Gastrointestinal tract transfer track acquisition system

Abstract

The invention relates to a gastrointestinal tract transport track acquisition system, which comprises a wearable positioning device provided with a magnetic sensor array of at least two magnetic sensors; a single administrable positioning marker comprising a magnetic substance and capable of moving within the gastrointestinal tract; m is greater than or equal to 1, and each auxiliary positioning marker comprises a magnetic substance and is used for being arranged at a fixed position of a human body; and a terminal device. The magnetic sensor can collect the magnetic field intensity of the wearable positioning marker and the auxiliary positioning marker through magnetic induction, and the terminal equipment receives the magnetic field intensity data to determine the real-time position of the positioning marker in the gastrointestinal tract of the human body. Wherein the terminal equipment calculates the position direction parameters of the administrable positioning marker and the auxiliary positioning marker according to the received magnetic field intensity data, corrects the position direction parameters of the administrable positioning marker by adopting a motion artifact interference elimination method, and determines the real-time position of the positioning marker in the gastrointestinal tract of the human body.

Description

Technical Field

[0001] The invention relates to a gastrointestinal tract motility detection system, in particular to a gastrointestinal tract transport trajectory collection system. Background Art

[0002] Gastrointestinal motility refers to the peristaltic ability of the gastrointestinal organs, which is of great significance to human health. Traditional gastrointestinal motility examinations usually require the use of invasive examination equipment, such as endoscopes, which bring certain pain and inconvenience to patients. Sometimes, in order to reduce the pain of patients, anesthesia is also required, which brings certain risks.

[0003] CN202011344388.9 discloses a gastrointestinal motility examination system, including a gastrointestinal motility capsule, a positioning device and a terminal device, wherein the gastrointestinal motility capsule includes a shell and a magnet, and the positioning device includes at least two magnetic sensors; the gastrointestinal motility capsule is used to generate a changing magnetic field when it moves with the peristalsis of the gastrointestinal tract in the body of the subject; the magnetic sensor is used to collect the intensity data of the magnetic field and send the intensity data to the terminal device. The terminal device determines the position, posture information and movement frequency of the gastrointestinal motility capsule in the gastrointestinal tract based on the intensity data.

[0004] However, the positioning device in this system is fixed on the abdomen of the subject, and the subject needs to wear it continuously for 24-72 hours. During this process, the change in the subject's body posture will cause the relative position and posture between the positioning device and the subject's body to change, that is, motion artifact interference. As a result, the position, posture and movement frequency information of the gastrointestinal motility capsule estimated by the system includes not only the normal movement of the gastrointestinal motility capsule caused by gastrointestinal peristalsis, but also the relative movement caused by motion artifact interference. Therefore, the positioning data obtained in this way has high error and inaccurate precision. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a gastrointestinal transport trajectory acquisition system that reduces errors caused by motion artifact interference and accurately locates.

[0006] The present invention is a gastrointestinal transport trajectory acquisition system, which comprises: a wearable positioning device, on which a magnetic sensor array including at least two magnetic sensors is arranged, a single ingestible positioning marker, the positioning marker contains a magnetic substance, and the ingestible positioning marker can move in the gastrointestinal tract; m auxiliary positioning markers, m≥1, the auxiliary positioning marker contains a magnetic substance, and the auxiliary positioning marker is used to be set at a fixed position on the human body; and a terminal device, wherein the magnetic sensor can collect the magnetic field strength of the ingestible positioning marker and the auxiliary positioning marker by magnetic induction, and the terminal device receives the magnetic field strength data of the magnetic sensor to determine the real-time position of the ingestible positioning marker in the human gastrointestinal tract, wherein the terminal device calculates the position direction parameters of the ingestible positioning marker and the auxiliary positioning marker according to the received magnetic field strength data, and uses a motion artifact interference elimination method to correct the position direction parameters of the ingestible positioning marker according to the position direction parameters of the auxiliary positioning marker, thereby determining the real-time position of the ingestible positioning marker in the human gastrointestinal tract.

[0007] Preferably, the magnetic sensor can collect magnetic field strength at multiple time points, so as to determine the real-time position of the ingestible positioning marker at each time point, thereby obtaining the transport trajectory of the ingestible positioning marker in the gastrointestinal tract.

[0008] Preferably, the terminal device obtains the initial magnetic field strength of the ingestible positioning marker and the auxiliary positioning marker collected by the magnetic sensor array at an initial time, and calculates the initial position and direction parameters of the ingestible positioning marker and the auxiliary positioning marker based on the initial magnetic field strength to determine the initial position of the ingestible positioning marker; the terminal device obtains the magnetic field strength generated by the ingestible positioning marker and the auxiliary positioning marker collected by the magnetic sensor array at the i-th moment, and calculates the position and direction parameters of the ingestible positioning marker and the auxiliary positioning marker at the i-th moment based on the magnetic field strength, and determines the position and direction of the ingestible positioning marker based on the difference between the position and direction parameters of the auxiliary positioning marker and its initial position and direction parameters.

[0009] Preferably, there are at least three auxiliary positioning markers.

[0010] Preferably, the fixing position of the human body may be the anterior superior iliac spine, the posterior superior iliac spine and / or the coccyx.

[0011] Preferably, the wearable positioning device is a belt or waist protector suitable for being fixed to the waist or abdomen of a human body.

[0012] Preferably, the position direction parameters of the edible positioning marker and the auxiliary positioning marker are calculated according to the received magnetic field strength data, and are obtained by a multi-magnetic target positioning method based on a neural network and an equivalent magnetizing current magnetic field model, and the multi-magnetic target positioning method comprises the following steps:

[0013] S1. Generate a data set based on the equivalent magnetizing current magnetic field model, including:

[0014] S11, determining a space to be located for each magnetic target according to an expected motion range of each magnetic target relative to the sensor array;

[0015] S12, performing uniform distribution sampling of positions and directions within the positioning space to obtain a series of labels of magnet positions and directions;

[0016] S13, using an equivalent magnetizing current magnetic field model, generating corresponding sensor array prediction values ​​for the position and direction of each magnet;

[0017] S14. Combine the sensor array prediction values ​​and the magnet position and orientation labels to form a labeled data set.

[0018] S2. Train the pose estimation network, which includes:

[0019] S21, dividing the generated data set into a training set and a test set;

[0020] S22, iteratively training the pose estimation network based on back gradient propagation and training data set;

[0021] S23, in each round of iterative training, the positioning accuracy of the pose estimation network is verified using the test set;

[0022] S24. When the positioning accuracy of the pose estimation network in the test set is lower than a certain threshold or the maximum number of training times is reached, the training is terminated.

[0023] S3, real-time positioning based on pose estimation network, including:

[0024] S31, acquiring the measurement value of each magnetic sensor in real time from the actual sensor array;

[0025] S32, performing pre-processing operations such as filtering on the measured values;

[0026] S33, passing the preprocessed sensor array measurement values ​​to the posture estimation network to perform real-time position and direction estimation on each magnetic target.

[0027] Preferably, in step S1, the magnetic sensors in the sensor array are magnetic sensors based on the Hall effect, anisotropic magnetoresistive (AMR) sensors, giant magnetoresistive (GMR) sensors and / or tunnel magnetoresistive (TMR) sensors.

[0028] Preferably, the dimension of data obtained by each sensor measuring the magnetic field may be 1-dimensional, 2-dimensional or 3-dimensional.

[0029] Preferably, in step S2, the basic network layer of the pose estimation network adopts a convolutional neural network (CNN) to better capture the local spatial correlation between adjacent sensors. The pose estimation network is an end-to-end convolutional neural network model, which maps the sensor measurement values ​​to the position and direction of the magnet. The pose estimation network is trained using a combined loss function that integrates position loss and direction loss. The loss function is:

[0030]

[0031] L p represents the position regression loss, L o Represents the directional regression loss, using the scaling factor β i To balance the regression losses of different dimensions of position and direction, the two losses are defined as follows

[0032]

[0033] Among them, ||·|| δ represents the δ norm, R(·) transforms the posture Euler angle mapping into a rotation matrix, ||·|| F represents the Frobenius norm, symbol Represents the output of the pose estimation network PoseNet.

[0034] Preferably, uniformly distributed sampling is performed in the working space of the positioning system to generate a data set, the position sampling interval is 1-50 mm, and the angle interval is 1-50°.

[0035] Preferably, the position sampling interval is 10 mm, and the angle interval is 10°.

[0036] Preferably, during the training process, noise is randomly added to the sensor measurements in the training data set to improve the robustness of the network.

[0037] Preferably, the motion artifact interference elimination method comprises: taking the ingestible positioning marker as a single target point and the auxiliary positioning marker as multiple reference points, performing the following steps:

[0038] A1. Data collection and sorting steps: Collect motion data sequences including single target points and multiple reference points;

[0039] A2. Analyze the correlation between a single target point and multiple reference points: calculate the correlation index between each reference point and the single target point, and determine the time delay matching point and the degree of correlation;

[0040] A3. Construction of multi-reference point weighted model: A weighted model is constructed based on correlation, and the stability index is introduced to correct the weighted coefficient, and a single target point data estimation model based on multiple reference points is constructed;

[0041] A4. Model optimization and artifact elimination: The weighted model is optimized using the least squares method to obtain a single target point data sequence after eliminating motion artifact interference.

[0042] Preferably, the motion artifact interference elimination method further includes:

[0043] A5. Data verification and quality assessment: Use indicators such as root mean square error and correlation coefficient to evaluate the quality of single target point data before and after processing, verify the effectiveness of the algorithm and adjust parameters or models as needed.

[0044] Preferably, in step A1, a motion data sequence including a single target point and multiple reference points is collected, wherein the collected single target point data sequence is S(t)={s1,s2,...,s n}, where t represents the time series, s i is the data value of the edible positioning marker (2) at time i; at the same time, the data sequence of m reference points is collected, and the data sequence of the jth reference point is recorded as R j (t) = {r j1 ,r j2 , ..., r jn}, j = 1, 2, ..., m; perform preliminary sorting on the collected data and remove obviously abnormal data points.

[0045] Preferably, the obviously abnormal data points are judged and eliminated by setting a data threshold range.

[0046] Preferably, in step A2: a cross-correlation function is used to measure the correlation. For the j-th reference point and the single target point, the cross-correlation function C s，Rj (τ) is calculated as follows:

[0047]

[0048] in is the mean of the single target point data sequence, is the mean value of the jth reference point data sequence, τ is the time delay parameter, and its value range is [-τ max ,τ max], by analyzing the peak position and size of the cross-correlation function, the best matching point and the degree of correlation between each reference point and the single target point in time delay are determined.

[0049] Preferably, in step A3: let the weighting coefficient be ω j , and introduce the stability index St j , the stability index can be obtained by calculating the variance of the reference point data sequence , that is,

[0050] And the weighting coefficient is corrected, the corrected weighting coefficient ω' j The calculation formula is:

[0051]

[0052] Construct a single target point data estimation model based on multiple reference points:

[0053]

[0054] Where Δt j is the time delay compensation value between the jth reference point and the single target point determined according to the cross-correlation function.

[0055] Preferably, the initial value is preliminarily allocated according to the degree of association. The reference point with a higher degree of association has a higher weight coefficient ω. j The bigger.

[0056] Preferably, in step A4: the weighted model is optimized using the least squares method to minimize the error between the estimated value S′(t) and the original single target point data S(t), wherein the error function is assumed to be By calculating the weighting coefficient ω' j Find the partial derivative and set it equal to zero to get a set of equations. Solving the set of equations can get the optimized weighting coefficient ω" j , substitute the optimized weighting coefficient into the data estimation model to obtain the single target point data sequence S"(t) after eliminating motion artifact interference.

[0057] Preferably, in step A5: the root mean square error (RMSE) and correlation coefficient (CC) indicators are used to evaluate the quality of the single target point data before and after eliminating motion artifact interference, wherein the RMSE calculation formula is:

[0058]

[0059] The relevant CC coefficient calculation formula is:

[0060]

[0061] The effectiveness of the algorithm is verified based on the evaluation results. If the results do not meet the expected standards, return to step A2 or step A3 to adjust parameters or rebuild the model.

[0062] Advantages of the present invention:

[0063] 1. No invasive examination equipment is required, it is easy to use, can monitor the gastrointestinal motility of the subject in real time, and can accurately reflect the intestinal motility status.

[0064] 2. By wearing a magnetic induction positioning device and swallowing a magnetic positioning marker (capsule), the examinee can understand his or her gastrointestinal health status at any time and detect and prevent potential gastrointestinal problems in a timely manner.

[0065] 3. By fixing auxiliary positioning markers on specific parts of the human body, errors caused by changes in human posture can be corrected and motion artifact interference can be eliminated.

[0066] 4. The positioning method adopts an end-to-end pose estimation network based on deep learning to perform real-time positioning of ingestible magnetic markers and auxiliary magnetic markers. Compared with the traditional optimization-based positioning method, it has stronger robustness (does not rely on initial values ​​and avoids the risk of falling into local optimality) and lower calculation delay (the pose of the magnetic marker is directly regressed from the feedforward pose estimation network instead of iterative optimization).

[0067] 5. The training data set is generated and acquired based on the equivalent magnetizing current and magnetic field model, rather than collecting data using a physical system, which reduces the time and labor cost of data collection.

[0068] 6. A more optimized algorithm for eliminating motion artifact interference is adopted to further improve positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a schematic diagram of an embodiment of the gastrointestinal transit trajectory collection system of the present invention.

[0070] Figure 2 It is a side view of an embodiment of the gastrointestinal transit trajectory collection system of the present invention.

[0071] Figure 3 It is a rear view of an embodiment of the gastrointestinal transit trajectory collection system of the present invention.

[0072] Figure 4 Accurately model the magnetic field of a cylindrical permanent magnet, where the X-axis is along the length of the magnet, the origin is at the center of the magnet, the magnet coordinate system, and the current direction of the equivalent magnetizing current model.

[0073] Figure 5 A framework for training and predicting pose estimation networks.

[0074] Figure 6 The figure is a flow chart of the method for eliminating motion artifact interference of the present invention.

[0075] To facilitate understanding, identical reference numerals have been used, where possible, to designate identical elements that are common to the figures. It is contemplated that elements disclosed in one embodiment may be beneficially utilized on other embodiments without specific recitation. DETAILED DESCRIPTION

[0076] See also Figure 1-3 A gastrointestinal transit trajectory collection system of the present invention comprises a wearable positioning device 1, on which a plurality of magnetic sensors 3 are arranged. The positioning device 1 can be fixed to the waist or abdomen of a human body in a manner similar to a belt or a waist protector. The magnetic sensors 3 are arranged in an array, for example, in 1×2, 2×2, 3×3, 3×4, 5×6, etc.

[0077] The system of the present invention also includes an ingestible positioning marker 2, which contains a magnetic substance. The positioning marker 2 is, for example, made into an oral capsule that can move in the gastrointestinal tract C as the gastrointestinal tract C moves, and contains a permanent magnet. When the positioning marker 2 moves in the gastrointestinal tract, its position and direction will change, and cause the intensity of the magnetic field to change. The magnetic sensor 3 can detect the change in the magnetic field intensity of the positioning marker 2 through magnetic induction.

[0078] The system of the present invention also includes at least one auxiliary positioning marker, preferably three auxiliary positioning markers P1, P2, and P3. The auxiliary positioning marker also contains a magnetic material, such as a magnet. The auxiliary positioning markers P1, P2, and P3 are arranged at a fixed position of the human body, such as the anterior superior iliac spine, the posterior superior iliac spine, and / or the coccyx. The magnetic sensor 3 can detect the magnetic field strength of the auxiliary positioning markers P1, P2, and P3 by magnetic induction.

[0079] The system also includes a terminal device (not shown in the figure), such as a computer, a mobile device such as a mobile phone or an iPad, or a dedicated control instrument. The terminal device is connected to the magnetic sensor 3, and is used to provide power for the magnetic sensor 3 to work, and receive the magnetic field strength data of the positioning marker 2 and the auxiliary positioning markers P1, P2, and P3 collected by the magnetic sensor 3. Based on the magnetic field strength of the ingestible positioning marker 2 and the auxiliary positioning markers P1, P2, and P3, the terminal device executes instructions using an algorithm to determine the relative position relationship between the auxiliary positioning markers P1, P2, and P3 and the ingestible positioning marker 2, so as to obtain the real-time position of the ingestible positioning marker 2.

[0080] In one embodiment, the system includes an ingestible positioning marker 2, namely magnetic target 1#; and three auxiliary positioning markers P1, P2, and P3, namely magnetic target 2#, magnetic target 3#, and magnetic target 4#, respectively.

[0081] By measuring the magnetic field strength (in three-dimensional directions) at the first time (initial time) by each magnetic sensor 3 in the magnetic sensor array, the four real-time position direction parameters of the magnetic targets 1#-4# at the first time are obtained according to the algorithm, and the initial positions of the edible positioning marker 2 and the three auxiliary positioning markers P1, P2, and P3 can be obtained.

[0082] Then the magnetic field intensity at the second time (jth time) is measured, and the four real-time position direction parameters of the magnetic target 1#-4# at the second time (jth time) are obtained according to the algorithm. Because the three auxiliary positioning markers P1, P2, and P3 are set at fixed positions of the human body, such as the anterior superior iliac spine, the posterior superior iliac spine, and / or the coccyx, these positions, regardless of how the human body posture changes, the actual position direction of the auxiliary positioning markers P1, P2, and P3 remains basically unchanged. Based on the difference between the position direction parameters of the three auxiliary positioning markers of the magnetic target 2#-4# at the jth time and the position direction parameters thereof at the initial time, the position direction changes of each magnetic sensor 3 due to the human body posture can be reversed, and then the position direction parameters of the obtained magnetic target 1# (i.e., the edible positioning marker 2) are corrected to obtain the second (jth) real-time position of the positioning marker 2. Through the above correction, the second real-time position of the positioning marker 2 obtained is more accurate and has a small error.

[0083] By repeating the above steps, a series of real-time positions of the ingestible positioning marker 2 can be obtained, thereby obtaining a series of time-related trajectories thereof, which can be displayed or recorded on a terminal device for further diagnosis reference.

[0084] The position and direction parameters of the edible positioning marker 2 and the auxiliary positioning markers P1, P2, and P3 are calculated based on the received magnetic field strength data, which is obtained by a multi-magnetic target positioning method based on a neural network and an equivalent magnetizing current magnetic field model.

[0085] In one embodiment, the above-mentioned multi-magnetic target positioning method comprises the following steps:

[0086] S1. Generate a data set based on the equivalent magnetizing current magnetic field model, including:

[0087] S11, determining a space to be located for each magnetic target according to an expected motion range of each magnetic target relative to the sensor array;

[0088] S12, performing uniform distribution sampling of positions and directions within the positioning space to obtain a series of labels of magnet positions and directions;

[0089] S13, using an equivalent magnetizing current magnetic field model, generating corresponding sensor array prediction values ​​for the position and direction of each magnet;

[0090] S14. Combine the sensor array prediction values ​​and the magnet position and orientation labels to form a labeled data set.

[0091] S2. Train the pose estimation network, which includes:

[0092] S21, dividing the generated data set into a training set and a test set;

[0093] S22, iteratively training the pose estimation network based on back gradient propagation and training data set;

[0094] S23, in each round of iterative training, the positioning accuracy of the pose estimation network is verified using the test set;

[0095] S24. When the positioning accuracy of the pose estimation network in the test set is lower than a certain threshold or the maximum number of training times is reached, the training is terminated.

[0096] S3, real-time positioning based on pose estimation network, including:

[0097] S31, acquiring the measurement value of each magnetic sensor in real time from the actual sensor array;

[0098] S32, performing pre-processing operations such as filtering on the measured values;

[0099] S33, passing the preprocessed sensor array measurement values ​​to the pose estimation network to perform real-time pose estimation on each magnetic target.

[0100] In the above step S1, the magnetic sensors in the sensor array are magnetic sensors based on the Hall effect, and may also be anisotropic magnetoresistance (AMR) sensors, giant magnetoresistance (GMR) sensors or tunnel magnetoresistance (TMR) sensors.

[0101] The data dimension obtained by each sensor measuring the magnetic field can be 1-dimensional, 2-dimensional or 3-dimensional.

[0102] In the above step S2, the basic network layer of the pose estimation network adopts a convolutional neural network (CNN) to better capture the local spatial correlation between adjacent sensors. The pose estimation network is an end-to-end convolutional neural network model that maps sensor measurements to the position and direction of the magnet. During the training process, noise is randomly added to the sensor measurements in the training data set to improve the robustness of the network. A combined loss function that integrates position loss and orientation loss is used to train the pose estimation network. The loss function will be described in detail below.

[0103] The specific description of permanent magnet magnetic field modeling is as follows.

[0104] like Figure 4 As shown in the figure, a cylindrical permanent magnet is used as the target to be tracked, with a length and radius of 2l and r respectively. Several equivalent models used to calculate the spatial magnetic field distribution of permanent magnets include the magnetic dipole model, the equivalent magnetizing current model and the equivalent magnetic charge model. The magnetic dipole model is widely used in magnetic tracking systems due to its simple expression, but it will bring approximate errors to non-spherical magnets, especially in the near field region. The equivalent magnetizing current model assumes that the magnetic field is generated by surface currents, and the direction of the current is indicated by a blue arrow in the figure. According to the Biot-Savart law, it can be calculated that the cylindrical permanent magnet is The position is determined by the surface current I = J s Magnetic field generated by dx dB c , where J s is the current density (A / m). (x,y,z) T is a position on the current. Position The magnetic field at can be estimated by the following formula

[0105]

[0106] where dB c It can be calculated by the Biot-Savart law as follows:

[0107]

[0108] Where C is the path of current flow, dl = (0, -sinθ, cosθ)rdθ. According to the superposition principle, the magnetic field strength generated by the equivalent current of the permanent magnet can be calculated as

[0109]

[0110] Where K c = rJ s μ0 / 4π, y=r cosθ, z=r sinθ. Therefore, through numerical integration, we can get the expression of the magnetic field generated by the cylindrical permanent magnet.

[0111] In this system, a magnetic sensor array is used to sense the magnetic field generated by a permanent magnet in space. Figure 4 As shown, a system coordinate frame {s} is established at the center of the magnetic sensor array. In formula (1.3), the magnetic induction intensity model of the magnet is established on the magnet coordinate system {m}. The position of the permanent magnet describes the translation and rotation of the coordinate system {m} relative to the coordinate system {s}.

[0112] Assume that v = [x, y, z, α, β, γ] is the six-degree-of-freedom position (position + attitude) of the permanent magnet in the system coordinate system {s}, where t sm =[x,y,z] T is the translation vector, and (α, β, γ) are the Euler angles of the permanent magnet rotating around the X, Y, and Z axes in sequence. Such a rotation sequence can be expressed as

[0113] R sm =R x (α)R y (β)R z (γ)(0.4)

[0114] Click P i The sensor position at can be expressed in coordinate system {s} and coordinate system {m} as and and The relationship between them is as follows:

[0115]

[0116] When the position of the permanent magnet is v, the magnetic induction intensity at the i-th magnetic sensor in the coordinate system {s} (assuming that the sensor array has N magnetic sensors) is

[0117]

[0118] The above formula can be used to predict the magnetic induction intensity at any position in the space around the permanent magnet.

[0119] It should be noted that the radially magnetized cylindrical permanent magnet is wound around X m When the axis (i.e., the SN axis of the magnet) rotates, due to the symmetry of the magnetic field, the magnetic field strength it generates in space remains unchanged. Therefore, α can be fixed to 0.

[0120] More generally, when the number of permanent magnet targets to be tracked is M, M ≥ 2, according to the superposition theorem, the magnetic field strength generated by the M permanent magnets at the i-th sensor is

[0121] B=B i (1)+B i(2)+...+B i (M)(0.7)

[0122] Given an arbitrary position and posture of the permanent magnet, the measurement values ​​of all magnetic sensors on the magnetic sensor array can be predicted using the magnetic field model (1.7).

[0123] The specific description of the multi-target pose estimation network is as follows.

[0124] The process of magnet pose estimation is essentially to deduce and solve the corresponding position and posture of the magnet based on the measurement values ​​of the sensor array, that is, to find the inverse solution of formula (1.7). Since the calculation of the magnetic field model involves integral operations, it is difficult to derive an analytical solution. Although swarm intelligence optimization algorithms (such as particle swarm optimization algorithm PSO, whale optimization algorithm WOA, etc.) can realize the estimation of magnet pose through swarm search strategies, the convergence of the search cannot be guaranteed, and the large amount of calculation makes it difficult to meet real-time requirements. Deep neural networks have powerful nonlinear modeling capabilities. The present invention adopts an end-to-end multi-target pose estimation network PoseNet, such as Figure 5 shown.

[0125] All parameters in the model are learned end-to-end through a combined loss function that combines the position loss and the orientation loss as:

[0126]

[0127] L p represents the position regression loss, L o represents the direction regression loss. Since the position and direction are represented in different dimensions, the scaling factor β is used i To balance the two regression tasks, the two losses are defined as follows

[0128]

[0129] Among them, ||·|| δ represents the δ norm, R(·) transforms the posture Euler angle mapping into a rotation matrix, ||·|| F represents the Frobenius norm, symbol Represents the output of the pose estimation network PoseNet.

[0130] Dataset generation: Uniformly distributed sampling is performed within the working space of the positioning system. The sampling interval of position and angle is a trade-off. If the sampling interval is smaller, the amount of data will be larger, and the accuracy of the trained pose estimation network model will be higher. Generally, the position sampling interval can be 1-50mm, preferably 10mm, and the angle interval is 1-50°, preferably 10°. The predicted value of the magnetic sensor array is generated through the magnetic field model of formula (1.7), thereby obtaining a data set for training the pose estimation network.

[0131] See also Figure 6 , the following algorithm is used to correct motion artifact interference using auxiliary positioning markers.

[0132] Step A1. Data collection and sorting: collect motion data sequences including a single target point and multiple reference points, sort the data preliminarily, and remove abnormal data points;

[0133] Step A2. Correlation analysis between reference points and target points: Calculate the correlation index between each reference point and a single target point, and determine the time delay matching point and the degree of correlation;

[0134] Step A3. Construction of multi-reference point weighted model: construct a weighted model based on correlation, introduce stability index to correct weighted coefficient, and construct a single target point data estimation model based on multi-reference points;

[0135] Step A4. Model optimization and artifact elimination: The weighted model is optimized using the least square method to obtain a single target point data sequence after eliminating motion artifact interference;

[0136] Step A5. Data verification and quality assessment: Use indicators such as root mean square error and correlation coefficient to evaluate the quality of single target point data before and after processing, verify the effectiveness of the algorithm and adjust parameters or models as needed.

[0137] In the above steps, the single target point can be the monitoring point of the displacement of the solution in the colon, which can be the positioning marker 2 in the present invention, and the reference point can be the monitoring point of other relatively stable parts of the body, which can be the auxiliary positioning markers P1, P2, and P3 in the present invention.

[0138] In step A1: collect motion data sequences including a single target point and multiple reference points, wherein the collected single target point data sequence is S(t)={s1,s2,...,s n}, where t represents the time series, s i is the data value of the edible positioning marker 2 at time i; at the same time, the data sequence of m reference points is collected, and the data sequence of the jth reference point is recorded as R j (t) = {r j 1,r j2, ..., r j n}, j = 1, 2, ..., m; preliminarily organize the collected data and remove obviously abnormal data points (such as data mutations caused by sudden noise interference of abnormal data points caused by equipment jitter, electromagnetic interference, etc. during data collection). The obviously abnormal data points can be judged and removed by setting the data threshold range (which can be determined based on prior knowledge or previous experimental data).

[0139] In step A2: calculate the correlation index between each reference point and the single target point. The cross-correlation function is used to measure this correlation. For the jth reference point and the single target point, the cross-correlation function C s，Rj (τ) is calculated as follows:

[0140]

[0141] in is the mean of the single target point data sequence, is the mean value of the jth reference point data sequence, τ is the time delay parameter, and its value range is [-τ max ,τ max ], by analyzing the peak position and size of the cross-correlation function, the best matching point and correlation degree of each reference point and single target point in time delay are determined. For example, when calculating the cross-correlation function, τ max It is a certain proportion (such as 10%) of the data sequence length to comprehensively search for the best time delay matching point.

[0142] Step A3: construct a weighted model based on the correlation between the reference point and the single target point: let the weight coefficient be ω j , its initial value is preliminarily allocated according to the degree of association. The higher the degree of association, the higher the weight coefficient ω j At the same time, considering the stability of data at different reference points, the stability index St is introduced j , the stability index can be obtained by calculating the variance of the reference point data sequence , that is, And the weighting coefficient is corrected, the corrected weighting coefficient ω' j The calculation formula is:

[0143]

[0144] Construct a single target point data estimation model based on multiple reference points:

[0145]

[0146] Where Δt jis the time delay compensation value between the jth reference point and the single target point determined according to the cross-correlation function.

[0147] Step A4, model optimization and artifact elimination: the weighted model is optimized using the least squares method to minimize the error between the estimated value S′(t) and the original single target point data S(t), where the error function is By calculating the weight coefficient ω′ of E j Find the partial derivative and set it equal to zero to get a set of equations. Solving this set of equations can get the optimized weight coefficient ω″ j , substitute the optimized weighting coefficient into the data estimation model to obtain the single target point data sequence S"(t) after eliminating motion artifact interference.

[0148] In step A5, data verification and quality assessment: the root mean square error (RMSE) and correlation coefficient (CC) indicators are used to evaluate the quality of the single target point data before and after eliminating motion artifact interference, where the RMSE calculation formula is:

[0149]

[0150] The relevant CC coefficient calculation formula is:

[0151]

[0152] The effectiveness of the algorithm is verified based on the evaluation results. If the results do not meet the expected standards, return to step A2 or step A3 to adjust parameters or rebuild the model.

[0153] In the embodiment of the above method, during the data collection and collation stage, the accuracy and stability of the data collection equipment are ensured, and data of a single target point and multiple reference points are collected at a predetermined sampling frequency. For example, in biomedical monitoring, a single target point may be a monitoring point for the displacement of dissolved matter in the colon, and a reference point may be a monitoring point in other relatively stable parts of the body. For the collected data, by setting a reasonable data threshold (which can be determined based on prior knowledge or previous experimental data), abnormal data points caused by equipment jitter, electromagnetic interference, etc. are removed.

[0154] In the reference point and target point correlation analysis step, for each reference point, the cross-correlation function between it and the single target point at different time delays is calculated. For example, when calculating the cross-correlation function, τ can be set max is a certain proportion of the data sequence length (such as 10%) to comprehensively search for the best time delay matching point. By analyzing the peak value of the cross-correlation function, Δt j and the degree of association, providing a basis for constructing a weighted model.

[0155] When constructing a multi-reference point weighted model, first roughly assign the weight coefficient ωj according to the degree of association. For example, the initial weight coefficient of the reference point with the highest degree of association is set to 0.5, and the other reference points are assigned between 0 and 0.5 according to the relative degree of association. Then calculate the stability index Stj of the reference point, and correct the weight coefficient ω′ according to the formula j Substitute the corrected weighting coefficient and time delay compensation value into the data estimation model to obtain a preliminary single target point data estimate.

[0156] In the model optimization and artifact elimination stage, the least squares method is used to solve the optimized weight coefficient ω" j By constructing the error function and finding the partial derivative, we get a set of equations, which can be solved by matrix operations and other methods. Substituting the optimized weighted coefficients into the model, we get the final single target point data sequence after eliminating motion artifacts.

[0157] Finally, in the data verification and quality assessment stage, calculate indicators such as RMSE and CC. If RMSE is too large or CC is too small, it means that the processing effect is not ideal. At this time, check whether the reference point selection is reasonable, adjust the parameters in the correlation analysis (such as τ max ), rebuild the weighted model or optimize the model parameters until a satisfactory processing result is obtained.

[0158] Those skilled in the art will appreciate that the foregoing examples are illustrative and not restrictive. It will be apparent to those skilled in the art upon reading the specification and studying the drawings that all permutations, enhancements, equivalents, and improvements are included within the true spirit and scope of the present invention. Therefore, the appended claims are intended to include all such modifications, permutations, and equivalents that fall within the true spirit and scope of these teachings.

[0159] Description of Reference Numerals

[0160] 1 Wearable positioning device

[0161] 2 Positioning markers

[0162] 3 Magnetic sensor

[0163] C. Gastrointestinal tract

[0164] P1 Auxiliary positioning mark

[0165] P2 auxiliary positioning mark

[0166] P3 Auxiliary positioning marker.

Claims

1. A gastrointestinal tract transport trajectory collection system, characterized in that: include: A wearable positioning device (1), wherein the positioning device is provided with a magnetic sensor array comprising at least two magnetic sensors (3), A single ingestible positioning marker (2), wherein the positioning marker comprises a magnetic substance, and the ingestible positioning marker (2) is movable in the gastrointestinal tract; m auxiliary positioning markers (P1, P2, P3), m≥1, the auxiliary positioning markers comprising magnetic material, the auxiliary positioning markers (P1, P2, P3) being used to be set at fixed positions on a human body; as well as The terminal device comprises: a magnetic sensor (3) capable of collecting the magnetic field strength of the ingestible positioning marker (2) and the auxiliary positioning markers (P1, P2, P3) by magnetic induction, and the terminal device receives the magnetic field strength data of the magnetic sensor (3) to determine the real-time position of the ingestible positioning marker (2) in the gastrointestinal tract of the human body, The terminal device calculates the position and direction parameters of the ingestible positioning marker (2) and the auxiliary positioning markers (P1, P2, P3) based on the received magnetic field strength data, and uses a motion artifact interference elimination method to correct the position and direction parameters of the ingestible positioning marker (2) based on the position and direction parameters of the auxiliary positioning markers (P1, P2, P3), thereby determining the real-time position of the ingestible positioning marker (2) in the human gastrointestinal tract.

2. The digestive tract transport trajectory collection system according to claim 1, characterized in that: The magnetic sensor (3) is capable of collecting magnetic field strength at multiple time points, thereby determining the real-time position of the ingestible positioning marker (2) at each time point, thereby obtaining the transport trajectory of the ingestible positioning marker (2) in the gastrointestinal tract.

3. The digestive tract transport trajectory collection system according to claim 1, characterized in that: The terminal device obtains the initial magnetic field strength of the ingestible positioning marker (2) and the auxiliary positioning markers (P1, P2, P3) collected by the magnetic sensor array at an initial time, and calculates the initial position direction parameters of the ingestible positioning marker (2) and the auxiliary positioning markers (P1, P2, P3) according to the initial magnetic field strength to determine the initial position of the ingestible positioning marker (2); The terminal device obtains the magnetic field strength generated by the ingestible positioning marker (2) and the auxiliary positioning markers (P1, P2, P3) collected by the magnetic sensor array at the i-th moment, and calculates the position and direction parameters of the ingestible positioning marker (2) and the auxiliary positioning markers (P1, P2, P3) at the i-th moment based on the magnetic field strength, and determines the position and direction of the ingestible positioning marker (2) based on the difference between the position and direction parameters of the auxiliary positioning markers (P1, P2, P3) and their initial position and direction parameters.

4. The digestive tract transport trajectory collection system according to claim 1, characterized in that: The number of the auxiliary positioning markers is at least three.

5. The digestive tract transport trajectory collection system according to claim 4, characterized in that: The fixing position of the human body may be the anterior superior iliac spine, the posterior superior iliac spine and / or the coccyx.

6. The digestive tract transport trajectory collection system according to claim 1, characterized in that: The wearable positioning device is a belt or waist protector suitable for being fixed to the waist or abdomen of a human body.

7. The digestive tract transport trajectory collection system according to claim 1, characterized in that: The position and direction parameters of the edible positioning marker (3) and the auxiliary positioning markers (P1, P2, P3) are calculated according to the received magnetic field strength data, and are obtained by a multi-magnetic target positioning method based on a neural network and an equivalent magnetizing current magnetic field model, wherein the multi-magnetic target positioning method comprises the following steps: S1. Generate a data set based on the equivalent magnetizing current magnetic field model, including: S11, determining a space to be located for each magnetic target according to an expected motion range of each magnetic target relative to the sensor array; S12, performing uniform distribution sampling of positions and directions within the positioning space to obtain a series of labels of magnet positions and directions; S13, using an equivalent magnetizing current magnetic field model, generating corresponding sensor array prediction values ​​for the position and direction of each magnet; S14. Combine the sensor array prediction values ​​and the magnet position and orientation labels to form a labeled data set. S2. Train the pose estimation network, which includes: S21, dividing the generated data set into a training set and a test set; S22, iteratively training the pose estimation network based on back gradient propagation and training data set; S23, in each round of iterative training, the positioning accuracy of the pose estimation network is verified using the test set; S24. When the positioning accuracy of the pose estimation network in the test set is lower than a certain threshold or the maximum number of training times is reached, the training is terminated. S3, real-time positioning based on pose estimation network, including: S31, acquiring the measurement value of each magnetic sensor in real time from the actual sensor array; S32, performing pre-processing operations such as filtering on the measured values; S33, passing the preprocessed sensor array measurement values ​​to the posture estimation network to perform real-time position and direction estimation on each magnetic target.

8. The digestive tract transport trajectory collection system according to claim 7, characterized in that: In step S1 , the magnetic sensors in the sensor array are magnetic sensors based on the Hall effect, anisotropic magnetoresistive (AMR) sensors, giant magnetoresistive (GMR) sensors and / or tunnel magnetoresistive (TMR) sensors.

9. The digestive tract transport trajectory collection system according to claim 7, characterized in that: The data dimension obtained by each sensor measuring the magnetic field can be 1-dimensional, 2-dimensional or 3-dimensional.

10. The digestive tract transport trajectory collection system according to claim 7, characterized in that: In step S2, the basic network layer of the pose estimation network adopts a convolutional neural network (CNN) to better capture the local spatial correlation between adjacent sensors. The pose estimation network is an end-to-end convolutional neural network model that maps the sensor measurement values ​​to the position and direction of the magnet. The pose estimation network is trained using a combined loss function that integrates position loss and direction loss. The loss function is: L p represents the position regression loss, L o Represents the directional regression loss, using the scaling factor β i To balance the regression losses of different dimensions of position and direction, the two losses are defined as follows Among them, ||·|| δ represents the δ norm, R(·) transforms the posture Euler angle mapping into a rotation matrix, ||·|| F represents the Frobenius norm, symbol Represents the output of the pose estimation network PoseNet.

11. The digestive tract transport trajectory collection system according to claim 10, characterized in that: Evenly distributed sampling is performed in the working space of the positioning system to generate a data set. The position sampling interval is 1-50 mm, and the angle interval is 1-50°.

12. The digestive tract transport trajectory collection system according to claim 11, characterized in that: The position sampling interval is 10 mm and the angle interval is 10°.

13. The digestive tract transport trajectory collection system according to claim 11, characterized in that: During the training process, noise is randomly added to the sensor measurements in the training dataset to improve the robustness of the network.

14. The digestive tract transport trajectory collection system according to claim 1, characterized in that: The motion artifact interference elimination method comprises: taking the edible positioning marker (2) as a single target point and taking the auxiliary positioning markers (P1, P2, P3) as multiple reference points, and performing the following steps: A1. Data collection and sorting steps: Collect motion data sequences including single target points and multiple reference points; A2. Analyze the correlation between a single target point and multiple reference points: calculate the correlation index between each reference point and the single target point, determine the time delay matching point and the degree of correlation; A3. Construction of multi-reference point weighted model: A weighted model is constructed based on correlation, and the stability index is introduced to correct the weighted coefficient to construct a single target point data estimation model based on multiple reference points; A4. Model optimization and artifact elimination: The weighted model is optimized using the least squares method to obtain an optimized single target point data sequence after eliminating motion artifact interference.

15. The digestive tract transport trajectory collection system according to claim 14, characterized in that: The motion artifact interference elimination method also includes: A5. Data verification and quality assessment: Use root mean square error and correlation coefficient indicators to evaluate the quality of single target point data before and after processing, verify the effectiveness of the algorithm and adjust parameters or models as needed.

16. The digestive tract transport trajectory collection system according to claim 14, characterized in that: Step A1, collect motion data sequence including a single target point and multiple reference points, wherein the collected single target point data sequence is S(t)={s1,s2,...,s n }, where t represents the time series, s i is the data value of the edible positioning marker (2) at time i; at the same time, the data sequence of m reference points is collected, and the data sequence of the jth reference point is recorded as R j (t) = {r j1 ,r j2 , ..., r jn }, j = 1, 2, ..., m; perform preliminary sorting on the collected data and remove obviously abnormal data points.

17. The digestive tract transport trajectory collection system according to claim 16, characterized in that: The data points with obvious abnormality are judged and eliminated by setting the data threshold range.

18. The digestive tract transport trajectory collection system according to claim 14, characterized in that: In step A2: the cross-correlation function is used to measure the correlation. For the jth reference point and the single target point, the cross-correlation function C s，Rj (τ) is calculated as follows: in is the mean of the single target point data sequence, is the mean value of the jth reference point data sequence, τ is the time delay parameter, and its value range is [-τ max ,τ max ], by analyzing the peak position and size of the cross-correlation function, the best matching point and the degree of correlation between each reference point and the single target point in time delay are determined.

19. The digestive tract transport trajectory collection system according to claim 14, characterized in that: In step A3: let the weighting coefficient be ω j , and introduce the stability index St j , the stability index can be obtained by calculating the variance of the reference point data sequence , that is, And the weighted coefficient is corrected, and the calculation formula of the corrected weighted coefficient is: Construct a single target point data estimation model based on multiple reference points: Where Δt j is the time delay compensation value between the jth reference point and the single target point determined according to the cross-correlation function.

20. The digestive tract transport trajectory collection system according to claim 19, characterized in that: Its initial value is preliminarily allocated according to the degree of association. The reference point with a higher degree of association has a higher weight coefficient ω j The bigger.

21. The digestive tract transport trajectory collection system according to claim 19, characterized in that: In step A4: the weighted model is optimized using the least squares method to minimize the estimated value S ＇ (t) and the original single target point data S(t), where the error function is By calculating the weight coefficient ω′ of E j Find the partial derivative and set it equal to zero to get a set of equations. Solving the set of equations can get the optimized weighting coefficient ω" j , substitute the optimized weighting coefficient into the data estimation model to obtain the single target point data sequence S"(t) after eliminating motion artifact interference.

22. The digestive tract transport trajectory collection system according to claim 15, characterized in that: In step A5: the root mean square error (RMSE) and correlation coefficient (CC) indicators are used to evaluate the quality of the single target point data before and after eliminating motion artifact interference, where the RMSE calculation formula is: The relevant CC coefficient calculation formula is: The effectiveness of the algorithm is verified based on the evaluation results. If the results do not meet the expected standards, return to step A2 or step A3 to adjust parameters or rebuild the model.

Images