the new hyperparameter combinations generated in step 330 (such as a specific number of convolutional filters, fully connected units, etc.), with each individual corresponding to a CNN model. Performance data are multi-dimensional indicators for measuring model performance. In the context of the puncture scenario, these typically include: primary objective: classification accuracy (the proportion of puncture stages correctly identified by the model, such as the probability of correctly judging "puncture of the ligamentum flavum"); secondary objective: inference time (the time taken for the model to output a prediction of a single puncture stage, which must meet clinical real-time requirements); and other auxiliary indicators (such as model stability, error rate, etc.).

[0090] In a feasible embodiment, when acquiring the performance data of each individual, the CNN model corresponding to the candidate individual can be tested on the puncture data validation set (preprocessed standard time series samples), and the specific values ​​of the above performance indicators (such as accuracy of 89% and inference time of 65ms) can be recorded.

[0091] In a feasible embodiment, step 420 primarily involves decomposing and integrating the differences between multiple performance indicators into a single comprehensive value based on an ideal target (reference point), quantifying the degree of deviation of an individual from the ideal state. The multi-target decomposition method refers to an algorithm (such as Tchebycheff decomposition) that transforms the differences between multiple targets, such as "classification accuracy" and "inference time," into a single comprehensive value. The comprehensive gap refers to the single numerical value obtained after decomposing the differences between multiple targets and the reference point, reflecting the overall degree of deviation of the individual from the ideal state.

[0092] In a feasible embodiment, taking Tchebycheff decomposition as an example, for a certain candidate individual, let its performance data be [f1,f2] (f1 is the actual accuracy and f2 is the actual inference time), then the comprehensive gap formula is as shown in Equation (2), and the maximum value of the two weighted gaps, max{·}, can be taken as the comprehensive gap.

[0093] In one feasible implementation, the core function of the fitness value is to quantify the quality of individuals. The smaller the overall difference, the closer the individual's performance is to the reference point (ideal state), and the better the fitness value; conversely, the larger the overall difference, the worse the fitness value. For example, if the overall difference of individual A is 0.05 and the overall difference of individual B is 0.12, then A has a better fitness than B and is more likely to be retained in subsequent population updates.

[0094] In a feasible embodiment, after obtaining the target model parameter combination (the optimal CNN hyperparameters obtained through MOEA / D optimization, such as 3 convolutional layers, 64 filters per layer, 32 fully connected layer units, and a learning rate of 5e4), an initial puncture progress recognition model can be constructed based on the target model parameter combination. This initial puncture progress recognition model is a CNN network structure built according to the target parameter combination. At this stage, the model's parameters (such as convolutional kernel weights and fully connected layer weights) are randomly initialized values, and it has not yet learned the characteristic patterns of the puncture data, thus lacking the ability to recognize puncture progress (i.e., unable to determine the tissue layer where the puncture needle is located). The specific construction process includes: constructing the CNN network layer by layer according to the structural hyperparameters (such as the number of convolutional layers, filters, and fully connected layer units) in the target parameter combination. The CNN network comprises: an input layer that receives standard time-series samples (e.g., puncture force data with 20 time steps and dimensions [20,1]); a pooling layer that reduces dimensionality by a window size (e.g., 2×1) while preserving key features (e.g., peak force); a fully connected layer that integrates the features output from the convolutional layers with a set number of units (e.g., 32) to prepare for classification; and an output layer that uses the Softmax activation function to output the probability distributions for four categories (0 and 3, corresponding to different puncture stages). All learnable parameters in the network (e.g., convolutional kernel weights and bias terms) are randomly initialized (e.g., using a normal distribution) to ensure the model has learning potential but is unbiased in its initial state.

[0095] In one feasible embodiment, preprocessed standard time series samples are input into the initial model, and the model parameters are iteratively optimized so that the model gradually learns the correlation between puncture data features and puncture severity categories, and finally has the ability to accurately identify puncture progress.

[0096] In a feasible embodiment, step 250 includes the following steps: First, the multiple standard time series samples are divided into a training set, a validation set, and a test set. The training set is used for learning and updating model parameters, the validation set is used to evaluate model performance during training to assist in optimization, and the test set is used to finally evaluate the generalization ability of the trained model. During the training phase, the standard time series samples in the training set are input into the initial puncture progress recognition model, and the model's predicted labels for the samples are calculated. The error between the predicted labels and the true labels of the samples is calculated using the cross-entropy loss function. Based on this error and a preset loss function, the gradients of the network parameters of each layer of the model are derived in reverse, and the network parameters are updated based on the gradients to obtain the optimized puncture progress recognition model. After each iteration, the classification accuracy of the optimized model is evaluated using the validation set to determine whether the model is overfitting or tending to converge. When the number of iterations reaches a preset number of training rounds, or the classification accuracy on the validation set does not exceed the historical benchmark value for a consecutive preset number of rounds (i.e., the model converges), the iteration stops. After stopping the iteration, the model at this point is taken as the trained puncture progress recognition model. Furthermore, the classification accuracy can be finally evaluated using the test set to verify the model's generalization ability on data not used in training.

[0097] Specifically, multiple standard time series samples are first divided into training, testing, and validation sets, with the training set typically accounting for 70%-80% and the testing and validation sets accounting for 20%-30%. Each of these standard time series samples contains puncture data from multiple consecutive time steps (such as normalized force values ​​with dimensions [20,1], where 20 represents the time step) and corresponding labels (0-3 classes, corresponding to puncture stages such as before skin penetration, subcutaneous tissue penetration, ligamentum flavum penetration, and target cavity penetration, respectively).

[0098] The labels of these samples need to be converted into one-hot vectors using formula (8):

[0099]

[0100] Where δ(·) is an indicator function, with the label corresponding to the category position set to 1 and the rest to 0, providing an adaptation format for subsequent cross-entropy calculation.

[0101] Entering the training loop, the forward propagation stage sequentially undergoes multiple layers of feature extraction and transformation: the first convolutional layer (Equation 9): 16 filters (kernel size 3) process the input time series. Perform convolution to extract local force features (such as abrupt changes or trends):

[0102]

[0103] Among them, W k (1)[d] represents the convolution kernel weights (covering three consecutive time steps). ReLU activation introduces nonlinearity, highlighting key force changes such as tissue penetration.

[0104] The final model, once trained, can take in new puncture data (such as real-time force sequences) and output the corresponding puncture degree category (or probability), thus enabling automatic identification of the tissue layer (such as skin, ligamentum flavum, or target cavity) where the puncture needle is located.

[0105] First pooling layer (Formula 10): Max pooling with a window size of 2 and a stride of 2, reducing the dimensionality of the convolutional output while preserving core features (such as peak force).

[0106] Z (1,pool) [t,k]=max p∈[0,1] Z (1) [2t+p,k] (10)

[0107] First batch normalization (Formula 11): using the batch mean μ B ,variance and learnable parameter γ (1) β (1) Stabilizing the input distribution of the layer accelerates convergence.

[0108]

[0109] The second convolutional layer (Equation 12), pooling and normalization, and the third convolutional layer (Equation 13, 64 filters to deepen features) are executed sequentially to gradually extract complex force patterns:

[0110]

[0111] The output of the third convolutional layer, after pooling and normalization, enters the flattening layer (Equation 14), which flattens the three-dimensional feature map into a one-dimensional vector:

[0112]

[0113] The fully connected layer (Equation 15) compresses features through two network layers (two Dense layers), with 64 and 32 units respectively, using ReLU activation:

[0114] h1 = ReLU(W1z) flat +b1),h2=ReLU(W2h1+b2) (15) The Softmax output layer (Equation 16) generates the probability distributions for the four types of puncture stages:

[0115]

[0116] For example, p1 = 0.95 means that the model predicts that the current puncture is in the "subcutaneous tissue" with a probability of 95%.

[0117] After the prediction is completed, the difference between the prediction and the true label is quantified by cross-entropy loss (Formula 17):

[0118]

[0119] Where N is the number of batch samples, the smaller the loss value, the closer the prediction is to the actual puncture stage.

[0120] Subsequently, backpropagation is used to calculate the gradient, and the model parameters θ (including convolutional kernels, fully connected layer weights, etc.) are updated using the Adam optimizer (Equation 18):

[0121]

[0122] in, For the first moment estimate of the gradient, b t The second moment of the gradient is estimated, and α is the learning rate, which drives the model to optimize in the direction of reducing loss.

[0123] After each iteration, the classification accuracy is evaluated using the validation set (the percentage of samples whose predictions match the true labels) to determine if the model is overfitting. Training terminates when the preset number of iterations is reached, or when the validation set accuracy fails to surpass the historical peak for several consecutive iterations (e.g., 5 iterations) (model convergence). The final saved model is the puncture progress recognition model. This model can take a new puncture force sequence as input, and output the stage probability through the aforementioned forward propagation process to determine the tissue layer (e.g., skin, ligamentum flavum, target cavity) where the puncture needle is located.

[0124] In a feasible embodiment, the trained puncture progress recognition model (which already has the ability to extract features and classify them from time-series data) can be further invoked to perform recognition processing based on the time series, thereby obtaining the corresponding puncture state recognition result. Specifically, when invoked, after the time series is input into the model, forward propagation is performed (convolutional layers extract local force features, pooling layers enhance key features, and fully connected layers integrate information). Finally, the Softmax layer outputs the probability distribution of four puncture stages (e.g., class 0: before skin penetration, class 1: subcutaneous tissue, class 2: ligamentum flavum, class 3: target cavity) (e.g., output [0.02, 0.95, 0.02, 0.01]). The class with the highest probability can be selected as the puncture state recognition result, directly indicating the tissue layer where the puncture needle is currently located, providing real-time reference for the doctor's operation.

[0125] See Figure 5 , Figure 5This is a flowchart of the overall algorithm for needle progression recognition provided in one embodiment of this application. The process includes: firstly, preprocessing the original time-series data such as puncture force and needle depth; extracting continuous data segments of a fixed length (e.g., 20 time steps) through a sliding window; eliminating dimensional differences through normalization; and dividing the processed data into training and testing datasets to support model training and generalization capability verification, respectively. Simultaneously, for the hyperparameters of the CNN model (e.g., the number of filters in convolutional layers, the number of units in fully connected layers), the MOEA / D multi-objective optimization algorithm is introduced: first, the population is initialized, and the hyperparameter search space and multi-objective reference points are defined; the individual fitness is calculated using the Tchebycheff decomposition method to evaluate the comprehensive performance of the hyperparameter combination on multiple objectives; then, a neighbor or global population strategy is used to perform crossover operations, combined with mutation to generate new individuals; after comparing fitness, a better solution is replaced; the process is iterated to a preset number of generations, and the optimal hyperparameter combination at the Pareto front is output. Next, a CNN model is constructed based on the optimized hyperparameters: convolutional and pooling layers extract temporal features of the puncture force (such as force abrupt changes and trend changes during tissue penetration), and fully connected layers integrate the features. A Softmax classifier is then used to learn the mapping relationship between "mechanical pattern → puncture stage (skin, ligamentum flavum, etc.)". The model parameters are iteratively optimized using the training set, while the validation set is used to monitor the risk of overfitting. The trained model can be deployed to actual puncture scenarios. For example, for force data collected in real time by a puncture robot, the preprocessing steps of filtering, normalization, and time step division are reproduced, input into the model to perform forward inference, and the output is the identification result of the tissue layer where the puncture needle is located (such as "puncture of the ligamentum flavum" or "entering the target cavity").

[0126] like Figure 6 As shown in the illustration, this application also discloses a puncture system applicable to the aforementioned spinal canal puncture needle progression recognition method. This system includes a data processing module 610, a data monitoring module 620, a puncture robot 630, and a wearable positioning mechanism 640. The wearable positioning mechanism 640 is worn on the user's waist to assist in fixing the puncture position. Figure 7 As shown, the puncture robot 630 is equipped with an end effector 710, which is responsible for not only performing puncture operations on the user's waist, but also simultaneously collecting real-time puncture data. The data processing module 610 filters and normalizes the collected real-time puncture data to generate standardized real-time puncture data, and then further divides it into time series by time steps. The data monitoring module 620 is equipped with a pre-trained puncture progress recognition model, which can complete the recognition processing based on the above time series and output the corresponding puncture status recognition result.

[0127] In one feasible embodiment, such as Figure 7As shown, the puncture robot 630 also includes a carriage housing equipped with a control device. The control device is electrically connected to the data monitoring module 620 to achieve command interaction. A first robotic arm 720 is mounted on the carriage housing, which is electrically connected to the control device. An end effector 710 is mounted on the first robotic arm 720. The first robotic arm 720 can flexibly adjust the spatial position of the end effector 710 under the control of the control device. In addition, a second robotic arm 730 is also mounted on the carriage housing, with a display device 740 mounted at its end. The second robotic arm 730 can adjust the orientation and height of the display device 740, which is used to display the puncture status recognition results in real time for easy observation by the operator.

[0128] In one feasible embodiment, such as Figure 8 As shown, the end-effector puncture actuator 710 is equipped with a force sensor 810. The sensor 810 establishes a communication connection with the data processing module 610, and can collect real-time puncture data such as mechanical data during the puncture process and transmit it to the data processing module 610.

[0129] In one feasible embodiment, such as Figure 9 As shown, the wearable positioning mechanism 640 includes a lumbar support unit 910, a needle insertion limiting cylinder 920, a ratchet-like telescopic connecting device, a telescopic arm 940, a locking contact unit 950, and a waist belt mounting component 960. The needle insertion limiting cylinder 920 is fixed to the locking grooves of the lumbar support unit 910 via locking points at both ends. The waist belt mounting component 960 is symmetrically distributed on both sides of the lumbar support unit 910, facilitating the fixation of the device to the user's waist. The telescopic arm 940 is connected to the lumbar support unit 910 via the ratchet-like telescopic connecting device. The ratchet-like telescopic connecting device is equipped with a ratchet-like device switch 930, which allows adjustment of the telescopic arm 940's extension length and angle. The locking contact unit 950 is installed at the tail of the telescopic arm 940 for precise positioning of the spine, ensuring accurate puncture.

[0130] In practical applications, users (such as doctors) can first press the ratchet-like switch 930 to adjust the telescopic arm 940 of the wearable positioning mechanism 640 to its maximum position. After the patient places the lumbar support unit 910 against their back spine, both telescopic arms 940 simultaneously tighten inward until they lock onto the spinous processes on both sides of the lumbar region. For patients with different body shapes, the locking contact unit 950 can be adjusted to fit the position of the lumbar spinous processes. After locking, the waist belt is further secured by fastening the waist belt mounting components 960 on both sides of the lumbar support unit 910. The needle insertion cylinder 920 is replaceable. After the doctor determines the puncture location, it is installed at the corresponding needle insertion position of the lumbar support unit 910 and locked in place by the locking points at both ends and the locking grooves of the lumbar support unit 910. Subsequently, the doctor moves the trolley to a suitable position according to the puncture site and locks the wheels; checks whether the internal passage between the puncture needle 830 and the connector 820 in the end-effector puncture actuator 710 is unobstructed, loosens the hand-operated chuck 840, inserts the rear cylindrical rod of the connector into the chuck for clamping and fixation, and can install a medical barometer at the passage opening. Next, the doctor grasps the hand-operated sleeve at the front end of the first robotic arm 720 and adjusts the position of the first robotic arm 720 by dragging it to position the end-effector puncture actuator 710 in a suitable puncture posture. After preparation, the system is started, the data monitoring module 620 loads the pre-trained puncture progress recognition model, and zeroes the force sensor 810; the second robotic arm 730 adjusts the display device 740 to a position for easy observation. When executing the puncture procedure, the first robotic arm 720 clamps the puncture needle at a 90-degree angle to the patient's back skin to ensure that the needle enters the spinal canal in a straight line along the spine, and then advances it at a preset speed (e.g., 3 mm / s) in servo mode. During the puncture, the force sensor 810 of the end-effector puncture actuator 710 collects puncture force data in real time (step 110) and transmits it to the data processing module 610. The data processing module 610 filters and normalizes the real-time force data to obtain standardized real-time puncture data, and then generates a time series by dividing the data into time steps (steps 120 and 130). The data monitoring module 620 calls the loaded puncture progress recognition model and identifies the tissue layer where the puncture needle is located based on the time series (step 140). The recognition result, along with the real-time force data and the pressure data from the barometer, is displayed on the display device 740 for easy observation by the doctor. When the model recognition result is Class 3 (puncture of the ligamentum flavum to reach the target cavity), the system sends a stop puncture command to the first robotic arm 720 to complete the puncture operation.

[0131] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the progression of a spinal canal puncture needle, characterized in that, include: Obtain real-time puncture data; The real-time puncture data is filtered and normalized to obtain standardized real-time puncture data; The standardized real-time puncture data is divided into time steps to obtain a time series; The pre-trained puncture progress recognition model is invoked, and the recognition process is performed based on the time series to obtain the corresponding puncture status recognition result.

2. The method for identifying the progression of a spinal canal puncture needle according to claim 1, characterized in that, The method further includes: Obtain puncture data samples; The puncture data samples are preprocessed to obtain multiple standard time series samples. The standard time series samples include multiple time step puncture data samples and corresponding label samples. The label samples are used to indicate the puncture degree category corresponding to the time step puncture data samples. Obtain the target model parameter combination; Based on the target model parameter combination, an initial puncture progress recognition model is constructed; The initial puncture progress recognition model is trained based on the multiple standard time series samples to obtain the puncture progress recognition model.

3. The method for identifying the progression of a spinal canal puncture needle according to claim 2, characterized in that, The puncture data samples are preprocessed to obtain multiple standard time series samples, including: The puncture data samples are filtered and labeled to obtain puncture-labeled data samples; The puncture marker data samples are divided into multiple time series samples, and each time series sample is normalized to obtain multiple standard time series samples.

4. The method for identifying the progression of a spinal canal puncture needle according to claim 2, characterized in that, The acquisition of the target model parameter combination includes: Based on a preset hyperparameter search space and reference point, an initial population containing several sets of hyperparameter combinations is generated, wherein the parameters of each individual in the initial population satisfy the constraint range of the hyperparameter search space. Using the reference point as a benchmark, determine the fitness value of each individual; Based on the fitness values ​​of each individual, parent individuals are selected from the initial population, and candidate individuals are obtained by performing crossover and mutation operations on the parent individuals based on the hyperparameter search space. Using the reference point as a benchmark, determine the candidate fitness value of the candidate individual; The initial population is updated based on the candidate fitness values ​​and the fitness values ​​of each individual in the parent population to obtain the updated population. The updated population is used as the initial population for a new round of iterations and iterative calculations are performed until the preset number of generations is reached. Target individuals are then selected from the population as the target model parameter combination.

5. The method for identifying the progression of a spinal canal puncture needle according to claim 4, characterized in that, The determination of the fitness value of each individual based on the reference point includes: Obtain the performance data of each of the individuals; Using the reference point as a benchmark, a multi-objective decomposition method is used to calculate the comprehensive difference between the performance data of each individual and the reference point; The overall gap is used as the fitness value for the corresponding individual.

6. The method for identifying the progression of a spinal canal puncture needle according to claim 2, characterized in that, The step of training the initial puncture progress recognition model based on the multiple standard time series samples to obtain the puncture progress recognition model includes: The multiple standard time series samples are divided into a training set and a validation set; The standard time series samples in the training set are input into the initial puncture progress recognition model. The predicted labels of the model for the standard time series samples are calculated. The error between the predicted labels and the corresponding label samples of the standard time series samples is calculated using the cross-entropy loss function. The gradients of the network parameters of each layer of the model are calculated based on the error and the preset loss function. The network parameters are updated based on the gradients to obtain the optimized initial puncture progress recognition model. The classification accuracy of the optimized initial puncture progress recognition model is evaluated using the validation set. The iteration stops when the number of iterations reaches a preset number of training rounds, or when the classification accuracy does not exceed the historical benchmark value for a preset number of consecutive rounds. The model at the point where iteration stops is used as the puncture progress recognition model after training is complete.

7. A puncture system, characterized in that, The method for identifying the progression of a spinal canal puncture needle as described in any one of claims 1-6, the system comprising a data processing module, a data monitoring module, a puncture robot, and a wearable positioning mechanism; The wearable positioning mechanism is worn on the user's waist; The puncture robot is equipped with an end effector, which is used to puncture the user's waist and acquire real-time puncture data. The data processing module is used to filter and normalize the real-time puncture data to obtain standardized real-time puncture data; and to divide the standardized real-time puncture data into time steps to obtain a time series. The data monitoring module is equipped with a puncture progress recognition model, which is used to perform recognition processing based on the time series to obtain the corresponding puncture status recognition result.

8. The puncture system according to claim 7, characterized in that, The end-effector is equipped with a force sensor, which communicates with the data processing module. The force sensor is used to collect real-time puncture data during the puncture process and send the real-time puncture data to the data processing module.

9. The puncture system according to claim 7, characterized in that, The puncture robot also includes: The trolley body is equipped with a control device, which is electrically connected to the data monitoring module; The first robotic arm is mounted on the trolley body and electrically connected to the control device. The end-effector is mounted on the first robotic arm, and the first robotic arm is used to adjust the position of the end-effector under the control of the control device. The second robotic arm is mounted on the trolley body; A display device is mounted on the second robotic arm, which is used to adjust the position of the display device. The display device is used to display the puncture status recognition result.

10. The puncture system according to claim 7, characterized in that, The wearable positioning mechanism includes a lumbar support unit, a needle insertion limiting cylinder, a ratchet-like telescopic connecting device, a telescopic arm, a locking contact unit, and a waist belt mounting component. The needle insertion limiting cylinder is locked into the locking groove of the lumbar support unit via locking points at both ends. The waist belt mounting component is symmetrically arranged on both sides of the lumbar support unit. The telescopic arm is connected to the lumbar support unit via the ratchet-like telescopic connecting device, which includes a ratchet-like device switch for adjusting the position of the telescopic arm. The locking contact unit is located at the tail of the telescopic arm for spinal positioning.

Images