An AI positioning method combined with key frame processing of ultrasonic image

Abstract

The application discloses a kind of combination ultrasonic image key frame processing sunroof hole AI positioning method, it is related to computer vision technical field, including acquisition sunroof hole static ultrasonic contrast image and dynamic video, keep probe and skin adhesion pressure 20-30kPa, moving speed is not more than 5mm / s;760x600 pixel gray scale chart is preprocessed, denoising enhancement and normalization, eliminate invalid data;Extract key frame;ResNet-50+FPN is trained as the integrated model of main stem;Inference output boundary frame, center point and confidence;Confidence is not less than 0.7 labeling visualization, lower then prompt to rescan.The positioning accuracy of the application is superior to traditional palpation and single AI model, real-time meets clinical guidance demand;Generalization ability is strong, adapts multiple ultrasonic equipment, convenient operation improves physician satisfaction, helps acupuncture precision.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an AI-based method for locating the Tianchuang acupoint by combining keyframe processing of ultrasound images. Background Technology

[0002] In acupuncture treatment, the Tianchuang acupoint, as an important acupoint on the Hand-Taiyang Small Intestine Meridian, directly affects the treatment effect and the safety of the procedure due to its accurate location. This acupoint is surrounded by blood vessels, nerves, and other vital tissues; improper operation can easily lead to complications such as bleeding and nerve damage. However, current clinical practice still primarily relies on traditional palpation to locate the Tianchuang acupoint. This method is highly dependent on the physician's clinical experience and tactile sense, is subjective, and has poor repeatability. Obese patients, due to their thicker subcutaneous fat, have unclear surface landmarks for the acupoint, easily leading to location errors; slender patients may have thinner muscle layers, making it easy to misjudge the depth of the acupoint during palpation. These individual differences increase the risk of location errors, thus affecting the treatment effect or causing safety hazards. Although some clinical practice uses CT or MRI-guided location, the equipment is expensive, the operation process is complex, and CT guidance involves ionizing radiation, making it unsuitable for special populations such as pregnant women and children. MRI guidance cannot achieve real-time dynamic location and is difficult to adapt to actual clinical scenarios such as probe position adjustments and slight changes in patient position during acupuncture procedures, limiting its widespread application in routine acupuncture treatment.

[0003] With the increasing penetration of artificial intelligence (AI) technology into the medical field, some studies have attempted to achieve acupoint localization using AI. However, existing solutions still have significant limitations. On the one hand, most AI localization methods only process static ultrasound images, ignoring the continuous changes in acupoint areas in dynamic ultrasound videos. Relying solely on a single static image can easily lead to incomplete localization results due to probe scanning angle deviations and dynamic tissue deformation. On the other hand, existing models often employ a single deep learning architecture (such as using only YOLO or Faster R-CNN), failing to consider the differences in how different models capture the texture and shape features of ultrasound images. This can easily result in insufficient localization accuracy due to inherent model biases. Furthermore, existing solutions lack the ability to adapt to individual patient body shape differences and do not incorporate the anatomical features around the Tianchuang acupoint (such as the distribution of blood vessels and nerves) for localization optimization. They also cannot actively avoid key tissues during localization, still posing operational safety risks and failing to meet the clinical triple requirements of "accuracy, real-time performance, and safety" for acupoint localization.

[0004] The current medical industry has an increasingly urgent need for acupuncture intelligence. Especially in primary medical institutions, the experience levels of physicians vary, and it is more difficult to locate acupoints with deeper positions and complex surrounding structures such as Tianchuang acupoint. There is an urgent need for a technical solution that takes into account operation convenience, positioning reliability, and clinical safety. The subjectivity of traditional methods, the limitations of existing imaging guidance technologies, and the insufficient adaptability of existing AI solutions have jointly led to a slow intelligent process for locating the Tianchuang acupoint, which cannot fully support the standardized and precise development of acupuncture treatment. Therefore, developing a positioning method that combines the dynamic characteristics of ultrasound images and AI integrated learning has become the key to solving the above problems. Summary of the Invention

[0005] An AI positioning method for the Tianchuang acupoint that combines the processing of key frames of ultrasound images is proposed in this invention to solve the problems mentioned in the above existing technologies.

[0006] To achieve the above objectives, this invention adopts the following technical solutions: An AI positioning method for the Tianchuang acupoint that combines the processing of key frames of ultrasound images includes the following steps: Data acquisition step: Collect static contrast-enhanced ultrasound images and dynamic contrast-enhanced ultrasound videos of the Tianchuang acupoint area, and keep the ultrasound probe close to the skin during collection; Data preprocessing step: Perform standardization processing on the ultrasound image data, uniformly adjust it to a grayscale image of 760×600 pixels, remove noise through Gaussian filtering with a 5×5 convolution kernel, enhance texture features with CLAHE, normalize the pixel values to the range of 0-1, and剔除 invalid data; Key frame extraction step: Extract candidate frames from the dynamic video at 3 frames per second, and calculate the grayscale value difference between frames through the Sobel operator; Model training step: Build a deep learning model based on ResNet-50+FPN, configure 9 anchor boxes with different scales and aspect ratios, unify the proposed region sizes through ROI alignment, and fuse the prediction results of three models using the weighted voting method; Use a specified loss function, AdamW optimizer, and CosineAnnealingLR to schedule the learning rate, and divide the dataset according to 3:1:1; AI positioning inference step: Input the preprocessed static image and key frames into the integrated model, output the bounding box and confidence score, and calculate the center point coordinates through the bounding box; Result output and visualization step: Real-time display the bounding box, center point coordinates, and confidence score. When the confidence is lower than 0.7, a text prompt is issued; when it is not lower than 0.7, draw a bounding box with a red solid line, mark the center point with a blue dot, and display the confidence score in white text on the image.

[0007] Furthermore, it also includes a key frame optimization and screening step, which calculates the comprehensive quality score of the candidate frames through the frame quality evaluation formula. The formula is Where This represents the overall quality score of the candidate frames. This represents the sharpness weighting coefficient. Indicates the weighting coefficients for inter-frame differences; Indicates the image height. Indicates the image width; Indicates the image in OK The gradient value of the column; Indicates the number of adjacent frames to the candidate frame; Indicates the candidate frame and the first The difference in grayscale values ​​between adjacent frames; This represents the average difference between three adjacent frames. This represents the maximum difference among three adjacent frames.

[0008] Furthermore, it also includes a localization error correction step, which compensates for errors in the bounding box coordinates output by the model, and adjusts the localization results based on the individual patient's body shape characteristics. The compensation formula is as follows: in , These represent the corrected center points. coordinates and coordinate; , Represents the original center point coordinates output by the model; This represents the width direction error correction coefficient. This represents the correction factor for height direction error; This represents the average width of the Tianchuang acupoint region in patients with standard body types. This indicates the width of the Tianchuang acupoint area for the current patient; Indicates the image width; This represents the mean height of the Tianchuang acupoint region in patients with standard body type, and the statistical sample is compared with... Consistent; This indicates the current height of the Tianchuang acupoint area in the patient; Indicates the image height.

[0009] Furthermore, the data augmentation step is included. Before model training, the preprocessed image data is augmented by generating new training samples through random flipping, rotation, scaling, and brightness adjustment. The random flipping uses a 50% probability of horizontal flipping and a 30% probability of vertical flipping, with the horizontal flipping axis being the vertical midline of the image and the vertical flipping axis being the horizontal midline of the image. The rotation angle is randomly selected from -15° to 15° in a uniform distribution, and bilinear interpolation is used to fill the image edges after rotation. The scaling ratio is taken from 0.8 to 1.2 in a normal distribution, with a mean of 1.0 and a standard deviation of 0.1. Brightness adjustment is completed through Gamma correction. Each data augmentation generates three times the number of new samples as the original samples.

[0010] Furthermore, it also includes a dynamic adjustment step for model ensemble weights, calculating the localization accuracy of each model for different types of ultrasound images in the training set, and assigning dynamic weights based on the accuracy. The weight allocation formula is as follows: in Indicates the first Weights of the model; Indicates the first The positioning accuracy of this model.

[0011] Furthermore, it also includes real-time optimization steps to accelerate the model inference process. The TensorRT tool is used to convert the model parameters from 32-bit floating-point numbers to 16-bit floating-point numbers. At the same time, the inference engine is optimized by packaging three adjacent key frames into a batch for inference calculation. Through optimization, the localization inference time of a single frame image is controlled within 80-95 milliseconds, which meets the clinical real-time guidance requirement of less than 100 milliseconds latency.

[0012] Furthermore, it also includes a confidence threshold adaptive adjustment step, dynamically setting the confidence threshold based on the quality level of the ultrasound image; the image quality level is calculated using the image signal-to-noise ratio (SNR). The signal area is a 20×20 pixel area around the manually marked skylight hole, and the noise area is a 10×10 pixel unorganized area at the edge of the image; the signal-to-noise ratio is greater than 30dB for high-quality images; 25-30dB for good images; 20-25dB for average images; and less than 20dB for poor images.

[0013] Furthermore, it also includes a multimodal feature fusion step, which introduces the anatomical structure features of the tissues around the Tianchuang acupoint during model training; marks the location of key tissues using medical image annotation tools to generate anatomical structure annotation maps; uses the ResNet-18 model to extract features from the anatomical structure annotation maps and outputs a 256-dimensional feature vector; performs a stitching operation on the 256-dimensional deep image features output by the backbone network to form a 512-dimensional fused feature; the fused feature is input into the subsequent region proposal network and classification regression layer to automatically avoid the annotated blood vessel and nerve regions.

[0014] Furthermore, the process includes an online model update step, collecting new ultrasound imaging data and positioning results feedback during clinical applications. The new data must include basic patient information, ultrasound scanning parameters, and positioning coordinates manually corrected by the physician. After preprocessing the new data, it is mixed with the original training set data at a 1:1 ratio for incremental training. During training, the model's historical optimal weights are retained, and the mean positioning error and mean confidence score are evaluated using a validation set. When the mean positioning error of the incrementally trained model decreases by more than 5% or the mean confidence score increases by more than 5%, the model is updated and deployed.

[0015] Furthermore, it also includes a location result verification step, which compares the location results output by the Euclidean distance calculation model with the results manually annotated by the physician, using Euclidean distance... ,in The coordinates of the center point after model correction. The center point coordinates were jointly annotated by three physicians with the title of associate chief physician or above; simultaneously, the average accuracy of the localization results was calculated, with the average accuracy calculated based on the intersection-union ratio (IU) of the predicted bounding box and the true bounding box. The pass rate and exact match rate are calculated monthly. Precise matching rate When the pass rate is below 90% or the exact match rate is below 85%, the model retraining process is triggered, using all historical data and newly collected data during retraining.

[0016] Compared with existing technologies, the beneficial effects of this invention are: In terms of positioning accuracy, this invention integrates the prediction results of multiple deep learning models through an integrated learning strategy, avoiding the limitations of a single model in capturing ultrasound image features. At the same time, it introduces a keyframe optimization and screening mechanism to select frames with high clarity and complete dynamic features for inference, reducing the interference of blurry or redundant frames on the positioning results. Combined with error correction steps based on the individual body shape characteristics of patients, the positioning results are more in line with the actual anatomical position of patients with different body shapes. The overall positioning accuracy is significantly better than that of traditional palpation methods and single AI models, which can effectively reduce the problem of poor treatment effect caused by positioning deviation.

[0017] In terms of real-time performance, this invention significantly shortens the localization inference time of a single frame image through model quantization acceleration and batch inference optimization, meeting the real-time guidance requirements of clinical acupuncture operations. The key frame extraction strategy of dynamic ultrasound video reduces the amount of data processing while preserving complete dynamic features, avoiding inference delays caused by excessive data volume. Physicians can obtain localization results instantly during the operation without waiting for too long, thus improving the efficiency of clinical operations.

[0018] In terms of safety, this invention introduces a multimodal feature fusion step, which integrates the anatomical features of blood vessels, nerves and other structures around the Tianchuang acupoint into the model training and inference process. This enables the model to actively identify and avoid key tissues when locating acupoints, reducing the risk of blood vessel and nerve damage during acupuncture. At the same time, the confidence threshold adaptive adjustment mechanism increases the threshold when the ultrasound image quality is poor, avoiding the output of unreliable positioning results, further ensuring operational safety and reducing the probability of complications.

[0019] In terms of generalization ability, this invention expands the diversity of training samples through data augmentation, covering ultrasound image data of patients of different genders, ages and physical conditions, so that the model can adapt to a wide range of patients. For ultrasound images under different scanning angles and lighting conditions, the model can also maintain stable positioning performance through deep feature learning, without the need for retraining for specific scenarios, and adapt to different clinical environments such as primary healthcare institutions and large hospitals.

[0020] In terms of clinical adaptability, this invention supports both static images and dynamic videos as data input formats to meet different clinical scanning habits; the ultrasound image visualization guidance method allows physicians to intuitively observe the location of acupoints and their relationship with surrounding tissues, reducing the learning cost of operation, and is especially suitable for inexperienced primary care physicians; the image quality adaptation mechanism can ensure positioning reliability through threshold adjustment in both high-quality and poor-quality image scenarios, and can adapt to ultrasound equipment of different quality without additional equipment upgrades.

[0021] Regarding model sustainability, the online model update step of this invention can continuously optimize model performance based on newly collected clinical data, enabling the model to adapt to new ultrasound equipment parameters and new patient group characteristics, and preventing the model from becoming outdated due to changes in clinical data. The regular positioning result verification and model retraining mechanism can correct model deviations in a timely manner, ensuring the stability of positioning performance during long-term use, and providing continuous support for the standardization and precision development of acupuncture treatment. Attached Figure Description

[0022] Figure 1 This is a schematic block diagram of an AI-based localization method for the Tianchuang acupoint that combines keyframe processing of ultrasound images, as proposed in this invention. Figure 2 A comparison chart of the core performance indicators of different positioning methods; Figure 3 This is a graph showing the variation in localization error for different patient body types. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0025] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0026] Reference Figures 1 to 3 A method for AI-based localization of the Tianchuang acupoint by combining keyframe processing of ultrasound images, comprising the following steps: Data acquisition steps: Acquire ultrasound image data of the Tianchuang acupoint area. The data format includes static ultrasound contrast images and dynamic ultrasound contrast videos. The static images are single-focused grayscale images of the Tianchuang acupoint area, and the dynamic videos are continuous scanning ultrasound contrast streams of the Tianchuang acupoint area. During the acquisition process, keep the pressure between the ultrasound probe and the skin stable at 20-30 kPa, and the probe movement speed should not exceed 5 mm / s to reduce image blurring and artifacts. Data preprocessing steps: Standardize the acquired ultrasound image data, adjust static images and dynamic video frames to 760×600 pixel grayscale images, use Gaussian filtering with 5×5 convolution kernel for noise reduction, enhance image texture features through adaptive histogram equalization, normalize pixel values ​​to make the values ​​range between 0 and 1, and remove invalid data with grayscale standard deviation less than 5 and format that does not conform to PNG / JPEG standards. Keyframe extraction steps: Keyframe screening is performed on the dynamic ultrasound contrast video. Candidate frames are initially extracted at a frequency of 3 frames per second. The gray value difference between the candidate frame and the adjacent frame is calculated by the Sobel operator. The gray value difference threshold is set to 15. Frames with a difference value greater than 15 are retained as keyframes, covering the complete dynamic change of the Tianchuang acupoint area from probe contact to stable scanning. Model training steps: A deep learning model based on a convolutional neural network is constructed, using ResNet-50+FPN as the backbone network. A region proposal network is configured to generate 9 anchor boxes, with 3 scales (8×8, 16×16, 32×32) and 3 aspect ratios (1:1, 1:2, 2:1). ROI alignment unifies the proposed regions of different sizes to 224×224 pixels. An ensemble learning strategy is used to fuse the prediction results of Faster R-CNN, YOLO, and SSD models. The fusion method is a weighted voting method, with weights allocated proportionally to the accuracy of each model on the validation set. During training, CIoULoss is used as the bounding box regression loss, FocalLoss as the classification loss, and binary cross-entropy as the confidence loss. The AdamW optimizer is used, with an initial learning rate of [missing information]. Cosine AnnealingLR was used for learning rate scheduling, with a batch size of 16, 300 training epochs, and a weight decay coefficient of [value missing]. The maximum gradient clipping value was set to 5.0. The preprocessed image data was divided into training, testing, and validation sets in a 3:1:1 ratio. ImageNet pre-trained weights were loaded for model training. Every 10 rounds, the mean localization error and mean confidence score were calculated on the validation set, and the anchor box scale and learning rate were adjusted based on the results. AI localization inference steps: The preprocessed static image and extracted keyframes were input into the trained ensemble model. The model outputs a bounding box representing the location of the Tianchuang acupoint. The bounding box format is... The coordinates of the center point are calculated by the intersection of the diagonals of the bounding box. It also outputs the confidence score of the model for the prediction results, with the confidence score ranging from 0 to 1, and the confidence threshold set at 0.7. Results output and visualization steps: The bounding box, center point coordinates and confidence score of the Tianchuang acupoint are displayed in real time. When the confidence score is lower than 0.7, a text prompt "Location result is unreliable, please rescan" is issued, and no positioning mark is marked on the ultrasound image. When the confidence score is not lower than 0.7, a bounding box is drawn on the ultrasound image with a red solid line, the center point is marked with a blue dot, and the confidence score is displayed in white text in the upper right corner of the bounding box, realizing the visualization guidance of the positioning process.

[0027] This invention also includes a keyframe optimization and screening step, which calculates the comprehensive quality score of candidate frames using a frame quality evaluation formula, and selects the top 30% of frames from highest to lowest score as the optimal keyframes. The formula is as follows: in This represents the overall quality score of the candidate frames. This represents the sharpness weighting coefficient, with a value of 0.7. This represents the inter-frame difference weighting coefficient, with a value of 0.3, and ; This represents the image height, with a value of 760 pixels. Indicates the image width, with a value of 600 pixels; Indicates the image in OK The gradient values ​​of the columns are calculated using the Sobel operator with a 3×3 convolution kernel; This indicates the number of adjacent frames to the candidate frame, with a value of 3, which is the frame before the current candidate frame and the two frames after it. Indicates the candidate frame and the first The difference in grayscale values ​​between adjacent frames is calculated by summing the absolute values ​​of the differences in grayscale values ​​of corresponding pixels in the two frames. This represents the average difference between three adjacent frames. This represents the maximum difference among three adjacent frames. This formula is used to filter out keyframes with high clarity that reflect dynamic changes, reducing the interference of blurry frames on localization inference. This invention also includes a localization error correction step, which compensates for errors in the bounding box coordinates output by the model. The localization results are adjusted based on the individual patient's body shape characteristics. The compensation formula is as follows: in , These represent the corrected x and y coordinates of the center point, in pixels; , Represents the original center point coordinates of the model output, in pixels; This represents the width direction error correction factor, with a value of 0.2. This represents the height direction error correction factor, with a value of 0.25; The average width of the Tianchuang acupoint region in patients with standard body type is 40 mm, which was obtained through statistical analysis of ultrasound data from 100 healthy adults with a height of 165-175 cm and a weight of 60-70 kg. The width of the Tianchuang acupoint area of ​​the current patient is measured by ultrasound imaging, and the distance between the muscles and fascia on both sides of the Tianchuang acupoint is measured in mm. Indicates the image width, with a value of 600 pixels; This represents the mean height of the Tianchuang acupoint region in patients with standard body type, taken as 35 mm. The statistical sample and... Consistent; The height of the Tianchuang acupoint area of ​​the current patient is determined by measuring the distance between the blood vessels at the upper and lower ends of the Tianchuang acupoint using ultrasound imaging, and the unit is mm. This represents the image height, with a value of 760 pixels. This correction method can adapt to individual differences among patients of different body types, making the positioning results more consistent with the patient's actual anatomical location.

[0028] This invention also includes a data augmentation step, which augments the preprocessed image data before model training by generating new training samples through random flipping, rotation, scaling, and brightness adjustment. Random flipping uses a 50% probability of horizontal flipping and a 30% probability of vertical flipping, with horizontal flipping centered on the vertical midline of the image and vertical flipping centered on the horizontal midline of the image. The rotation angle is randomly selected from -15° to 15° in a uniform distribution, and bilinear interpolation is used to fill the image edges after rotation. The scaling ratio is normally distributed between 0.8 and 1.2, with a mean of 1.0 and a standard deviation of 0.1, and the image is cropped to 760×600 pixels after scaling. Brightness adjustment is achieved through Gamma correction, with a Gamma value ranging from 0.7 to 1.3; a Gamma value less than 1 increases brightness, and a value greater than 1 decreases brightness. Each data augmentation generates three times the number of new samples as the original samples, increasing the diversity of training samples, reducing the risk of overfitting the model under a single image feature, and improving the model's adaptability to ultrasound images under different scanning angles and lighting conditions.

[0029] This invention also includes a model ensemble weight dynamic adjustment step, which calculates the localization accuracy of each model in the training set for different types of ultrasound images, and assigns dynamic weights based on the accuracy. The weight allocation formula is as follows: in Indicates the first The weights of the model The values ​​are 1, 2, and 3, corresponding to the Faster R-CNN, YOLO, and SSD models, respectively. Indicates the first The positioning accuracy of this model was tested, and the localization error was less than 2%. The weights are calculated as the proportion of the number of samples to the total number of samples. For example, when Faster R-CNN has an accuracy of 92% on the test set, YOLO has an accuracy of 88%, and SSD has an accuracy of 85%, the weights for the three are 0.35, 0.33, and 0.32, respectively. The dynamic weights are updated in real time after each training round based on the accuracy of the validation set and applied to the ensemble learning process, so that the model ensemble results are more in line with the characteristics of different image data and reduce the prediction error of a single model under a specific image type.

[0030] This invention also includes a real-time optimization step to accelerate the model inference process. The TensorRT tool is used to convert model parameters from 32-bit floating-point numbers to 16-bit floating-point numbers, while retaining the 32-bit floating-point output layer during quantization to ensure positioning accuracy. Simultaneously, the inference engine is optimized by packaging three adjacent keyframes into a batch for inference calculations. During batch inference, the calculation results of the feature extraction layer are shared, reducing redundant computations. Through optimization, the positioning inference time for a single frame image is controlled within 80-95 milliseconds, meeting the clinical real-time guidance requirement of less than 100 milliseconds of latency, ensuring that physicians can obtain positioning results instantly during operation.

[0031] This invention also includes a confidence threshold adaptive adjustment step, which dynamically sets the confidence threshold based on the quality level of the ultrasound image; the image quality level is calculated using the image signal-to-noise ratio (SNR). The signal region is a 20×20 pixel area around the manually marked skylight hole, and the noise region is a 10×10 pixel unorganized area at the image edge. An image with a signal-to-noise ratio (SNR) greater than 30 dB is considered a high-quality image, with a corresponding confidence threshold of 0.6; 25-30 dB is considered a good image, with a threshold of 0.7; 20-25 dB is considered a fair image, with a threshold of 0.75; and less than 20 dB is considered a poor image, with a threshold of 0.8. Through adaptive threshold adjustment, the threshold is lowered when the image quality is good to reduce missed detections, and the threshold is raised when the image quality is poor to reduce false detections, thus adapting to image data under different scanning conditions.

[0032] This invention also includes a multimodal feature fusion step, which introduces the anatomical structure features of the tissues surrounding the Tianchuang acupoint during model training; marks the locations of key tissues such as blood vessels and nerves using medical image annotation tools to generate anatomical structure annotation maps; extracts features from the anatomical structure annotation maps using a ResNet-18 model, outputting a 256-dimensional feature vector; performs a stitching operation on the 256-dimensional deep image features output by the backbone network (ResNet-50+FPN) to form a 512-dimensional fused feature; inputs the fused feature into the subsequent region proposal network and classification regression layer, enabling the model to combine image texture features and anatomical structure features simultaneously when locating the Tianchuang acupoint, automatically avoiding the marked blood vessel and nerve areas, reducing the risk of acupuncture operation damaging key tissues, and improving the safety of localization.

[0033] This invention also includes an online model update step, which collects new ultrasound imaging data and positioning results feedback during clinical application. The new data must include basic patient information (age, gender, height, weight), ultrasound scanning parameters (probe frequency, gain), and manually corrected positioning coordinates by the physician. The preset update cycle is 3 months, with at least 50 new data points collected in each cycle. After preprocessing the new data, it is mixed with the original training set data at a 1:1 ratio for incremental training. The batch size for incremental training is set to 8, the number of training rounds is 50, and the initial learning rate is set to... The cosine annealing learning rate scheduling is adopted; the model's historical best weights are retained during training, and the mean localization error and mean confidence score are evaluated through the validation set; when the mean localization error of the incrementally trained model decreases by more than 5% or the mean confidence score increases by more than 5%, the model is updated and deployed; otherwise, the original model is retained, and the model is adapted to new clinical data features through online updates, thereby continuously improving localization accuracy and generalization ability.

[0034] This invention also includes a positioning result verification step, which verifies the difference between the positioning result output by the Euclidean distance calculation model and the result manually annotated by the physician, using the Euclidean distance... ,in The coordinates of the center point after model correction. The center point coordinates, jointly marked by three physicians with the title of associate chief physician or above, are considered qualified if the difference is less than 2 mm. Simultaneously, the average accuracy of the positioning results is calculated, based on the intersection-union ratio (IUR) of the predicted bounding box and the true bounding box. An intersection-union ratio (IUU) greater than 0.7 is considered an exact match. The pass rate and exact match rate are calculated monthly. Precise matching rate When the pass rate is below 90% or the accurate matching rate is below 85%, the model retraining process is triggered. During retraining, all historical data and newly collected data are used to adjust the basic size of the anchor box and the weight of the loss function to ensure the long-term stability and reliability of the localization method.

[0035] The following two examples further illustrate the specific implementation of this system: Example 1: Location of the Tianchuang acupoint in the acupuncture clinic of a tertiary hospital This embodiment was applied to the acupuncture outpatient department of a tertiary hospital, which sees an average of 30-40 acupuncture patients per day. Among them, about 15% of the cases require location operation of the Tianchuang acupoint. The patients cover different genders, ages 18-75 years and body weights 45-90kg. Clinically, it is necessary to balance positioning accuracy, real-time guidance and operation safety. The ultrasound equipment is a desktop high-definition ultrasound machine that supports static image and dynamic video acquisition. The method of this invention has passed the hospital's ethics review, and all patients have signed informed consent forms.

[0036] I. Data Collection Steps The patient sits upright, slowly turning their head to the opposite side to expose the Tianchuang acupoint area on the neck (posterior border of the sternocleidomastoid muscle, behind the Futu acupoint, level with the Adam's apple). A linear array ultrasound probe is used, with a frequency set to 7-10MHz. After applying medical ultrasound coupling gel, the pressure between the probe and the skin is maintained at a stable 20-30kPa, monitored in real time by the built-in pressure sensor of the ultrasound machine. The probe movement speed is controlled at 3-5mm / s to avoid artifacts caused by excessive movement. The acquired data includes two types: static ultrasound contrast images, with three focused images taken from different angles for each patient: frontal, left 45°, and right 45°; and dynamic ultrasound contrast video, each video segment 10 seconds long, with a frame rate of 25 frames / second, covering the complete scanning process of the probe slowly moving from above to below the Tianchuang acupoint. The video is stored in MP4 format, and the static images are stored in PNG format.

[0037] II. Data Preprocessing Steps The acquired image data underwent standardization processing: First, static images and dynamic video frames were uniformly adjusted to 760×600 pixel grayscale images using OpenCV. The grayscale conversion employed a weighted average method, with grayscale value = 0.299×R + 0.587×G + 0.114×B. Noise denoising was achieved using a Gaussian filter with a 5×5 convolution kernel and a Gaussian standard deviation of 1.2, smoothing image noise through convolution operations. Texture features were enhanced using Adaptive Histogram Equalization (CLAHE), with a clipping limit of 2.0 and a grid size of 8×8, improving the contrast between the acupoint area and surrounding tissues. Pixel values ​​were normalized using the formula: Normalized pixel value = (Original pixel value - Minimum pixel value) / (Maximum pixel value - Minimum pixel value), ensuring the value range was between 0 and 1. Finally, invalid data was filtered using the grayscale standard deviation. The grayscale standard deviation of each frame was calculated, and blurry images with a standard deviation less than 5 were removed. Simultaneously, file formats were checked to exclude abnormal files not in PNG / JPEG formats.

[0038] III. Keyframe Extraction Steps Candidate frames were initially extracted from a 10-second dynamic video (250 frames in total) at a frequency of 3 frames per second, resulting in 30 candidate frames (frames 1, 8, 15, 22…245). The grayscale difference between the candidate frames and adjacent frames was calculated using a Sobel operator with a 3×3 convolution kernel. The Sobel operator is divided into horizontal (Gx) and vertical (Gy) directions, calculating the gradient value for each pixel separately. Then, the absolute value of the gradient difference between corresponding pixels in the two frames is taken and summed to obtain the grayscale difference D. The grayscale difference threshold is set to 15, and candidate frames with D>15 are retained, resulting in 22 keyframes. These frames cover the complete dynamic changes of the probe from contacting the skin, stabilizing the scan, to leaving the skin, ensuring that the features of the Tianchuang acupoint area under different scanning angles are included.

[0039] IV. Model Training Steps A deep learning model based on convolutional neural networks was constructed: the backbone network used ResNet-50+FPN, where ResNet-50 contains 50 convolutional layers and FPN improves the detection capability of small objects through multi-scale feature fusion; a Region Proposal Network (RPN) was configured to generate 9 anchor boxes with scales of 8×8, 16×16, and 32×32 and aspect ratios of 1:1, 1:2, and 2:1, with the basic anchor size set to [8, 16, 32]; ROI alignment was used to unify the proposed regions of different sizes to 224×224 pixels to avoid feature shift caused by interpolation.

[0040] An ensemble learning strategy was employed to fuse three models: Faster R-CNN, YOLO, and SSD. First, the localization accuracy of each model was tested on the validation set. Faster R-CNN achieved 92% accuracy, YOLO 88%, and SSD 85%. The accuracy was then determined using a weight allocation formula. Calculate the weights, where =92%, =88%, =85%, therefore =92 / (92+88+85)=0.35、 =88 / 265=0.33、 =85 / 265=0.32, the fusion method is weighted voting, the bounding box prediction results of the three models are weighted according to this weight, and the bounding box with the highest weight is taken as the final output.

[0041] Training parameter settings: The loss function uses CIoULoss (bounding box regression loss) + FocalLoss (classification loss) + binary cross-entropy (confidence loss), where the focusing parameters of FocalLoss are α=0.25 and γ=2.0; the optimizer is AdamW, and the initial learning rate is set to... The weight decay coefficient is Cosine AnnealingLR learning rate scheduling is used. =300 (number of training rounds) The batch size was set to 16, the number of training epochs to 300, and the maximum gradient clipping value to 5.0, implemented using `torch.nn.utils.clip_grad_norm_`. The preprocessed image data was divided into a 3:1:1 ratio for training (600 examples), a test (200 examples), and a validation (200 examples). ImageNet pre-trained weights were loaded to initialize the model. Every 10 epochs, the mean localization error and mean confidence score were calculated on the validation set. When the mean error on the validation set increased in the 50th epoch, the anchor box scale was adjusted to [10, 20, 35]. After the 100th epoch, the learning rate was halved to ensure model convergence.

[0042] V. AI Localization and Reasoning Steps The ensemble model was trained by inputting three preprocessed still images and 22 keyframes: the model first extracts deep features of the images using ResNet-50+FPN, RPN generates proposed regions, ROIs are aligned to a uniform size, and the classification and regression layers output bounding boxes. , For example, the bounding box of a patient's static image is (200, 180, 350, 320); the coordinates of the center point are calculated from the intersection of the diagonals of the bounding box using the formula: , Simultaneously, the confidence score is output. The confidence score of the patient's static image is 0.85, and the average confidence score of the keyframe is 0.82, both of which are higher than the set threshold of 0.7.

[0043] VI. Results Output and Visualization Steps The localization results are displayed in real time on the 19-inch screen built into the ultrasound machine: a red solid line (2px wide) is used to draw the bounding box, and a blue dot (5px in diameter) is used to mark the center point. The confidence score is displayed in white Arial font (size 14) in the upper right corner of the bounding box. When the confidence score of a key frame is 0.68 (below 0.7), a black background text message "Localization results are unreliable, please rescan" pops up at the bottom of the screen, and no localization mark is marked on the image of that frame. The physician adjusts the acupuncture needle insertion angle according to the stable localization mark, and the screen displays the relative position of the needle insertion path and the acupoint simultaneously to achieve visual guidance.

[0044] VII. Additional Steps 1. Keyframe Optimization and Filtering Through frame quality evaluation formula Calculate 22 candidate frames Value, of which =760、 =600, =0.7、 =0.3, =3; Select a candidate frame, its The sum is 12500, so the sharpness value is 0.7 × (12500 / (760 × 600)) = 0.7 × 0.0274 = 0.0192; They are 22, 19, and 25 respectively. =22, =25, so the difference term is 0.3×(1-(|22-22|+|19-22|+|25-22|) / (3×25))=0.3×(1-6 / 75)=0.3×0.92=0.276; total =0.0192 + 0.276 = 0.2952. (According to...) The top 30% (7 frames) of values ​​from high to low are selected as the optimal keyframes for subsequent inference, reducing interference from blurry frames.

[0045] 2. Positioning error correction For an obese patient weighing 85kg ( ), calculated using the error correction formula: =40mm =600, =0.2, therefore =275+0.2×(40-45) / 40×600=275+0.2×(-5) / 40×600=275-15=260; =35mm =760, =0.25, therefore =250+0.25×(35-38) / 35×760=250+0.25×(-3) / 35×760≈250-16.29=233.71, the corrected center point coordinates are closer to the patient's actual anatomical position.

[0046] 3. Verification of positioning results The difference between the model output and the annotations of three associate chief physicians was calculated: the physicians' annotation center point was (262, 235), and the corrected model output was (260, 233.71). The Euclidean distance = √[(260-262)²+(233.71-235)²] = √[4+1.66] = √5.66≈2.38mm (the error is slightly higher than 2mm due to the patient's obesity, but it is still considered acceptable). The intersection-union ratio of the predicted bounding box and the actual bounding box = (120×130) / (150×140) = 15600 / 21000≈0.74, which is greater than 0.7, and is judged as an exact match.

[0047] VIII. Data Representation and Interpretation Table 1: Comparison of the positioning effects of different positioning methods for the Tianchuang acupoint

[0048] Table 1 shows that traditional palpation methods, due to their high subjectivity, have an average error of 3.8 mm and are the most time-consuming, requiring physicians to repeatedly confirm acupoint locations, resulting in a satisfaction rate of only 65%. While a single YOLO model improves accuracy, it does not integrate the advantages of multiple models, still resulting in an error of 2.5 mm and a complication rate of 2.1%. This invention, through ensemble learning and error correction, reduces the average error to 1.8 mm, with a complication rate of only 0.5%. Due to real-time guidance and visualization functions, the operation time is shortened to 1.2 minutes, and physician satisfaction increases to 96%. The data demonstrates that this invention significantly outperforms traditional methods and single models in terms of accuracy, safety, and efficiency, making it more suitable for the high-volume, high-requirement clinical scenarios of tertiary hospitals.

[0049] Example 2: Location scenario of Tianchuang acupoint in a primary community health service center This embodiment is applied to a traditional Chinese medicine clinic in a community health service center, which sees an average of 10-15 acupuncture patients per day. The patients are mainly elderly people aged 50-75, many of whom have underlying diseases such as hypertension and diabetes. This places higher demands on the safety of the operation. The ultrasound equipment is a portable ultrasound instrument, weighing about 3kg, which needs to be moved and used in multiple scenarios such as the clinic and ward. The doctors have relatively limited clinical experience and rely on intelligent tools to assist in positioning.

[0050] I. Data Collection Steps The patient lies supine with shoulders elevated 15° and head turned to one side, exposing the Tianchuang acupoint area on the neck. A convex array probe is used with a portable ultrasound device at a frequency of 5-8MHz, suitable for deep tissue imaging. After applying water-soluble ultrasound coupling gel, the contact pressure is controlled by the pressure feedback light on the probe handle (green light indicates 20-30kPa). The probe movement speed is controlled at 2-3mm / s to avoid image blurring due to excessive movement. Data acquisition: Two static images (front and side) are acquired for each case; each dynamic video segment is 8 seconds long (20 frames / second). Videos are stored in MOV format, and static images are in JPEG format. Data is temporarily stored in the ultrasound device's built-in 128GB storage module.

[0051] II. Data Preprocessing Steps Due to the low resolution of portable ultrasound equipment, preprocessing was enhanced with noise reduction and enhancement: after adjusting the image size to 760×600 pixels, Gaussian filtering with a 5×5 convolution kernel (standard deviation 1.5) was used for noise reduction; CLAHE clipping was set to 2.5 to improve the contrast of weak texture areas; normalization was performed as in Example 1; additional median filtering (3×3 kernels) was added to remove salt-and-pepper noise and avoid image graininess caused by equipment performance; invalid data with a grayscale standard deviation less than 4 were removed to ensure the image quality of the input model.

[0052] III. Keyframe Extraction Steps An 8-second dynamic video (160 frames) was used to extract candidate frames at 3 frames per second, for a total of 24 frames (frames 1, 7, 13...157). Gray-scale differences were calculated using the Sobel operator. Due to the small inter-frame differences in portable ultrasound, a threshold of 12 was set, and 18 keyframes with D>12 were retained, covering the scanning process of the Tianchuang acupoint area from superficial to deep layers, ensuring the inclusion of anatomical features such as blood vessels and nerves.

[0053] IV. Model Training Steps Considering the limited amount of data at the grassroots level (only data from 150 patients were collected), transfer learning was adopted: the model weights trained in Example 1 were loaded and fine-tuned on the local dataset; the ensemble learning weights were adjusted to... =0.34、 =0.33、 =0.33, due to improved performance of the SSD model on local data; training epochs reduced to 150, initial learning rate set to The batch size was set to 8 due to the computing power limitations of portable devices; the model was evaluated on the validation set (30 cases) every 5 rounds, and the learning rate was decayed by 0.9 times after the 80th round to ensure that the model did not overfit on small datasets.

[0054] V. AI Localization and Reasoning Steps Inputting two static images and 18 keyframes, the model outputs a bounding box for an elderly patient (180, 160, 320, 300), with center point coordinates (250, 230). The confidence score for the static images is 0.81, and the mean confidence score for the keyframes is 0.78. Because the patient has a history of hypertension, the model identifies the carotid artery 5mm below the Tianchuang acupoint (marked with a red dashed box) through multimodal feature fusion. The inference results simultaneously display the text prompt "Pay attention to the carotid artery below, it is recommended to adjust the needle insertion angle to 15°".

[0055] VI. Results Output and Visualization Steps The portable ultrasound device is equipped with a 10-inch touchscreen. The positioning results are displayed in the same way as in Example 1, but with the addition of a "one-click freeze" function. Doctors can freeze the current positioning marker by clicking the screen to avoid image jitter caused by mobile devices. When the ward light is strong, the screen brightness is automatically increased from 300 cd / m² to 500 cd / m² to ensure that the positioning marker is clearly visible. It supports connection to a tablet via Bluetooth to synchronize the positioning results to the tablet, which is convenient for doctors to view when operating on the other side of the patient.

[0056] VII. Additional Steps 1. Confidence threshold adaptive adjustment The signal-to-noise ratio (SNR) of a patient's ultrasound image was calculated as follows: the signal area was 20×20 pixels around the Tianchuang acupoint (mean grayscale 200), and the noise area was 10×10 pixels at the image edge (standard deviation grayscale 12). The SNR was approximately 200 / 12 ≈ 16.67 dB (poor image). The confidence threshold was automatically adjusted to 0.8. The patient's static image had a confidence level of 0.79. The system prompted "Please adjust the probe angle and rescan." After rescanning, the SNR increased to 22 dB (normal image), the threshold decreased to 0.75, and the confidence level was 0.78, meeting the requirements and displaying the positioning marker.

[0057] 2. Online model updates Every 3 months, 50 new data points are collected (e.g., including 20 elderly patients and 15 obese patients). After preprocessing, these data points are mixed with the existing 150 data points at a 1:1 ratio for incremental training: batch size 8, training epochs 50, initial learning rate... Before incremental training, the average model localization error was 2.2 mm, which decreased to 1.9 mm after training (a reduction of 13.6%). The average confidence score increased from 0.78 to 0.83, meeting the update conditions. The new model weights were imported into the portable ultrasound machine via USB interface. The update process took about 5 minutes and did not affect daily patient care.

[0058] VIII. Data Representation and Interpretation V. AI Localization and Reasoning Steps Inputting two static images and 18 keyframes, the model outputs a bounding box for an elderly patient (180, 160, 320, 300), with center point coordinates (250, 230). The confidence score for the static images is 0.81, and the mean confidence score for the keyframes is 0.78. Because the patient has a history of hypertension, the model identifies the carotid artery 5mm below the Tianchuang acupoint (marked with a red dashed box) through multimodal feature fusion. The inference results simultaneously display the text prompt "Pay attention to the carotid artery below, it is recommended to adjust the needle insertion angle to 15°".

[0059] VI. Results Output and Visualization Steps The portable ultrasound device is equipped with a 10-inch touchscreen. The positioning results are displayed in the same way as in Example 1, but with the addition of a "one-click freeze" function. Doctors can freeze the current positioning marker by clicking the screen to avoid image jitter caused by mobile devices. When the ward light is strong, the screen brightness is automatically increased from 300 cd / m² to 500 cd / m² to ensure that the positioning marker is clearly visible. It supports connection to a tablet via Bluetooth to synchronize the positioning results to the tablet, which is convenient for doctors to view when operating on the other side of the patient.

[0060] VII. Additional Steps 1. Confidence threshold adaptive adjustment The signal-to-noise ratio (SNR) of a patient's ultrasound image was calculated as follows: the signal area was 20×20 pixels around the Tianchuang acupoint (mean grayscale 200), and the noise area was 10×10 pixels at the image edge (standard deviation grayscale 12). The SNR was approximately 200 / 12 ≈ 16.67 dB (poor image). The confidence threshold was automatically adjusted to 0.8. The patient's static image had a confidence level of 0.79. The system prompted "Please adjust the probe angle and rescan." After rescanning, the SNR increased to 22 dB (normal image), the threshold decreased to 0.75, and the confidence level was 0.78, meeting the requirements and displaying the positioning marker.

[0061] 2. Online model updates Every 3 months, 50 new data points are collected (e.g., including 20 elderly patients and 15 obese patients). After preprocessing, these data points are mixed with the existing 150 data points at a 1:1 ratio for incremental training: batch size 8, training epochs 50, initial learning rate... Before incremental training, the average model localization error was 2.2 mm, which decreased to 1.9 mm after training (a reduction of 13.6%). The average confidence score increased from 0.78 to 0.83, meeting the update conditions. The new model weights were imported into the portable ultrasound machine via USB interface. The update process took about 5 minutes and did not affect daily patient care.

[0062] VIII. Data Representation and Interpretation Table 2: Comparison of the positioning performance of the present invention under different ultrasonic devices

[0063] Table 2 shows that the desktop high-definition ultrasound machine has the highest positioning accuracy due to its high resolution, but its multi-scenario adaptability is only 92%, making it unsuitable for small wards. The portable ultrasound machine, while having a slightly lower accuracy (93%), has an inference time controlled at 92 milliseconds and a multi-scenario adaptability of 98%, making it suitable for use in clinics, wards, and home-based care. The older ultrasound machines used in primary care facilities (over 5 years old) still achieve an accuracy of 90% with a mean confidence level of 0.78, meeting clinical needs. The data indicates that this invention can work stably on ultrasound devices with different performance levels, especially adapting to the equipment conditions of primary healthcare institutions, addressing the problem of insufficient positioning experience among primary care physicians, and promoting the standardized application of acupuncture techniques at the primary care level.

[0064] Reference Figure 2The figure clearly illustrates the performance gap between different positioning methods: traditional palpation, due to its strong subjectivity, has the largest positioning error and the longest operation time, with a complication rate as high as 4.2% and the lowest physician satisfaction; while the single model is superior to the traditional method, the positioning error still exceeds 2mm, and the complication rate is not lower than 1.8%; the method of this invention, through ensemble learning and error correction, reduces the average positioning error to 1.8mm, shortens the operation time to 1.2 minutes, reduces the complication rate to only 0.5%, and increases physician satisfaction to 96%. The data differences intuitively demonstrate the comprehensive advantages of this invention in terms of accuracy, efficiency, safety, and user experience, especially in controlling the complication rate and improving satisfaction, highlighting its adaptability to actual clinical needs.

[0065] Reference Figure 3 This figure illustrates the adaptability of this invention to patients of different body types: the positioning error is smallest (1.6mm) in patients with normal body type, gradually increasing with BMI, and only 2.2mm in severely obese patients, with the overall fluctuation range controlled within 0.6mm. Traditional methods often result in errors exceeding 3mm in obese patients, while this invention effectively suppresses the increase in error by combining an individual body type error correction formula with anatomical structural features. Data shows that even in obese patients with thick subcutaneous fat and blurred acupoint landmarks, this invention can still maintain a low error, significantly improving the clinical universality of the positioning method and avoiding uneven treatment effects caused by body type differences.

[0066] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for AI-based localization of the Tianchuang acupoint combining ultrasound image keyframe processing, characterized in that, Includes the following steps: Data acquisition steps: Acquire static ultrasound contrast images and dynamic ultrasound contrast videos of the Tianchuang acupoint area, keeping the ultrasound probe in contact with the skin during acquisition; Data preprocessing steps: The ultrasound image data is standardized and uniformly adjusted to 760×600 pixel grayscale images. Noise is removed by 5×5 convolution kernel Gaussian filtering, texture features are enhanced by CLAHE, pixel values ​​are normalized to the range of 0-1, and invalid data is removed. Keyframe extraction steps: Extract candidate frames from the dynamic video at 3 frames per second, and calculate the difference in grayscale values ​​between frames using the Sobel operator; Model training steps: Construct a deep learning model based on ResNet-50+FPN, configure 9 anchor boxes with different scales and aspect ratios, unify the proposed region size by ROI alignment, and fuse the prediction results of the three models using a weighted voting method; use a specified loss function, AdamW optimizer, and CosineAnnealingLR to schedule the learning rate, and split the dataset in a 3:1:1 ratio. AI localization inference steps: Input the preprocessed static image and keyframes into the integrated model, output bounding boxes and confidence scores, and calculate the center point coordinates through the bounding boxes; Results output and visualization steps: Real-time display of bounding box, center point coordinates, and confidence score. A text prompt is issued when the confidence score is below 0.7; when it is not below 0.7, a bounding box is drawn on the image with a solid red line, the center point is marked with a blue dot, and the confidence score is displayed in white text.

2. The AI ​​localization method for the Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes a keyframe optimization and screening step, which calculates the overall quality score of candidate frames using a frame quality evaluation formula, the formula being: in This represents the overall quality score of the candidate frames. This represents the sharpness weighting coefficient. Indicates the weighting coefficients for inter-frame differences; Indicates the image height. Indicates the image width; Indicates the image in OK The gradient value of the column; Indicates the number of adjacent frames to the candidate frame; Indicates the candidate frame and the first The difference in grayscale values ​​between adjacent frames; This represents the average difference between three adjacent frames. This represents the maximum difference among three adjacent frames.

3. The AI ​​localization method for the Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes a localization error correction step, which compensates for errors in the bounding box coordinates output by the model. The localization results are adjusted based on the individual patient's body shape characteristics. The compensation formula is as follows: in , These represent the corrected center points. coordinates and coordinate; , Represents the original center point coordinates output by the model; This represents the width direction error correction coefficient. This represents the correction factor for height direction error; This represents the average width of the Tianchuang acupoint region in patients with standard body types. This indicates the width of the Tianchuang acupoint area for the current patient; Indicates the image width; This represents the mean height of the Tianchuang acupoint region in patients with standard body type, and the statistical sample is compared with... Consistent; This indicates the current height of the Tianchuang acupoint area in the patient; Indicates the image height.

4. The AI ​​localization method for the Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes a data augmentation step, which expands the preprocessed image data before model training by generating new training samples through random flipping, rotation, scaling, and brightness adjustment. The random flipping uses a 50% probability of horizontal flipping and a 30% probability of vertical flipping, with the horizontal flipping axis being the vertical midline of the image and the vertical flipping axis being the horizontal midline of the image. The rotation angle is randomly selected from -15° to 15° in a uniform distribution, and bilinear interpolation is used to fill the image edges after rotation. The scaling ratio is taken from 0.8 to 1.2 in a normal distribution, with a mean of 1.0 and a standard deviation of 0.

1. Brightness adjustment is completed through Gamma correction. Each data augmentation generates three times the number of new samples as the original samples.

5. The AI ​​localization method for Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes a model ensemble weight dynamic adjustment step, which calculates the localization accuracy of each model for different types of ultrasound images in the training set, and assigns dynamic weights based on the accuracy. The weight allocation formula is as follows: in Indicates the first Weights of the model; Indicates the first The positioning accuracy of this model.

6. The AI ​​localization method for Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes real-time optimization steps to accelerate the model inference process. The TensorRT tool is used to convert the model parameters from 32-bit floating-point numbers to 16-bit floating-point numbers. At the same time, the inference engine is optimized by packaging three adjacent key frames into a batch for inference calculation. Through optimization, the localization and inference time of a single frame image is controlled within 80-95 milliseconds, which meets the clinical real-time guidance requirement of less than 100 milliseconds latency.

7. The AI ​​localization method for Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes a confidence threshold adaptive adjustment step, which dynamically sets the confidence threshold based on the quality level of the ultrasound image; the image quality level is calculated using the image signal-to-noise ratio (SNR). The signal area is a 20×20 pixel area around the manually marked skylight hole, and the noise area is a 10×10 pixel unorganized area at the edge of the image; the signal-to-noise ratio is greater than 30dB for high-quality images; 25-30dB for good images; 20-25dB for average images; and less than 20dB for poor images.

8. The AI ​​localization method for Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes a multimodal feature fusion step, which introduces the anatomical structure features of the tissues around the Tianchuang acupoint during model training; marks the location of key tissues using medical image annotation tools to generate anatomical structure annotation maps; uses the ResNet-18 model to extract features from the anatomical structure annotation maps and outputs a 256-dimensional feature vector; performs a stitching operation on the 256-dimensional deep image features output by the backbone network to form a 512-dimensional fused feature; and inputs the fused feature into the subsequent region proposal network and classification regression layer to automatically avoid the annotated blood vessel and nerve regions.

9. The AI ​​localization method for Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes an online model update step, collecting new ultrasound imaging data and positioning results feedback during clinical application. The new data must include basic patient information, ultrasound scanning parameters, and positioning coordinates manually corrected by the physician. After preprocessing the new data, it is mixed with the original training set data at a 1:1 ratio for incremental training. During training, the model's historical optimal weights are retained, and the mean positioning error and mean confidence score are evaluated through the validation set. When the mean positioning error of the incrementally trained model decreases by more than 5% or the mean confidence score increases by more than 5%, the model is updated and deployed.

10. The AI ​​localization method for the Tianchuang acupoint combining ultrasound image keyframe processing according to claim 1, characterized in that, It also includes a location result verification step, which compares the location results output by the Euclidean distance calculation model with the results manually annotated by the physician, using Euclidean distance... ,in The coordinates of the center point after model correction. The center point coordinates were jointly annotated by three physicians with the title of associate chief physician or above; simultaneously, the average accuracy of the localization results was calculated, with the average accuracy calculated based on the intersection-union ratio (IU) of the predicted bounding box and the true bounding box. The pass rate and exact match rate are calculated monthly. Precise matching rate When the pass rate is below 90% or the exact match rate is below 85%, the model retraining process is triggered, using all historical data and newly collected data during retraining.

Images