阅读说明：本技术 一种基于剪切结束判断模型的控制方法和系统 (Control method and system based on cutting end judgment model ) 是由 王福源 姚龙洋 李阳 刘振中 骆威 于 2021-08-30 设计创作，主要内容包括：本发明揭示了一种基于剪切结束判断模型的控制方法和系统,系统由耦接的超声刀杆和换能器组成,并通过线缆连接到发生器；使用机器学习算法模型来获得剪切状态变化特征,包括但不限于工作反馈参数、物理结构特征参数、环境参数的一种或多种组合,有效判断剪切结束点,控制输出功率并提醒医生结束剪切,从而减小医生操作压力,在达到最佳手术效果的同时保护刀具,降低刀具磨损,延长刀具寿命,降低潜在高温风险。(The invention discloses a control method and a system based on a shearing end judgment model, wherein the system consists of an ultrasonic cutter bar and an energy converter which are coupled and is connected to a generator through a cable; the machine learning algorithm model is used for obtaining the shearing state change characteristics, including but not limited to one or more combinations of work feedback parameters, physical structure characteristic parameters and environmental parameters, effectively judging the shearing end point, controlling the output power and reminding a doctor to finish shearing, so that the operating pressure of the doctor is reduced, the cutter is protected while the optimal operation effect is achieved, the cutter abrasion is reduced, the service life of the cutter is prolonged, and the potential high-temperature risk is reduced.)

1. A control method based on a cutting end judgment model is characterized by comprising the following steps,

s1, saving the cutting end judgment model and at least one threshold;

s2, inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

s3, comparing the model output result with the threshold value;

and S4, according to the comparison result, adjusting the power level applied to the ultrasonic blade transducer to control the current output of the ultrasonic blade, and when the ultrasonic blade is judged to reach the tissue shearing end point, reducing the control current of the ultrasonic blade at the time point and generating a shearing end prompt.

2. The method according to claim 1, wherein the threshold is a set of category values corresponding to different clipping states, and the model in step S2 outputs one of the set of category values.

3. The method as claimed in claim 1, wherein the threshold is at least one decimal value within 0-1, the decimal value representing a predetermined end probability threshold, and the real-time clipping end probability value outputted in step S2 is compared with the probability threshold in step S3.

4. The method according to any one of claims 1 to 3, wherein the cutting-end judgment model is a neural network algorithm model, and comprises one or more algorithm model combinations of a feedforward neural network, a memory neural network and an attention neural network, and the model training method is one or more combinations of supervised learning, semi-supervised learning, unsupervised learning and reinforcement learning.

5. The method as claimed in claim 4, wherein the model training method is specifically to take input features from a training data set, input the features into a neural network model to calculate a median value and a gradient value of each neuron, the loss function of the model may be a loss function such as cross entropy, Mean Square Error (MSE), etc., and update weights by using a gradient descent method, and repeat the above processes until the model reaches a predetermined stopping condition, and stop training and save the model after the stopping condition is reached.

6. The method according to claim 5, wherein all input feature information and training labels of the model need to be collected periodically during the model training process, and the training labels are either a category value set, are divided into at least two types of state categories, and are represented by integer values; the training labels are either probability values within 0-1 sampled from a probability distribution, which may be a uniform distribution or a normal distribution.

7. The method of claim 5, wherein the input features of the cropping end judgment model comprise one or more combinations of working feedback parameters, physical structure feature parameters, and environmental parameters.

8. The method of claim 7, wherein the operational feedback parameters include, but are not limited to, real-time voltage U, real-time current I, real-time resonant frequency f, first derivative of frequency df, power P, impedance R, voltage current phase difference Φ; the physical structure characteristic parameters include but are not limited to ultrasonic cutter bar material, length and shape; the environmental parameters include, but are not limited to, ambient temperature, ambient humidity.

9. The method according to claim 8, wherein the real-time voltage U and the real-time current I are sampled by the generator in real time through a current-voltage sensor, and the real-time resonant frequency f, the first derivative of frequency df, the power P, the impedance R, and the voltage-current phase difference Φ are obtained by FFT mathematical operation of the sampled values: the material, the length and the shape of the ultrasonic knife bar, the type of the ultrasonic knife and the tissue to be cut are stored in the ultrasonic knife or a storage chip of a generator, and the generator directly reads the corresponding storage chip to obtain the ultrasonic knife bar material, the length and the shape; the environmental parameters are obtained by real-time measurement through a sensor.

10. The method of claim 1, wherein the clipping-complete determination model is composed of layers and corresponding neurons and weights, the weight parameters and the application program are stored in a generator memory, the memory is Flash, EEPROM or other non-volatile storage device, the application program runs in a processor, and the processor is either an ARM, DSP, FPGA, CPU, GPU or ASIC chip existing in the generator, or a remote server connected through a network.

11. A control system based on a cutting end judgment model is characterized by comprising

A storage unit for storing a cutting end judgment model and at least one threshold;

the processing unit is used for inputting corresponding input characteristics to the cutting end judgment model and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

a comparison unit for comparing the model output result with the threshold value;

and an adjusting unit for adjusting a power level applied to the ultrasonic blade transducer according to the comparison result to control the ultrasonic blade current output, and when it is judged that the ultrasonic blade has reached a tissue cutting end point, reducing the ultrasonic blade control current at the time point and generating a cutting end prompt.

12. A generator for performing control based on a cut-end determination model is characterized by comprising

A control circuit coupled to a memory, the control circuit configured to be capable of:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

comparing the model output result with the threshold value;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output, and when it is determined that the ultrasonic blade has reached the tissue cutting end point, the ultrasonic blade control current is decreased at that point and an end-of-cut cue is generated.

13. The generator of claim 12, wherein the input features input by the control circuit to the crop-end decision model comprise one or more combinations of operational feedback parameters, physical structure feature parameters, and environmental parameters.

14. A generator according to claim 13, wherein the operational feedback parameters include, but are not limited to, real-time voltage U, real-time current I, real-time resonant frequency f, first derivative of frequency df, power P, impedance R, voltage current phase difference Φ; the physical structure characteristic parameters include but are not limited to ultrasonic cutter bar material, length and shape; the environmental parameters include, but are not limited to, ambient temperature, ambient humidity.

15. An ultrasonic surgical instrument based on a shear-over judgment model, comprising

An ultrasonic electromechanical system comprising an ultrasonic transducer coupled to an ultrasonic blade via an ultrasonic waveguide; and

a generator configured to supply power to the ultrasound transducer, wherein the generator comprises a control circuit configured to be capable of:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

comparing the model output result with the threshold value;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output, and when it is determined that the ultrasonic blade has reached the tissue cutting end point, the ultrasonic blade control current is decreased at that point and an end-of-cut cue is generated.

16. The ultrasonic surgical blade instrument of claim 15 wherein the input characteristics of the control circuit configured to input to the shear termination assessment model include one or more combinations of operational feedback parameters including, but not limited to, real-time voltage U, real-time current I, real-time resonant frequency f, first derivative of frequency df, power P, impedance R, voltage current phase difference Φ; the physical structure characteristic parameters include but are not limited to ultrasonic cutter bar material, length and shape; the environmental parameters include, but are not limited to, ambient temperature, ambient humidity.

17. An ultrasonic blade system based on a shear termination determination model comprising a processor and a non-volatile storage device, wherein the non-volatile storage device contains an application that, when executed by the processor, causes the processor to:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

comparing the model output result with the threshold value;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output, and when it is determined that the ultrasonic blade has reached the tissue cutting end point, the ultrasonic blade control current is decreased at that point and an end-of-cut cue is generated.

18. The ultrasonic blade system of claim 17, wherein the clipping-complete determination model is composed of layers and corresponding neurons and weights, the weight parameters and the application program are stored in a generator memory, the memory is Flash, EEPROM or other non-volatile storage device, the application program runs in a processor, the processor is either an ARM, DSP, FPGA, CPU, GPU or ASIC chip present in the generator, or a remote server connected through a network.

Technical Field

The invention relates to the field of medical instruments, in particular to a control method and a control system of an ultrasonic scalpel, and particularly relates to a control method and a control system based on a shearing end judgment model, a generator provided with the system, an ultrasonic scalpel surgical instrument and an ultrasonic scalpel system.

Background

An ultrasonic surgical system for cutting and hemostasis of soft tissue (called ultrasonic knife system for short) is an instrument which further amplifies ultrasonic vibration obtained by a piezoelectric converter (electric energy is transmitted to the piezoelectric converter through an energy generator and is converted into mechanical energy by the piezoelectric converter), and uses the amplified ultrasonic vibration for cutting and coagulating the soft tissue by an ultrasonic knife rod. Clinical use of such devices allows for focal resection with lower temperatures and less bleeding, and ensures minimal lateral thermal tissue damage. With the popularization of minimally invasive surgery, an ultrasonic surgical knife has become a conventional surgical instrument.

The ultrasonic blade system is mainly composed of a generator, a transducer and an ultrasonic blade bar, as shown in fig. 1, the transducer 11 of the ultrasonic blade is coupled with an ultrasonic blade housing 12, a sleeve 13 is located at the distal end of the ultrasonic blade housing 12, an ultrasonic blade bar 14 located at the most distal end is coupled with the transducer 11 inside the sleeve 13, and the transducer 11 is connected with the generator (not shown) through a cable 15. The current of ultrasonic frequency in the generator is conducted to the transducer, the transducer converts the electric energy into mechanical energy of back and forth vibration, the transmission and amplification of the ultrasonic cutter bar enable the tail end (also called an ultrasonic cutter head) of the ultrasonic cutter bar to vibrate at a certain frequency (such as 55.6kHz), the heat generated by friction causes the water in tissue cells contacted with the cutter point to be vaporized, the protein hydrogen bonds are broken, the cells are disintegrated and fused again, and the tissue is cut after being solidified; when the blood vessel is cut, the ultrasonic cutter bar is contacted with tissue protein, heat is generated through mechanical vibration, the collagen structure in the tissue is damaged, protein coagulation is caused, the blood vessel is further sealed, and the purpose of hemostasis is achieved.

Generally, the working principle of the ultrasonic scalpel is that the working frequency of the transducer is changed by using a phase-locking algorithm in real time according to factors such as actual impedance change and temperature change of a piezoelectric crystal, so that the transducer works at the maximum working efficiency. The existing ultrasonic knife is used for observing and judging whether the shearing of tissues by the ultrasonic knife is finished or not by a doctor, so that a large error is caused: if the cutting is finished in advance, the operation fails, and if the cutting is finished in delay, the jaws can wear the gasket all the time, the temperature of a cutter head can be very high, the service life of the cutter is shortened, and the potential scalding risk is caused.

The prior art CN201910124064.5 discloses a self-adaptive cutting hemostasis control method, which includes "acquiring a feedback signal generated by the target biological tissue collected by a signal collector, calculating the biological impedance of the target biological tissue, determining a target current value or a target voltage value of a driving signal required to be generated by a cutter driving module according to the biological impedance of the target biological tissue, and adjusting the driving signal generated by the cutter driving module according to the target current value or the target voltage value. "the prior art aims to improve the cutting efficiency of the cutter by adjusting the excitation current based on the impedance value, and is implemented by querying a biological tissue database according to the impedance value, that is, by looking up a table, and the control method has a low degree of intelligence and has a large error. In view of the above, the present applicant has previously proposed a method and system for performing data processing and adaptively determining a tissue shear point based on a real-time operating frequency or a real-time impedance.

However, when the ultrasonic knife head clamps tissue during operation, the type and pressure of the clamped tissue, the clamping position and the tissue area all affect the impedance, and the impedance value R changes along with the changes of the clamping pressure, the clamping position and the tissue area during the cutting process of the ultrasonic knife. A curve of the impedance of the ultrasonic blade during operation is shown in fig. 2, wherein the impedance value is normalized and can be in any unit. The total shearing time in the figure is approximately 16s, and the impedance changes continuously along with the shearing process: 0-4 s is the initial stage of tissue shearing, the tissue is gradually dried due to the temperature rise, and the impedance is changed violently; 4-14 s are the tissue shearing and separating stage, and the impedance change is slow; 14 s-16 s are cutting end stages, and the impedance value is gradually reduced in the cutting end stages, so that the cutting end stages have obvious change characteristics. In the actual operation process, the impedance change in the cutting process of the ultrasonic knife is not limited to the change trend in the figure, and for different ultrasonic knife heads, different types of cut tissues, different cutting environments, different cutting modes and the like, the impedance can present complex and various change characteristics.

In the shearing working process of the ultrasonic knife, the generator can continuously adjust the working frequency of the ultrasonic knife through a frequency control algorithm, so that the transducer always works in a resonance state. The transducer operating frequency is affected by factors such as impedance and temperature. One frequency change and frequency first derivative change during operation of the ultrasonic blade is shown in fig. 3. The first derivative df of the frequency may represent the rate of change of the frequency, which may be the first difference of the frequency f, and the calculation formula is:

df[k]＝f[k]-f[k-1] (1)

where df [ k ] is the first derivative value and fk, fk-1 are the frequency values.

In the figure, both the frequency and the first derivative of the frequency are normalized, and the unit can be any unit. The shearing duration is approximately 13s, and it can be seen that both the frequency and the first derivative of the frequency have significant changes during shearing: 0-4 s is the initial stage of shearing, and the first derivative of the frequency changes obviously due to rapid temperature rise and violent impedance change; 4-11 s are the shearing separation stage, the frequency basically changes linearly, and the first derivative of the frequency basically does not change; in the 11-13 s shearing end stage, due to tissue separation and temperature reduction, the frequency has a small-amplitude rising process, the first-order derivative of the frequency is rapidly increased and then rapidly reduced, and the change characteristic is obvious. In the actual operation process, the frequency and the first derivative change are not limited to the change trend in the graph, the working frequency can be influenced by factors such as the type of a cutter, the type of sheared tissue, real-time impedance and temperature, and the like, and complex and various change characteristics can also be presented.

In addition to impedance, frequency, and the first derivative of frequency, the voltage-current phase difference also changes with the shearing process, and these characteristics are influenced by the profile material of the ultrasonic blade holder, the type of tissue to be sheared, the shearing environment, and other factors.

Disclosure of Invention

In order to solve the problems of the prior art, the invention provides a control method and a control system based on a shearing end judgment model, and a generator, an ultrasonic scalpel surgical instrument and an ultrasonic scalpel system provided with the system.

In order to solve the technical problems, the technical scheme of the invention is as follows:

a control method based on a cutting end judgment model comprises the following steps,

s1, saving the cutting end judgment model and at least one threshold;

s2, inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

s3, comparing the model output result with the threshold value;

and S4, according to the comparison result, adjusting the power level applied to the ultrasonic blade transducer to control the current output of the ultrasonic blade, and when the ultrasonic blade is judged to reach the tissue shearing end point, reducing the control current of the ultrasonic blade at the time point and generating a shearing end prompt.

Preferably, the threshold is a category value set corresponding to different clipping states, and the output result of the model in step S2 is one of the category values in the category value set.

Preferably, the threshold is at least one decimal value within 0-1, the decimal value representing a predetermined end probability threshold, and the real-time cut end probability value output in step S2 is compared with the probability threshold in step S3.

Preferably, the cutting end judgment model is a neural network algorithm model, and comprises one or more algorithm model combinations of a feedforward neural network, a memory neural network and an attention neural network, and the model training method is one or more combinations of supervised learning, semi-supervised learning, unsupervised learning and reinforcement learning.

Preferably, the model training method specifically includes the steps of acquiring input features from a training data set, inputting the input features into a neural network model, calculating a middle value and a gradient value of each neuron, updating weights by using a gradient descent method, repeating the above processes until the model reaches a preset stopping condition, stopping training after the stopping condition is reached, and storing the model, wherein the loss function of the model can be a loss function such as cross entropy, Mean Square Error (MSE) and the like.

Preferably, all input characteristic information and training labels of the model need to be acquired periodically in the model training process, and the training labels are or a category value set, are divided into at least two types of state categories, and are represented by integer values; the training labels are either probability values within 0-1 sampled from a probability distribution, which may be a uniform distribution or a normal distribution.

Preferably, the input characteristics of the cropping end judgment model comprise one or more combinations of working feedback parameters, physical structure characteristic parameters and environmental parameters.

Preferably, the working feedback parameters include, but are not limited to, real-time voltage U, real-time current I, real-time resonance frequency f, first derivative of frequency df, power P, impedance R, voltage current phase difference Φ; the physical structure characteristic parameters include but are not limited to ultrasonic cutter bar material, length and shape; the environmental parameters include, but are not limited to, ambient temperature, ambient humidity.

Preferably, the real-time voltage U and the real-time current I are obtained by real-time sampling by a current-voltage sensor through a generator, and the real-time resonant frequency f, the first derivative df of frequency, the power P, the impedance R, and the voltage-current phase difference phi are obtained by performing FFT mathematical operation on sampling values: the material, the length and the shape of the ultrasonic knife bar, the type of the ultrasonic knife and the tissue to be cut are stored in the ultrasonic knife or a storage chip of a generator, and the generator directly reads the corresponding storage chip to obtain the ultrasonic knife bar material, the length and the shape; the environmental parameters are obtained by real-time measurement through a sensor.

Preferably, the clipping ending judgment model is composed of layers and corresponding neurons and weights, weight parameters and an application program are stored in a generator memory, the memory is Flash, EEPROM or other nonvolatile storage devices, the application program runs in a processor, and the processor is either an ARM, DSP, FPGA, CPU, GPU or ASIC chip existing in the generator or a remote server connected through a network.

The invention also discloses a control system based on the cutting end judgment model, which comprises the following components:

a storage unit for storing a cutting end judgment model and at least one threshold;

the processing unit is used for inputting corresponding input characteristics to the cutting end judgment model and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

a comparison unit for comparing the model output result with the threshold value;

and an adjusting unit for adjusting a power level applied to the ultrasonic blade transducer according to the comparison result to control the ultrasonic blade current output, and when it is judged that the ultrasonic blade has reached a tissue cutting end point, reducing the ultrasonic blade control current at the time point and generating a cutting end prompt.

The invention also discloses a generator for controlling based on the shearing ending judgment model, which comprises:

a control circuit coupled to a memory, the control circuit configured to be capable of:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

comparing the model output result with the threshold value;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output, and when it is determined that the ultrasonic blade has reached the tissue cutting end point, the ultrasonic blade control current is decreased at that point and an end-of-cut cue is generated.

Preferably, the input characteristics input to the cropping end judgment model by the control circuit comprise one or more combinations of working feedback parameters, physical structure characteristic parameters and environmental parameters.

Preferably, the working feedback parameters include, but are not limited to, real-time voltage U, real-time current I, real-time resonance frequency f, first derivative of frequency df, power P, impedance R, voltage current phase difference Φ; the physical structure characteristic parameters include but are not limited to ultrasonic cutter bar material, length and shape; the environmental parameters include, but are not limited to, ambient temperature, ambient humidity.

The invention also discloses an ultrasonic scalpel surgical instrument based on the shearing end judgment model, which comprises:

an ultrasonic electromechanical system comprising an ultrasonic transducer coupled to an ultrasonic blade via an ultrasonic waveguide; and

a generator configured to supply power to the ultrasound transducer, wherein the generator comprises a control circuit configured to be capable of:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

comparing the model output result with the threshold value;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output, and when it is determined that the ultrasonic blade has reached the tissue cutting end point, the ultrasonic blade control current is decreased at that point and an end-of-cut cue is generated.

Preferably, the input characteristics input to the cutting end judgment model by the control circuit include one or more combinations of working feedback parameters, physical structure characteristic parameters and environmental parameters, wherein the working feedback parameters include, but are not limited to, real-time voltage U, real-time current I, real-time resonance frequency f, frequency first derivative df, power P, impedance R and voltage-current phase difference Φ; the physical structure characteristic parameters include but are not limited to ultrasonic cutter bar material, length and shape; the environmental parameters include, but are not limited to, ambient temperature, ambient humidity.

The invention also discloses an ultrasonic blade system based on a shear over judge model, comprising a processor and a non-volatile storage device, wherein the non-volatile storage device contains an application program which, when executed by the processor, causes the processor to:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the cutting end judgment model, and outputting a model output result at least comprising a real-time cutting end probability value or a real-time state category;

comparing the model output result with the threshold value;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output, and when it is determined that the ultrasonic blade has reached the tissue cutting end point, the ultrasonic blade control current is decreased at that point and an end-of-cut cue is generated.

Preferably, the clipping ending judgment model is composed of layers and corresponding neurons and weights, weight parameters and an application program are stored in a generator memory, the memory is Flash, EEPROM or other nonvolatile storage devices, the application program runs in a processor, and the processor is either an ARM, DSP, FPGA, CPU, GPU or ASIC chip existing in the generator or a remote server connected through a network.

The invention has the following beneficial effects: the machine learning algorithm model is used for extracting the shearing state change characteristics, effectively judging the shearing end point, controlling the output power and reminding a doctor to finish shearing, so that the operating pressure of the doctor is reduced, the cutter is protected while the optimal operation effect is achieved, the cutter abrasion is reduced, the service life of the cutter is prolonged, and the potential high-temperature risk is reduced.

Drawings

FIG. 1 is a schematic view of a prior art ultrasonic blade configuration;

FIG. 2 is a schematic diagram showing a variation of impedance during a shearing operation of the ultrasonic blade;

FIG. 3 is a schematic diagram of a variation curve of frequency and the first derivative of frequency during the shearing operation of the ultrasonic blade;

FIG. 4 is a schematic diagram of an exemplary multi-layer feedforward neural network architecture;

FIG. 5 is a diagram of a typical convolutional neural network architecture;

FIG. 6 is a schematic diagram of a convolutional neural network architecture employed in the present invention;

FIG. 7 is a schematic diagram of a model building training storage method according to the present invention;

FIG. 8 is a flow chart of the present invention for model-based cut-end determination;

FIG. 9 is a flowchart of a first embodiment of the present invention for controlling based on a cropping end decision model;

FIG. 10 is a flowchart of a second embodiment of the present invention for controlling based on a cropping end decision model;

fig. 11 is a flowchart of the subsequent control according to the present invention based on the cut end determination model.

Detailed Description

The present invention will be described in detail below with reference to specific embodiments shown in the drawings. These embodiments are not intended to limit the present invention, and structural, methodical, or functional changes that may be made by one of ordinary skill in the art in light of these embodiments are intended to be within the scope of the present invention.

The invention discloses a machine learning algorithm model, in particular to a neural network algorithm model, wherein the artificial neural network algorithm model is a mathematical model which is inspired by the human brain nervous system and is constructed, similar to biological neurons, by connecting a plurality of nodes (artificial neurons) with each other, and can be used for modeling complex relationships among data. Connections between different nodes are given different weights, each weight representing the magnitude of the effect of one node on another node. Each node represents a specific function, and information from other nodes is input into an activation function through the corresponding weight comprehensive calculation and obtains a new activity value. The activation function is used for introducing nonlinear elements and increasing the expression capability of the neural network, and commonly used activation functions include Sigmoid, Tanh, ReLU and the like.

From a system perspective, an artificial neuron is an adaptive nonlinear dynamical system composed of a large number of neurons connected by extremely rich and perfect connections. At present, the most common neural network learning algorithm is a back propagation algorithm, and the optimization method is a gradient descent algorithm. Theoretically, a two-layer neural network can approach any function, and the increase of the network layer number can enable the neural network to have stronger expression capability under the same neuron number. The neural network models which are commonly used at present include a feedforward neural network model, a memory neural network model, an attention neural network model and the like: a Multilayer Perceptron (MLP) and a Convolutional Neural Network (CNN) are feedforward Neural Network models; a Recurrent Neural Network (RNN) is a Memory Neural Network model, and commonly used RNN models include a gated Neural Unit (GRU) and a Long-Short Term Memory Neural Network (LSTM); the attention neural network model includes a Transformer and the like.

A typical multi-layer feedforward neural network model is shown in fig. 4, and includes an input layer, a plurality of hidden layers, and an output layer, where input information X propagates forward through each layer to obtain an output y.

The information propagation formula of each layer is as follows:

a^l＝f_l(W^la^l-1+b^l) (2)

wherein a is^l-1For the first layer input, a^lFor the output of the first layer, f_lAs a function of activation of layer I neurons, W^lIs a weight matrix from layer l-1 to layer l, b^lIs the bias of the l-1 th to l-th layers.

The memory neural network model increases the memory capacity on the basis of a feedforward neural network, is commonly used for processing time sequence data, and commonly used memory neural networks comprise RNN, GRU, LSTM and the like. GRU and LSTM have long-term memory and are capable of handling long-term sequences.

The number of neurons of the conventional MLP model of the fully-connected neural network increases sharply with the increase of the input feature dimension, which finally results in very inefficient training of the whole neural network and is prone to overfitting. The convolution neural network can be used for replacing a fully-connected neural network to extract effective features of the image, and the convolution neural network simulates a receptive field in a biological visual system by using a convolution kernel. The conventional convolutional neural network structure is shown in fig. 5, and is generally formed by stacking a plurality of convolutional layers and a plurality of fully-connected layers.

The cropping end judgment model of the invention can comprise one or more algorithm model combinations in a neural network algorithm model based on a machine learning algorithm model. As shown in fig. 6, the neural network model of the preferred embodiment employed in the present invention is constructed based on a convolution unit and a residual junction unit structure. The convolution layer adopts two convolution kernels which are respectively a 1 × 1 convolution kernel and a 3 × 3 convolution kernel, an activation function is activated after convolution, the activation function can adopt a ReLU function, a convergence function can use maximum convergence, features extracted by the two convolution kernels are merged after passing through the activation function and the convergence layer, residual operation is performed with the input of the convolution layer, the features are repeatedly performed with three times of convolution, merging and residual operation, then the features are input into a full connection layer, and the features are output after passing through three full connection layers and a softmax function. Of course, the implementation of the model is not limited to the above convolutional neural network structure, and a recurrent neural network structure (RNN, GRU, LSTM) or an attention model may be adopted, or based on various combinations of the above algorithm models, the model output may be the probability of the end of clipping or the state class.

The input characteristics of the model comprise one or more combinations of working feedback parameters, physical structure characteristic parameters and environment parameters. The working feedback parameters include, but are not limited to, real-time voltage U, real-time current I, real-time resonant frequency f, frequency first derivative df, power P, impedance R, voltage current phase difference φ; the physical structure characteristic parameters include but are not limited to ultrasonic cutter bar material, length and shape; the environmental parameters include, but are not limited to, ambient temperature, ambient humidity.

The more complete the input features, the stronger the approximation capability of the neural network model. In the model, the real-time voltage U and the real-time current I are obtained by the generator through real-time sampling of the current-voltage sensor, and the sampling frequency of the voltage-current sensor can be 64 times or 128 times of the actual signal frequency; parameters such as real-time resonant frequency f, frequency first derivative df, power P, impedance R, voltage current phase difference phi and the like are obtained by sampling values through mathematical operations such as FFT and the like:

the real-time power P and the impedance R can be calculated by the following equations:

P＝U×I (3)

the voltage-current phase difference phi can be calculated by the following formula:

φ＝φ_U-φ_I (5)

wherein the voltage phase is phi_UCurrent phase of phi_I；

The real-time resonant frequency f is calculated by the following formula:

f＝k×(φ-φ₀) (6)

wherein k is determined by a functional relationship between the real-time voltage U and the current I:

k＝K(U，I) (7)

voltage phase phi_UCurrent phase phi_IIs obtained by real-time sampling of a generator₀Is a constant.

Physical structure characteristic parameters such as ultrasonic knife bar material, length and shape, the type of the ultrasonic knife and the tissue to be cut can be stored in a memory chip of the ultrasonic knife or a generator, and the generator can directly read the corresponding memory chip to obtain the characteristic parameters; environmental parameters such as environmental temperature and environmental humidity can be obtained by real-time measurement through the sensor.

The model training method can be in modes of supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning and the like. The supervised learning needs to collect all input feature information and training labels of the model, and the input feature information and the training labels can be collected at certain time intervals, and the time intervals can be 1ms or 10 ms.

As shown in fig. 7, a neural network model training process implemented by model supervised learning includes obtaining input features from a training data set S, inputting the input features into a neural network model, calculating a median value and a gradient value of each neuron, updating weights by using a gradient descent method, repeating the above processes until the model reaches a predetermined stop condition, for example, the prediction accuracy reaches a target value or the loss is not reduced, stopping training and storing the model after the stop condition is reached.

The training labels in the model may use the state class or end probability for each sampling time point: the state classification can be two-classification or multi-classification, wherein the multi-classification can set the initial shearing stage as class 0, the shearing separation stage as class 1 and the shearing end stage as class 2; the end probability may be a probability distribution along the clipping time, may be a uniform distribution or a normal distribution, etc.

The trained model is composed of each layer and corresponding neurons and weights, weight parameters and an application algorithm program are stored in a generator memory, the memory can be Flash, EEPROM or other nonvolatile storage devices, the application program runs in a processor, the processor can be an ARM, DSP, FPGA, CPU, GPU or ASIC chip which is stored in the generator, and the processor can also be a remote server which is connected through a network.

Fig. 8 shows a processing method of the cropping end determination model according to the present invention.

And inputting the real-time ultrasonic blade characteristic parameter set X into the model, wherein the model can obtain the real-time ending probability or state category according to the input characteristic set.

The control method based on the cutting end judgment model of the invention is to perform ultrasonic knife output control according to the neural network prediction result, and as shown in fig. 11, the control method specifically comprises the following steps:

s1, saving the cutting end judgment model and at least one threshold;

s2, inputting corresponding input characteristics to the cutting end judgment model, and outputting a real-time cutting end probability value;

s3, comparing the cutting end probability value with the threshold value;

and S4, adjusting the power level applied to the ultrasonic blade transducer according to the comparison result to control the current output of the ultrasonic blade. When it is determined that the ultrasonic blade has reached the tissue cutting end point, the ultrasonic blade control current is reduced at the time point and an end cutting cue is generated.

In the present invention, the threshold may be a category value set corresponding to different clipping states, for example, the starting clipping stage is set as category 0, the clipping separation stage is set as category 1, the clipping end stage is set as category 2, and the model output result in step S2 is one of the category values in the category value set. For example, when the output value of the model is 2, which indicates that the time reaches the cutting end stage, the ultrasonic knife control current is reduced and an end cutting prompt is generated.

Alternatively, in the present invention, the threshold may also be a decimal value within 0-1, for example, the decimal value may be set to be a fixed value of 0.9 or 0.95, the probability value output by the neural network model conforms to probability distribution such as uniform distribution or normal distribution, the real-time clipping end probability value output by the neural network model is compared with the threshold in step S3, and when the real-time clipping end probability value is greater than the threshold, it may be determined that clipping is ended.

In view of this, there are many ways to implement, and one of the implementation methods of the present invention is as shown in fig. 9, a first end probability threshold P1 and a second end probability threshold P2 are preset. When the ending probability value P predicted by the model is lower than the threshold value P1, controlling the output according to a first control algorithm K1, wherein K1 can be used for keeping constant current output; when the ending probability value P is between the threshold values P1 and P2, the output is controlled according to a second control algorithm K2, and K2 can be used for keeping constant power output; when the ending probability value P is higher than the threshold value P2, the output is controlled according to the third control algorithm K3, K3 may be to reduce the current to 10% of the original current in 100 ms.

The output control algorithm of the present invention may include a machine learning algorithm.

The ending probability corresponds to a probability value within 0-1, and the ending probability may include one or more probability threshold values. The output control algorithm K can adjust the output power by adjusting the output voltage or current, the power adjustment direction may be increased, decreased or kept unchanged, and the adjustment time interval and the adjustment amplitude value may be any values meeting the conditions, or may be any other self-defined power adjustment modes.

Besides adjusting the output power, the preset ending condition may be reached by prompting the doctor to end the surgical cutting operation through an audio prompt, and in another implementation method, as shown in fig. 10, the ending probability threshold P0 may be any value within 0-1, for example, may be set to 0.95, and when the predicted probability value is higher than 0.95, the cutting operation may be stopped. The prompting method can be audio prompting or generator display screen picture and text prompting and the like.

The invention also discloses a control system based on the cutting end judgment model, which comprises the following steps:

a storage unit for storing a cutting end judgment model and at least one threshold;

the processing unit is used for inputting corresponding input characteristics to the shearing ending judgment model and outputting a real-time shearing ending probability value;

a comparing unit, configured to compare the clipping end probability value with the threshold;

and an adjusting unit for adjusting the power level applied to the ultrasonic blade transducer to control the ultrasonic blade current output according to the comparison result.

The invention also discloses a generator for controlling based on the shearing ending judgment model, which comprises:

a control circuit coupled to a memory, the control circuit configured to be capable of:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the shearing ending judgment model, and outputting a real-time shearing ending probability value;

comparing the cut-out probability value to the threshold;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output.

The invention also discloses an ultrasonic scalpel surgical instrument based on the shearing end judgment model, which comprises:

an ultrasonic electromechanical system comprising an ultrasonic transducer coupled to an ultrasonic blade via an ultrasonic waveguide; and

a generator configured to supply power to the ultrasound transducer, wherein the generator comprises a control circuit configured to be capable of:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the shearing ending judgment model, and outputting a real-time shearing ending probability value;

comparing the cut-out probability value to the threshold;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output.

The invention also discloses an ultrasonic blade system based on a shear over judge model, comprising a processor and a non-volatile storage device, wherein the non-volatile storage device contains an application program which, when executed by the processor, causes the processor to:

saving the cutting end judgment model and at least one threshold;

inputting corresponding input characteristics to the shearing ending judgment model, and outputting a real-time shearing ending probability value;

comparing the cut-out probability value to the threshold;

based on the comparison, the power level applied to the ultrasonic blade transducer is adjusted to control the ultrasonic blade current output.

The above is only a preferred embodiment of the present invention, and it should be noted that the above preferred embodiment should not be considered as limiting the present invention, and the protection scope of the present invention should be subject to the scope defined by the claims. It will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the spirit and scope of the invention, and these modifications and adaptations should be considered within the scope of the invention.