A handheld parathyroid gland detection device and method

Abstract

This invention relates to the field of medical device technology, specifically to a handheld parathyroid gland detection device and detection method. The detection device includes a housing, with a disposable optical fiber consumable at one end of the housing. Inside the housing are a miniature optical path module, a photoelectric sensor, and a miniature laser, all connected in sequence to the disposable optical fiber consumable. The housing also contains a main control board and a lithium battery. A pressure sensor is located at the front end of the disposable optical fiber consumable. The invention also includes a detection method based on this device, which innovatively proposes a dynamic noise compensation algorithm based on IMU motion data. This algorithm quantifies probe displacement, posture, and vibration to establish a mapping model between motion interference and background noise, and subtracts this noise from the original fluorescence signal in real time. This significantly suppresses interference introduced by operational jitter and changes in ambient light, eliminating the need for localized lighting shutdown in the operating room during the detection process, simplifying the surgical procedure, and shortening the detection time.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, specifically to a handheld parathyroid gland detection device and detection method. Background Technology

[0002] Intraoperative parathyroid gland detection is a crucial step in clinical surgery, and this application requires extremely high sterility of the operating handle of the detection equipment used directly on the operating table. Current parathyroid gland detection equipment mainly consists of an integrated optical path, a fiber optic probe, and an armored fiber optic cable approximately 5 meters long. During operation, the fiber optic probe transmits the internal laser light to the tissue being tested and receives the weak autofluorescence signal excited by the tissue, enabling intraoperative identification and localization of the parathyroid glands.

[0003] The existing equipment has significant shortcomings in practical use: First, the equipment structure requires doctors to manually adjust the probe position during the detection operation, relying entirely on the doctor's clinical experience, and it is not equipped with an automated calibration function, making it difficult to guarantee detection stability; Second, the optical fiber is prone to slight movement during the detection process, which generates background noise and signal interference. To ensure the detection effect, it is usually necessary to turn off the lights in the operating room, which significantly increases the complexity of the surgery and the operation time; In addition, the overall portability of the equipment is poor, and the preoperative preparation and postoperative storage process is cumbersome and time-consuming, making it unsuitable for high-intensity, real-time-critical surgical detection scenarios.

[0004] In summary, existing intraoperative parathyroid gland detection devices have significant shortcomings in terms of automation, anti-interference ability, and ease of use, making it difficult to meet the needs of efficient, accurate, and stable detection in clinical surgery. To address these issues, it is necessary to provide a new handheld parathyroid gland detection device that integrates sensor fusion technology and adopts a portable design to improve detection accuracy, simplify the operation process, and enhance the device's adaptability in the surgical environment. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention provides a handheld parathyroid gland detection device and detection method to solve the problem of poor anti-interference ability of existing detection devices.

[0006] To solve the above problems, the technical solution of the present invention is as follows: A handheld parathyroid gland detection device includes a shell, a disposable optical fiber consumable is detachably provided at the end of the shell, a miniature optical path module, a photoelectric sensor, and a miniature laser are provided inside the shell and connected in sequence with the disposable optical fiber consumable, a main control board is also provided inside the shell, and a lithium battery for powering the photoelectric sensor, the miniature laser, and the main control board, and a pressure sensor is provided at the front end of the disposable optical fiber consumable.

[0007] Furthermore, the micro optical path module includes an optical path housing, with an optical fiber consumable connector on one side and a signal recovery optical fiber connector connected to a photoelectric sensor and a laser optical fiber connector connected to a micro laser on the other side. The laser emitted by the micro laser passes sequentially through a first lens, a first filter, a first dichroic mirror, a second dichroic mirror, and a second lens to reach the disposable optical fiber consumable. A second filter, a third filter, and a third lens are sequentially arranged between the second dichroic mirror and the signal recovery optical fiber connector.

[0008] Furthermore, it also includes a speaker for voice prompts, a tri-color LED for light prompts, and a button switch on the casing, as well as an IMU sensor inside the casing, and a battery compartment sealing ring and a battery compartment cover at the bottom of the lithium battery.

[0009] The detection method using the aforementioned handheld parathyroid gland detection device includes the following steps: S1: The probe end of the disposable fiber optic consumable is attached to the patient's normal thyroid tissue. When the contact pressure detected by the pressure sensor reaches the preset threshold, the reference acquisition command is triggered to collect and store the fluorescence signal data of the tissue as the reference signal. S2: Move the probe end to the suspected tissue to be detected and attach it. When the contact pressure reaches a preset threshold, trigger the detection command and simultaneously collect the fluorescence detection signal of the current tissue and the IMU motion data of the probe end. S3: Based on the IMU motion data, perform dynamic noise compensation on the fluorescence detection signal, and extract fluorescence wavelength features, fluorescence intensity features, and fluorescence signal time-domain variation features from the compensated signal to form a feature vector; S4: Input the feature vector into the pre-trained neural network model to obtain preliminary tissue category identification results; S5: Compare and verify the preliminary identification results with the features corresponding to the reference signal, perform weighted processing based on the weighting coefficients of probe fit and posture, and output the final parathyroid tissue identification conclusion based on the verification results.

[0010] Furthermore, in S2, the IMU motion data includes triaxial acceleration. , , Triaxial angular velocity , , And spatial attitude angle; when the probe attitude deviates from the preset detection angle range, the operator is prompted to adjust the attitude through a three-color LED light or a speaker.

[0011] Furthermore, S3 includes: S31. Perform low-pass filtering and integration processing on the IMU motion data to extract the linear displacement. Attitude offset and vibration intensity Calculate the comprehensive motion interference characteristics: ,in, , , These are the system calibration coefficients; The system calibration coefficients Obtained through the following offline calibration method: S311. Conduct experiments on a standard reference body (such as a uniformly scattering material) that does not show fluorescence change; S312, Control the probe to perform three types of independent disturbances: linear displacement scanning, acquiring... Attitude deflection scan to acquire High-frequency vibration excitation to obtain ; S313, Synchronous acquisition of fluorescence signal changes ; S314. Constructing least squares: ; S315. Obtain the optimal coefficients through linear regression. , , ; S32. Establish a mapping model between background noise and motion characteristics generated by fiber optic movement: , Where K and b are the compensation coefficients calibrated offline by the system; S33, convert the original fluorescence detection signal Subtract motion interference components The output signal after compensation is: ; S34. Extract fluorescence wavelength features, fluorescence intensity features, and fluorescence signal time-domain variation features from the compensated output signal to form the feature vector.

[0012] Furthermore, in S34, the extraction of the fluorescence wavelength features includes: acquiring spectral data x1, x2, x3, ... x n Based on the emission wavelength range of parathyroid autofluorescence, the maximum fluorescence intensity within that wavelength range is extracted as the wavelength characteristic value. The extraction of the time-domain variation features of the fluorescence signal includes: obtaining the output signal function when the laser is on. and the output signal function when the laser is extinguished The time-domain data sequence is obtained. Outlier removal was performed on the data sequence using the quartile method: the lower quartile was calculated. and the upper quartile The interquartile range is obtained. It will be less than Q1 Data with a value of 1.5IQR or greater than Q3+1.5IQR are identified as outliers and removed. The data after outlier removal is smoothed using a median filtering algorithm, and the median of the data within the window is used to replace the current point to obtain the time-domain sequence change characteristics. The extraction of fluorescence intensity features includes: repeatedly performing the above-mentioned outlier removal and smoothing processing steps on the signals when the laser is on and when the laser is off, respectively, eliminating the device background noise by using the differential method, and obtaining the final fluorescence intensity feature value.

[0013] Furthermore, in S4, the neural network model is a multi-feature fusion deep learning model based on a one-dimensional convolutional neural network and a long short-term memory network, including: a 1D-CNN layer for extracting local waveform features, an LSTM layer for encoding temporal dependencies, an MLP layer for single-value feature mapping of fluorescence intensity, a fusion layer for concatenating the CNN-LSTM output and the MLP output to generate a joint feature vector, and a Sigmoid activation function layer for outputting classification probabilities between 0 and 1. The training process of the neural network model includes: standardizing and preprocessing the multimodal input features; dividing the labeled dataset into training, validation, and test sets in a 7:2:1 ratio; initializing the weight parameters of the 1D-CNN, LSTM, and fully connected layers using a He normal distribution, with the bias term initialized to 0; and setting the initial learning rate to 1×10⁻⁶. -4 The batch size is 16, the maximum number of iterations is 100, the Adam adaptive optimizer is used, and the loss function is binary cross-entropy. Forward and backward propagation are performed in each iteration, and the weights and bias parameters are dynamically updated using the Adam optimizer. An early stopping strategy is adopted during training, and training is automatically terminated when the validation set loss does not decrease for 10 consecutive rounds.

[0014] Furthermore, S5 includes: The currently detected feature vector is compared in real time with the feature corresponding to the reference signal. The weighting coefficients of the signal are weighted according to the different fit and posture of the probe to obtain the weighted detection value. The detected feature vector is compared with the baseline signal features, and the detection results are weighted and fused by combining the probe's contact state and motion stability to obtain the final detection value. The weighted fusion model is defined as follows: , in, The classification probability is the output of the neural network. For IMU-based stability evaluation metrics; Stability evaluation metrics based on IMU Defined as: , in: For attitude deviation, For vibration intensity, For the probe's micro-displacement, For calibration parameters, weights Determined based on optimized clinical data; If the weighted detection value If the value is greater than a preset threshold, it is determined to be parathyroid tissue, and the parathyroid gland is announced via a speaker, with the three-color LED light displaying red. If the weighted detection value is less than or equal to the preset threshold, it is determined to be non-parathyroid tissue, and the non-parathyroid gland is announced via a speaker.

[0015] Furthermore, S3 also includes an adaptive correction step: the system automatically adjusts the compensation coefficients K and b according to the real-time signal variance to achieve adaptive dynamic compensation.

[0016] Compared with existing technologies, this invention has the following advantages: It innovatively proposes a dynamic noise compensation algorithm based on IMU motion data. This algorithm establishes a mapping model between motion interference and background noise by quantifying probe displacement, posture and vibration, and subtracts it from the original fluorescence signal in real time. This significantly suppresses interference introduced by operation jitter and changes in ambient light. This eliminates the need for local light extinguishing in the operating room during the detection process, simplifies the surgical procedure and shortens the detection time.

[0017] Abandoning the traditional single intensity threshold judgment method, at the signal processing level, through dynamic compensation and feature extraction steps, a multi-dimensional feature vector is constructed by comprehensively utilizing the wavelength features, intensity features, and temporal variation features of the fluorescence signal. At the decision level, this feature vector is input into a pre-trained deep neural network model for classification, achieving more complex and robust intelligent recognition. This multi-feature fusion + AI intelligent decision-making mode can more accurately identify the features of the parathyroid gland from complex biological tissue signals, effectively reducing the false judgment rate.

[0018] By highly integrating the laser, optical path, sensor, main control board, and battery into a single handheld housing, the device is freed from the constraints of the long, armored optical fiber of traditional equipment. This design makes the device compact, ready to use right out of the box, and highly mobile, greatly simplifying the preoperative preparation and postoperative storage process. It is perfectly suited for clinical scenarios with extremely high requirements for real-time performance and efficiency, such as thyroid surgery, and improves the overall smoothness of the surgical procedure. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the appearance of the present invention; Figure 2 This is a schematic diagram of the cross-sectional structure of the present invention; Figure 3 This is a schematic diagram of the miniature optical path of the present invention; Figure 4 This is a summary diagram of the parathyroid gland and other tissue mapping data of the present invention.

[0020] In the diagram: 1. Disposable fiber optic consumable; 2. Consumable sealing ring; 3. Tri-color LED; 4. Miniature optical path; 401. Fiber optic consumable connector; 402. First lens; 403. First filter; 404. First dichroic mirror; 405. Second dichroic mirror; 406. Second lens; 407. Second filter; 408. Third filter; 409. Third lens; 410. Signal recovery fiber optic connector; 411. Laser fiber optic connector; 5. Miniature photoelectric sensor; 6. Main control board; 7. Speaker sealing ring; 8. Speaker; 9. Lithium battery; 10. Battery compartment sealing ring; 11. Battery compartment cover; 12. Silicone coating layer; 13. Push button switch; 14. Miniature laser; 15. Housing; 16. Optical path housing. Detailed Implementation

[0021] This invention discloses a handheld parathyroid gland detection device and method. The device features a handheld, integrated, sterile detection structure. Through disposable fiber optic consumables, a miniature optical path module, multi-sensor fusion, a fiber optic motion dynamic compensation algorithm, and CNN-LSTM neural network recognition, it achieves rapid, accurate, and stable intraoperative detection of the parathyroid glands. The entire process is simple to operate, highly resistant to interference, and offers high aseptic safety, making it suitable for high-intensity, real-time surgical scenarios. The specific solution is as follows: like Figure 1 , 2 As shown in Figure 3, the present invention provides a handheld parathyroid gland detection device, including a housing 15. A disposable optical fiber consumable 1 is detachably provided at the end of the housing 15. A consumable sealing ring 2 is provided at the connection position between the disposable optical fiber consumable 1 and the device body to achieve a sealed connection between the disposable optical fiber consumable 1 and the micro optical path module 4, while ensuring the waterproof performance of the optical path system in the surgical environment. Inside the housing 15, there is a micro optical path module 4, a photoelectric sensor 5, and a micro laser 14 connected in sequence to the disposable optical fiber consumable 1. Inside the housing 15, there is also a main control board 6 and a lithium battery 9 that powers the photoelectric sensor 5, the micro laser 14, and the main control board 6. A pressure sensor is provided at the front end of the disposable optical fiber consumable 1.

[0022] like Figure 2As shown, the outer shell 15 adopts an upper and lower snap-fit ​​structure, and sealing rings are provided at the connection of the outer shell, the battery compartment and the speaker position to improve the waterproof and pollution-proof capabilities of the device. The grip part of the outer shell 15 is provided with a silicone coating layer 12 for anti-slip treatment of the handle grip part. At the bottom of the lithium battery 9, there is a battery compartment sealing ring 10 and a battery compartment cover 11. The battery compartment cover 11 allows quick release and can be rotated to perform battery replacement.

[0023] like Figure 3 As shown, the micro optical path module 4 includes an optical path housing 16. A disposable optical fiber consumable 1 connection connector 401 is provided on one side of the optical path housing 16, and a signal recovery optical fiber connector 410 connected to the photoelectric sensor 5 and a laser optical fiber connector 411 connected to the micro laser 14 are provided on the other side.

[0024] like Figure 3 As shown, the miniature laser 14 is used to generate a laser signal with a wavelength of 785nm. This laser signal enters the miniature optical path module 4 through the laser fiber connector 411, and passes sequentially through the first lens 402, the first filter 403, the first dichroic mirror 404, the second dichroic mirror 405, and the second lens 406 to reach the disposable fiber optic consumable 1. The fiber optic probe then irradiates the tissue to be tested. When the laser irradiates the parathyroid gland tissue, it can excite the generation of an autofluorescence signal in the approximately 820nm band.

[0025] like Figure 3 As shown, a second filter 407, a third filter 408, and a third lens 409 are sequentially arranged between the second dichroic mirror 405 and the signal recovery fiber optic connector 410. The 820nm autofluorescence signal is transmitted back to the micro optical path module 4 through the disposable optical fiber consumable 1, and then sequentially through the second lens 406, the second dichroic mirror 405, the second filter 407, and the third filter 408, before being transmitted to the photoelectric sensor 5 through the signal recovery fiber optic connector 410.

[0026] like Figure 3 As shown, the first lens 402, the second lens 406, and the third lens 409 are used to collimate the light beam or focus the reflected light onto the detector, respectively; the first dichroic mirror 404 and the second dichroic mirror 405 are placed at an angle to change the light path and guide light of a specific wavelength to the top channel; the first filter 403, the second filter 406, and the third filter 408 are placed close to the lens to filter stray light and allow only specific wavelengths (such as 785nm or 820nm lasers) to pass through, so as to improve the signal-to-noise ratio.

[0027] The miniature optical path module 4 achieves efficient transmission of 785nm laser and accurate recovery of 820nm autofluorescence signal through integrated design. The multi-level filter structure effectively filters stray light and significantly improves the detection signal-to-noise ratio. The modular design facilitates maintenance and replacement.

[0028] The photoelectric sensor 5 is used to convert the received autofluorescence signal into an electrical signal and transmit the electrical signal to the main control board 6 for signal processing. The main control board 6 integrates a signal processing module and an intelligent recognition module to realize the automatic recognition of parathyroid tissue.

[0029] It also includes a speaker 8 for voice prompts, a tri-color LED 3 for light prompts, and a button switch 13, all mounted on the housing 15. It also includes an IMU sensor inside the housing 15, which can also be installed inside a disposable fiber optic consumable 1. The IMU sensor is used to collect probe attitude information in real time, including spatial attitude angle, angular velocity, and acceleration. When the probe attitude deviates from the preset detection angle range, the system prompts the operator to adjust the attitude through the tri-color LED 3 or the speaker 8. The speaker sealing ring 7 is used for waterproof sealing of the speaker 8.

[0030] The tri-color LED 3 can switch between red, blue, and green states: green indicates the device is in standby mode; blue indicates the reference tissue acquisition status; and red indicates the tissue detection status. The speaker 8 is used to provide voice prompts for the test results. The button switch 13 can perform single-click, long-press (2s), and long-press (10s) operations: when the button 13 performs a long-press (10s) operation, the device is turned on or off, the tri-color LED 3 indicates green, and the speaker 8 prompts whether the device is turned on or off; when the disposable fiber optic consumable 1 is applied to the thyroid gland for calibration, the button 13 performs a long-press (2s) operation, the speaker 8 alarms to indicate calibration, and the tri-color LED 3 indicates blue; when the disposable fiber optic consumable 1 is applied to the parathyroid gland or other tissues for testing, the button 13 performs a single-click operation, the speaker 8 alarms to indicate testing, and the tri-color LED 3 indicates red; the corresponding test results are alarmed via the speaker 8 for parathyroid or non-parathyroid glands.

[0031] The detection method using the above-mentioned handheld parathyroid gland detection device includes the following steps: S1: Reference signal acquisition The probe end of the disposable fiber optic consumable 1 is attached to the patient's normal thyroid tissue. When the contact pressure detected by the pressure sensor reaches the preset threshold, a reference acquisition command is triggered to collect and store the fluorescence signal data of the tissue as a reference signal.

[0032] S2: Simultaneous acquisition of fluorescence detection signals and IMU motion data The probe end of the disposable fiber optic consumable 1 is moved to the suspected tissue to be detected and attached. When the contact pressure reaches a preset threshold, a detection command is triggered, and the fluorescence detection signal of the current tissue and the IMU motion data of the probe end are collected simultaneously.

[0033] The IMU motion data includes triaxial acceleration a x(t), a (t), a u (t), triaxial angular velocity ω x (t), ω (t), ω u (t) and spatial attitude angle; when the probe attitude deviates from the preset detection angle range, the operator is prompted to adjust the attitude by the three-color LED light 3 or the speaker 8.

[0034] S3: Dynamic Noise Compensation and Multi-Dimensional Feature Extraction Based on the IMU motion data, dynamic noise compensation is performed on the fluorescence detection signal, and fluorescence wavelength features, fluorescence intensity features, and fluorescence signal temporal variation features are extracted from the compensated signal to form a feature vector.

[0035] Specifically, it includes the following sub-steps: S31. Motion Feature Extraction and Comprehensive Motion Disturbance Calculation The IMU motion data is low-pass filtered and integrated to extract the linear displacement d(t), attitude offset θ(t), and vibration intensity A. vi (t), calculate the comprehensive motion disturbance characteristics: ,in, , , These are the system calibration coefficients; System calibration coefficients Obtained through the following offline calibration method: S311. Conduct experiments on a standard reference body (such as a uniformly scattering material) that does not show fluorescence change; S312, Control the probe to perform three types of independent disturbances: linear displacement scanning, acquiring... Attitude deflection scan to acquire High-frequency vibration excitation to obtain ; S313, Synchronous acquisition of fluorescence signal changes ; S314. Constructing least squares: ; S315. Obtain the optimal coefficients through linear regression. , , ; S32. Establishment of Noise Mapping Model Establish a mapping model between background noise generated by fiber optic movement and motion characteristics: Where K and b are the compensation coefficients calibrated offline by the system; S33, Real-time signal compensation The original fluorescence detection signal S av (t) Subtract the motion disturbance component to obtain the compensated output signal: ; S34, Multi-dimensional Feature Extraction The fluorescence wavelength features, fluorescence intensity features, and fluorescence signal time-domain variation features are extracted from the compensated output signal to form the feature vector.

[0036] The extraction of fluorescence wavelength features includes: acquiring spectral data x1, x2, x3, ..., x n Based on the emission wavelength range of parathyroid autofluorescence (820nm-830nm), the maximum fluorescence intensity within this wavelength range is extracted as the wavelength characteristic value, i.e., O(t1) = max(x1,x2,x3,…,x…). n Then, the data is compensated according to the compensation algorithm to obtain the current value.

[0037] The extraction of the time-domain variation features of the fluorescence signal includes: obtaining the output signal function O(t) when the laser is on and the output signal function when the laser is off. The time-domain data sequence is obtained. Outlier removal was performed on the data sequence using the quartile method: the lower quartile was calculated. and the upper quartile The interquartile range is obtained. It will be less than Q1 Data with a value of 1.5IQR or greater than Q3+1.5IQR are identified as outliers and removed. The data after outlier removal is smoothed using a median filtering algorithm, and the median of the data within the window is used to replace the current point to obtain the time-domain sequence change characteristics.

[0038] The extraction of fluorescence intensity features includes: repeatedly performing the above-mentioned outlier removal and smoothing processing steps on the signals when the laser is on and when the laser is off, respectively, eliminating the device background noise by using the differential method, and obtaining the final fluorescence intensity feature value.

[0039] S35, Adaptive Correction The system automatically adjusts the compensation coefficients K and b based on the real-time signal variance to achieve adaptive dynamic compensation, which is suitable for different vibration environments.

[0040] The specific steps are as follows: Calculate the real-time variance of the fluorescence signal within the sliding time window: , , in: It is a fluorescent signal. The mean of the sliding window. This is the window length.

[0041] Disturbance deviation calculation: , in: The reference variance obtained from system calibration; Compensation parameters are updated adaptively: , , in: Adaptive step size; When the signal variance increases, the system increases the compensation strength; when the signal tends to stabilize, the compensation strength decreases. Through the above adaptive mechanism, when the probe movement causes an increase in signal fluctuation, the system automatically enhances the noise compensation capability; when the signal is stable, it automatically reduces the compensation intensity, thereby maintaining a better signal quality under different vibration environments.

[0042] S4: Neural Network-Based Tissue Identification The feature vector is input into a pre-trained neural network model to obtain preliminary tissue category identification results.

[0043] The neural network model is a multi-feature fusion deep learning model based on a one-dimensional convolutional neural network and a long short-term memory network, including: a 1D-CNN layer for extracting local waveform features, an LSTM layer for encoding temporal dependencies, an MLP layer for single-value feature mapping of fluorescence intensity, a fusion layer for concatenating the CNN-LSTM output and the MLP output to generate a joint feature vector, and a Sigmoid activation function layer for outputting classification probabilities between 0 and 1.

[0044] The training process of the neural network model includes: firstly, standardizing the multimodal input features to unify the numerical distribution of wavelength sequences, time-domain sequences and fluorescence intensity features, and eliminating dimensional differences; secondly, dividing the labeled dataset into training set, validation set and test set in a 7:2:1 ratio, where the training set is used for model parameter learning, the validation set is used for real-time monitoring of overfitting risk and adjustment of hyperparameters, and the test set is used for final performance evaluation.

[0045] Before training begins, the weight parameters of the 1D-CNN, LSTM, and fully connected layers are initialized using a He normal distribution, with the bias term initialized to 0; the initial learning rate is set to 1×10. -4The batch size is 16, the maximum number of epochs is 100, the optimizer is the Adam adaptive algorithm, and the loss function is binary cross-entropy.

[0046] Binary cross-entropy is used for binary classification problems, and its definition is: , Where: N is the number of samples; ∈{0,1} represents the true label of the sample; ∈(0,1) represents the model's predicted probability.

[0047] Then, it is mapped to class probability using the Sigmoid function: , The binary cross-entropy loss function is used to measure the difference between the predicted results and the true labels. The core idea is to calculate the log-likelihood loss of the positive class and the negative class separately, and then sum the losses of individual samples in a weighted manner.

[0048] The single-sample loss is: , For a batch of data, the average loss of all samples is taken as the overall loss value of the current batch. The sample mean within the batch is: , The gradient of the loss function with respect to the model parameters is calculated by backpropagation algorithm, and the network weights are iteratively updated by Adam optimizer, thereby continuously reducing prediction error and improving the classification performance of the model.

[0049] In each iteration, the model performs forward propagation: sequence features are extracted by 1D-CNN to extract local waveform features, and then temporal dependencies are encoded by LSTM; single-value fluorescence intensity features are mapped by MLP; the two outputs are concatenated in the fusion layer to generate a joint feature vector, and finally the classification probability between 0 and 1 is output by the Sigmoid activation function.

[0050] The loss value is calculated based on the model's predicted values ​​and the true binary classification labels (0 - non-parathyroid, 1 - parathyroid). The loss gradient is backpropagated layer by layer to each layer of the network using the backpropagation algorithm. The Adam optimizer is used to dynamically update the weights and bias parameters to gradually minimize the classification loss. An early stopping strategy is adopted during training. Training is automatically terminated when the validation set loss does not decrease for 10 consecutive rounds to avoid model overfitting. After multiple rounds of iteration and convergence, the optimal weight parameters are saved to complete the training of the parathyroid classification model based on multi-feature fusion.

[0051] By deeply fusing wavelength features, temporal sequence features, and fluorescence intensity features through a lightweight neural network model, intelligent classification of parathyroid glands and non-parathyroid glands is achieved, significantly improving recognition accuracy and stability. The edge deployment design does not rely on external computing resources, meeting the needs of real-time surgical scenarios.

[0052] S5: Real-time comparison of intraoperative data and output of results The preliminary identification results are compared and verified with the features corresponding to the reference signal. The weighting coefficients of the signal are weighted according to the different fit and posture of the probe to obtain the weighted detection value.

[0053] The detected feature vector is compared with the baseline signal features, and the detection results are weighted and fused by combining the probe's contact state and motion stability to obtain the final detection value. The weighted fusion model is defined as follows: , in, The classification probability is the output of the neural network. For IMU-based stability evaluation metrics; Stability evaluation metrics based on IMU Defined as: , in: For attitude deviation, For vibration intensity, For the probe's micro-displacement, For calibration parameters, weights Determined based on optimized clinical data; If the weighted detection value If the value is greater than a preset threshold, it is determined to be parathyroid tissue, and the parathyroid gland is announced via a speaker, with the three-color LED light displaying red. If the weighted detection value is less than or equal to the preset threshold, it is determined to be non-parathyroid tissue, and the non-parathyroid gland is announced via a speaker.

[0054] If the weighted detection value is greater than a preset threshold (e.g., 200), it is determined to be parathyroid tissue, and the parathyroid gland is announced via speaker 8, with the tri-color LED 3 displaying red; if the weighted detection value is less than or equal to the preset threshold, it is determined to be non-parathyroid tissue, and the non-parathyroid gland is announced via speaker 8.

[0055] By comparing the data with the patient's tissue in real time during the operation, the feature values ​​of this test are compared with the signals of normal thyroid tissue collected during the operation to further verify the judgment results and improve accuracy. The weighted processing mechanism takes into account the probe fit and posture factors to reduce the impact of the operation variables on the results and ensure the reliability of the test conclusions.

[0056] Clinically validated data showed that the difference between parathyroid gland values ​​and thyroid gland calibration values ​​both exceeded the threshold of 200, while the values ​​for other tissues (i.e., perithyroid tissues such as muscle, fat, and lymph nodes) were all less than 200. Figure 4 As shown, this reflects the comparison of measurement data from multiple tissues (parathyroid gland and surrounding tissues) during clinical trials, verifying the effectiveness and accuracy of the detection method of this invention.

[0057] The handheld parathyroid gland detection device and method provided by this invention integrate disposable optical fiber consumables 1, miniature optical path 4, multi-sensor fusion system and lightweight intelligent recognition algorithm to achieve rapid, accurate and stable intraoperative identification of parathyroid tissue. The device adopts a waterproof and sealed design to meet the sterilization requirements of aseptic surgery. The detection method significantly improves the recognition accuracy and ease of operation by extracting multiple features of fluorescence signals, intelligent classification of neural networks, dynamic noise compensation and real-time comparison and verification during surgery, and reduces the complexity of surgery, showing good prospects for clinical application.

[0058] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A handheld parathyroid gland detection device, comprising a housing, characterized in that: The disposable optical fiber consumable is detachably provided at the end of the housing. Inside the housing are a miniature optical path module, a photoelectric sensor, and a miniature laser connected in sequence with the disposable optical fiber consumable. The housing also contains a main control board and a lithium battery that powers the photoelectric sensor, the miniature laser, and the main control board. A pressure sensor is provided at the front end of the disposable optical fiber consumable.

2. The handheld parathyroid gland detection device of claim 1, wherein: The micro optical path module includes an optical path housing. On one side of the optical path housing is a fiber optic consumable connector, and on the other side are a signal recovery fiber optic connector connected to a photoelectric sensor and a laser fiber optic connector connected to a micro laser. The laser emitted by the micro laser passes sequentially through a first lens, a first filter, a first dichroic mirror, a second dichroic mirror, and a second lens to reach the disposable fiber optic consumable. Between the second dichroic mirror and the signal recovery fiber optic connector are a second filter, a third filter, and a third lens.

3. The handheld parathyroid gland detection device according to claim 1, characterized in that: It also includes a speaker for voice prompts, a tri-color LED for light prompts, and a button switch on the casing, as well as an IMU sensor inside the casing, and a battery compartment sealing ring and a battery compartment cover at the bottom of the lithium battery.

4. A detection method using a handheld parathyroid gland detection device according to any one of claims 1-3, characterized in that, Includes the following steps: S1: The probe end of the disposable fiber optic consumable is attached to the patient's normal thyroid tissue. When the contact pressure detected by the pressure sensor reaches the preset threshold, the reference acquisition command is triggered to collect and store the fluorescence signal data of the tissue as the reference signal. S2: Move the probe end to the suspected tissue to be detected and attach it. When the contact pressure reaches a preset threshold, trigger the detection command and simultaneously collect the fluorescence detection signal of the current tissue and the IMU motion data of the probe end. S3: Based on the IMU motion data, perform dynamic noise compensation on the fluorescence detection signal, and extract fluorescence wavelength features, fluorescence intensity features, and fluorescence signal time-domain variation features from the compensated signal to form a feature vector; S4: Input the feature vector into the pre-trained neural network model to obtain preliminary tissue category identification results; S5: Compare and verify the preliminary identification results with the features corresponding to the reference signal, perform weighted processing based on the weighting coefficients of probe fit and posture, and output the final parathyroid tissue identification conclusion based on the verification results.

5. The method of claim 4, wherein: In S2, the IMU motion data includes triaxial acceleration. , , Triaxial angular velocity , , And spatial attitude angle; when the probe attitude deviates from the preset detection angle range, the operator is prompted to adjust the attitude through a three-color LED light or a speaker.

6. The detection method according to claim 5, characterized in that S3 include: S31, low-pass filtering and integral processing are performed on the IMU motion data to extract linear displacement , attitude offset and vibration intensity , and a comprehensive motion interference feature is calculated: , wherein , , is a system calibration coefficient S32, a mapping model of background noise generated by fiber movement and motion features is established: , Where K and b are the compensation coefficients calibrated offline by the system; S33, subtracting the motion interference component from the original fluorescence detection signal Subtracting the motion interference component to obtain a compensated output signal ； S34. Extract fluorescence wavelength features, fluorescence intensity features, and fluorescence signal time-domain variation features from the compensated output signal to form the feature vector.

7. The method of claim 6, wherein: In S34, the extraction of the fluorescence wavelength feature includes: acquiring the spectral data x1, x2, x3, … x n , according to the emission wavelength range of the parathyroid gland autofluorescence, extracting the fluorescence intensity maximum value in the wavelength range as the wavelength feature value; The extraction of the time-domain variation features of the fluorescence signal includes: obtaining the output signal function when the laser is on. and the output signal function when the laser is extinguished The time-domain data sequence is obtained. Outlier removal was performed on the data sequence using the quartile method: the lower quartile was calculated. and the upper quartile The interquartile range is obtained. It will be less than Q1 Data with a value of 1.5IQR or greater than Q3+1.5IQR are identified as outliers and removed. The data after outlier removal is smoothed using a median filtering algorithm, and the median of the data within the window is used to replace the current point to obtain the time-domain sequence change characteristics. The extraction of fluorescence intensity features includes: repeatedly performing the above-mentioned outlier removal and smoothing processing steps on the signals when the laser is on and when the laser is off, respectively, eliminating the device background noise by using the differential method, and obtaining the final fluorescence intensity feature value.

8. The detection method according to claim 7, characterized in that: In S4, the neural network model is a multi-feature fusion deep learning model based on a one-dimensional convolutional neural network and a long short-term memory network, including: a 1D-CNN layer for extracting local waveform features, an LSTM layer for encoding temporal dependencies, an MLP layer for single-value feature mapping of fluorescence intensity, a fusion layer for concatenating the CNN-LSTM output and the MLP output to generate a joint feature vector, and a Sigmoid activation function layer for outputting classification probabilities between 0 and 1. The training process of the neural network model includes: standardizing and preprocessing the multimodal input features; dividing the labeled dataset into training, validation, and test sets in a 7:2:1 ratio; initializing the weight parameters of the 1D-CNN, LSTM, and fully connected layers using a He normal distribution, with the bias term initialized to 0; and setting the initial learning rate to 1×10⁻⁶. -4 The batch size is 16, the maximum number of iterations is 100, the Adam adaptive optimizer is used, and the loss function is binary cross-entropy. Forward and backward propagation are performed in each iteration, and the weights and bias parameters are dynamically updated using the Adam optimizer. An early stopping strategy is adopted during training, and training is automatically terminated when the validation set loss does not decrease for 10 consecutive rounds.

9. The detection method according to claim 8, characterized in that S5 include: The currently detected feature vector is compared in real time with the feature corresponding to the reference signal. The weighting coefficients of the signal are weighted according to the different fit and posture of the probe to obtain the weighted detection value. If the weighted detection value is greater than a preset threshold, it is determined to be parathyroid tissue, and the parathyroid gland is announced via a speaker, with the tri-color LED light displaying red. If the weighted detection value is less than or equal to the preset threshold, it is determined to be non-parathyroid tissue, and the non-parathyroid gland is announced via a speaker.

10. The detection method of claim 9, wherein: S3 also includes an adaptive correction step: the system automatically adjusts the compensation coefficients K and b according to the real-time signal variance to achieve adaptive dynamic compensation.

Images