Clinical ultrasonic image assisted positioning method

Abstract

The invention provides a clinical ultrasonic image assisted positioning method, which belongs to the technical field of clinical ultrasonic image processing, and comprises the following steps: acquiring an ultrasonic image sequence of a target part of a patient and carrying out image preprocessing; extracting anatomical structure feature points in the ultrasonic image sequence; identifying and classifying the anatomical structure feature points; calculating contribution values of the anatomical structure feature points to target part positioning; constructing a three-dimensional positioning coordinate system; calculating spatial position parameters of the target part in the three-dimensional positioning coordinate system; generating a target part positioning track; acquiring real-time spatial position information of the medical instrument; matching the real-time spatial position information of the medical instrument with the positioning track of the target part; establishing a relative position relationship between the medical instrument and the screened anatomical structure feature points; calculating a position deviation value of the medical instrument; and generating a medical instrument position adjustment instruction for adjustment. The technical problem of low accuracy of medical instrument positioning under ultrasonic guidance in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of clinical ultrasonic image processing, and more particularly, relates to a method for clinically assisting in ultrasonic image positioning. Background Art

[0002] The technology of positioning medical devices under ultrasonic guidance has important applications in clinical minimally invasive surgeries. This technology provides positioning and navigation information for medical devices through real-time ultrasonic imaging. Traditional ultrasonic guidance positioning mainly relies on the doctor's experience judgment. By manually adjusting the position of the ultrasonic probe, cross-sectional and longitudinal-sectional images of the target site are obtained, and the positioning operation of the medical device is completed under the guidance of the images. With the development of computer vision and image processing technologies, image-assisted positioning technologies have gradually been applied to clinical practice. By performing real-time analysis and processing on ultrasonic images, more accurate spatial position information is provided for the positioning of medical devices. Currently, commonly used image-assisted positioning methods include methods based on image registration, methods based on feature point tracking, and methods based on deep learning. The method based on image registration realizes the tracking and positioning of medical devices by establishing the spatial correspondence relationship between ultrasonic images at different times. The method based on feature point tracking estimates the spatial position of the medical device by extracting and matching significant feature points in the image. The method based on deep learning directly learns the position information of the medical device from the ultrasonic image using a neural network.

[0003] However, there are still some problems in the existing image-assisted positioning technologies in practical applications. First, the quality of ultrasonic images is easily affected by various factors, such as tissue movement, beam attenuation, scattering noise, etc. These factors will cause unstable extraction and tracking of feature points in the image. Second, traditional feature point selection methods often only consider the local features of the image, ignoring the spatial and temporal relationships between feature points, which easily leads to randomness and uncertainty in feature point selection. Third, most of the existing positioning methods use a single evaluation index to measure the positioning accuracy, lacking comprehensive consideration of multiple influencing factors during the positioning process. In addition, the movement of medical devices in tissues has non-linear and time-varying characteristics, and traditional position adjustment methods are difficult to achieve fast and accurate position compensation.

[0004] In summary, there is a technical problem of low accuracy in positioning medical devices under ultrasonic guidance in the prior art. Summary of the Invention

[0005] In view of this, the present invention provides a method for clinically assisting in ultrasonic image positioning, which can solve the technical problem of low accuracy in positioning medical devices under ultrasonic guidance in the prior art.

[0006] The present invention is implemented as follows: The present invention provides a method for clinically assisted ultrasound image localization, which includes the following steps: collecting an ultrasound image sequence of a patient's target site; performing image preprocessing on the ultrasound image sequence; extracting anatomical structure feature points from the ultrasound image sequence; using a first machine learning model to identify and classify the anatomical structure feature points; calculating the contribution value of the anatomical structure feature points to the localization of the target site; screening the anatomical structure feature points based on the contribution value; constructing a three-dimensional localization coordinate system based on the screened anatomical structure feature points; calculating the spatial position parameters of the target site in the three-dimensional localization coordinate system; generating a localization trajectory of the target site; obtaining the real-time spatial position information of the medical device; matching the real-time spatial position information of the medical device with the localization trajectory of the target site; establishing the relative position relationship between the medical device and the screened anatomical structure feature points; calculating the position deviation value of the medical device; generating a medical device position adjustment instruction; and performing real-time adjustment on the needle insertion direction and needle insertion depth of the medical device according to the medical device position adjustment instruction.

[0007] Among them, the step of collecting an ultrasound image sequence of a patient's target site is specifically: first, preheating and calibrating the ultrasound probe; then setting the working frequency of the ultrasound probe within the range of 4 MHz to 15 MHz; adjusting the total gain parameter of the ultrasound device between 45 dB and 75 dB; applying a coupling agent to the ultrasound probe; then placing the probe on the surface of the target site; keeping the probe at a 90-degree angle to the skin; slowly moving the probe position for scanning; controlling the scanning speed between 1 cm and 2 cm per second; and setting the image acquisition frame rate at 25 frames to 30 frames per second.

[0008] Among them, the step of performing image preprocessing on the ultrasound image sequence is specifically: using adaptive median filtering for noise suppression; performing histogram equalization to enhance image contrast; using an improved Sobel operator for edge enhancement; performing normalization on the processed image; and the preprocessed image sequence needs to meet a peak signal-to-noise ratio of not less than 35 dB.

[0009] Among them, the contribution value is obtained by calculating through a contribution value equation set. The contribution value equation set includes a feature point spatial distribution equation, a feature point temporal stability equation, and a feature point correlation intensity equation. The inputs of the feature point spatial distribution equation include the three-dimensional coordinate values of the anatomical structure feature points, the Euclidean distance values between the anatomical structure feature points and the target site, the gray gradient values in the areas around the anatomical structure feature points, and the anatomical layer depth values where the anatomical structure feature points are located.

[0010] Among them, the inputs of the characteristic point temporal stability equation include the displacement amount of the anatomical structure characteristic points between adjacent ultrasound images, the change amplitude of the gray value of the anatomical structure characteristic points, the shape change rate of the anatomical structure characteristic points, and the contrast change value of the anatomical structure characteristic points.

[0011] Among them, the inputs of the characteristic point correlation intensity equation include the distance matrix value between the anatomical structure characteristic points, the gray similarity value between the anatomical structure characteristic points, the direction consistency value between the anatomical structure characteristic points, and the morphological similarity value between the anatomical structure characteristic points.

[0012] Among them, the step of matching the real-time spatial position information of the medical device with the target site positioning trajectory is specifically: using the iterative closest point algorithm to calculate the registration matrix between the real-time position and the expected trajectory; calculating the motion trend of the medical device; performing predictive compensation on the trajectory matching result based on the motion trend; calculating the matching accuracy evaluation index.

[0013] Among them, the steps of generating the medical device position adjustment instruction include: constructing a motion compensation matrix and an attitude adjustment matrix; generating an adjustment instruction based on the motion compensation matrix and the attitude adjustment matrix; evaluating the system stability using a feedback stability evaluation mathematical model; constructing a real-time optimization objective function.

[0014] Among them, the steps of real-time adjustment of the needle insertion direction and depth of the medical device are specifically: converting the position adjustment instruction into a motor control signal; controlling the needle insertion direction of the medical device, with an adjustment range of plus or minus 30 degrees; controlling the needle insertion depth of the medical device, with an adjustment accuracy of 0.1 mm; monitoring the adjustment effect in real time, and when the position deviation value is less than 0.5 mm and remains stable for more than 1 second, the position adjustment process is completed.

[0015] Among them, the spatial position parameters include a depth position value, a lateral position offset value, and a longitudinal position offset value. The uncertainty of the depth position value is less than 1 mm, and the uncertainties of the lateral position offset value and the longitudinal position offset value are less than 0.5 mm.

[0016] Compared with the prior art, a clinical ultrasound image assisted positioning method provided by the present invention. The clinical ultrasound image assisted positioning method proposed by the present invention constructs a characteristic point contribution value equation set, comprehensively evaluates the anatomical structure characteristic points from three dimensions of spatial distribution, temporal stability, and correlation intensity, and realizes reliable screening and precise positioning of the characteristic points. This method adopts an adaptive image preprocessing strategy, effectively improving the quality of ultrasound images and the stability of characteristic point extraction. By establishing the relative position relationship between the characteristic points and the medical device, and combining a real-time optimized position adjustment strategy, the accuracy and robustness of the medical device positioning process are ensured.

[0017] The method of the present invention has significant advantages in solving the problems of the prior art: First, by introducing a contribution value equation system, the multi-dimensional characteristics of feature points are systematically considered, avoiding the randomness and uncertainty of feature point selection in traditional methods; Second, a position matching strategy based on prediction compensation is adopted, effectively overcoming the influence of tissue movement and image noise on the positioning accuracy; Finally, by constructing a feedback stability evaluation model, real-time and precise adjustment of the position of the medical device is achieved.

[0018] The present invention solves the problems of the accuracy and stability of the positioning of medical devices under ultrasound guidance, which are mainly reflected in the following aspects: A scientific feature point evaluation system is established to ensure the reliability of the positioning reference points; The precise tracking and prediction compensation of the position of the medical device are realized, improving the real-time performance of the positioning process; A complete position adjustment feedback mechanism is constructed to ensure the stability of the positioning result; The technical problem of low accuracy in the positioning of medical devices under ultrasound guidance in the prior art is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] In order to make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] As Figure 1 shown, it is a flowchart of a clinical ultrasound image assisted positioning method provided by the present invention, and this method includes the following steps:

[0022] S01. Collect an ultrasound image sequence of the target part of the patient, and the ultrasound image sequence includes transverse section ultrasound images and longitudinal section ultrasound images;

[0023] S02. Perform image preprocessing on the ultrasound image sequence, and the image preprocessing includes ultrasound image noise reduction, ultrasound image contrast enhancement, and ultrasound image edge enhancement;

[0024] S03. Extract anatomical structure feature points from the ultrasound image sequence and establish an anatomical structure feature point database;

[0025] S04. Use a first machine learning model to identify and classify the anatomical structure feature points to obtain an anatomical structure feature point classification result;

[0026] S05. Calculate the contribution value of the anatomical structure feature points to the positioning of the target part by using a contribution value equation system, and the contribution value is used to characterize the importance of each anatomical structure feature point to the positioning of the target part;

[0027] S06. Screen the anatomical structure feature points based on the contribution value, and mark the anatomical structure feature points with the contribution value greater than the first preset threshold as key positioning feature points;

[0028] S07. Construct a three-dimensional positioning coordinate system based on the key positioning feature points, and map the target part into the three-dimensional positioning coordinate system;

[0029] S08. Calculate the spatial position parameters of the target part in the three-dimensional positioning coordinate system, where the spatial position parameters include a depth position value, a lateral position offset value, and a longitudinal position offset value;

[0030] S09. Generate a target part positioning trajectory according to the spatial position parameters, and the target part positioning trajectory is used to guide a medical device to reach the target part;

[0031] S10. Obtain the real-time spatial position information of the medical device;

[0032] S11. Match the real-time spatial position information of the medical device with the target part positioning trajectory;

[0033] S12. Establish a relative position relationship between the real-time spatial position information of the medical device and the key positioning feature points;

[0034] S13. Calculate the position deviation value between the real-time spatial position information of the medical device and the target part positioning trajectory based on the relative position relationship;

[0035] S14. When the position deviation value is greater than the second preset threshold, generate a medical device position adjustment instruction;

[0036] S15. Adjust the needle insertion direction and needle insertion depth of the medical device in real time according to the medical device position adjustment instruction;

[0037] The contribution value equation set includes a feature point spatial distribution equation, a feature point temporal stability equation, and a feature point correlation intensity equation;

[0038] The feature point spatial distribution equation is used to calculate the spatial distribution weight of the anatomical structure feature points in the three-dimensional positioning coordinate system. The inputs of the feature point spatial distribution equation include the three-dimensional coordinate values of the anatomical structure feature points, the Euclidean distance value between the anatomical structure feature points and the target part, the gray level gradient value of the area around the anatomical structure feature points, and the depth value of the anatomical layer where the anatomical structure feature points are located. The output of the feature point spatial distribution equation is the spatial distribution coefficient of the anatomical structure feature points;

[0039] The feature point temporal stability equation is used to evaluate the stability degree of the anatomical structure feature points in the ultrasound image sequence. The inputs of the feature point temporal stability equation include the displacement amount between adjacent ultrasound images of the anatomical structure feature points, the amplitude of the gray value change of the anatomical structure feature points, the shape change rate of the anatomical structure feature points, and the contrast change value of the anatomical structure feature points. The output of the feature point temporal stability equation is the temporal stability coefficient of the anatomical structure feature points;

[0040] The feature point association intensity equation is used to calculate the association degree between the anatomical structure feature points and adjacent anatomical structure feature points. The inputs of the feature point association intensity equation include the distance matrix value between the anatomical structure feature points, the gray similarity value between the anatomical structure feature points, the direction consistency value between the anatomical structure feature points, and the morphological similarity value between the anatomical structure feature points. The output of the feature point association intensity equation is the association intensity coefficient of the anatomical structure feature points.

[0041] The first machine learning model adopts a convolutional neural network (CNN) structure, specifically a deep network composed of multiple convolutional layers, batch normalization layers, and pooling layers. Each convolutional layer uses a 3×3 convolutional kernel for feature extraction and image enhancement. The batch normalization layer is used to improve the network convergence speed and stability, and the pooling layer is used for downsampling and feature screening. Finally, the classification result is obtained through the fully connected layer and the softmax output layer.

[0042] The second machine learning model adopts a support vector machine (SVM) structure, specifically using a radial basis function (RBF) kernel to achieve non-linear classification. This model first performs PCA dimensionality reduction processing on the input features, and then trains the SVM classifier to obtain the optimal hyperplane and support vectors.

[0043] The steps for establishing the training data set of the first machine learning model are specifically as follows: Randomly extract 1000 images from the original ultrasound image sequence, manually label the anatomical structure feature points and their categories in each image, and construct an initial training sample set. Perform data augmentation on the training sample set, including image rotation, scaling, and translation transformations, and expand the number of training samples to 5000. Finally, divide the training sample set into a training set and a validation set for model training and performance evaluation.

[0044] The steps for establishing the training data set of the second machine learning model are specifically as follows: According to the anatomical structure feature points extracted by the first model, calculate the spatial distribution feature, temporal stability feature, and association intensity feature of each feature point, and form a 3D feature vector. Collect 1000 feature vector samples and label their categories as the SVM training data set. Perform normalization and PCA dimensionality reduction preprocessing on the training data set.

[0045] The steps for training the first machine learning model are specifically as follows: First, initialize the network weights and biases, and use the mini-batch stochastic gradient descent method to optimize the network parameters, with the cross-entropy loss function as the optimization objective. During the training process, an early stopping strategy is adopted, and the training is terminated in advance when the performance on the validation set no longer improves. Finally, a CNN model with the highest classification accuracy is obtained.

[0046] The steps for training the second machine learning model are specifically as follows: First, use the grid search method to optimize the soft margin parameter C and the kernel function parameter γ of the SVM to find the optimal hyperparameter combination. Then, use the training set samples to train the SVM classifier to obtain the optimal hyperplane and support vectors. Finally, evaluate the classification performance of the SVM model on the validation set to ensure the generalization ability.

[0047] The specific implementation manners of the above steps are described in detail below. The specific implementation manner of step S01 is to use an ultrasonic imaging device to scan the target part of the patient. The scanning frequency of the probe of the ultrasonic imaging device is set between 2 megahertz and 15 megahertz, the scanning depth is adjusted according to the anatomical position of the target part, generally set between 2 centimeters and 20 centimeters, and the acquisition frame rate is set between 15 frames per second and 60 frames per second. During the acquisition process, first scan along the transverse direction of the target part to obtain continuous transverse ultrasonic images. After the transverse scanning is completed, then scan along the longitudinal direction of the target part to obtain continuous longitudinal ultrasonic images. The resolution of the acquired ultrasonic image sequence is not less than 1024×768 pixels, and the image bit depth is 8 bits. During the acquisition process, it is necessary to keep the probe in close contact with the skin surface to avoid generating acoustic shadow artifacts, and at the same time maintain an appropriate probe pressure to avoid excessive deformation of the tissue. The purpose of this step is to obtain high-quality original ultrasonic image data, providing a basis for subsequent image processing and feature extraction.

[0048] The specific implementation manner of step S02 is to first perform adaptive median filtering noise reduction processing on the acquired ultrasonic image sequence. The size of the filtering window is dynamically adjusted according to the image noise level, generally set between 3×3 and 7×7 pixels, and a larger filtering window is used for areas with heavier noise. After noise reduction, perform contrast enhancement processing. Use the histogram equalization method to redistribute the image gray values within the range of 0 to 255 to improve the contrast of the image. Then perform edge enhancement processing. Use the improved Sobel operator to calculate the gradient values of the image, and set the gradient threshold to 1.5 times the average gradient of the image. Enhance the edge areas greater than the threshold. Finally, perform image normalization processing to scale the image pixel values to the range of 0 to 1. The function of this step is to improve the quality of the ultrasonic image, highlight the edge features of the target structure, and provide clear image data for subsequent feature point extraction.

[0049] The specific implementation of step S03 is to detect feature points in the preprocessed ultrasonic image sequence using the Scale-Invariant Feature Transform (SIFT) algorithm. First, a Difference of Gaussian (DoG) pyramid is constructed. The number of scale space levels is set to 4, each level contains 5 sub-levels, and the sampling interval ratio is set to 1.5. Local extreme points are detected in the scale space, and the extreme points need to satisfy the local maximum or minimum conditions in both the spatial domain and the scale domain. The detected candidate feature points are screened, removing points with a contrast lower than 30% of the average image contrast and points with a principal curvature ratio greater than 10. For the remaining feature points, feature descriptors are calculated. Using the Histogram of Oriented Gradients (HOG) method, the neighborhood of the feature point is divided into 4×4 sub-regions, and 8-direction gradient histograms are calculated for each sub-region, generating a 128-dimensional feature vector. The extracted feature points and their descriptors are stored in a database, and an index structure is established for fast retrieval. The purpose of this step is to extract stable anatomical structure feature points from ultrasonic images, providing a data basis for subsequent feature point recognition and classification.

[0050] The specific implementation of step S04 is to use a convolutional neural network to identify and classify the extracted anatomical structure feature points. The network structure includes 5 convolutional layers and 3 fully connected layers, and the input is an image patch of 31×31 pixels around the feature point. The first convolutional layer uses 32 5×5 convolutional kernels, the second convolutional layer uses 64 5×5 convolutional kernels, and the third to fifth convolutional layers use 128 3×3 convolutional kernels. After each convolutional layer, a max-pooling layer and a batch normalization layer are connected. The number of neurons in the fully connected layers is 1024, 512, and the number of feature point categories respectively, and the rectified linear unit function is used as the activation function. During training, a stochastic gradient descent optimizer is used, the initial learning rate is set to 0.01, the cosine annealing strategy is used for learning rate adjustment, the batch size is set to 64, and the number of training epochs is 100. The cross-entropy loss function is used to evaluate the model performance, and the ratio of training data to validation data is 4:1. The role of this step is to classify the anatomical structure feature points into different anatomical landmark types, providing semantic information for subsequent localization.

[0051] The specific implementation of step S05 is to calculate the positioning contribution values of each anatomical structure feature point using a contribution value equation system. The feature point spatial distribution equation takes into account the spatial position information of the feature points, calculates the Euclidean distance between the feature points and the target part, and the closer the distance, the greater the weight. The Gaussian function is used to map the distance to the range of 0 to 1, and the standard deviation parameter is set to 0.5 times the diameter of the target part. Calculate the gray gradient value in the area around the feature point. The larger the gradient value, the more significant the feature, and the gradient value is obtained by calculating with the Sobel operator. Consider the depth value of the anatomical layer where the feature point is located, which is obtained from the depth information of the ultrasonic image and normalized. The spatial distribution coefficient is obtained by weighted summation, and the weight coefficient is determined by cross-validation. The feature point temporal stability equation evaluates the stability of the feature points between consecutive frames, calculates the displacement amount between adjacent frames of the feature points, and the feature points with a displacement amount less than 1% of the image size are considered stable. Calculate the change amplitude of the gray value of the feature points. The feature points with a change amplitude less than the threshold show better temporal stability, and the threshold is set to 10% of the average gray value. Calculate the shape change rate of the feature points, which is obtained by analyzing the structure tensor in the neighborhood of the feature points. Calculate the contrast change value of the feature points, and the contrast change is calculated by the standard deviation of the local area. The feature point association strength equation calculates the association relationship between the feature points, constructs a feature point distance matrix, and the distance is calculated using the Euclidean distance. Calculate the gray similarity between the feature points, using the normalized cross-correlation method. Calculate the direction consistency between the feature points, which is obtained by comparing the gradient direction histograms. Calculate the morphological similarity between the feature points, using the shape context descriptor for comparison. The purpose of this step is to quantify the importance of each feature point for the positioning of the target part and provide a basis for feature point screening.

[0052] The specific implementation of step S06 is to screen the anatomical structure feature points according to the calculated contribution values. Arrange the contribution values in descending order, set the first preset threshold to 60% of the maximum contribution value, and retain the feature points with contribution values greater than this threshold as key positioning feature points. Conduct a spatial distribution uniformity analysis on the selected key positioning feature points, calculate the minimum distance between the feature points, and if the distance between some feature points is less than 5% of the image size, only retain the feature point with the largest contribution value. Calculate the spatial distribution entropy of the feature points to ensure sufficient dispersion of the feature points in space. Calculate the reliability score for the retained feature points. The score takes into account the stability coefficient and the association strength coefficient of the feature points, and the weight coefficients are set to 0.6 and 0.4 respectively. The role of this step is to select the most representative and reliable feature points for subsequent positioning calculations.

[0053] The specific implementation of step S07 is to construct a three-dimensional positioning coordinate system based on the selected key positioning feature points. First, calculate the centroid of all key positioning feature points and set the centroid as the coordinate origin. Select three non-collinear feature points farthest from the centroid to construct a reference plane, and use the singular value decomposition method to calculate the normal vector of the reference plane as the z-axis direction. Select two feature points farthest apart on the reference plane to determine the x-axis direction, and determine the y-axis direction through cross product operation, thus establishing a right-handed coordinate system. Calculate the three-dimensional coordinates of each feature point in this coordinate system, and perform coordinate transformation to map the target part into this coordinate system. Optimize the coordinate system, and use the least squares method to adjust the axis directions to minimize the projection error of the feature points. The purpose of this step is to establish a stable spatial reference system and provide a coordinate basis for the positioning of the target part.

[0054] The specific implementation of step S08 is to calculate the spatial position parameters of the target part in the constructed three-dimensional positioning coordinate system. First, calculate the projection value of the target part in the z-axis direction as the depth position value, and the depth position value represents the vertical distance from the target part to the body surface. Calculate the projection position of the target part on the xy plane, the projection value in the x-axis direction as the lateral position offset value, and the projection value in the y-axis direction as the longitudinal position offset value. To improve the positioning accuracy, use the weighted least squares method to optimize the position parameters, and the weights are determined according to the contribution values of each key positioning feature point. Calculate the uncertainty of the position parameters, and use the Monte Carlo method for error propagation analysis to generate a 95% confidence interval. The role of this step is to obtain the accurate spatial coordinates of the target part and provide reference data for subsequent positioning trajectory planning.

[0055] The specific implementation of step S09 is to generate a positioning trajectory for the target part based on the calculated spatial position parameters. Use an improved rapidly-exploring random tree algorithm for trajectory planning. First, randomly sample in three-dimensional space to generate path nodes, and the sampling density decreases as the distance from the target part increases. Perform collision detection on the sampled nodes and remove the nodes that interfere with the known anatomical structures. Use the dynamic programming method to find the optimal path, and the cost function considers three factors: path length, smoothness, and safety distance, and the weight coefficients are set to 0.4, 0.3, and 0.3 respectively. Perform cubic spline interpolation on the obtained path to generate a smooth and continuous positioning trajectory. Calculate the tangent vector and normal vector of each point on the trajectory as a reference for subsequent medical device attitude adjustment. The purpose of this step is to plan a safe and effective operation path to guide the medical device to reach the target part.

[0056] The specific implementation of step S10 is to use an electromagnetic navigation system to obtain the real-time spatial position information of the medical device. The transmitter of the navigation system is fixed on the operating table, and the receiver is installed on the medical device. The Kalman filtering algorithm is used to filter the position information in real time to remove the influence of measurement noise. The measurement frequency is set to 100 Hz to ensure the real-time performance of position tracking. At the same time, the attitude angle information of the medical device is recorded, including yaw angle, pitch angle, and roll angle. A real-time position display interface is established to visually display the position and attitude information of the medical device. The function of this step is to monitor the spatial position of the medical device in real time and provide data support for position matching and adjustment.

[0057] The specific implementation of step S11 is to use the iterative closest point algorithm to match the real-time spatial position information of the medical device with the target site positioning trajectory. First, calculate the shortest distance from the medical device position point to the positioning trajectory to find the corresponding point on the trajectory. Calculate the angle between the movement direction vector of the medical device and the tangent vector of the trajectory to evaluate the consistency of the movement direction. Calculate the deviation between the attitude angle of the medical device and the normal vector of the trajectory to evaluate the matching degree of the attitude. The sliding time window method is used to track the movement trend of the medical device, and the length of the time window is set to 1 second. The purpose of this step is to establish the correspondence between the real-time position of the medical device and the expected trajectory and provide a basis for position adjustment.

[0058] The specific implementation of step S12 is to establish the relative position relationship between the real-time spatial position of the medical device and the key positioning feature points. First, calculate the distance vector from the medical device position point to each key positioning feature point. Construct a relative position constraint matrix, and the matrix elements are the ratios of the distance vectors to the preset safety distance. Calculate the angle between the movement direction of the medical device and the connection direction of the feature points to evaluate the influence of the movement on the feature points. Establish a feature point visibility graph to record the set of feature points that can be observed at the current position. The function of this step is to monitor the spatial relationship between the medical device and the surrounding anatomical structures to ensure operation safety.

[0059] The specific implementation of step S13 is to calculate the position deviation value between the real-time position of the medical device and the positioning trajectory based on the established relative position relationship. The weighted Euclidean distance is used to calculate the position deviation, and the weight coefficients are set according to the importance of different directions. Calculate the attitude deviation, including two components: direction deviation and angle deviation. Combine the position deviation and the attitude deviation to obtain a comprehensive deviation value. When combining, consider the importance of both, and set the weight of the position deviation to 0.7 and the weight of the attitude deviation to 0.3. Consider the influence of the movement speed of the medical device on the deviation, and appropriately increase the weight of the position deviation when the speed is large. Set a deviation smoothing factor to avoid drastic fluctuations in the deviation value. The function of this step is to quantify the deviation degree between the real-time position of the medical device and the expected trajectory and provide a basis for subsequent position adjustment.

[0060] The specific implementation of step S14 is to set the second preset threshold to 15% of the diameter of the target site, and trigger the position adjustment mechanism when the position deviation value exceeds this threshold. The generation of the position adjustment instruction adopts a fuzzy control algorithm. The input variables are the position deviation value and the deviation change rate, and the output variable is the adjustment amount. The fuzzy rule base contains 25 rules, covering various deviation situations. The linguistic variable of the position deviation value is divided into five levels: very small, small, medium, large, and very large. The linguistic variable of the deviation change rate is divided into five levels: large negative, small negative, zero, small positive, and large positive. The centroid method is used for defuzzification to obtain the specific adjustment amount. According to the adjustment amount, the needle insertion direction adjustment instruction and the needle insertion depth adjustment instruction are generated. The adjustment instruction includes the adjustment direction and the adjustment amplitude. The purpose of this step is to generate a suitable adjustment strategy when the medical device deviates from the expected trajectory.

[0061] The specific implementation of step S15 is to perform real-time adjustment of the position of the medical device according to the generated position adjustment instruction of the medical device. The proportional-integral control algorithm is used to achieve smooth adjustment. The proportional coefficient is set to 0.6, and the integral coefficient is set to 0.2. The adjustment of the needle insertion direction adopts a step-by-step adjustment strategy, and the maximum angle of each adjustment does not exceed 5 degrees. The adjustment of the needle insertion depth adopts a progressive method, and the adjustment speed decreases as the distance from the target site decreases. During the adjustment process, the acoustic echo signal of the surrounding tissue is continuously monitored, and the adjustment is immediately stopped when an abnormal echo is detected. The position trajectory during the adjustment process is recorded for subsequent analysis and optimization. The role of this step is to achieve precise adjustment of the position of the medical device and ensure that it reaches the target site along the expected trajectory.

[0062] The specific implementation method of the feature point spatial distribution equation is in the form of weighted summation. The three-dimensional coordinate value weight, Euclidean distance weight, gray level gradient weight, and anatomical hierarchy weight are calculated respectively. The three-dimensional coordinate value weight is normalized, and the coordinate value is mapped to the range from 0 to 1. The Euclidean distance weight is mapped using a negative exponential function, and the closer the distance, the greater the weight. The gray level gradient weight is mapped using a sigmoid function, and the area with a large gradient value obtains a higher weight. The anatomical hierarchy weight considers the importance of the layer where the feature point is located, and the weight of the deep feature point is relatively higher. The final spatial distribution coefficient is obtained through the linear combination of these weights, and the combination coefficient is determined by an optimization algorithm.

[0063] The specific implementation method of the characteristic point time series stability equation is to comprehensively evaluate four indicators: displacement, gray level change, shape change, and contrast change. The displacement is calculated by matching characteristic points between consecutive frames, and the optical flow method is used for characteristic point tracking. The change amplitude of the gray level value is obtained by calculating the change in the average gray level of the neighborhood of the characteristic point. The shape change rate is obtained by calculating the ratio of the eigenvalues of the structure tensor in the neighborhood of the characteristic point. The contrast change value is obtained by calculating the local contrast change in the neighborhood of the characteristic point. After normalizing each indicator, they are weighted and combined to obtain the time series stability coefficient.

[0064] The specific implementation method of the characteristic point correlation intensity equation is to construct a correlation matrix between characteristic points, and the matrix elements are composed of distance similarity, gray level similarity, direction similarity, and morphological similarity. The distance similarity is calculated using a Gaussian kernel function, and the gray level similarity is calculated using the normalized cross-correlation coefficient. The direction similarity is obtained by calculating the Bhattacharyya distance of the gradient direction histogram, and the morphological similarity is obtained by calculating the matching degree of the shape context descriptor. The correlation intensity coefficient is obtained by performing eigenvalue decomposition on the correlation matrix, and the eigenvector component corresponding to the dominant eigenvalue is used as the final correlation intensity measure. The purpose of this step is to evaluate the degree of mutual correlation between characteristic points and provide group information support for the selection of characteristic points.

[0065] Next, the functions or calculation processes involved in the present invention will be described in detail.

[0066] First, the image preprocessing in step S02 involves the following calculation processes:

[0067] The mathematical model of adaptive median filtering is specifically expressed as follows:

[0068] g(x, y) = median{f(x + s, y + t), (s, t) ∈ W};

[0069] In the formula, g(x, y) is the gray level value of the filtered image at point (x, y); f(x, y) is the gray level value of the original image at point (x, y); W is the filtering window; (s, t) is the relative coordinate value within the window.

[0070] The mathematical model of adaptive adjustment of the window size is specifically expressed as follows:

[0071] W size = 3 + 2 × floor(σ local / σ global );

[0072] In the formula, W size is the filtering window size; σ local is the standard deviation of the local area; σ global is the standard deviation of the entire image; floor represents the floor function.

[0073] The mathematical model of histogram equalization is specifically expressed as follows:

[0074]

[0075] In the formula, s k is the output gray level; r k is the input gray level; n j is the number of pixels with gray level j; n is the total number of pixels in the image; the value range of k is from 0 to 255.

[0076] The mathematical model for calculating the gradient by the improved Sobel operator is specifically expressed as follows:

[0077]

[0078] In the formula, G x is the gradient in the x direction; G y is the gradient in the y direction; G is the total gradient; f(x, y) is the image gray function; * represents the convolution operation.

[0079] The mathematical model of image normalization is specifically expressed as follows:

[0080]

[0081] In the formula, f norm (x, y) is the normalized image; f(x, y) is the original image; f min is the minimum gray value of the image; f max is the maximum gray value of the image.

[0082] The feature point extraction process in step S03 involves the following calculations:

[0083] The mathematical model for constructing the Gaussian difference pyramid is specifically expressed as follows:

[0084] L(x, y, σ) = G(x, y, σ) * I(x, y);

[0085] D(x, y, σ) = L(x, y, kσ) - L(x, y, σ);

[0086] In the formula, L(x, y, σ) is the image after Gaussian blur; G(x, y, σ) is the Gaussian kernel function; I(x, y) is the input image; D(x, y, σ) is the Gaussian difference image; k is the adjacent scale ratio factor, with a value of 1.5; σ is the standard deviation of the Gaussian kernel.

[0087] The Gaussian kernel function is specifically expressed as follows:

[0088]

[0089] Wherein, (x, y) are pixel coordinates; σ is the standard deviation of the Gaussian kernel, and the initial value is set to 1.6.

[0090] The mathematical model for calculating the principal curvature ratio of feature points is specifically expressed as follows:

[0091]

[0092]

[0093] Wherein, H is the Hessian matrix; D xx , D xy , D yy are second-order partial derivatives; r is the principal curvature ratio; Tr represents the trace of the matrix; Det represents the determinant of the matrix.

[0094] The mathematical model for calculating the feature descriptor is specifically expressed as follows:

[0095]

[0096] Wherein, m(x, y) is the gradient magnitude; θ(x, y) is the gradient direction; and are the image gradients in the x and y directions respectively.

[0097] The convolutional neural network calculation in step S04 involves the following model:

[0098] The mathematical model of the convolutional layer is specifically expressed as follows:

[0099]

[0100] Wherein, is the weighted input of the j-th feature map in the l-th layer; is the convolutional kernel of the l-th layer; is the input feature map of the (l - 1)-th layer; is the bias term; is the activated feature map; f is the activation function; M l-1 is the number of feature maps in the previous layer.

[0101] The mathematical model of the batch normalization layer is specifically expressed as follows:

[0102]

[0103] Wherein, x (k) is the input feature; μ B is the batch mean; is the batch variance; ∈ is the numerical stability constant, and the value is 10 -5 ; γ and β are learnable parameters.

[0104] The mathematical model of the cross-entropy loss function is specifically expressed as follows:

[0105]

[0106] Where N is the number of samples; C is the number of categories; y ij is the true label; p ij is the predicted probability.

[0107] The specific calculation model of the contribution value equations in step S05 is as follows:

[0108] The mathematical model of the feature point spatial distribution equation is specifically expressed as follows:

[0109]

[0110] Where w spatial is the spatial distribution weight; d is the Euclidean distance from the feature point to the target part; σ d is the distance standard deviation; is the gray gradient value; z is the anatomical depth value; z max is the maximum depth value; α1, α2, α3, α4 are weight coefficients and satisfy

[0111] The mathematical model of the feature point temporal stability equation is specifically expressed as follows:

[0112]

[0113] Where w temporal is the temporal stability weight; Δp is the displacement; Δg is the change amplitude of the gray value; Δs is the shape change rate; Δc is the contrast change value; λ1, λ2, λ3, λ4 are attenuation coefficients; β1, β2, β3, β4 are weight coefficients and satisfy

[0114] The mathematical model of the feature point correlation intensity equation is specifically expressed as follows:

[0115]

[0116] Where w correlation is the correlation intensity matrix; w ij is the correlation intensity between feature points i and j; d ij is the distance between feature points; ρ ij is the gray similarity; θ ij is the direction difference; s ij is the morphological similarity; γ1, γ2, γ3, γ4 are weight coefficients and satisfy

[0117] The calculation model of the final contribution value is specifically expressed as follows:

[0118] C = ω1w spatial + ω2w temporal + ω3w correlation ;

[0119] In the formula, C is the total contribution value; ω1, ω2, ω3 are combined weights and satisfy

[0120] The construction of the coordinate system in step S07 involves the following calculations:

[0121] The mathematical model for calculating the normal vector of the reference plane is specifically expressed as follows:

[0122] V = [v1, v2, v3] = USV T ;

[0123] In the formula, V is the eigenpoint coordinate matrix; U and V T are orthogonal matrices; S is the singular value diagonal matrix; the normal vector takes the right singular vector corresponding to the minimum singular value of V.

[0124] The mathematical model for optimizing the coordinate axes is specifically expressed as follows:

[0125]

[0126] In the formula, E is the optimization objective function; P i is the actual position of the eigenpoint; is the projection position; R i is the actual rotation matrix; is the estimated rotation matrix; μ is the trade-off coefficient; n is the number of eigenpoints.

[0127] The calculation of the spatial position parameters in step S08 involves the following model:

[0128] The mathematical model for weighted least squares optimization is specifically expressed as follows:

[0129]

[0130] In the formula, θ is the parameter vector to be optimized; w i is the eigenpoint weight; y i is the actual observed value; f(x i ; θ) is the predicted value; η is the regularization coefficient; n is the number of eigenpoints.

[0131] The mathematical model for calculating the uncertainty of the position parameters is specifically expressed as follows:

[0132]

[0133] In the formula, is the covariance matrix of position parameters; J is the Jacobian matrix; ∑ is the covariance matrix of observations.

[0134] The trajectory planning in step S09 involves the following calculations:

[0135] The mathematical model of the path cost function is specifically expressed as follows:

[0136]

[0137] In the formula, p i is the path node; is the change in the velocity vector; dist(p i , O) is the minimum distance from the node to the obstacle; κ1, κ2, κ3 are the weight coefficients; m is the number of path nodes.

[0138] The mathematical model of cubic spline interpolation is specifically expressed as follows:

[0139] S i (t) = a i + b i (t - t i ) + c i (t - t i ) 2 + d i (t - t i ) 3 ;

[0140] In the formula, S i (t) is the i-th spline function; t is the parameter variable; a i , b i , c i , d i are the coefficients; t i is the node parameter value.

[0141] The position tracking in step S10 involves the following calculations:

[0142] The mathematical model of Kalman filtering is specifically expressed as follows:

[0143] x k = Ax k-1 + Bu k + w k ;

[0144] z k = Hx k + v k ;

[0145] In the formula, x k is the state vector; z k is the observation vector; A is the state transition matrix; B is the control matrix; H is the observation matrix; uk For the control input; w k and v k are the process noise and the observation noise, respectively.

[0146] The mathematical model for state estimation update is specifically expressed as follows:

[0147]

[0148] In the formula, is the state estimation value; K k is the Kalman gain; P k is the estimation error covariance matrix; I is the identity matrix.

[0149] The position matching in step S11 involves the following calculations:

[0150] The mathematical model of the iterative closest point algorithm is specifically expressed as follows:

[0151]

[0152] In the formula, E(R, t) is the error function; R is the rotation matrix; t is the translation vector; x i is the source point set; y i is the target point set; N is the number of point pairs.

[0153] The mathematical model for motion trend tracking is specifically expressed as follows:

[0154]

[0155] In the formula, v(t) is the current speed estimation; W is the time window length; p i is the position sequence; Δt is the sampling time interval.

[0156] The relative position constraint in step S12 involves the following calculations:

[0157] The mathematical model of the constraint matrix is specifically expressed as follows:

[0158]

[0159] In the formula, M constraint is the constraint matrix; r ij is the relative distance ratio; p i , p j are the position vectors; d safe is the safety distance threshold.

[0160] The mathematical model for visibility graph construction is specifically expressed as follows:

[0161]

[0162] Where, V ij is the visibility matrix element; θ ij is the line-of-sight angle; d ij is the spatial distance; θ threshold is the angle threshold; d threshold is the distance threshold.

[0163] The position deviation calculation in step S13 involves the following calculations:

[0164] The mathematical model of the weighted Euclidean distance is specifically expressed as follows:

[0165]

[0166] Where, D is the weighted distance; w i is the weight in each direction; x i is the actual position component; is the expected position component.

[0167] The mathematical model of the attitude deviation calculation is specifically expressed as follows:

[0168]

[0169] Where, Δθ is the direction deviation angle; is the actual direction vector; is the expected direction vector.

[0170] The mathematical model of the comprehensive deviation value calculation is specifically expressed as follows:

[0171] E total = φ1D + φ2Δθ + φ3v;

[0172] Where, E total is the comprehensive deviation value; D is the position deviation; Δθ is the attitude deviation; v is the motion speed; φ1, φ2, φ3 are weight coefficients and satisfy

[0173] The position adjustment instruction generation in step S14 involves the following calculations:

[0174] The mathematical model of the motion compensation matrix is specifically expressed as follows:

[0175]

[0176] Where, C motion is the motion compensation matrix; Δx i , Δy i , Δz i (i = 1, 2, 3) are the compensation components in each direction; v x , v y , v z are the current speed components; ξx , ξ y , ξ z is a random disturbance term, following a normal distribution with a mean of 0 and a standard deviation of 0.1; Δ pos is the position compensation vector.

[0177] The mathematical model of the attitude adjustment matrix is specifically expressed as follows:

[0178]

[0179] In the formula, R adjust is the attitude adjustment matrix; φ, θ, and ψ are the roll angle, pitch angle, and yaw angle respectively.

[0180] The mathematical model of the adjustment instruction generation is specifically expressed as follows:

[0181]

[0182] In the formula, CMD adjust is the adjustment instruction vector; E is the position error; p, v, and a are the position, velocity, and acceleration vectors respectively; η1, η2, η3, and η4 are control gain coefficients and satisfy t is the integration time.

[0183] The mathematical model of the feedback stability evaluation is specifically expressed as follows:

[0184]

[0185] In the formula, S(t) is the system stability function; λ1, λ2, and λ3 are weight coefficients; α, β, and γ are attenuation coefficients; ω is the oscillation frequency; is the phase angle; t is the time variable.

[0186] The mathematical model of the real-time optimization objective function is specifically expressed as follows:

[0187]

[0188] In the formula, J is the optimization objective function; x k , v k are the position and velocity at the k-th step; are the desired position and velocity; u k is the control input; ρ1, ρ2, and ρ3 are weight coefficients; N is the prediction step number.

[0189] The mathematical model of the constraint condition function is specifically expressed as follows:

[0190]

[0191] In the formula, g(x, u) is the constraint condition vector; x min , xmax is the position constraint; u min , u max is the control input constraint; h(x, u) is the state-control coupling constraint.

[0192] The principle and significance of the above equations are as follows: The motion compensation matrix adopts a 3×3 matrix form, considering the compensation components in three directions, and introducing a random perturbation term to simulate the uncertainty in the actual system; the attitude adjustment matrix is based on the Euler angle representation method, and the precise description of the spatial attitude is realized through the combination of three rotation angles; the adjustment instruction generation adopts the PID control idea, combining the position term, velocity term, acceleration term and integral term of the error to achieve comprehensive error compensation; the feedback stability evaluation function includes a decaying sine term, a decaying cosine term and a pure decay term, which can describe the oscillation and convergence characteristics of the system; the real-time optimization objective function considers the position error, velocity error and control cost, and balances various indicators through weight coefficients; the constraint condition function ensures that the system operates within a safe range, including position constraints, control input constraints and coupling constraints.

[0193] Specifically, the principle of the present invention is: The technical principle of the present invention is based on multi-dimensional feature fusion and real-time feedback control. In terms of feature point evaluation, the distribution characteristics of feature points in three-dimensional space and their geometric relationship with the target part are investigated through the spatial distribution equation, the stability degree of feature points in the image sequence is evaluated by the temporal stability equation, and the topological relationship between feature points is analyzed by the correlation strength equation. The comprehensive evaluation in three dimensions ensures the scientificity and reliability of feature point selection. In terms of position tracking, the iterative closest point algorithm is adopted to achieve the precise matching of the real-time position of the medical device and the expected trajectory, and prediction compensation is carried out through motion trend analysis to overcome the influence of system delay. In terms of position adjustment, a comprehensive control strategy including motion compensation and attitude adjustment is constructed, and the convergence of the adjustment process is ensured through feedback stability evaluation.

[0194] The solution of the present invention has a complete logical chain: First, the reliability of feature point extraction is improved through image preprocessing, then a feature point evaluation system is established based on the contribution value equations, and then a stable three-dimensional positioning coordinate system is constructed. Finally, precise positioning is achieved through real-time position tracking and feedback adjustment. Each technical link adopts corresponding solutions for the specific problems of the existing technology. For example, in order to solve the randomness problem of feature point selection, a multi-dimensional evaluation system is introduced; in order to overcome the delay problem of position tracking, a prediction compensation mechanism is adopted; in order to ensure the stability of position adjustment, a feedback control strategy is constructed. The organic combination of these technical means forms a complete medical device positioning solution.

[0195] Summary of the technical solution of the present invention: After collecting the ultrasonic image sequence, preprocessing is carried out, anatomical structure feature points are extracted and identified and classified by a machine learning model. The contribution value of the feature points to the positioning is calculated by the contribution value equation set and the key positioning feature points are screened. A three-dimensional positioning coordinate system is constructed and the spatial position parameters of the target part are calculated. A positioning trajectory is generated and the real-time position information of the medical device is obtained. Position matching is carried out and the relative position relationship is established. The position deviation is calculated and an adjustment instruction is generated. Finally, the real-time position of the medical device is adjusted to achieve precise positioning.

[0196] A specific Embodiment 1 of the present invention is provided below. The specific implementation manners of each step in this Embodiment 1 are described in detail as follows.

[0197] The specific implementation manner of step S01 is based on the basic principle of medical ultrasonic imaging and includes a series of image acquisition and processing operations: First, the ultrasonic probe needs to be preheated and calibrated. The working frequency of the probe is set within the range of 4 MHz to 15 MHz. The specific frequency is selected according to the anatomical depth of the target part. High frequency is used for superficial tissues and low frequency is used for deep tissues. Secondly, the gain parameters of the ultrasonic device are adjusted. The total gain is controlled between 45 dB and 75 dB, and the time gain compensation curve adopts an S-shaped curve. Then, the ultrasonic probe is smeared with a coupling agent, and the thickness of the coupling agent between the probe and the skin is controlled between 2 mm and 5 mm. Then, the probe is placed on the surface of the target part, and the probe is kept at a 90-degree angle to the skin, and the probe position is slowly moved for scanning. The scanning speed is controlled between 1 cm and 2 cm per second. At the same time, the image acquisition frame rate is set to 25 frames to 30 frames per second, and the storage depth is 12 bits. During the acquisition process, the spatial position information of the probe needs to be recorded simultaneously, including the coordinate values and attitude angle values of the probe in the three-dimensional space. Finally, the collected transverse and longitudinal ultrasonic image sequences are saved as DICOM format files respectively, and each sequence contains 50 frames to 100 frames of image data.

[0198] The specific implementation manner of step S02 is to perform a series of preprocessing operations on the collected ultrasonic image sequence: First, adaptive median filtering is used for noise suppression. The size of the filtering window is adaptively adjusted according to the local noise intensity of the image. The window size range is 3×3 to 11×11 pixels, and the specific calculation adopts the adaptive median filtering mathematical model given above. Then, histogram equalization processing is carried out to enhance the image contrast, and the calculation process follows the histogram equalization mathematical model given above. Then, an improved Sobel operator is used for edge enhancement, and the specific calculation adopts the gradient calculation mathematical model given above. Finally, the processed image is normalized, and the pixel values are mapped to the range of 0 to 1, and the calculation process follows the image normalization mathematical model given above. The preprocessed image sequence needs to meet the quality requirements that the peak signal-to-noise ratio is not less than 35 dB and the structural similarity index is not less than 0.85.

[0199] The specific implementation of step S03 is to extract anatomical structure feature points from the preprocessed ultrasound image sequence: First, construct a Difference of Gaussian (DoG) pyramid, using the DoG pyramid mathematical model given above. The number of pyramid layers is set to 4 to 6 layers. Then, detect local extreme points in each layer of the image, and screen the feature points by calculating the principal curvature ratio. The specific calculation uses the principal curvature ratio mathematical model of feature points given above, and the principal curvature ratio threshold is set to 10. Then, calculate the feature descriptor for the initially selected feature points, using the feature descriptor calculation mathematical model given above, and the dimension of the descriptor is 128. Finally, store the extracted feature points and their descriptor information in the feature point database. The database is organized using a KD-tree structure to improve the efficiency of subsequent feature point matching.

[0200] The specific implementation of step S04 is to use a convolutional neural network to identify and classify feature points: First, construct a deep convolutional neural network. The network structure includes 5 convolutional layers and 3 fully connected layers. The specific calculation uses the convolutional layer mathematical model given above. After each convolutional layer, a batch normalization layer and a ReLU activation function are connected. The calculation of the batch normalization layer uses the batch normalization layer mathematical model given above. The network training uses a cross-entropy loss function, and the specific calculation uses the cross-entropy loss function mathematical model given above. During the training process, the Adam optimizer is used, and the initial learning rate is set to 0.001, and it decays to 0.1 times the original value every 50 epochs. The training dataset contains 10,000 labeled ultrasound images, the validation set contains 2,000 images, and the test set contains 1,000 images. The classification accuracy on the test set after network training needs to reach more than 90%.

[0201] The specific implementation of step S05 is to calculate the contribution value of feature points to the localization of the target part: First, calculate the spatial distribution weight of feature points, using the feature point spatial distribution equation given above, where the distance standard deviation σ d is set to 30 mm, and the weight coefficients α1 to α4 are set to 0.4, 0.3, 0.2, and 0.1 respectively. Then, calculate the temporal stability weight of feature points, using the feature point temporal stability equation given above, where the decay coefficients λ1 to λ4 are set to 0.1, 0.2, 0.3, and 0.4 respectively, and the weight coefficients β1 to β4 are set to 0.35, 0.3, 0.2, and 0.15 respectively. Then, calculate the correlation intensity weight of feature points, using the feature point correlation intensity equation given above, where the correlation intensity standard deviation σ c is set to 20 mm, and the weight coefficients γ1 to γ4 are set to 0.3, 0.3, 0.2, and 0.2 respectively. Finally, calculate the total contribution value of feature points, using the final contribution value calculation model given above, where the combination weights ω1 to ω3 are set to 0.4, 0.3, and 0.3 respectively.

[0202] The specific implementation of step S06 is to screen feature points based on the calculated contribution values: First, normalize the contribution values of all feature points, mapping the contribution values to the range of 0 to 1; then set the first preset threshold to 0.7, and mark the feature points with contribution values greater than this threshold as key positioning feature points; then perform a spatial distribution uniformity analysis on the marked key positioning feature points, calculate the minimum distance between the feature points, and if the minimum distance is less than 5 millimeters, retain the feature point with a larger contribution value; finally, sort the screened key positioning feature points according to the contribution value size, and select the 30 feature points with the largest contribution values as the final set of key positioning feature points.

[0203] The specific implementation of step S07 is to construct a three-dimensional positioning coordinate system: First, calculate the reference plane based on the spatial distribution of the key positioning feature points, and use the reference plane normal vector calculation mathematical model given above; then define the coordinate axis directions on the reference plane, use the coordinate axis optimization mathematical model given above, and set the trade-off coefficient μ in the optimization objective function to 0.5; then project the target part onto the constructed three-dimensional coordinate system and calculate the three-dimensional coordinate values of the target part; finally, perform a stability evaluation on the constructed coordinate system, calculate the orthogonality error between the coordinate axes, and ensure that the orthogonality error is less than 1 degree.

[0204] The specific implementation of step S08 is to calculate the spatial position parameters of the target part: First, use the weighted least squares optimization method to calculate the depth position value of the target part. The specific calculation uses the weighted least squares optimization mathematical model given above, where the regularization coefficient η is set to 0.1; then calculate the lateral position offset value and the longitudinal position offset value of the target part, using the position parameter uncertainty calculation mathematical model given above; then perform an uncertainty analysis on the calculated position parameters to ensure that the uncertainty of the depth position value is less than 1 millimeter, and the uncertainty of the lateral and longitudinal position offset values is less than 0.5 millimeters; finally, convert the calculated spatial position parameters into depth values and angle values relative to the body surface.

[0205] The specific implementation of step S09 is to generate the positioning trajectory of the target part: First, construct a path cost function, using the path cost function mathematical model given above, where the weight coefficients κ1 to κ3 are set to 0.4, 0.3, and 0.3 respectively; then use the A* algorithm for path planning, discretize the path into several path nodes; then use the cubic spline interpolation method to smooth the discrete path nodes. The specific calculation uses the cubic spline interpolation mathematical model given above; finally, generate the motion trajectory for guiding the medical device, and set the trajectory sampling interval to 1 millimeter.

[0206] The specific implementation of step S10 is to obtain the real-time spatial position information of the medical device: First, an optical positioning system or an electromagnetic positioning system is used to track the spatial position of the medical device, and the sampling frequency of the positioning system is set to 60 Hz; Then, the Kalman filter algorithm is used to filter the original position data, and the specific calculation uses the Kalman filter mathematical model given above, where the covariance matrices of the process noise and the observation noise are obtained through experimental calibration; Then, the filtered position data is subjected to coordinate system transformation, and the position data is transformed into the constructed three-dimensional positioning coordinate system; Finally, the motion speed and acceleration information of the medical device are calculated to provide a basis for subsequent trajectory matching.

[0207] The specific implementation of step S11 is to match the real-time position information of the medical device with the target part positioning trajectory: First, the iterative closest point algorithm is used to calculate the registration matrix between the real-time position and the expected trajectory, and the specific calculation uses the iterative closest point algorithm mathematical model given above; Then, the motion trend of the medical device is calculated, and the motion trend tracking mathematical model given above is used, where the time window length W is set to 10; Then, based on the motion trend, the trajectory matching result is predicted and compensated, and the compensation time is set to 100 milliseconds; Finally, the matching accuracy evaluation indexes are calculated, including the average registration error and the maximum registration error, ensuring that the average registration error is less than 0.5 mm.

[0208] The specific implementation of step S12 is to establish the relative position relationship between the medical device and the key positioning feature points: First, a constraint matrix is constructed, and the constraint matrix mathematical model given above is used, where the safety distance threshold d safe is set to 10 mm; Then, a visibility graph is established, and the visibility graph construction mathematical model given above is used, where the angle threshold θ threshold is set to 60 degrees, and the distance threshold d threshold is set to 50 mm; Then, based on the constraint matrix and the visibility graph, the spatial relationship between the medical device and each key positioning feature point is calculated; Finally, a topological relationship network between the feature points is established for subsequent position deviation calculation.

[0209] The specific implementation of step S13 is to calculate the position deviation value of the medical device: First, the weighted Euclidean distance is used to calculate the spatial position deviation, and the specific calculation uses the weighted Euclidean distance mathematical model given above, where the weights w1 to w3 in each direction are set to 0.4, 0.3, and 0.3 respectively; Then, the attitude deviation angle is calculated, and the attitude deviation calculation mathematical model given above is used; Then, the comprehensive deviation value is calculated, and the comprehensive deviation value calculation mathematical model given above is used, where the weight coefficients φ1 to φ4 are set to 0.4, 0.3, 0.2, and 0.1 respectively; Finally, the calculated deviation value is compared with the second preset threshold, and the second preset threshold is set to 1 mm.

[0210] The specific implementation of step S14 is to generate a medical device position adjustment instruction: First, construct a motion compensation matrix using the motion compensation matrix mathematical model given above; then calculate the attitude adjustment matrix using the attitude adjustment matrix mathematical model given above; then generate an adjustment instruction using the adjustment instruction generation mathematical model given above, where the control gain coefficients η1 to η4 are set to 0.4, 0.3, 0.2, and 0.1 respectively; evaluate the system stability using the feedback stability evaluation mathematical model given above; construct a real-time optimization objective function using the real-time optimization objective function mathematical model given above; and finally consider various constraint conditions using the constraint condition function mathematical model given above.

[0211] The specific implementation of step S15 is to execute the medical device position adjustment: First, convert the position adjustment instruction into a motor control signal; then control the needle insertion direction of the medical device, with an adjustment range of ±30 degrees; then control the needle insertion depth of the medical device, with an adjustment accuracy of 0.1 mm; and finally, monitor the adjustment effect in real time. When the position deviation value is less than 0.5 mm and remains stable for more than 1 second, the position adjustment process is completed.

[0212] To better understand and implement the present invention, the following provides an embodiment 2 of a specific application scenario of the present invention:

[0213] The research team of a certain medical unit first collected the imaging data of ultrasonic-guided surgeries for liver and kidney tumors, heart valve diseases, etc. carried out by the medical group over the years, including ultrasonic image sequences in cross-section and sagittal plane, as well as corresponding surgical records and image annotation information. To ensure the extensiveness and representativeness of the data, they selected patients of different genders and age groups, covering various common target organ lesion types. After preliminary statistical analysis, this ultrasonic image database contains more than 100,000 ultrasonic images from more than 2,000 patients.

[0214] Next, the research team preprocesses the collected original ultrasonic image data. First, use the adaptive median filtering method to eliminate the speckle noise in the image, and at the same time adaptively adjust the filtering window size according to the gradient information of the local area to retain the image details to the greatest extent. Then, use the histogram equalization method to enhance the image contrast and highlight the texture features of the region of interest. Finally, apply the improved Sobel operator to detect the image edges and further enhance the contour information of the target organ. After these preprocessing steps, the noise of the original ultrasonic image is effectively suppressed, and at the same time, the image contrast and edge sharpness are also improved, laying a good foundation for subsequent feature extraction and pattern recognition.

[0215] Table 1 Parameter settings of ultrasonic image preprocessing algorithm

[0216] Algorithm Parameter Adaptive median filtering Initial window size: 3×3, Local area standard deviation threshold: 0.2 Histogram equalization - Improved Sobel operator Convolution kernel size: 3×3, Gradient threshold: 0.3

[0217] After the image preprocessing was completed, the research team proceeded to automatically extract and identify the anatomical structure feature points in the ultrasound images. They first constructed a feature point database containing 10 major anatomical structures (such as the liver, kidneys, heart valves, etc.). Then, they used a Convolutional Neural Network (CNN) model to perform semantic segmentation on the ultrasound images to identify and locate these feature points. Specifically, this CNN model consisted of 5 convolutional layers, 3 pooling layers, and 2 fully connected layers. The input was an ultrasound image of 256×256 pixels, and the output was the corresponding probability distribution of the feature point categories. During the training process, the research team used image enhancement techniques to expand the training samples and adopted an early stopping strategy to optimize the network parameters. Finally, they achieved a 95% feature point recognition accuracy on the validation set.

[0218] Table 2 CNN model network structure and parameter settings

[0219] Layer type Input size Output size Parameter Convolutional layer 1 256×256×3 128×128×32 Kernel size: 3×3, Stride: 2, Padding: 1 Pooling layer 1 128×128×32 64×64×32 Kernel size: 2×2, Stride: 2 Convolutional layer 2 64×64×32 32×32×64 Kernel size: 3×3, Stride: 2, Padding: 1 Pooling layer 2 32×32×64 16×16×64 Kernel size: 2×2, Stride: 2 Convolutional layer 3 16×16×64 8×8×128 Kernel size: 3×3, Stride: 2, Padding: 1 Pooling layer 3 8×8×128 4×4×128 Kernel size: 2×2, Stride: 2 Fully connected layer 1 4×4×128 512 - Fully connected layer 2 512 10 - Output layer 10 10 Softmax activation

[0220] Training optimization: Mini-batch SGD, Learning rate: 10 -4 , Momentum: 0.9, Regularization: L2, Batch size: 32, Epoch: 100

[0221] Based on the automatic extraction and identification of the feature points, the research team further developed a machine learning-based three-dimensional localization method. Specifically, they first calculated the contribution value of each feature point to the target organ localization according to its spatial distribution, temporal stability, and correlation degree. This involved the following three equations:

[0222] Spatial distribution equation:

[0223] Temporal stability equation:

[0224] Correlation strength equation:

[0225] where, w spatial represents the spatial distribution weight of the feature point, w temporal represents the temporal stability weight of the feature point, w ij represents the correlation strength between feature points i and j. α i , β i , γ i are the corresponding weight coefficients, satisfying

[0226] After calculating the contribution value of each feature point based on the above three equations, the research team marked the feature points with contribution values greater than the preset threshold as key positioning feature points. Then, based on these key feature points, a three-dimensional positioning coordinate system was constructed, and the spatial position of the target organ was mapped into this coordinate system. Specifically, they first calculated the normal vector of the key feature points as the z-axis of the coordinate system; then they optimized the coordinate axes to minimize the position error and rotation error of the feature points on the projection plane, thereby determining the x-axis and y-axis. The finally obtained three-dimensional positioning coordinate system can accurately describe the positional relationship of the target organ in space.

[0227] Table 3 Parameter Settings of the Three-Dimensional Positioning Coordinate System Construction Algorithm

[0228] Parameter Value Normal vector calculation Singular value decomposition Axis optimization Weighted least squares Weight coefficient μ = 0.5 Error threshold <![CDATA[ε pos = 2 mm, ε θ = 2°]]>

[0229] After mapping the target organ to the three-dimensional positioning coordinate system, the research team calculated its depth position value, lateral offset value, and longitudinal offset value in the coordinate system, and constructed a three-dimensional positioning trajectory of the target site. At the same time, they also real-time tracked the spatial position of the medical device (such as a puncture needle) and established the relative positional relationship between the medical device and the key positioning feature points.

[0230] By continuously matching the real-time position information of the medical device with the positioning trajectory of the target site, the research team can real-time calculate the position deviation between the two. When the deviation exceeds the preset threshold, the system will automatically generate a medical device position adjustment instruction, including the real-time optimization of the needle insertion direction and the needle insertion depth. This automatic navigation method based on feedback control can effectively make up for the uncertainties in the surgical operation process and ensure that the medical device accurately reaches the target area.

[0231] Table 4 Parameter Settings of the Medical Device Real-Time Navigation Algorithm

[0232]

[0233]

[0234] In summary, the research team proposed an innovative solution based on deep learning and machine learning for the positioning accuracy and stability problems of traditional ultrasound-guided surgery. This method can make full use of the rich anatomical structure information in the ultrasound image, realize the precise positioning of the surgical target organ by automatically extracting and identifying key feature points and constructing a three-dimensional positioning coordinate system. At the same time, this method also integrates real-time tracking and automatic navigation functions, and can perform closed-loop control according to the real-time position information of the medical device to ensure that the device accurately reaches the target area. Compared with the traditional positioning methods that rely on a single anatomical marker or manual measurement, the proposed solution of the present invention can significantly improve the positioning accuracy and stability of ultrasound-guided surgery, reduce the surgical risk, and enhance the clinical efficacy.

[0235] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 5 and Table 6 below.

[0236] Table 5 Variable Explanation Table (First Part)

[0237]

[0238]

[0239] Table 6 Variable Explanation Table (Second Part)

[0240]

[0241]

[0242] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A clinical ultrasound image-assisted positioning method, characterized in that: The following steps are involved: Acquiring an ultrasound image sequence of a target part of a patient; performing image preprocessing on the ultrasound image sequence; extracting anatomical structure feature points from the ultrasound image sequence; and identifying and classifying the anatomical structure feature points using a first machine learning model; Calculating the contribution value of the anatomical structure feature point to the positioning of the target part; Screening the anatomical structure feature points based on the contribution values; Constructing a three-dimensional positioning coordinate system based on the screened anatomical structure feature points; calculating the spatial position parameters of the target part in the three-dimensional positioning coordinate system; Generate a target part positioning trajectory; obtain real-time spatial position information of a medical device; match the real-time spatial position information of the medical device with the target part positioning trajectory; establish a relative position relationship between the medical device and the screened anatomical structure feature points; calculate the position deviation value of the medical device; generate a medical device position adjustment instruction; and adjust the needle insertion direction and depth of the medical device in real time according to the medical device position adjustment instruction.

2. The clinical ultrasound image-assisted positioning method according to claim 1, characterized in that: The steps of acquiring an ultrasound image sequence of a target part of a patient are specifically as follows: first, preheating and calibrating the ultrasound probe; then setting the ultrasound probe operating frequency within the range of 4 MHz to 15 MHz; adjusting the total gain parameter of the ultrasound device within the range of 45 decibels to 75 decibels; applying coupling agent to the ultrasound probe; then placing the probe on the surface of the target part; maintaining the probe at a 90-degree angle to the skin; slowly moving the probe position for scanning; controlling the scanning speed within the range of 1 cm to 2 cm per second; and setting the image acquisition frame rate within the range of 25 to 30 frames per second.

3. The clinical ultrasound image-assisted positioning method according to claim 1, characterized in that: The steps of image preprocessing the ultrasound image sequence are specifically: using adaptive median filtering to suppress noise; performing histogram equalization to enhance image contrast; using an improved Sobel operator to perform edge enhancement; normalizing the processed image; the preprocessed image sequence needs to meet a peak signal-to-noise ratio of not less than 35 decibels.

4. The clinical ultrasound image-assisted positioning method according to claim 1, characterized in that: The contribution value is obtained by calculating a contribution value equation group, which includes a feature point spatial distribution equation, a feature point temporal stability equation and a feature point association strength equation. The input of the feature point spatial distribution equation includes the three-dimensional coordinate value of the anatomical structure feature point, the Euclidean distance value between the anatomical structure feature point and the target part, the grayscale gradient value of the area around the anatomical structure feature point, and the depth value of the anatomical layer where the anatomical structure feature point is located.

5. The clinical ultrasound image-assisted positioning method according to claim 4, characterized in that: The input of the feature point temporal stability equation includes the displacement of the anatomical structure feature point between adjacent ultrasound images, the gray value change amplitude of the anatomical structure feature point, the shape change rate of the anatomical structure feature point, and the contrast change value of the anatomical structure feature point.

6. The clinical ultrasound image-assisted positioning method according to claim 4, characterized in that: The input of the feature point association strength equation includes the distance matrix value between the anatomical structure feature points, the grayscale similarity value between the anatomical structure feature points, the direction consistency value between the anatomical structure feature points, and the morphological similarity value between the anatomical structure feature points.

7. The clinical ultrasound image-assisted positioning method according to claim 1, characterized in that: The step of matching the real-time spatial position information of the medical device with the positioning trajectory of the target part is specifically: using an iterative closest point algorithm to calculate the registration matrix between the real-time position and the expected trajectory; calculating the motion trend of the medical device; and predicting and compensating the trajectory matching result based on the motion trend; Calculate the matching accuracy evaluation index.

8. The clinical ultrasound image-assisted positioning method according to claim 1, characterized in that: The steps of generating the medical device position adjustment instruction include: constructing a motion compensation matrix and a posture adjustment matrix; generating an adjustment instruction based on the motion compensation matrix and the posture adjustment matrix; evaluating system stability using a feedback stability evaluation mathematical model; and constructing a real-time optimization objective function.

9. The clinical ultrasound image-assisted positioning method according to claim 1, characterized in that: The real-time adjustment steps of the needle insertion direction and needle insertion depth of the medical device are specifically: converting the position adjustment instruction into a motor control signal; controlling the needle insertion direction of the medical device with an adjustment range of plus or minus 30 degrees; controlling the needle insertion depth of the medical device with an adjustment accuracy of 0.1 mm; real-time monitoring of the adjustment effect, and completing the position adjustment process when the position deviation value is less than 0.5 mm and remains stable for more than 1 second.

10. The clinical ultrasound image-assisted positioning method according to claim 1, characterized in that: The spatial position parameters include a depth position value, a lateral position offset value and a longitudinal position offset value, the uncertainty of the depth position value is less than 1 mm, and the uncertainties of the lateral position offset value and the longitudinal position offset value are less than 0.5 mm.

Images