Accurate acupoint positioning system and method based on multi-modal sensor fusion

Abstract

The invention discloses an accurate acupoint positioning system and method based on multi-modal sensor fusion, and relates to the technical field of traditional Chinese medicine diagnosis and treatment. The system integrates ultrasonic imaging, infrared thermal imaging, bioelectrical impedance and flexible pressure tactile sensing, and synchronously collects multi-dimensional physiological information of an acupuncture point area. And through a deep learning cross-modal attention fusion algorithm, three-dimensional coordinates and confidence coefficients of the acupuncture points are output, and the positioning error is smaller than or equal to 2 mm. The Monte Carlo Dropout uncertainty quantification technology is adopted, needle insertion is guided when the confidence coefficient is larger than or equal to 80%, the signal change rate is verified in real time after needling, and it is confirmed that positioning is effective when the comprehensive change rate is larger than or equal to 15%, so that an intelligent closed loop is formed. According to the system, the positioning precision and repeatability (ICC is larger than or equal to 0.95) are remarkably improved, the gas yield is increased by 40%, the positioning time is shorter than 30 seconds, the system is suitable for patients with special body types, and precise, standardized and intelligent development of acupuncture diagnosis and treatment is promoted.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine diagnosis and treatment technology, and in particular to a precise acupoint positioning system and method based on multimodal sensor fusion. Background Technology

[0002] Acupuncture, as an important component of traditional Chinese medicine, has demonstrated unique efficacy in pain management, neurological rehabilitation, and chronic disease management. The World Health Organization (WHO) has officially recognized the effectiveness of acupuncture for various diseases and has established international guidelines for the location of 361 standard acupoints. However, the consistent effectiveness of acupuncture depends heavily on the precise location of acupoints. Both traditional acupoint selection methods and modern electronic assistive devices have significant technical bottlenecks, severely restricting the standardized promotion and clinical reproducibility of acupuncture treatment.

[0003] Current clinical practice relies primarily on two main methods for acupoint location: one is the age-old empirical surface location method, including bone measurement, surface landmarks, and finger measurement; the other is electronic acupoint instruments based on electrical detection principles. Traditional methods heavily depend on the practitioner's experience and tactile judgment, exhibiting significant subjectivity and uncertainty. Research data shows that the consistency coefficient (ICC) for different acupuncturists locating the same acupoint is only 0.45–0.65, far below the clinically acceptable threshold of 0.80. Even with the same physician operating at different times, the location error can reach 5–10 millimeters. For obese patients (BMI>30), those with edema, muscle atrophy, or anatomical variations (approximately 15–20% of the population), the failure rate of traditional methods is as high as 40–60%. Furthermore, deep acupoints (such as Huantiao and Zhibian, reaching depths of 8–10 cm) cannot be palpated through the body surface, and their location relies entirely on rough anatomical deductions, further increasing treatment risks and uncertainty in efficacy.

[0004] To overcome the shortcomings of manual location, various electronic acupoint detection devices have emerged on the market, with low-frequency bioelectrical impedance analysis (BIA) being the mainstream technology. These devices are based on the assumption that "the resistance at acupoints is lower than that of surrounding tissues," using skin electrodes to detect impedance changes to locate acupoints. However, this technology has fundamental limitations: First, external factors such as skin humidity, sweat secretion, and electrode contact pressure can significantly interfere with measurement results, leading to a false positive rate as high as 30-40%, resulting in generally low trust from clinicians. Second, the penetration depth of low-frequency current (usually <10kHz) is limited, only reflecting the electrical characteristics of the superficial 2-3 millimeters of skin, and cannot detect the true anatomical location and physiological state of deep acupoints. Third, local pressure in non-acupoint areas can also induce changes in electrical characteristics, making it difficult for the device to distinguish between genuine and false acupoints. Finally, the widespread presence of devices such as electrocardiogram monitors and electrosurgical units in hospital environments generates electromagnetic interference, severely affecting the stability and accuracy of low-frequency impedance measurements.

[0005] In the era of personalized medicine, existing technologies still reveal two major problems: insufficient adaptability and a lack of verification mechanisms. On the one hand, standard acupoint coordinate maps are based on a "standard human body" model, ignoring individual differences caused by age (the difference between children and adults can be as high as 40%), gender, body type (subcutaneous fat thickness can increase 5-10 times in obese individuals), disease state, and even changes in body position. On the other hand, existing equipment only provides positioning suggestions "before needle insertion," and cannot provide real-time objective verification of whether the acupoints are actually punctured "after needle insertion" or whether the expected "deqi" physiological response (such as local electromyographic activation and increased blood flow) is triggered. Treatment effects can only rely on the patient's subjective description and the doctor's experience in retrospect, lacking closed-loop feedback and quantitative evaluation methods. This has kept acupuncture treatment in a "black box" state for a long time, seriously hindering its development as evidence-based medicine and its international recognition.

[0006] In summary, traditional acupoint selection methods are highly subjective and have poor repeatability. Existing electronic devices detect superficial acupoints, are susceptible to interference, cannot adapt to individual differences, and lack real-time verification capabilities after needle insertion. Therefore, there is an urgent need for an intelligent acupoint positioning system that can integrate multi-dimensional physiological information, combine superficial and deep acupoint selection, provide individualized adaptation, and possess closed-loop verification capabilities to promote the development of acupuncture diagnosis and treatment towards precision, standardization, and repeatability. Summary of the Invention

[0007] The purpose of this invention is to overcome the aforementioned shortcomings of existing technologies and provide a precise acupoint localization system and method based on multimodal sensor fusion. By integrating multimodal technologies such as ultrasound imaging, infrared thermal imaging, bioelectrical impedance analysis, and pressure tactile sensing, combined with deep learning fusion algorithms and 3D modeling, the system achieves precise, objective, and repeatable 3D spatial localization of acupoints (error <2mm). The system can verify and adapt to individual anatomical differences in real time, providing a standardized and intelligent closed-loop solution for acupuncture treatment, and promoting the precision and modernization of traditional Chinese medicine diagnosis and treatment.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a precise acupoint positioning system based on multimodal sensor fusion, comprising: The multimodal sensor module is used to simultaneously acquire ultrasound images, infrared thermograms, multi-frequency bioelectrical impedance spectra, and pressure distribution data of acupoint areas; The data processing module is used to extract and fuse features from the multimodal data and output the three-dimensional coordinates and confidence scores of the acupoints. The 3D modeling and registration module is used to construct individualized 3D body surface models of patients and map multimodal data to a unified coordinate system; The verification feedback module is used to retest multimodal signals before and after acupuncture, and verify the effectiveness of positioning by the rate of change of signals.

[0009] Furthermore, the multimodal sensor module includes: Ultrasonic imaging unit, frequency range 10–20MHz, axial resolution ≤0.1mm; Infrared thermal imaging unit, thermal sensitivity ≤0.05℃, spatial resolution ≥640×480 pixels; The bioelectrical impedance measurement unit supports multi-frequency scanning from 1 kHz to 1 MHz and employs a four-electrode method. A flexible pressure tactile sensor array with a spatial resolution of ≤2mm supports real-time pressure distribution acquisition.

[0010] Furthermore, the data processing module includes: The feature extraction submodule uses a convolutional neural network to extract high-dimensional feature vectors for each modality. The multimodal fusion submodule implements feature weighted fusion based on a cross-modal attention mechanism; The localization output submodule outputs the three-dimensional coordinates (x, y, z) of the acupoint center and the confidence score P. The confidence score P is calculated based on the Bayesian posterior probability, as shown in the following formula: ; in, For the number of samples taken in Monte Carlo, For neural network prediction functions, This is the Sigmoid activation function.

[0011] Secondly, the present invention provides a method for precise acupoint localization based on multimodal sensor fusion, comprising the following steps: S1: Acquire three-dimensional point cloud data of the patient's body surface and identify at least 10 bony landmarks; S2: Calculate the initial estimated coordinates of the target acupoint based on the bone measurement method; S3: Simultaneously collect ultrasound, infrared, electrical impedance and pressure data within a 3cm×3cm area around the estimated coordinates; S4: Preprocess and extract features from each modality of data; S5: Input multimodal features into a pre-trained deep learning model and output a heatmap of acupoint probabilities; S6: Extract the location with the highest probability as the final acupoint coordinates and calculate the confidence level; S7: If the confidence level is ≥80%, guide the operator to insert the needle; S8: After needle insertion, retest the multimodal signal. If the signal change rate is ≥15%, the positioning is confirmed to be effective.

[0012] Furthermore, the preprocessing of the bioelectrical impedance data in step S4 includes: Based on the Cole-Cole model, the impedance spectrum is fitted, and the characteristic frequency fc and phase angle ϕ are extracted. Calculate the impedance difference ratio between the acupoint and the surrounding tissues. : ; in, Resistance at acupoints This represents the average impedance of the surrounding area.

[0013] Furthermore, the processing of the infrared thermal imaging data in step S4 includes: Calculate the temperature difference between the target area and the surrounding 5cm annular area. : ; Analysis of the recovery time constant of the temperature recovery curve after pressurization : ; Among them, acupoints The value is typically 1.5–2 times smaller than that of the surrounding tissue.

[0014] Furthermore, the processing of the pressure tactile data in step S4 includes: Calculate the tissue stiffness index H: ; in, To apply maximum pressure, The maximum deformation depth; The hardness index at acupoints is usually 20–40% lower than that of surrounding tissues.

[0015] Furthermore, the deep learning model described in step S5 employs a cross-modal attention fusion mechanism, and its attention weights are calculated as follows: ; in, , , These represent the query, key, and value matrices, respectively, derived from feature vectors of different modalities.

[0016] Furthermore, the confidence calculation in step S6 is based on an uncertainty quantification method, which involves multiple inferences using Monte Carlo Dropout to calculate the variance of the coordinate prediction. : ; in, For the number of samples, For single-time predicted coordinates, The coordinates are averaged.

[0017] Furthermore, the signal change rate mentioned in step S8 The calculation formula is: ; in, and These are the characteristic values ​​of a certain modal signal (such as impedance, temperature, or pressure) before and after acupuncture.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: Significantly improved positioning accuracy: average positioning error ≤2mm, far lower than the 5-10mm of traditional methods, and repeatability ICC≥0.95, significantly improving the consistency and reliability of acupoint positioning.

[0019] Multimodal information fusion: For the first time, four orthogonal physical quantities, namely ultrasound (anatomical structure), infrared (thermal distribution), electrical impedance (electrical properties) and pressure (mechanical response), are integrated to form a "multidimensional fingerprint" of acupoints. Even when a single sensor fails, the accuracy rate can still be maintained at over 80%.

[0020] Full coverage of depth and superficial layers: Ultrasonic detection of deep layers (up to 4cm), impedance detection of middle layers, and thermal imaging and pressure detection of the surface layer achieve full-layer coverage of 0-40mm, solving the problem of locating deep acupoints (such as Huantiao acupoint).

[0021] Strong individualized adaptive capability: Based on the patient's three-dimensional scan model, it automatically identifies bony landmarks and maps standard acupoints, completing individualized modeling within 2-3 minutes, adapting to different body types (BMI 18-35 accuracy rate >92%), age, and positional changes.

[0022] Real-time verification and feedback closed loop: Multimodal signals are retested in real time before and after acupuncture. When the change rate is >15%, the positioning is automatically confirmed to be effective, forming an intelligent closed loop of "positioning-verification-optimization", which improves the gas delivery rate by 40%.

[0023] High level of intelligence: Based on a deep learning-based cross-modal attention fusion algorithm, it supports uncertainty quantification output, and the model is continuously optimized through online learning; the system integrates AR navigation and force feedback guidance, and the single acupoint positioning time is less than 30 seconds.

[0024] Highly clinically applicable: It seamlessly integrates with electroacupuncture equipment, supports standardized data collection and management, provides a technical foundation for remote guidance, efficacy evaluation, and big data analysis, and promotes the standardization of acupuncture diagnosis and treatment and the development of evidence-based medicine. Attached Figure Description

[0025] Figure 1 This is a block diagram of the overall system structure of the present invention; Figure 2 A flowchart illustrating the method for locating acupoints; Figure 3 This is a multimodal data diagram of the location of the Hegu acupoint in an embodiment of the present invention; Figure 4 This is a signal comparison diagram from the verification stage of an embodiment of the present invention; Figure 5 This is a graph showing the comparative experimental data of the positioning accuracy of the present invention. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0027] Example 1 This embodiment provides a precise acupoint positioning system based on multimodal sensor fusion, including: The multimodal sensor module is used to simultaneously acquire ultrasound images, infrared thermograms, multi-frequency bioelectrical impedance spectra, and pressure distribution data of acupoint areas; The data processing module is used to extract and fuse features from the multimodal data and output the three-dimensional coordinates and confidence scores of the acupoints. The 3D modeling and registration module is used to construct individualized 3D body surface models of patients and map multimodal data to a unified coordinate system; The verification feedback module is used to retest multimodal signals before and after acupuncture, and verify the effectiveness of positioning by the rate of change of signals.

[0028] Specifically, such as Figure 1 As shown in the overall system structure diagram, a precise acupoint positioning system based on multimodal sensor fusion is implemented as follows: Multimodal sensor module: This module integrates an ultrasound imaging probe, an infrared thermal imaging camera, a bioelectrical impedance electrode, and a flexible pressure sensor array. The probe housing is made of medical-grade ABS plastic, and the internal sensor spacing is precisely calibrated to ensure alignment of the data acquisition centers for each modality.

[0029] Data processing module: Deployed on a dedicated industrial computer, it adopts a modular design. Data from each sensor is transmitted via a USB 3.2 Gen2 interface, with a total bandwidth of up to 10Gbps. The processing module runs CentOS 7.9 and uses PyTorch 1.12 as the deep learning framework.

[0030] 3D modeling module: Employs an Einstar handheld structured light scanner with a precision of 0.1mm and a scanning speed of 3 million points / second. After scanning, the system completes point cloud registration and bony landmark recognition within 10 seconds.

[0031] The verification feedback module includes a Natus NeuroWorks electromyography (EMG) acquisition system (2kHz sampling rate) and a Perimed PeriFlux 5000 laser Doppler flowmeter (32Hz sampling rate). Data is transmitted to the main control system in real time via gigabit Ethernet.

[0032] In this embodiment, the system is calibrated in a laboratory environment, and the spatiotemporal synchronization accuracy of each sensor is verified to be 0.1mm / 1ms using a standard phantom.

[0033] Example 2 Furthermore, the multimodal sensor module includes: Ultrasonic imaging unit, frequency range 10–20MHz, axial resolution ≤0.1mm; Infrared thermal imaging unit, thermal sensitivity ≤0.05℃, spatial resolution ≥640×480 pixels; The bioelectrical impedance measurement unit supports multi-frequency scanning from 1 kHz to 1 MHz and employs a four-electrode method. A flexible pressure tactile sensor array with a spatial resolution of ≤2mm supports real-time pressure distribution acquisition.

[0034] Specific parameters and implementation details of multimodal sensors: Implementation of the ultrasound imaging unit: The L14-5W linear array probe was used, with a center frequency of 15MHz and a bandwidth of 5-20MHz. The probe incorporates 128 piezoelectric crystals, and the acoustic beam focusing depth is adjustable (1-4cm). Image acquisition was performed using a Verasonics Vantage256 system with an RF data sampling rate of 62.5MHz. During the procedure, an ultrasonic coupling agent was applied to the probe surface, and the probe was moved across the skin at a speed of 50mm / s to acquire B-mode image sequences.

[0035] Infrared thermal imaging unit implementation: The FLIRA655sc thermal imager was used, with a resolution of 640×480 and a NETD < 30mK. The lens focal length was 25mm, and the working distance was adjustable from 10-50cm. During implementation, the ambient temperature was controlled at 22±1℃, and the relative humidity was 40-60%. The system automatically corrected the emissivity (0.98 for skin by default) and eliminated the influence of environmental reflections.

[0036] Bioelectrical impedance unit implementation: The ADIAD5933 impedance analysis chip is used, with a frequency range of 1kHz-1MHz. The electrodes are configured with four terminals: 20mm spacing between the outer electrodes and 10mm spacing between the inner electrodes. During implementation, a 1μARMS AC current is applied, and the voltage measurement accuracy is ±0.5%. The system incorporates a 50Hz / 60Hz power frequency notch filter.

[0037] Implementation of pressure-sensitive haptic unit: The flexible sensor is based on the capacitive principle, with an 8×8 array covering a 25mm×25mm area. Each unit contains upper and lower copper electrodes and a middle PDMS dielectric layer. It has a pressure sensitivity of 0.8pF / kPa and a linearity >95%. In practice, data is read at a rate of 500Hz via an I²C interface, and a built-in temperature compensation algorithm is included.

[0038] Example 3 Furthermore, the data processing module includes: The feature extraction submodule uses a convolutional neural network to extract high-dimensional feature vectors for each modality. The multimodal fusion submodule implements feature weighted fusion based on a cross-modal attention mechanism; The localization output submodule outputs the three-dimensional coordinates (x, y, z) of the acupoint center and the confidence score P. The confidence score P is calculated based on the Bayesian posterior probability, as shown in the following formula: ; in, For the number of samples taken in Monte Carlo, For neural network prediction functions, This is the Sigmoid activation function.

[0039] The specific implementation process of the data processing module: Hardware configuration: Processor: NVIDIA Jetson AGXXavier (32GB RAM) Storage: 1TB NVMe SSD Interfaces: 4 x USB 3.1, 2 x Gigabit Ethernet Software implementation: The feature extraction submodule is deployed based on TensorRT 8.5, and each network model is quantized to FP16 precision: Ultrasound branch: ResNet34 input size 256×256, inference time 8ms Infrared branch: EfficientNet-B2 input size 320×240×3, inference time 12ms Impedance branch: 1D-CNN+LSTM, 21×4 input, inference time 3ms Stress branch: MobileNetV2, 64×64 input, inference time 5ms Implementation of the fusion algorithm: The multimodal fusion submodule uses a custom Transformer layer with the following parameters: Embedding dimension: 1024 Attention count: 8 Feedforward network dimensions: 2048 Dropout rate: 0.1 The confidence level is calculated as follows: The system performs N=20 Monte Carlo Dropout samplings, with Dropout enabled for each inference (p=0.3). The confidence score P is calculated using the following formula: ; in, For the Sigmoid function, This is a neural network model. During implementation, when... The system displays a green highlight. The system will prompt you to remeasure.

[0040] Output module implementation: The positioning results are pushed to the GUI interface in real time via WebSocket and simultaneously output to the electrocautery device via serial port. The coordinate format is (x, y, z) millimeter values, accompanied by a 3×3 covariance matrix.

[0041] Example 4 This embodiment provides a method for precise acupoint localization based on multimodal sensor fusion, including the following steps: S1: Acquire three-dimensional point cloud data of the patient's body surface and identify at least 10 bony landmarks; S2: Calculate the initial estimated coordinates of the target acupoint based on the bone measurement method; S3: Simultaneously collect ultrasound, infrared, electrical impedance and pressure data within a 3cm×3cm area around the estimated coordinates; S4: Preprocess and extract features from each modality of data; S5: Input multimodal features into a pre-trained deep learning model and output a heatmap of acupoint probabilities; S6: Extract the location with the highest probability as the final acupoint coordinates and calculate the confidence level; S7: If the confidence level is ≥80%, guide the operator to insert the needle; S8: After needle insertion, retest the multimodal signal. If the signal change rate is ≥15%, the positioning is confirmed to be effective.

[0042] like Figure 2 The following is a flowchart illustrating the complete method for locating the Hegu acupoint on the hand: S1 Implementation Details: The patient sits with their right hand resting flat on the scanning platform. A three-dimensional point cloud of the back of the hand is acquired using an Einstar scanner. The system automatically fits the hand's curvature and identifies the following bony landmarks: Radial styloid process (RS) Ulnar styloid process (US) Base of the first metacarpal bone (MC1) Base of the second metacarpal bone (MC2) The RANSAC algorithm is used for recognition, with a recognition accuracy of 0.3mm.

[0043] S2 Implementation Details: Based on the standards of "Acupuncture and Moxibustion", the formula for locating the Hegu acupoint is as follows: ; in It automatically adjusts based on hand length. This example calculates... .

[0044] S3 Data Acquisition Implementation: probe with Centered on the target area, scan along a spiral path (radius 15mm, step size 2mm). Acquisition parameters: Ultrasound: Gain 50dB, Depth 25mm Infrared: 30Hz frame rate, 2ms integration time Impedance: Scan 21 frequency points, averaging 10 scans per point. Pressure: Contact force control 200±20kPa S4 preprocessing implementation: Ultrasound image: CLAHE enhancement, 3×3 median filtering Heatmap: Non-uniformity correction, 3×3 Gaussian smoothing Impedance: Remove contact resistance, Cole-Cole fitting Pressure: Bilinear interpolation to 1mm grid S5 reasoning implementation: Input the preprocessed data into the model of Example 3, and output a 64×64 probability heatmap. The center peak coordinates of the heatmap are (45.2, 32.8, 8.1).

[0045] S6 confidence level calculation implementation: Based on 20 Monte Carlo samplings, calculate: ; Confidence .

[0046] S7 needle insertion guidance implementation: The system projects a red crosshair (2mm in diameter) onto the skin surface using a projector, while simultaneously providing a voice prompt: "Please insert the needle to a depth of 15mm, at a perpendicular angle."

[0047] S8 Validation Implementation: Measure immediately after needle insertion and compare signal changes: ; If the value exceeds the threshold by 15%, the system will display "Location valid".

[0048] Example 5 Furthermore, the preprocessing of the bioelectrical impedance data in step S4 includes: Based on the Cole-Cole model, the impedance spectrum is fitted, and the characteristic frequency fc and phase angle ϕ are extracted. Calculate the impedance difference ratio between the acupoint and the surrounding tissues. : ; in, Resistance at acupoints This represents the average impedance of the surrounding area.

[0049] In step S4 of Example 4, combined with Figure 3 (The impedance spectrum portion of the example - multimodal data diagram for locating the Hegu acupoint; detailed processing of the impedance data:) Measurement Protocol Implementation: Frequency point selection: Thirteen points were measured at frequencies of 1k, 1.78k, 3.16k, 5.62k, 10k, 17.8k, 31.6k, 56.2k, 100k, 178k, 316k, 562k, and 1MHz. Each point was measured 10 times and the average was taken.

[0050] Cole-Cole fitting implementation: The Levenberg-Marquardt algorithm is used for fitting, with the objective function as follows: ; Initial value: , , , .

[0051] Feature parameter extraction implementation: Characteristic frequency calculation: ; Impedance difference ratio calculation implementation: Measure the impedance at the center point of the acupoint =65.3kΩ@1kHz Measure the average impedance at 8 points around the perimeter. =84.7kΩ@1kHz ; Phase angle measurement: exist Location, acupoint phase =11.8°, surrounding points =7.9°, the difference is significant.

[0052] Quality control implementation: Goodness-of-fit requirements >0.98, residual standard deviation <5%. (Example) =0.991, passed quality control.

[0053] Example 6 Furthermore, the processing of the infrared thermal imaging data in step S4 includes: Calculate the temperature difference between the target area and the surrounding 5cm annular area. : ; Analysis of the recovery time constant of the temperature recovery curve after pressurization : ; Among them, acupoints The value is typically 1.5–2 times smaller than that of the surrounding tissue.

[0054] In step S4 of Example 4, combined with Figure 3 In the infrared thermogram portion of the example (multimodal data map for locating the Hegu acupoint), infrared thermographic processing is implemented as follows: Data collection implementation: Environmental conditions: Room temperature 23.5℃, no wind, avoid direct sunlight. The patient sat quietly for 10 minutes to acclimatize. The thermal imager was held 20cm away from the skin, with automatic focus adjustment.

[0055] Static temperature difference analysis implementation: by Centered on, define: Target area: radius Circular shape, approximately 200 pixels Control area: Inner diameter , outer diameter Circular shape, approximately 1800 pixels Temperature calculation implementation: ; Actual measurement: , ; ; Dynamic test implementation: Record baseline temperature before applying pressure ; Apply 300 kPa pressure for 3 seconds using a pressure sensor. After releasing the pressure, the recovery curve was recorded at 10Hz for 60 seconds. Fitting exponential model: ; Using the least squares method for fitting, we obtain: Acupoints: =8.3s, ; Surrounding tissues: =15.8s, ; Recovery rate ratio = =1.90 Implementation of spatial gradient analysis: Calculate the two-dimensional temperature gradient: , ; At the center of the acupoint, |G| = 1.2℃ / cm, and the direction is consistent with the direction of the Large Intestine Meridian of Hand Yangming (towards the index finger).

[0056] Example 7 Furthermore, the processing of the pressure tactile data in step S4 includes: Calculate tissue stiffness index : ; in, To apply maximum pressure, The maximum deformation depth; The hardness index at acupoints is usually 20–40% lower than that of surrounding tissues.

[0057] In step S4 of Example 4, combined with Figure 3 (In the example - multimodal data map of Hegu acupoint location), the pressure distribution map section, pressure data processing implementation: Hardness testing protocol implementation: Initial contact: Increase the pressure to 50 kPa at a rate of 10 kPa / s and hold for 2 seconds. Gradual pressurization: Increase the pressure to 300 kPa in 25 kPa increments, holding each increment for 3 seconds. Relaxation test: Maintain 300 kPa pressure for 10 seconds Unload: Unload at a rate of 50 kPa / s Hardness index calculation implementation: Depth of tissue deformation was measured under 300 kPa pressure. ; ; Average hardness index of 8 points around the perimeter ; Hardness reduction ratio calculation: ; Implementation of tenderness threshold test: Pressure increased in 25 kPa increments The patient holds the wireless button and presses it when they first feel pain. The system records the pressure value at that moment. ; Actual measurement: acupoints average surrounding ; Increased tenderness sensitivity: ; Stress relaxation analysis implementation: Record the pressure decay curve while maintaining a pressure of 300 kPa. Fit a standard linear solid model: ; Actual measurement: , , ; Relaxation rate: .

[0058] Example 8 Furthermore, the deep learning model described in step S5 employs a cross-modal attention fusion mechanism, and its attention weights are calculated as follows: ; in, , , These represent the query, key, and value matrices, respectively, derived from feature vectors of different modalities.

[0059] Implementation details of deep learning models: Model architecture implementation: Input layer: Quad-modal parallel input Encoder: Independent encoding for each mode Fusion Layer: Cross-modal Attention Mechanism Output layer: Coordinate regression + confidence classification Attention mechanism implementation: Taking ultrasonic modality as an example, attention calculation implementation: Query matrix: ; Key matrix: Others are similar; Value matrix: Others are similar; Implementation of attention weight calculation: ; In practice The scaling factor is 1 / √64 = 0.125.

[0060] For ultrasonic modalities, weighted fusion is implemented as follows: ; in These are the attention weights after Softmax normalization.

[0061] Training Implementation: Dataset: 850 patients, 30 acupoints, 50-100 samples per acupoint Data augmentation: rotation ±15°, scaling 0.8-1.2x, loss of random modalities. Loss function: ; in , , ; 4. Optimizer: AdamW(lr=1e-4, weight) decay =1e-5) 5. Training cycle: 200 epochs, with early stop and patience (20 epochs). Reasoning optimization implementation: The model was quantized using TensorRT, which improved inference speed from 45ms to 15ms. Memory usage was reduced from 350MB to 85MB.

[0062] Example 9 Furthermore, the confidence calculation in step S6 is based on an uncertainty quantification method, which involves multiple inferences using Monte Carlo Dropout to calculate the variance of the coordinate prediction. : ; in, For the number of samples, For single-time predicted coordinates, The coordinates are averaged.

[0063] In step S6 of Example 4, combined with Figure 5 The uncertainty analysis section of the (positioning accuracy comparison experiment data chart) includes the implementation of uncertainty quantification: Monte Carlo Dropout Implementation: During the inference phase, the Dropout layer remains active, and the inference is repeated N=20 times. Before each inference, 30% of the neurons are randomly dropped to simulate model uncertainty.

[0064] Variance calculation implementation: Record the three-dimensional coordinates of each inference output. ; Calculate the mean: ; Calculate the variance: ; During the implementation, the results of 20 samplings were as follows: :45.1,45.3,45.0,45.4,45.2,45.1,45.3,45.2,45.0,45.4,45.3,45.1,45.2,45.4,45.0,45.3,45.1,45.2,45.3,45.1; The calculation yields: , ; Similarly, the variances in the y and z directions are calculated to obtain the covariance matrix Σ: ; Uncertainty ellipsoid implementation: Eigenvalue decomposition of Σ: Σ=VΛV^T The eigenvalues ​​are calculated as follows: , , ; The ellipsoid direction is defined by the corresponding eigenvector.

[0065] The semi-axial length of the ellipsoid is calculated according to the 3σ principle: ; Confidence scoring implementation: Confidence calculation based on trace: ; in k=1.5 ; During implementation, when Displays "High confidence level". Displays "Medium confidence level". It is recommended to remeasure.

[0066] Example 10 Furthermore, the signal change rate mentioned in step S8 The calculation formula is: ; in, and These are the characteristic values ​​of a certain modal signal (such as impedance, temperature, or pressure) before and after acupuncture.

[0067] In step S8 of Example 4, combined with Figure 4 (Example - Signal Comparison Chart for Verification Phase) shows the verification feedback implementation: Data collection before and after acupuncture: Monitoring is initiated immediately after needle insertion and continues for 30 seconds. Sampling parameters: Impedance: 1kHz single-frequency continuous measurement, 10Hz sampling rate Temperature: Infrared thermal imager, 30Hz continuous shooting Electromyography: Surface electrodes were placed 1 cm away from the acupoint, and sampling was performed at 2 kHz. Implementation of signal change rate calculation: Calculation of the rate of change of electrical impedance: Before needle insertion: (10-second average); After needle insertion: (The average duration of obtaining Qi is 10 seconds); ; Calculation of temperature change rate: Before needle insertion: (10-second average); After needle insertion: (Peak temperature rise); ; Calculation of electromyographic rate of change: Before needle insertion (Restless state); After needle insertion (During the period of obtaining Qi); ; Implementation of comprehensive rate of change calculation: Weighting based on clinical validation: ; in , , ; ; Verify the implementation of decisions: The threshold setting is based on clinical trials: ; Effective positioning Suggestion: Fine-tuning It is recommended to reposition. This example The system determines that the location is valid and automatically records it: Acupoint coordinates: Hegu (45.2, 32.8, 8.1) Needle insertion parameters: Depth 15mm, Angle 90° Physiological response: impedance decreased by 10.1%, temperature increased by 1.07%, and EMG increased by 43.2%. Timestamp: 2024-05-20 14:30:25 The data is uploaded to a cloud database for subsequent model optimization and efficacy analysis.

[0068] Implementation of effect verification like Figure 5 (As shown in the experimental data chart comparing positioning accuracy), the clinical trial design is as follows: Participants: 60 individuals, aged 25-65 years, BMI 18-32 Acupoints: Hegu (LI4), Zusanli (ST36), and Shenshu (BL23), 20 cases each. Control group: 3 acupuncturists with over 15 years of experience manually positioned the needles. Evaluation indicators: positioning error, repeatability, gas yield, and operation time. Positioning accuracy verification: The distance between the actual needle insertion point and the system positioning point is measured using an optical tracking system (NDIPolaris).

[0069] System group average error: 1.8 ± 0.6 mm The average error of the traditional group was 5.2 ± 2.1 mm. The t-test showed p < 0.001, indicating a significant difference.

[0070] Repeatable implementation verification: The same operator repeated the positioning 10 times, and the intragroup correlation coefficient (ICC) was calculated.

[0071] System group ICC=0.95 (95% CI: 0.92-0.97) The conventional group had an ICC of 0.68 (95% CI: 0.61–0.74). Three different operators located the same patient and calculated the intergroup ICC.

[0072] System group ICC=0.89 (95% CI: 0.85-0.92) The conventional group had an ICC of 0.52 (95% CI: 0.43-0.60).

[0073] Verification of gas yield: Based on patient subjective reports and objective electromyography (EMG) testing.

[0074] System-guided group gas yield rate: 92% (55 / 60) Gas yield rate in the traditional method group: 68% (41 / 60) The test showed that p < 0.01, indicating a significant difference.

[0075] Operation time implementation verification: The time from the start of scanning to confirmation of location.

[0076] System group: 28.5 ± 3.2 seconds Traditional group: 45.8 ± 8.7 seconds The t-test showed p < 0.001, indicating that the system was significantly faster.

[0077] Applicability verification: Tests grouped by BMI: BMI < 25 (n=20): Accuracy 95% 25 ≤ BMI < 30 (n=20): Accuracy 93% BMI ≥ 30 (n=20): Accuracy 88% Traditional methods have an accuracy rate of only 62% in the BMI ≥ 30 group.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A precise acupoint positioning system based on multimodal sensor fusion, characterized in that, include: The multimodal sensor module is used to simultaneously acquire ultrasound images, infrared thermograms, multi-frequency bioelectrical impedance spectra, and pressure distribution data of acupoint areas; The data processing module is used to extract and fuse features from the multimodal data and output the three-dimensional coordinates and confidence scores of the acupoints. The 3D modeling and registration module is used to construct individualized 3D body surface models of patients and map multimodal data to a unified coordinate system; The verification feedback module is used to retest multimodal signals before and after acupuncture, and verify the effectiveness of positioning by the rate of change of signals.

2. The system according to claim 1, characterized in that, The multimodal sensor module includes: Ultrasonic imaging unit, frequency range 10–20MHz, axial resolution ≤0.1mm; Infrared thermal imaging unit, thermal sensitivity ≤0.05℃, spatial resolution ≥640×480 pixels; The bioelectrical impedance measurement unit supports multi-frequency scanning from 1 kHz to 1 MHz and employs a four-electrode method. A flexible pressure tactile sensor array with a spatial resolution of ≤2mm supports real-time pressure distribution acquisition.

3. The system according to claim 1, characterized in that, The data processing module includes: The feature extraction submodule uses a convolutional neural network to extract high-dimensional feature vectors for each modality. The multimodal fusion submodule implements feature weighted fusion based on a cross-modal attention mechanism; The localization output submodule outputs the three-dimensional coordinates (x, y, z) of the acupoint center and the confidence score P. The confidence score P is calculated based on the Bayesian posterior probability, as shown in the following formula: ; in, For the number of samples taken in Monte Carlo, For neural network prediction functions, This is the Sigmoid activation function.

4. A method for precise acupoint localization based on multimodal sensor fusion, characterized in that, Includes the following steps: S1: Acquire three-dimensional point cloud data of the patient's body surface and identify at least 10 bony landmarks; S2: Calculate the initial estimated coordinates of the target acupoint based on the bone measurement method; S3: Simultaneously collect ultrasound, infrared, electrical impedance and pressure data within a 3cm×3cm area around the estimated coordinates; S4: Preprocess and extract features from each modality of data; S5: Input multimodal features into a pre-trained deep learning model and output a heatmap of acupoint probabilities; S6: Extract the location with the highest probability as the final acupoint coordinates and calculate the confidence level; S7: If the confidence level is ≥80%, guide the operator to insert the needle; S8: After needle insertion, retest the multimodal signal. If the signal change rate is ≥15%, the positioning is confirmed to be effective.

5. The method according to claim 4, characterized in that, The preprocessing of bioelectrical impedance data in step S4 includes: Based on the Cole-Cole model, the impedance spectrum is fitted, and the characteristic frequency fc and phase angle ϕ are extracted. Calculate the impedance difference ratio between the acupoint and the surrounding tissues. : ; in, Resistance at acupoints This represents the average impedance of the surrounding area.

6. The method according to claim 4, characterized in that, The processing of infrared thermal imaging data in step S4 includes: Calculate the temperature difference between the target area and the surrounding 5cm annular area. : ; Analysis of the recovery time constant of the temperature recovery curve after pressurization : ; Among them, acupoints The value is typically 1.5–2 times smaller than that of the surrounding tissue.

7. The method according to claim 4, characterized in that, The processing of pressure tactile data in step S4 includes: Calculate tissue stiffness index : ; in, To apply maximum pressure, The maximum deformation depth; The hardness index at acupoints is usually 20–40% lower than that of surrounding tissues.

8. The method according to claim 4, characterized in that, The deep learning model described in step S5 employs a cross-modal attention fusion mechanism, and its attention weights are calculated as follows: ; in, , , These represent the query, key, and value matrices, respectively, derived from feature vectors of different modalities.

9. The method according to claim 4, characterized in that, The confidence calculation in step S6 is based on the uncertainty quantification method, which involves multiple inferences using Monte Carlo Dropout to calculate the variance of the coordinate prediction. : : in, For the number of samples, For single-time predicted coordinates, The coordinates are averaged.

10. The method according to claim 4, characterized in that, The rate of change of the signal in step S8 The calculation formula is: ; in, and These are the characteristic values ​​of a certain modal signal (such as impedance, temperature, or pressure) before and after acupuncture.

Images