Tongue tumor tissue detection system and analysis model training method thereof

Abstract

The application discloses a tongue tumor tissue detection system based on bioimpedance spectrum. A multi-frequency voltage signal emitted by a BIS device is used as an excitation signal of a sensor detection unit. The sensor detection unit detects tongue tissue to be detected through the excitation signal. The BIS device collects a voltage signal of the tongue tissue to be detected detected by the sensor detection unit and sends the collected voltage signal of the tongue tissue to be detected to a data processing and control unit for processing. The data processing and control unit first obtains a complex impedance value of the tongue tissue to be detected according to the received voltage signal of the tongue tissue to be detected. Then, the data processing and control unit analyzes the real part and the imaginary part of the complex impedance value of the tongue tissue to be detected by using an analysis model established and trained according to a logistic regression algorithm. The sensor detection unit is small in size, convenient to detect, fast in detection speed and high in precision.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning, and particularly relates to a tongue tumor tissue detection system and an analysis model training method thereof. BACKGROUND

[0002] Bioelectrical impedance spectroscopy is a detection technology for extracting biomedical information by using the electrical characteristics and change rules of biological tissues. The technology is non-invasive and harmless to the human body, and has a wide application prospect in the detection of cancer tissues.

[0003] At present, in the tongue tumor boundary detection method, the conventional technologies are CT, MRI and oral ultrasound methods. Since the tongue area is small and located inside the mouth, the existing methods have certain detection ability for the tongue tumor boundary, but there are problems of unable to guide in real time, slow detection speed, large equipment volume, high detection cost, fixed equipment, complicated operation, etc. SUMMARY

[0004] The application aims at the problems in the prior art, and provides a tongue tumor tissue detection system which is convenient to use and can quickly detect.

[0005] The technical scheme is as follows: in order to achieve the above-mentioned purpose, the application provides a tongue tumor tissue detection system based on bioelectrical impedance spectroscopy, which comprises a sensor detection unit, a BIS device and a data processing and control unit. The data processing and control unit controls the BIS device to send a multi-frequency voltage signal. The multi-frequency voltage signal sent by the BIS device is used as an excitation signal of the sensor detection unit. The sensor detection unit detects the tongue tissue to be detected through the excitation signal. The BIS device collects the voltage signal of the tongue tissue to be detected detected by the sensor detection unit, and sends the collected voltage signal of the tongue tissue to be detected to the data processing and control unit for processing.

[0006] The data processing and control unit first obtains the complex impedance value of the tongue tissue to be detected according to the received voltage signal of the tongue tissue to be detected. Then, the real part and the imaginary part of the complex impedance value of the tongue tissue to be detected are analyzed by using an analysis model established and trained according to a logistic regression algorithm. Finally, the detection result graph is obtained according to the analysis result.

[0007] Further, the sensor detection unit comprises a base body, a signal transmission port arranged at one end of the base body, and a microelectrode array arranged on the surface of the other end of the base body and located in the middle of the surface of the other end of the base body; the microelectrode array comprises a plurality of BIS detection electrodes, each of which is surrounded by a shielding ring, and the plurality of BIS detection electrodes are arranged in an equal-interval dot matrix, wherein the number of BIS detection electrodes in each row is the same as that in each column; every four adjacent BIS detection electrodes form an equal-interval quadrilateral, and a position sensor is arranged on two BIS detection electrodes at the top of any diagonal line in the microelectrode array. In this way, the volume of the detection device is effectively reduced, and the tongue body is more convenient to detect.

[0008] Further, the BIS device sequentially inputs an excitation signal to the square region surrounded by every four adjacent BIS detection electrodes; wherein the two BIS detection electrodes on one diagonal line of the quadrilateral region are the input ends of the excitation signal; and the two BIS detection electrodes on the other diagonal line are the output ends of the voltage signal of the tongue tissue to be detected. The impedance value in the quadrilateral region can be more accurately collected, and the error caused by adjacent other regions during detection is reduced.

[0009] Further, the distance between each two adjacent BIS detection electrodes is 3mm. On the basis of considering the actual structure of the tongue body, the mutual interference between the BIS detection electrodes is effectively avoided.

[0010] Further, the BIS device comprises a signal generation module, a mirror constant current source circuit and two differential circuits; wherein the signal generation module is used to generate a plurality of voltage signals with different frequencies; the mirror constant current source circuit converts the voltage signals output by the signal generation module into a pair of equal and opposite current signals, and inputs them into the two differential circuits; the first differential circuit is used to collect a first voltage signal between two ends of a resistor with a known resistance value; and the second differential circuit collects a second voltage signal of the tongue tissue to be detected.

[0011] Further, the signal generation module adopts a dragon fruit development board. The dragon fruit development board is a better multi-frequency signal generation platform, can emit a plurality of high-speed voltage signals with different frequencies, stability and small amplitude attenuation within 0.1s, and can greatly improve the detection efficiency.

[0012] Further, the detection result graph is a grid graph corresponding to the structure of the microelectrode array, wherein each small grid corresponds to a detection area, and different colors are given according to the result of each detection area output by the data analysis module. In this way, the detection result can be more intuitively displayed.

[0013] The application also provides an analysis model training method of a tongue tumor tissue detection system based on bioimpedance spectroscopy, comprising the following steps:

[0014] Step 1: Establish a data set: the data set includes a training set and a test set; wherein the training set is the impedance spectrum data sample of normal tongue tissue measured in the sensitive frequency range; the test set is the impedance spectrum data sample of tongue tissue of different tongue cancer types, different tongue cancer positions and different tongue cancer sizes measured in the sensitive frequency range; the sensitive frequency range is: 1000Hz-5MHz;

[0015] Step 2: Select multiple different frequencies in the training set, and separate the real part and the imaginary part of the complex impedance value corresponding to each frequency, respectively, as training features;

[0016] Step 3: The training set separated in step 2 is input into the analysis model established according to the logistic regression algorithm with the frequency as the input variable and the impedance imaginary part and the impedance real part as the dependent variables, to preliminarily analyze and train the model, to obtain multiple sets of weight coefficients class_weight and other parameters and the numerical value of the loss function corresponding to each parameter; the other parameters include the regularization strength C, the maximum number of iterations max_iter and the classification method selection parameter multi_class;

[0017] Step 4: Select multiple parameters with the minimum numerical value of the loss function corresponding to each parameter, and arrange and combine different parameter values corresponding to different parameter types in turn to generate multiple analysis models;

[0018] Step 5: Input the test set into the multiple analysis models obtained in step 4 for prediction, and calculate the prediction success rate of each test set;

[0019] Step 6: Select the analysis model with the highest prediction success rate of the test set as the final analysis model.

[0020] Further, the step 2 further includes preprocessing the real part and the imaginary part of the complex impedance value: calculating the mean μ1 and the standard deviation σ1 of the impedance imaginary part of each frequency; the mean μ2 and the standard deviation σ2 of the impedance real part of each frequency; retaining the complex impedance value data in the training set that satisfies [μ1-2σ1, μ1+2σ1] for the complex impedance value imaginary part and [μ2-2σ2, μ2+2σ2] for the complex impedance value real part. In this way, the quality of the training set is effectively improved, thereby improving the accuracy of the trained analysis model.

[0021] Working principle: The tongue tumor tissue detection system provided by the application detects the tongue complex impedance based on the bioimpedance spectrum technology, analyzes the tongue complex impedance at different frequencies by using the analysis model established according to the logistic regression algorithm, and constructs a detection result graph according to the analysis result.

[0022] Beneficial effects: compared with the prior art, the tongue tumor tissue detection system provided by the application not only has small sensor detection unit volume, convenient detection, but also has fast detection speed and high precision; users can quickly view the detection result graph on the client. The analysis model training method provided by the application has high precision and fast analysis speed. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A tongue tumor tissue detection system based on bioimpedance spectrum provided by the application is shown in the figure;

[0024] Figure 2 A structure diagram of the sensor detection unit in the application is shown in the figure;

[0025] Figure 3 A structure diagram of the microelectrode array in the application is shown in the figure;

[0026] Figure 4 An imaging principle diagram of the tongue tumor tissue detection system based on bioimpedance spectrum provided by the application is shown in the figure;

[0027] Figure 5 A judgment logic diagram in the image generation module in the application is shown in the figure;

[0028] Figure 6 A tongue tumor tissue boundary profile diagram output by the tongue tumor tissue detection system based on bioimpedance spectrum provided by the application is shown in the figure. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the examples of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0030] As shown in the figure, Figure 1 The embodiment discloses a tongue tumor tissue detection system based on bioimpedance spectrum, which comprises a sensor detection unit, a BIS device and a data processing and control unit. The sensor detection unit is connected to the BIS device through a data transmission line. The BIS device provides an excitation signal to the sensor detection unit, and the sensor detection unit transmits the collected data to the BIS device. The BIS device is connected to the data processing and control unit through a wireless network. The BIS device sends the received data to the data processing and control unit for processing.

[0031] Among them, as shown in the figure, Figure 2As shown, the sensor detection unit includes a substrate 1, a signal transmission port 2 at one end of the substrate 1, and a microelectrode array 3 at the other end; wherein, as... Figure 3 As shown, the microelectrode array 3 includes multiple shielding rings 31, BIS detection electrodes 32, and position sensors 33. Each pair of adjacent BIS detection electrodes 32 is spaced equidistantly, forming a dot matrix. Each pair of adjacent four BIS detection electrodes forms an equally spaced quadrilateral, typically square. The BIS detection electrodes are disposed in BIS detection electrode slots, and each BIS detection electrode 32 is surrounded by a shielding ring 31 located in a shielding ring slot. Both the BIS detection electrode slots and the shielding ring slots are formed on the surface of the substrate 1. Position sensors 33 are located on two diagonally opposite BIS detection electrodes of the microelectrode array 3. In this embodiment, the microelectrode array 3 uses 16 BIS detection electrodes 32, with a distance of 3mm between each pair of adjacent BIS detection electrodes 32. The microelectrode array 3 is a 4*4 dot matrix, meaning each row and each column has 4 BIS detection electrodes 32. To minimize contact resistance and improve detection accuracy, all BIS detection electrodes 32 are treated with immersion gold. The shielding ring 31 is made of copper electrodes and surrounds the outside of the BIS detection electrode 32.

[0032] The BIS device includes a signal generation module, a mirror constant current source circuit, and two differential circuits. The signal generation module generates multiple stable, high-speed voltage signals with small amplitude attenuation at different frequencies according to instructions from the control unit. In this embodiment, a dragon fruit development board is used as the signal generation module. The mirror constant current source circuit converts the single high-speed voltage signal U output by the signal generation module into a pair of equal-amplitude, reverse-biased high-speed current signals. as well as The signals are then input into two differential circuits. The first differential circuit is used to acquire the first voltage signal across a resistor with a known resistance value. In this embodiment, a high-precision resistor with a resistance of 100Ω is used. The second differential circuit acquires the second voltage signal of the tongue tissue to be tested. The BIS device will collect the first voltage signal Second voltage signal The value is transmitted back to the terminal for processing. Where I... + I represents the amplitude of the positive current signal output by the mirror constant current source circuit. - This represents the amplitude of the positive current signal output by the mirror constant current source circuit. U1 represents the phase of the positive current signal output by the mirror constant current source circuit; U2 represents the voltage amplitude of the first voltage signal. U1 represents the voltage phase of the first voltage signal, and U2 represents the voltage amplitude of the second voltage signal. The voltage phase representing the second voltage signal.

[0033] The sensor detection unit and the BIS device are mainly used for the collection of relevant data. In the collection process, as shown in the figure, each adjacent square area surrounded by four BIS detection electrodes 32 is an independent detection area, and each detection area sequentially completes the collection of the second voltage signal. Figure 4

[0034] The collection process of the second voltage signal for each detection area is as follows: taking the first detection area as an example, the area surrounded by the first, second, fifth and sixth BIS detection electrodes 32 collects the second voltage signal. First, the first and sixth BIS detection electrodes 32 are the input terminals of the excitation current, and the input excitation current is the current information input into the second differential circuit by the mirror constant current source circuit; the second and fifth BIS detection electrodes 32 are the collection terminals of the second voltage signal; then, the signal generation module generates a plurality of high-speed voltage signals with different frequencies, stability and small amplitude attenuation; the BIS device collects the second voltage signal of the detection area at different frequencies in sequence. When the second voltage signal of the first detection area at each frequency is collected, the second detection area collects the second voltage signal, and the process is repeated until the last detection area is collected. During the collection process, if the sensor data collection unit moves, a new round of data collection starts from the first detection area. The BIS device transmits the collected data to the data processing and control unit in real time for processing.

[0035] The data processing and control unit is generally set in a terminal such as a PC, and the data processing and control unit includes a control signal output module, a data processing module, a data analysis module and an image generation module. The control signal output module mainly sends control instructions to the BIS device; the data processing module calculates the impedance amplitude and impedance phase of the detection area according to the first voltage signal and the second voltage signal fed back by the BIS device; and transmits the calculated results to the data analysis module for analysis, and sends the analysis results to the image generation module, and the image generation module generates a detection result image.

[0036] The data processing module first calculates the current signal input into the first differential circuit according to the first voltage signal and Ohm's law The data processing module first calculates the current signal input into the first differential circuit according to the first voltage signal and Ohm's law Since the two pairs of current signals generated by the mirror constant current source circuit are equal and opposite, the current flowing into the second differential circuit is known at this time According to the second voltage signal and the impedance calculation formula ​Obtain impedance amplitude, impedance phase and other information of the tongue tissue to be detected; transmit the calculated impedance amplitude, impedance phase and other information of the detection area to the data analysis module for analysis.

[0037] The data analysis module mainly uses an analysis model established according to a logistic regression algorithm. The analysis model directly outputs an analysis result according to the input impedance amplitude, impedance phase and other information of the detection area. If the detection area is normal tissue, the analysis model outputs 1; if the detection area is diseased tissue, the analysis model outputs 0.

[0038] The analysis model needs to be trained first. In this embodiment, a trained analysis model is used, and the training method is as follows:

[0039] Step 1: Establish a data set: the data set includes a training set and a test set; the training set is impedance spectrum data samples of normal tongue tissue measured in a sensitive frequency range; the test set is impedance spectrum data samples of tongue tissue of different tongue cancer types, different tongue cancer positions and different tongue cancer sizes measured in the sensitive frequency range; the sensitive frequency range is 1000Hz-5MHz. The data set is obtained by using the tongue tumor tissue detection system based on bioimpedance spectrum disclosed in this embodiment.

[0040] Step 2: Select 100 different sampling frequencies in the training set, and separate the real part and the imaginary part of the complex impedance value corresponding to each frequency as training features. Before training, the real part and the imaginary part of the complex impedance value can be preprocessed: calculate the mean μ1 and the standard deviation σ1 of the impedance imaginary part at each frequency; calculate the mean μ2 and the standard deviation σ2 of the impedance real part at each frequency; retain the impedance data in the training set that satisfies [μ1-2σ1, μ1+2σ1] for the impedance imaginary part and [μ2-2σ2, μ2+2σ2] for the impedance real part; if either the imaginary part or the real part of the impedance does not meet the condition, delete the corresponding impedance data.

[0041] Step 3: Separate the training set in step 2, use frequency as input variable, and input impedance imaginary part and impedance real part as dependent variables into the analysis model established according to the logistic regression algorithm to perform preliminary analysis model training, obtain multiple sets of weight coefficients class_weight and other parameters and the numerical value of the loss function corresponding to each parameter; the other parameters include regularization strength C, maximum iteration number max_iter and classification method selection parameter multi_class.

[0042] Step 4: Select multiple parameters with the smallest loss function value corresponding to each parameter, and arrange and combine different parameter values corresponding to different parameter types in sequence to generate multiple analysis models.

[0043] Step 5: Input the test set into the multiple analysis models obtained in Step 4 for prediction, and calculate the prediction success rate of the test set for each test set.

[0044] Step 6: Select the analysis model with the highest prediction success rate on the test set as the final analysis model.

[0045] During the parameter tuning process, based on the characteristics of the logistic regression algorithm, the main parameters adjusted are the regularization strength (C) and the class weights. Regularization is not applicable in this invention, so C needs to be initially set to a value greater than 1. More importantly, the class weights (class_weight) are adjusted. In this embodiment, the two corresponding classes are the real and imaginary parts of impedance. A suitable value needs to be adjusted based on the performance of the trained model on the test set. In this embodiment, the weights obtained after training are 0.8 for the real part of impedance and 0.2 for the imaginary part of impedance.

[0046] The image generation module reconstructs the image based on the output of the data analysis module to obtain the detection result image. Specifically, a dot matrix is ​​obtained based on the structure of the microelectrode array. The dots in each row and column are connected to form a grid, where each small grid corresponds to a region to be detected. Different colors are assigned to each region to be detected based on the output of the data analysis module. If the data analysis module outputs 0, the corresponding small grid is light gray; if the data analysis module outputs 1, the corresponding small grid is dark gray.

[0047] Each time the sensor detection unit moves, it generates a corresponding detection result image. Because an absolute position coordinate sensor is set in the sensor detection unit, each detection result image has corresponding position information. The image generation module stitches together multiple detection result images to obtain a cross-sectional image of the tongue tumor tissue boundary.

[0048] like Figure 5 As shown, during the stitching process, the image generation module determines the coordinate relationship between two adjacent detection result images. If the coordinates of the two adjacent detection result images show no overlap, they are directly stitched together. If the coordinates of the two adjacent detection result images show partial overlap, the overlapping portion is deleted from the first measurement, the second measurement is retained, and the other non-overlapping portions are directly stitched together. If the coordinates of the two adjacent detection result images show complete overlap, all results from the first measurement are deleted, and all results from the second measurement are taken. The stitched result is used as the initial range image for the next stitching; this process continues until all detection result images are stitched together at the end of the detection process. Figure 6 As shown, a cross-sectional view of the boundary of the tongue tumor tissue is obtained, which is easy to observe.

Claims

1. A method for training an analytical model of a tongue tumor tissue detection system based on bioimpedance spectroscopy, the tongue tumor tissue detection system comprising: The sensor detection unit, the BIS device and the data processing and control unit; The data processing and control unit controls the BIS device to send a multi-frequency voltage signal, the multi-frequency voltage signal sent by the BIS device serving as an excitation signal of the sensor detection unit, the sensor detection unit detecting the tongue tissue to be detected through the excitation signal, the BIS device collecting a voltage signal of the tongue tissue to be detected detected by the sensor detection unit, and sending the collected voltage signal of the tongue tissue to be detected to the data processing and control unit for processing; the data processing and control unit first obtains a complex impedance value of the tongue tissue to be detected according to the received voltage signal of the tongue tissue to be detected; then analyzes the real part and the imaginary part of the complex impedance value of the tongue tissue to be detected by using an analysis model established and trained according to a logistic regression algorithm; and obtains a detection result graph according to the analysis result. It is characterized by comprising the following steps: Step 1: establishing a data set, the data set comprising a training set and a test set; wherein the training set is impedance spectrum data samples of normal tongue tissue measured in a sensitive frequency range; the test set is impedance spectrum data samples of tongue tissue of different tongue cancer types, different tongue cancer positions and different tongue cancer sizes measured in the sensitive frequency range; the sensitive frequency range is 1000 Hz-5 MHz; Step 2: selecting multiple different frequencies in the training set, and separating the real part and the imaginary part of the complex impedance value corresponding to each frequency, respectively, as training features; Step 3: inputting the training set separated in step 2 into an analysis model established according to a logistic regression algorithm with frequency as an input variable and impedance imaginary part and impedance real part as dependent variables, respectively, to perform preliminary training of the analysis model, to obtain multiple sets of weight coefficients class_weight and other parameters and the numerical values of the loss function corresponding to each parameter; the other parameters include regularization strength C, maximum iteration number max_iter and classification mode selection parameter multi_class; Step 4: selecting multiple parameters with the minimum numerical value of the loss function corresponding to each parameter, and sequentially arranging and combining different parameter values corresponding to different parameter types to generate multiple analysis models; Step 5: inputting the test set into the multiple analysis models obtained in step 4 to perform prediction, and calculating the prediction success rate of each test set; Step 6: selecting the analysis model with the highest prediction success rate of the test set as the final analysis model.

2. The method of claim 1, wherein the method is characterized by: The sensor detection unit comprises a base body, a signal transmission port provided at one end of the base body, and a microelectrode array provided on the surface of the other end of the base body, the microelectrode array being located in the middle of the surface of the other end of the base body; the microelectrode array comprises multiple BIS detection electrodes, each BIS detection electrode being provided with a shielding ring on the periphery, and the multiple BIS detection electrodes being arranged in an equidistant dot matrix, wherein the number of BIS detection electrodes in each row is the same as that in each column; every four adjacent BIS detection electrodes form an equidistant quadrilateral, and a position sensor is arranged on the two BIS detection electrodes at the top of any diagonal line in the microelectrode array.

3. The method of claim 2, wherein the method is characterized by: The BIS device sequentially inputs an excitation signal to each of four adjacent square regions formed by BIS detection electrodes; wherein two BIS detection electrodes on one diagonal of the square region are input terminals of the excitation signal; and two BIS detection electrodes on the other diagonal are output terminals of a voltage signal of the tongue tissue to be detected.

4. The method of claim 2, wherein the method is characterized by: The distance between each of the two adjacent BIS detection electrodes is 3mm.

5. The method of claim 1, wherein the method is characterized by: The BIS device comprises a signal generation module, a mirror constant current source circuit and two differential circuits; wherein the signal generation module is used to generate a plurality of voltage signals of different frequencies; the mirror constant current source circuit converts the voltage signals output by the signal generation module into a pair of equal and opposite current signals, and inputs them into the two differential circuits; the first differential circuit is used to collect a first voltage signal between two resistors of a known resistance; and the second differential circuit collects a second voltage signal of the tongue tissue to be detected.

6. The method of claim 5, wherein the method is characterized by: The signal generation module adopts a pitaya development board.

7. The method of claim 2, wherein the method is characterized by: The detection result map is a grid map corresponding to the structure of the microelectrode array, wherein each small grid corresponds to a detection area, and each detection area result output by the data analysis module is given a different color.

8. The analytic model training method of claim 2, wherein: In step 2, the real part and the imaginary part of the complex impedance value are also preprocessed: the impedance imaginary part mean µ1 and the standard deviation σ1 of each frequency are calculated; the impedance real part mean µ2 and the standard deviation σ2 of each frequency are calculated; the complex impedance value data satisfying [µ1-2σ1, µ1+2σ1] for the complex impedance value imaginary part and [µ2-2σ2, µ2+2σ2] for the complex impedance value real part in the training set are reserved.

Images