Machine learning based dynamic monitoring of tumor susceptibility method and system

Abstract

The application provides a tumor susceptibility dynamic monitoring method and system based on machine learning, and relates to the technical field of data processing.The method comprises the following steps: mapping each unit in a spatial discrete set to a corresponding grid heterogeneity division block to obtain a discrete attribution feature subset; extracting a distribution moment feature of a Doppler frequency shift feature parameter of each discrete attribution feature subset, and calculating a local heterogeneity adjustment parameter; weighting and aggregating all the local heterogeneity adjustment parameters to generate a global adaptive compensation quantity; separating the spatial discrete set according to the global adaptive compensation quantity to obtain a dynamic component feature subset; and inputting the Doppler frequency shift feature parameter in the dynamic component feature subset into a pre-trained improved dense connection network model after time sorting to obtain a tumor susceptibility dynamic monitoring result.The application effectively improves the reliability and distinguishability of the monitoring result.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for dynamic monitoring of tumor susceptibility based on machine learning. Background Technology

[0002] In applications of dynamic monitoring of tumor susceptibility based on three-dimensional spatiotemporal data of Doppler ultrasound echo signals, existing technologies typically employ the following processing flow: first, discrete sampling points are mapped onto a three-dimensional spatial grid; then, a trend hypersurface is constructed through global least squares fitting; and finally, the frequency shift statistical features (such as mean and variance) of the entire hypersurface are extracted within a fixed time window as the basis for classification.

[0003] However, the projection points of the actual collected 3D spatiotemporal data on the hyperplane of spatial trend features often exhibit a non-uniform cluster density distribution. For example, near the lesion area, due to abnormal hemodynamics, multiple sampling points with similar spatial coordinates but different times will form high-density clusters on the hyperplane projection; while in normal tissue areas, the projection points are relatively sparse. Existing technologies, when processing such data, generally adopt fixed-resolution global grid partitioning or determine the projection boundary based on a single threshold, failing to adaptively segment the boundary constraint region according to the local cluster density differences of the projection points. This leads to two problems: on the one hand, subtle frequency shift changes in high-density clusters are lost due to averaging by coarse-grained grids; on the other hand, noise in low-density areas is over-divided into independent blocks, causing spurious heterogeneity. Ultimately, the dynamic component feature subset separated from the data cannot effectively characterize the local frequency shift fluctuation patterns related to tumor susceptibility, reducing the monitoring accuracy of subsequent classification models. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for dynamic monitoring of tumor susceptibility based on machine learning, which can more closely solve the displacement field and surface shape changes of ultrathin crystal devices after coating, and effectively improve the convergence stability and numerical reliability of surface shape analysis.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for dynamic monitoring of tumor susceptibility based on machine learning, the method comprising: Step 1: Acquire the dynamic monitoring data of the subjects and transform the dynamic monitoring data into a spatial discrete set in three-dimensional space. Each unit in the spatial discrete set contains spatial coordinate parameters, time parameters, and Doppler frequency shift characteristic parameters. Step 2: Fit the spatial coordinate parameters of all units in the spatial discrete set to construct a spatial trend feature analytical hypersurface; Step 3: Determine a boundary-constrained four-way domain based on the projection boundaries of each unit in the spatial discrete set onto the spatial trend feature analytic hyperplane; and divide the boundary-constrained four-way domain into multiple grid point heterogeneous partitioning blocks based on the projection clustering density distribution of the spatial coordinate parameters of each unit in the spatial discrete set onto the spatial trend feature analytic hyperplane. Step 4: Map each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block to obtain discrete attribution feature subsets; extract the distribution moment features of the Doppler frequency shift feature parameter of each discrete attribution feature subset, and calculate the local heterogeneity adjustment parameter; Step 5: Weighted aggregation of all local heterogeneity adjustment parameters to generate a global adaptive compensation quantity; Based on the global adaptive compensation quantity, the spatial discrete set is separated to obtain a dynamic component feature subset; Step 6: After sorting the Doppler frequency shift feature parameters in the dynamic component feature subset by time, input them into the pre-trained improved dense connection network model to obtain the dynamic monitoring results of tumor susceptibility.

[0006] Secondly, a machine learning-based dynamic monitoring system for tumor susceptibility includes: The module is used to acquire dynamic monitoring data of the subjects and transform the dynamic monitoring data into a spatial discrete set in three-dimensional space. Each unit in the spatial discrete set contains spatial coordinate parameters, time parameters, and Doppler frequency shift feature parameters. The fitting module is used to fit the spatial coordinate parameters of all units in the spatial discrete set and construct a spatial trend feature analytical hypersurface. The partitioning module is used to determine a boundary-constrained four-way domain based on the projection boundaries of each unit in the spatial discrete set on the spatial trend feature analytic hyperplane; and to partition the boundary-constrained four-way domain based on the projection clustering density distribution of the spatial coordinate parameters of each unit in the spatial discrete set on the spatial trend feature analytic hyperplane, thereby obtaining multiple grid point heterogeneous partitioning blocks. The computation module is used to map each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block to obtain discrete attribution feature subsets; extract the distribution moment features of the Doppler frequency shift feature parameter of each discrete attribution feature subset, and calculate the local heterogeneity adjustment parameter; The separation module is used to weight and aggregate all local heterogeneity adjustment parameters to generate a global adaptive compensation quantity; based on the global adaptive compensation quantity, the spatial discrete set is separated to obtain a dynamic component feature subset; The output module is used to sort the Doppler frequency shift feature parameters in the dynamic component feature subset by time and input them into a pre-trained improved dense connection network model to obtain the dynamic monitoring results of tumor susceptibility.

[0007] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: By mapping spatial discrete set units to corresponding grid-based heterogeneous partitioning blocks, the Doppler frequency shift distribution moment features (mean, variance, skewness) of each block are extracted. Local heterogeneity adjustment parameters are calculated, and a global adaptive compensation quantity is generated through weighted aggregation. This achieves accurate separation of the spatial discrete set, resulting in a high-quality dynamic component feature subset. This process, through the combination of local heterogeneity adjustment and global adaptive compensation, effectively filters noise interference, enhances dynamic feature signals related to tumor susceptibility, and solves the problem that dynamic component feature subsets cannot effectively characterize local frequency shift fluctuation patterns. Furthermore, an improved dense connection network enables automated output of monitoring results, replacing the reliance on manual feature extraction and judgment, reducing manual intervention costs and improving monitoring efficiency. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the process of a machine learning-based dynamic monitoring method for tumor susceptibility provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a machine learning-based dynamic monitoring system for tumor susceptibility provided in an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a method for dynamic monitoring of tumor susceptibility based on machine learning, the method comprising the following steps: Step 1: Acquire the dynamic monitoring data of the subjects and transform the dynamic monitoring data into a spatial discrete set in three-dimensional space. Each unit in the spatial discrete set contains spatial coordinate parameters, time parameters, and Doppler frequency shift characteristic parameters. Step 2: Fit the spatial coordinate parameters of all units in the spatial discrete set to construct a spatial trend feature analytical hypersurface; Step 3: Determine a boundary-constrained four-way domain based on the projection boundaries of each unit in the spatial discrete set onto the spatial trend feature analytic hyperplane; and divide the boundary-constrained four-way domain into multiple grid point heterogeneous partitioning blocks based on the projection clustering density distribution of the spatial coordinate parameters of each unit in the spatial discrete set onto the spatial trend feature analytic hyperplane. Step 4: Map each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block to obtain discrete attribution feature subsets; extract the distribution moment features of the Doppler frequency shift feature parameter of each discrete attribution feature subset, and calculate the local heterogeneity adjustment parameter; Step 5: Weighted aggregation of all local heterogeneity adjustment parameters to generate a global adaptive compensation quantity; Based on the global adaptive compensation quantity, the spatial discrete set is separated to obtain a dynamic component feature subset; Step 6: After sorting the Doppler frequency shift feature parameters in the dynamic component feature subset by time, input them into the pre-trained improved dense connection network model to obtain the dynamic monitoring results of tumor susceptibility.

[0014] In this embodiment of the invention, by mapping spatial discrete set units to corresponding grid-based heterogeneous partitioning blocks, extracting the Doppler frequency shift distribution moment features (mean, variance, skewness) of each block, calculating local heterogeneity adjustment parameters, and generating a global adaptive compensation quantity through weighted aggregation, accurate separation of the spatial discrete set is achieved, resulting in a high-quality dynamic component feature subset. This process, through the combination of local heterogeneity adjustment and global adaptive compensation, effectively filters noise interference, strengthens dynamic feature signals related to tumor susceptibility, and solves the problem that dynamic component feature subsets cannot effectively characterize local frequency shift fluctuation patterns. Furthermore, by using an improved dense connection network to automate the output of monitoring results, it replaces the method of relying on manual feature extraction and manual judgment, reducing the cost of manual intervention and improving monitoring efficiency.

[0015] In a preferred embodiment of the present invention, step 1 involves acquiring dynamic monitoring data of the subject and converting the dynamic monitoring data into a spatial discrete set in three-dimensional space. Each unit in the spatial discrete set includes spatial coordinate parameters, time parameters, and Doppler frequency shift characteristic parameters, and may include: Step 101: Acquire dynamic monitoring data of the subject. This dynamic monitoring data consists of continuously acquired Doppler ultrasound time-series signals, with each acquisition moment corresponding to a set of spatial location codes. Specifically, this includes: selecting a high-resolution Doppler ultrasound detection device; adjusting the device parameters to match the ultrasound probe frequency and detection depth with the subject's target monitoring area (such as the breast, liver, lungs, or other areas prone to tumors); placing the ultrasound probe against the subject's skin in the target monitoring area; and starting the device for continuous, uninterrupted data acquisition. The acquisition duration is set according to monitoring needs, typically 30 minutes to 1 hour, with the acquisition frequency set to 10 to 20 frames per second to ensure the capture of subtle changes in hemodynamics. The dynamic monitoring data obtained is a Doppler ultrasound time-series signal, which is an ultrasound echo signal that changes continuously over time. Its signal amplitude is positively correlated with the blood flow velocity in the monitoring area, and can directly reflect the dynamic characteristics of blood flow in the monitoring area, including key information such as blood flow velocity, blood flow direction, and blood flow concentration. At each data acquisition moment, the Doppler ultrasound detection device will synchronously record a set of spatial position codes. These spatial position codes are generated by the device's built-in positioning component and correspond one-to-one with the acquisition moment. The code consists of a three-dimensional coordinate code and a timestamp. The three-dimensional coordinate code is used to accurately identify the specific location of the ultrasound detection point in three-dimensional space at that moment, and the timestamp is used to correspond to the specific acquisition moment.

[0016] Step 102: Based on the spatial location encoding, the spatial sampling points at each acquisition time are resampled into a grid to obtain uniformly distributed three-dimensional spatial grid points. Each three-dimensional spatial grid point carries the interpolation result of the original Doppler frequency shift. Based on the interpolation result, a sliding window energy detection is performed on the time axis to extract the peak value and mean value of the Doppler frequency shift within each sliding window. The combination of the peak value and mean value is used as the Doppler frequency shift feature parameter of the corresponding three-dimensional spatial grid point at the center time of the sliding window, resulting in a time-ordered sequence of Doppler frequency shift feature parameters. Specifically, this includes: the spatial location encoding of each acquisition time obtained in step 101. The location is encoded, and the three-dimensional coordinate code in the encoding is parsed to determine the three-dimensional spatial coordinates (x, y, z) of the ultrasound sampling point at each acquisition time. The x-axis and y-axis correspond to the horizontal direction, and the z-axis corresponds to the depth direction perpendicular to the skin. All spatial sampling points at each acquisition time are subjected to gridded resampling processing. Specifically, a uniform three-dimensional spatial grid is set, with the x-axis and y-axis spacing set to 0.1 to 0.3 mm and the z-axis spacing set to 0.2 to 0.4 mm. In this implementation, the preferred x-axis and y-axis spacing is 0.2 mm and the z-axis spacing is 0.3 mm. This spacing is determined according to the resolution of the ultrasound detection equipment.

[0017] Each spatial sampling point at each acquisition time is mapped to its corresponding grid node according to its three-dimensional spatial coordinates. For grid nodes not directly covered by the sampling point, an interpolation method is used to supplement the original Doppler frequency shift data corresponding to that grid point, obtaining the interpolated result of the original Doppler frequency shift carried by each three-dimensional spatial grid point. In the interpolation process, a linear weighted average method is used. Taking the grid point as the center, all adjacent sampling points within a 3×3×3 range are selected, and the spatial distance between each adjacent sampling point and the grid point is calculated. The closer the distance, the larger the weight coefficient. The weight coefficient is calculated by dividing the weight of the adjacent sampling point by the distance between the sampling point and the grid point. The spatial distance between points is weighted with a coefficient ranging from 0.05 to 0.8. The closer the distance, the closer the weight coefficient is to 0.8; the farther the distance, the closer the weight coefficient is to 0.05. The Doppler frequency shift data of each adjacent sampling point is multiplied by its corresponding weight coefficient. All products are summed, and then divided by the sum of all weight coefficients to obtain the interpolation result for that grid point. A fixed-length sliding window is set on the time axis, with a window length of 8 to 12 acquisition frames and a time step of 1 to 2 acquisition frames. In this implementation, a window length of 10 acquisition frames and a time step of 1 acquisition frame are preferred. The sliding window moves sequentially along the time axis according to the fixed time step. Each time the window moves, energy detection is performed on the Doppler frequency shift interpolation results of all 3D spatial grid points within the current sliding window. The specific calculation process for energy detection is as follows: first, the square value of the Doppler frequency shift interpolation result of each grid point is calculated; then, the square values ​​of all grid points within the sliding window are added together to obtain the total energy value within the sliding window. Based on this total energy value, the peak value and mean value of the Doppler frequency shift within each sliding window are further extracted. The peak value extraction process involves iterating through the Doppler frequency shift interpolation results of all grid points within the sliding window and selecting the interpolation result with the largest value, which is the peak value of the Doppler frequency shift within the sliding window. The mean value extraction process involves... The Doppler frequency shift interpolation results of all grid points within the window are summed to obtain the total interpolation result. This total is then divided by the total number of grid points within the sliding window to obtain the mean Doppler frequency shift within the sliding window. The peak value and mean value corresponding to each sliding window are combined, with the peak value as the first element and the mean value as the second element, to form a two-dimensional feature vector. This two-dimensional feature vector is used as the Doppler frequency shift feature parameter of the corresponding three-dimensional spatial grid point at the center time of the sliding window. The Doppler frequency shift feature parameters corresponding to each center time are recorded sequentially according to the order in which the sliding window moves along the time axis, finally obtaining a time-sorted sequence of Doppler frequency shift feature parameters.

[0018] Step 103: Based on the Doppler frequency shift feature parameter sequence, the three-dimensional spatial coordinates corresponding to each Doppler frequency shift feature parameter are used as spatial coordinate parameters, the corresponding sliding window center time is used as a time parameter, and the Doppler frequency shift feature parameter itself is used as the third dimension. The spatial coordinate parameters, time parameters, and Doppler frequency shift feature parameters are collectively encapsulated into a single unit. All units are organized according to spatial location and temporal order to obtain a spatial discrete set in three-dimensional space. Specifically, this includes: extracting each Doppler frequency shift feature parameter from the Doppler frequency shift feature parameter sequence obtained in step 102. Each feature parameter corresponds to a specific three-dimensional spatial grid point and a specific sliding window center time. For each Doppler frequency shift feature parameter, the three-dimensional spatial coordinates (x, y, z) of its corresponding three-dimensional spatial grid point are determined, and these three-dimensional spatial coordinates are directly used as spatial coordinate parameters. Simultaneously, the center time of the sliding window corresponding to the Doppler frequency shift feature parameter is determined, and the sliding window... The method for calculating the window center time is as follows: add half the length of the sliding window to the starting time of the sliding window. If the starting time of the sliding window is frame t and the window length is 10 frames, then the center time is the time point corresponding to frame t+5, and this center time is used as the time parameter. The Doppler frequency shift characteristic parameter itself is used as the third dimension data, which directly reflects the blood flow frequency shift characteristics at the corresponding spatial location and time. The spatial coordinate parameter, time parameter, and Doppler frequency shift characteristic parameter are integrated and encapsulated to form an independent unit, and each unit contains three core pieces of information. After the encapsulation of all units is completed, all units are organized and arranged according to the spatial position order of the three-dimensional spatial grid points and the order of the time parameters corresponding to each unit. The spatial position order is arranged in ascending order of the x-axis, y-axis, and z-axis, and the time parameters are arranged in ascending order of the time. All units are integrated to form a spatial discrete set in three-dimensional space.

[0019] This embodiment, by accurately acquiring Doppler ultrasound time-series signals and simultaneously recording spatial location codes, combined with gridded resampling and sliding window energy detection, constructs a well-structured three-dimensional spatial discrete set, effectively solving the problem of feature extraction deviation caused by uneven distribution of sampling points and chaotic data organization.

[0020] In a preferred embodiment of the present invention, step 2, fitting the spatial coordinate parameters of all units in the spatial discrete set to construct a spatial trend feature analytical hypersurface, may include: Step 201: Extract the spatial coordinate parameters of all units from the spatial discrete set. Each spatial coordinate parameter contains three-dimensional spatial coordinate values. Center all spatial coordinate parameters to obtain a centralized set of spatial coordinate parameters. Specifically, this includes: traversing all encapsulated units from the spatial discrete set constructed in step 103, extracting the spatial coordinate parameters contained in each unit. Each spatial coordinate parameter contains three-dimensional spatial coordinate values, corresponding to the x-axis, y-axis, and z-axis coordinates in three-dimensional space, which can accurately identify the spatial position of each unit. Due to differences in body shape and target monitoring area among different subjects, and the possibility of slight deviations in equipment placement during ultrasound acquisition, the results may vary. The spatial coordinate parameters of the subjects vary significantly in value range. For example, some subjects have an x-axis coordinate range of 0 to 50 mm, while others have an x-axis coordinate range of 20 to 70 mm. This difference in value range can seriously affect the accuracy of subsequent surface fitting, causing the fitted surface to fail to accurately reflect the spatial trend of the monitored area. Therefore, it is necessary to center all extracted spatial coordinate parameters. The specific method of centering is to calculate the average value of the coordinates of all spatial coordinate parameters in the x-axis direction. The calculation method is to add the x-axis coordinate values ​​of all units to obtain the sum of the x-axis coordinates, and then divide the sum of the x-axis coordinates by the total number of units in the spatial discrete set to obtain the average value of the coordinates in the x-axis direction, denoted as . Using the same method, calculate the average value of the coordinates of all spatial coordinate parameters in the y-axis direction. Calculate the average value of the coordinates of all spatial coordinate parameters in the z-axis direction. For each spatial coordinate parameter, perform a centering calculation, subtracting the x-axis coordinate value of each spatial coordinate parameter from the centering calculation. This yields the centered coordinates of the unit along the x-axis; the y-axis coordinates of each spatial coordinate parameter are then subtracted. This yields the centered coordinate value of the unit along the y-axis; the z-axis coordinate value of each spatial coordinate parameter is then subtracted from the centered coordinate value. We obtain the centralized coordinate value of the unit along the z-axis; we integrate the three centralized coordinate values ​​of each unit to obtain the centralized coordinate value corresponding to each spatial coordinate parameter, and then integrate the centralized coordinate values ​​of all units to form a centralized set of spatial coordinate parameters.

[0021] Step 202 involves fitting a three-dimensional quadratic surface to the centered set of spatial coordinate parameters. During the fitting process, two horizontal dimensions from the three dimensions of the spatial coordinate parameters are used as independent variables, and the third vertical dimension is used as the dependent variable to obtain the fitting coefficients. Specifically, this includes: performing a three-dimensional quadratic surface fitting process on the centered set of spatial coordinate parameters. The selection of independent and dependent variables is clearly defined during the fitting process. Two horizontal dimensions are selected as independent variables from the three dimensions of the spatial coordinate parameters, and the third vertical dimension is selected as the dependent variable. The horizontal dimensions are typically selected as the x-axis and y-axis coordinates, and the vertical dimension as the z-axis coordinate. This selection method closely reflects the actual situation in ultrasonic testing where the horizontal direction represents the plane of the monitoring area and the vertical direction represents the depth direction, ensuring that the fitted surface accurately reflects the spatial trend of the monitoring area. During the fitting process, all data points in the centered set of spatial coordinate parameters are combined, and the least squares method is used for fitting calculation. The specific calculation process of the least squares method is as follows: the function expression of the three-dimensional quadratic surface is set as follows: Where p, q, r, s, t, and u are the fitting coefficients to be solved, x and y are the independent variables, and z is the dependent variable; traverse each data point in the set of centered spatial coordinate parameters, each data point has corresponding centered coordinates. , each data point and Substituting the above surface function, we obtain the predicted value. ; Calculate the actual coordinates of each data point Predicted coordinate values ​​of fitted surface The difference between them, i.e., the error value All error values Perform the squaring operation to obtain the squared error value for each data point. The total error is obtained by summing the squared errors of all data points. , where m is the total number of data points in the centered spatial coordinate parameter set. By adjusting the parameters p, q, r, s, t, and u of the fitted surface, the total error S is minimized using the gradient descent method. Specifically, the gradient descent method calculates the partial derivatives of the total error S with respect to each parameter p, q, r, s, t, and u, and adjusts the value of each parameter according to the direction of the partial derivatives. The step size for each adjustment is set to 0.0005 to 0.0015; in this implementation, a step size of 0.001 is preferred. This adjustment process is repeated until the total error S reaches its minimum value. The error threshold is set to... to When the total error S is less than or equal to the threshold, the adjustment stops. The corresponding parameters p, q, r, s, t, and u are the fitting coefficients. The fitting coefficients include quadratic coefficients p, q, and r, linear coefficients s and t, and a constant term u. Each coefficient corresponds to a parameter in the fitted surface function. The values ​​of p and q range from -0.5 to 0.5, the value of r ranges from -0.3 to 0.3, the value of s and t ranges from -1.0 to 1.0, and the value of u ranges from -5.0 to 5.0.

[0022] Step 203: Based on the fitting coefficients, construct a continuous three-dimensional spatial surface function, which is the spatial trend feature analysis hypersurface. Specifically, this includes: constructing a continuous three-dimensional spatial surface function based on all the fitting coefficients p, q, r, s, t, and u obtained in step 202. This surface function uses two selected horizontal dimensions x and y as independent variables and the vertical dimension z as the dependent variable. The fitting coefficients p, q, r, s, t, and u are directly substituted into the preset three-dimensional quadratic surface function expression. In this process, a complete three-dimensional spatial surface function is obtained, which is the spatial trend feature analytical hyperface.

[0023] This embodiment eliminates the influence of individual differences and systematic errors in subjects by centering the spatial coordinate parameters. Combined with the three-dimensional quadratic surface fitting to construct a spatial trend feature analysis hypersurface, it can more accurately reflect the spatial distribution trend of the monitoring area and capture subtle changes in the spatial coordinate parameters.

[0024] In a preferred embodiment of the present invention, step 3, based on the projected boundaries of each unit in the spatial discrete set on the spatial trend feature analytic hyperplane, determines a boundary-constrained four-way domain; based on the projected clustering density distribution of the spatial coordinate parameters of each unit in the spatial discrete set on the spatial trend feature analytic hyperplane, the boundary-constrained four-way domain is partitioned to obtain multiple grid point heterogeneous partitioning blocks, which may include: Step 301: Using the spatial trend feature analysis hyperplane as the projection plane, the spatial coordinate parameters of each unit in the spatial discrete set are projected perpendicularly onto the spatial trend feature analysis hyperplane to obtain the projection point corresponding to each unit. Specifically, this includes: using the spatial trend feature analysis hyperplane constructed in step 203 as the projection plane, and determining the projection direction to be perpendicular to the hyperplane. The method for determining the projection direction is based on the function expression of the spatial trend feature analysis hyperplane. Calculate the normal vector of the hyperface at any point. The direction of the normal vector is perpendicular to the hyperface. Use this normal vector direction as the projection direction to ensure that the projection result accurately reflects the position of each element's spatial coordinate parameter on the hyperface, avoiding errors in the projection point position caused by projection direction deviation. Traverse all elements in the spatial discrete set constructed in step 103 and extract the spatial coordinate parameters contained in each element. The spatial coordinate parameters of each unit are projected onto the spatial trend feature analysis hyperplane according to a determined projection direction. The specific calculation process of the projection is as follows: the spatial coordinate points of the unit are... Draw a straight line parallel to the projection direction (normal vector direction). The intersection of this line and the spatial trend characteristic analytical hypersurface is the projection point corresponding to this element. ,in By and Substituting into the hypersurface function, we obtain the result. and The coordinates of the projection point in the horizontal direction of the hyperface are given. The position of each projection point is determined by the spatial coordinate parameters of the corresponding element and the shape of the hyperface. It can accurately represent the relative position of the element on the spatial trend hyperface. The distribution of projection points can reflect the spatial aggregation of discrete set elements in space.

[0025] Step 302: Extract the minimum bounding rectangle boundary of all projection points. Using the minimum bounding rectangle boundary as the constraint boundary, determine a rectangular region on the spatial trend feature analysis hyperplane. This rectangular region is the boundary constraint four-way domain. Specifically, this includes: extracting the coordinate information of all projection points obtained in step 301. Since all projection points lie on the spatial trend feature analysis hyperplane, can be and The only certainty is that the focus is on analyzing all projection points in the horizontal direction of the hyperplane. shaft and The distribution range of the axes; traversing all projection points. Axis coordinates, filter The maximum and minimum values ​​of the axis coordinates are denoted as follows: and Using the same method, traverse all projection points. Axis coordinates, filter The maximum and minimum values ​​of the axis coordinates are denoted as follows: and Based on these four extreme coordinates , , , Define a rectangular boundary, with the coordinates of its top-left corner being... The coordinates of the upper right corner are The coordinates of the lower left corner are The coordinates of the lower right corner are The rectangular boundary is the minimum bounding rectangle boundary of all projection points, which can completely contain all projection points inside the rectangle and ensure that no projection point exceeds the rectangular boundary. Using the minimum bounding rectangle boundary as the constraint boundary, a rectangular region is defined on the spatial trend feature analysis hyperplane. The range of the rectangular region is completely consistent with the minimum bounding rectangle boundary. This rectangular region is the boundary constraint four-way domain.

[0026] Step 303: Perform density-based spatial clustering on all projected points to obtain multiple projected point clusters. Each projected point cluster corresponds to a high-density clustering region. Calculate the minimum bounding closure region for each projected point cluster to obtain multiple minimum bounding closure regions. Specifically, this includes: performing density-based spatial clustering on all projected points obtained in step 301, using the DBSCAN algorithm as the density-based spatial clustering method. This algorithm can automatically divide clusters according to the clustering density of projected points, effectively identifying high-density clustering regions and low-density regions. The specific operation process involves setting two core parameters, one of which is the radius threshold. The other is the density threshold K, where the radius threshold... The value is set to 0.3 to 0.7 mm, with 0.5 mm being preferred in this implementation. This value is determined based on the distribution density of the projection points to ensure that the clustering relationship between adjacent projection points can be captured. The density threshold K is set to 3 to 7, with 5 being preferred in this implementation. That is, when the radius around a projection point is... When the number of projection points within the range reaches or exceeds a set density threshold, that projection point is designated as the core point; traverse all projection points, using each projection point as the center, and calculate the radius threshold... Define a range and count the number of other projection points within that range. If the number is greater than or equal to a set density threshold K, then that projection point is considered a core point; if the number is less than the density threshold K, then determine whether that projection point is within the radius of a core point. Within the specified range, if it is, it is treated as an edge point; otherwise, it is treated as a noise point, which will be subsequently removed. Points with a distance less than the radius threshold from the core point will be considered. All projection points (including core points and edge points) are grouped into the same cluster. All core points are traversed sequentially to avoid redundant clustering, resulting in multiple projection point clusters. Each projection point cluster corresponds to a high-density aggregation region. These high-density aggregation regions usually correspond to lesion areas within the monitoring area. Because the hemodynamics near the lesion area is abnormal, the blood flow velocity and concentration differ from those of normal tissue, causing the corresponding projection points to exhibit a high-density aggregation state.

[0027] After clustering, all noise points are removed. For each cluster of projected points, the minimum bounding closed region of all projected points within that cluster is calculated. This is done by traversing all projected points within the cluster to find the minimum bounding closed region of all projected points within that cluster. Maximum value of axis coordinates Minimum value of axis coordinates Maximum sum of axis coordinates The minimum values ​​of the axis coordinates are denoted as follows: , , , Based on these extreme coordinates, a rectangular closed region is constructed, and the coordinates of the four vertices of this rectangular closed region are as follows: If a closed region can completely contain all the projection points of the cluster, then the closed region is the smallest circumscribed closed region corresponding to the cluster of projection points.

[0028] Step 304: Perform intersection calculations between each minimum bounding closed region and the boundary constraint four-way domain to obtain the intersection closed region of each minimum bounding closed region within the boundary constraint four-way domain; determine the inclusion relationship between the intersection closed regions, and mark each intersection closed region as either a root closed region or a sub-closed region. Regions not included by any other intersection closed region are marked as root closed regions, and regions included by at least one other intersection closed region are marked as sub-closed regions. Specifically, this includes: defining the vertex sets of the two closed rectangular regions, with the boundary constraint four-way domain as the closed rectangular region A, and its vertex set denoted as... Arranged in clockwise order; each smallest circumscribed closed region is called a closed rectangular region B, and its vertex set is denoted as B0. The vertices are arranged in a clockwise order. Using closed rectangular region A as the clipping window, all vertices of closed rectangular region B are clipped. The final set of vertices is the set of vertices of the intersection region of the two closed rectangular regions. Specifically, the clipping result vertex set C is initialized by storing all vertices of closed rectangular region B into C sequentially. Each edge of closed rectangular region A is used as a clipping edge to clip the current clipping result set C, following the principle of retaining interior points, discarding exterior points, and finding intersection points across edges. An interior point is defined as a point located to the right of the clipping edge (in clockwise arrangement of closed rectangular regions, the right side is the interior of the region). The first edge of closed rectangular region A... For example, its linear equation is: During the clipping process, iterate through adjacent vertex pairs (P, Q) in set C. If both P and Q are interior points, then keep Q; if P is an interior point and Q is an exterior point, calculate the distance between line segment PQ and the clipping edge. If P is an external point and Q is an internal point, calculate the intersection point of line segment PQ and the trimming edge. If P and Q are both external points, keep the intersection point and Q; if P and Q are both external points, discard Q; then process the closed rectangular region A in turn. (Equation of a straight line) Interior point definition ), (Equation of a straight line) Interior point definition ), (Equation of a straight line) Interior point definition Repeat the above trimming operation on all four sides; after trimming, if the result set C is empty, it means that the two closed rectangular regions have no intersection, and the smallest bounding closed region is discarded; if set C contains at least 3 vertices, connect these vertices in clockwise order, and the resulting closed region is the intersection closed region of the two closed rectangular regions. The coordinate range of the axes is all vertices in set C. The minimum to maximum value of the coordinates, i.e. , The coordinate range of the axes is all vertices in set C. The minimum to maximum value of the coordinates, i.e. The result is consistent with the intersection calculation of the coordinate range. After completing the intersection calculation, analyze the inclusion relationship between all closed intersection regions. Traverse each closed intersection region and compare it with all other closed intersection regions to determine whether the closed intersection region is completely contained by other closed intersection regions. The determination method is as follows: if the closed intersection region... The axis range is entirely within another intersection closed region. Within the axis range, and its The axis range is also completely within the other intersection closed region. If the intersection is within the axis range, it means that the intersection closed region is completely contained by another intersection closed region; if the intersection closed region is not contained by any other intersection closed region, it is marked as the root closed region; if the intersection closed region is contained by at least one other intersection closed region, it is marked as a sub-closed region.

[0029] Step 305: Remove all directly embedded sub-closed regions from the root closed region to obtain a set of regions with a porous structure; using each removed sub-closed region as input, recursively perform the removal operation, that is, remove the innermost sub-closed regions directly embedded in the sub-closed region from the sub-closed region until no more removable sub-closed regions exist; collect all non-overlapping regions generated during the entire recursive process as grid heterogeneity partitioning blocks, specifically including: for all root closed regions, check one by one whether there are directly embedded sub-closed regions inside each root closed region, until... An embedded sub-closed region refers to a sub-closed region that is completely contained within the root closed region and is not contained by other sub-closed regions. The checking method is to compare each root closed region with all sub-closed regions, filter out the sub-closed regions that are completely contained within the root closed region and are not contained by other sub-closed regions, and remove all directly embedded sub-closed regions within each root closed region. After removal, each root closed region forms a region with a hole structure. The position of the hole is the position of the removed sub-closed region, and the size and shape of the hole are completely consistent with the removed sub-closed region.

[0030] Each removed sub-closed region is treated as a new processing object, and the above removal operation is recursively performed. That is, it is checked whether there is a directly embedded inner sub-closed region inside each removed sub-closed region. The checking method is the same as above. If there is, it is removed to form a new region with a hole structure. Then, the removed inner sub-closed region is treated as a processing object, and the removal operation is performed again. This process is repeated until there are no removable inner sub-closed regions inside any sub-closed region. At this time, all remaining sub-closed regions are the innermost sub-closed regions and no longer contain other sub-closed regions. All non-overlapping regions generated in the recursive process are collected. These regions include all regions with hole structures formed in the recursive process and the final remaining unremoved sub-closed regions. These regions are used as grid point heterogeneity partitioning blocks.

[0031] This embodiment projects spatial coordinate parameters onto a hyperplane, performs clustering and region partitioning based on the cluster density of the projected points, and obtains non-overlapping grid point heterogeneity partitioning blocks through recursive removal operations. It can adaptively adjust the size and number of blocks according to the density distribution of the projected points, effectively avoiding the problems of feature averaging in high-density regions and false heterogeneity in low-density regions caused by fixed grids.

[0032] In a preferred embodiment of the present invention, step 4, mapping each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block, obtains a discrete attribution feature subset; extracting the distribution moment features of the Doppler frequency shift feature parameter of each discrete attribution feature subset, and calculating the local heterogeneity adjustment parameter, may include: Step 401: Based on the heterogeneity of each grid point, divide the space into blocks, traverse all units in the spatial discrete set, and determine whether the projection point of the spatial coordinate parameter of each unit on the spatial trend feature analysis hyperplane falls within the grid point heterogeneity division block. If it falls within the block, the unit is assigned to the set corresponding to the block. After traversing all blocks, each block corresponds to a set of units, which is the discrete attribution feature subset. Specifically, this includes: first, obtaining all the grid point heterogeneity division blocks obtained in step 305, clarifying the boundary range of each grid point heterogeneity division block. The boundary range of each grid point heterogeneity division block is determined by its horizontal coordinate range on the spatial trend feature analysis hyperplane, that is, the boundary range of each block. Minimum to maximum values ​​of axis coordinates From minimum to maximum axis coordinates, and for each block... The axis coordinates can be derived from the corresponding shaft and The coordinate axes are substituted into the function expression of the spatial trend feature analysis hyperface to obtain the coordinates. A traversal operation is initiated, extracting each cell from the spatial discrete set constructed in step 103. For each extracted cell, the spatial coordinate parameters contained within the cell are first obtained. Then, according to the projection direction and projection method determined in step 301, the projection point coordinates of the cell's spatial coordinate parameters on the spatial trend feature analysis hyperface are recalculated to ensure the accuracy of the projection point coordinates and avoid incorrect attribution due to previous projection errors. After obtaining the projection point coordinates, the projection point is... axis coordinates and The axis coordinates, and the range of axis coordinates for each grid point's heterogeneous partitioning block. The coordinate range of the axes is compared to determine whether the projected point falls within a certain grid point heterogeneity partitioning block. The criterion for this judgment is the range of the projected point. The axis coordinate is greater than or equal to that of the block. The minimum value of the axis coordinate, which is less than or equal to the value of the block. Maximum value of axis coordinates, and simultaneously the projection point The axis coordinate is greater than or equal to that of the block. The minimum value of the axis coordinate, which is less than or equal to the value of the block. If the maximum value of the axis coordinates is satisfied simultaneously, it means that the projection point falls within the block, and the unit is then assigned to the temporary set corresponding to the block.

[0033] If a projection point does not fall within any grid heterogeneous partitioning block, the cell is discarded because the projection point corresponding to such a cell exceeds the boundary constraint four-dimensional domain, and the Doppler frequency shift characteristic parameter it carries is meaningless for subsequent local heterogeneous parameter calculations. Discarding it avoids redundant data interfering with the calculation results. Following the above method, all cells in the spatial discrete set are traversed to ensure that each cell that meets the conditions can be accurately assigned to the corresponding block set. After the traversal is completed, each grid heterogeneous partitioning block corresponds to an independent cell set. Each cell set contains all cells whose projection points fall within the block. These cell sets are the discrete attribution feature subsets.

[0034] Step 402: Extract the Doppler frequency shift feature parameters of all units in each discrete attribution feature subset, and calculate the first-order raw moment, second-order central moment, and third-order central moment of the Doppler frequency shift feature parameters, which are used as the mean feature, variance feature, and skewness feature, respectively. Specifically, for each discrete attribution feature subset, extract the Doppler frequency shift feature parameters carried by all units in that subset. The Doppler frequency shift feature parameters of each unit are two-dimensional feature vectors, containing two elements: peak value and mean value. To ensure the uniformity and accuracy of the calculation, this step uniformly selects the peak value in the two-dimensional feature vector as the basic data for calculating the distribution moment feature, because the peak value can more accurately represent the mean value. It intuitively reflects the extreme changes in blood flow frequency shift, which is more in line with the characteristics of hemodynamic abnormalities in the lesion area. After extraction, the total number of units in the discrete attribution feature subset is counted and recorded as the total number of units. The first-order origin moment (mean feature) is then calculated. The calculation method is to add the peak values ​​of the Doppler frequency shift feature parameters of all units in the subset in sequence to obtain the sum of the peak values ​​(where each value is the peak value of the corresponding unit). Then, the sum of the peak values ​​is divided by the total number of units in the subset. The result is the mean feature of the discrete attribution feature subset. The mean feature can reflect the overall average level of the blood flow frequency shift peak value in the region and reflect the overall intensity of the blood flow frequency shift in the region.

[0035] The second-order central moment (variance feature) is calculated by subtracting the mean feature of the subset from the peak value of the Doppler frequency shift characteristic parameter of each unit. This yields the difference between the peak value and the mean feature for each unit. Each difference is then squared. The squared differences of all units are summed sequentially to obtain a total sum of squared differences. This total sum of squared differences is divided by the total number of units in the subset. The result is the variance feature of this discrete subset. The variance feature reflects the dispersion of the blood flow frequency shift peak value within the region; a larger variance indicates a more dispersed region. The greater the difference in blood flow frequency shift peak values ​​among different units within a subset, the more uneven the blood flow distribution within the region. The third-order central moment (skewness feature) is calculated by subtracting the mean feature of the subset from the peak value of the Doppler frequency shift feature parameter of each unit in the subset, obtaining the difference between the peak value and the mean feature of each unit. Then, each difference is cubed to obtain the cubed value of the difference for each unit. The cubed values ​​of the differences of all units are added together to obtain the sum of the cubed differences. The sum of the cubed differences is divided by the total number of units in the subset, and the result is the skewness feature of the discrete attribution feature subset.

[0036] Step 403: Calculate the ratio of variance feature to mean feature to obtain the coefficient of variation; calculate the product of the absolute value of skewness feature and coefficient of variation to obtain the initial value of local heterogeneity; normalize the initial value of local heterogeneity with the area of ​​the grid heterogeneity partition corresponding to the discrete attribution feature subset to obtain the local heterogeneity adjustment parameter for each discrete attribution feature subset, specifically including: for each discrete attribution feature subset, obtain the variance feature of the subset calculated in step 402. and mean characteristics Calculate the coefficient of variation (CV) using the variance characteristics of the subset. Divide by the mean characteristic of the subset The result obtained is the coefficient of variation. (Note: If the mean characteristic) If the value is 0, the absolute value of the mean characteristic can be used, or a small constant can be introduced. (To avoid a denominator of 0), the coefficient of variation can eliminate the influence of the magnitude of the mean feature and more objectively reflect the relative dispersion of the blood flow frequency shift peak in the region. It avoids distortion in the comparison of dispersion caused by large differences in the mean features of different regions. For example, if two regions have the same variance features but different mean features, directly comparing the variance features cannot accurately determine the difference in dispersion. The coefficient of variation can effectively solve this problem.

[0037] The initial value of local heterogeneity H1 is calculated by taking the absolute value of the skewness characteristic S of the subset, obtaining the absolute skewness value |S|, and then multiplying the absolute skewness value |S| by the coefficient of variation CV of the subset. The result is the initial value of local heterogeneity H1. The initial value of local heterogeneity can comprehensively reflect the dispersion and asymmetry of the peak blood flow frequency shift in the region. The larger the value, the stronger the blood flow heterogeneity in the region, and the more likely there is a lesion area. This is because abnormal blood flow velocity and concentration near the lesion area will lead to a large dispersion and asymmetric distribution of the peak blood flow frequency shift. The area of ​​the grid point heterogeneity partition corresponding to this discrete subset of features is calculated. The calculation method is to obtain the block's... Maximum value of axis coordinates and minimum value ,use minus The horizontal width of the block is obtained. ; then obtain the block Maximum value of axis coordinates and minimum value ,use Subtracting this gives the vertical height of the block. Multiply the horizontal width W by the vertical height H to obtain the area of ​​the heterogeneous partition block at that grid point. The area size reflects the spatial extent of the region, providing a basis for subsequent normalization processing. The area value range is related to the grid spacing, typically 0.04 mm. 2 Up to 10mm 2 Normalization is performed by dividing the initial value of the local heterogeneity H1 of the subset by the area of ​​the corresponding grid heterogeneity partition. The result obtained is the local heterogeneity adjustment parameter of the discrete attribution feature subset. .

[0038] This embodiment solves the problem of confusion of blood flow features in different regions caused by the failure to partition and classify spatial discrete set units. By dividing and classifying units into blocks according to grid heterogeneity, blood flow features in different regions can be extracted in a targeted manner, avoiding calculation deviations caused by cross-interference of regional features.

[0039] In a preferred embodiment of the present invention, step 5 involves weighted aggregation of all local heterogeneity adjustment parameters to generate a global adaptive compensation quantity; based on the global adaptive compensation quantity, the spatial discrete set is separated to obtain a dynamic component feature subset, which may include: Step 501: Calculate the ratio of the area of ​​the grid heterogeneous partition block corresponding to each discrete attribution feature subset to the total area of ​​all grid heterogeneous partition blocks, and use this ratio as the weighting coefficient for each discrete attribution feature subset; calculate the weighted local heterogeneity adjustment parameter by combining the local heterogeneity adjustment parameter of each discrete attribution feature subset with the weight of each discrete attribution feature subset; sum all the weighted local heterogeneity adjustment parameters to generate the global adaptive compensation quantity, specifically including: obtaining all grid heterogeneous partition blocks obtained in step 305, and calculating the block area according to the method in step 403. The method involves calculating the area of ​​each grid point's heterogeneous partitioned block one by one. First, the maximum and minimum horizontal coordinate values ​​of each block are obtained, and the difference between them is the horizontal width of the block. Then, the maximum and minimum vertical coordinate values ​​of each block are obtained, and the difference between them is the vertical height of the block. The horizontal width and vertical height are multiplied to obtain the area of ​​each grid point's heterogeneous partitioned block. The areas of all grid point heterogeneous partitioned blocks are added together to obtain the total area of ​​all grid point heterogeneous partitioned blocks. This total area reflects the spatial sum of all effective monitoring areas within the boundary constraint four-way domain, providing a basis for the subsequent calculation of weighting coefficients.

[0040] The weighting coefficients for each discrete attribution feature subset are calculated. The weighting coefficients are based on the proportion of each block area to the total area. Specifically, the area of ​​the block divided by the grid heterogeneity corresponding to the discrete attribution feature subset is divided by the total area of ​​all grid heterogeneous blocks. The result is the weighting coefficient of the discrete attribution feature subset. The weighting coefficients range from 0 to 1, and the sum of the weighting coefficients of all discrete attribution feature subsets is 1. The larger the block area, the larger the corresponding weighting coefficient, indicating that the region contains more units and its blood flow heterogeneity has a more significant impact on the overall heterogeneity of the entire monitoring area. Weighting can highlight the heterogeneity contribution of large areas while taking into account the characteristics of small areas, avoiding the neglect of heterogeneity in some areas.

[0041] The weighted local heterogeneity adjustment parameter is calculated by multiplying the local heterogeneity adjustment parameter of each discrete attribution feature subset (calculated in step 403) by the weighting coefficient corresponding to that subset to obtain the weighted local heterogeneity adjustment parameter of that subset; the weighted local heterogeneity adjustment parameters of all discrete attribution feature subsets are summed sequentially, and the sum is the global adaptive compensation amount; the value range of the global adaptive compensation amount is usually from 0.01 to 5.0.

[0042] Step 502: Subtract the global adaptive compensation amount from the Doppler frequency shift feature parameter of each unit in the spatial discrete set to obtain the compensated Doppler frequency shift feature parameter of each unit; repackage the compensated Doppler frequency shift feature parameter of each unit with the spatial coordinate parameter and time parameter of each unit to obtain the compensated spatial discrete set. Specifically, this includes: obtaining the global adaptive compensation amount generated in step 501, clarifying a unified compensation standard, and ensuring that the compensation processing of all units is consistent; traversing all units in the spatial discrete set constructed in step 103, and extracting the Doppler frequency shift feature parameter carried by each unit one by one. This feature parameter is a two-dimensional feature vector containing two elements: peak value and mean value. To ensure the comprehensiveness of the compensation, the same compensation processing is performed on these two elements during the compensation process.

[0043] The specific compensation method involves subtracting the global adaptive compensation amount from the peak value of the Doppler frequency shift characteristic parameter of the unit to obtain the compensated peak value of the unit; subtracting the global adaptive compensation amount from the mean value of the Doppler frequency shift characteristic parameter of the unit to obtain the compensated mean value of the unit; and then recombining the compensated peak value and mean value to form the compensated Doppler frequency shift characteristic parameter of the unit. The core purpose of the compensation process is to suppress the interference caused by global blood flow heterogeneity, reduce the masking of abnormal features in the target area by blood flow fluctuations in normal tissue areas, and make the abnormal blood flow features in the target area more prominent, thus providing more accurate data support for subsequent dynamic component separation and tumor susceptibility feature analysis.

[0044] After compensation, each unit is repackaged in the same format as the unit in step 103. That is, the compensated Doppler frequency shift characteristic parameter of the unit is integrated with the original spatial coordinate parameter (three-dimensional spatial coordinate) and time parameter (center time of the sliding window) of the unit to form a new independent unit. After traversing all units for compensation and repackaging, all new units are integrated in accordance with the spatial discrete set organization method in step 103 (arranged in spatial position order and time parameter order) to obtain the compensated spatial discrete set.

[0045] Step 503: Calculate the mean and standard deviation of the compensated Doppler frequency shift characteristic parameters of all units in the compensated spatial discrete set, and use the sum of the mean and standard deviation by a preset multiple as the dynamic threshold; extract all units in the compensated spatial discrete set whose absolute value of the compensated Doppler frequency shift characteristic parameters is greater than the dynamic threshold, and use the set of the extracted units' compensated Doppler frequency shift characteristic parameters sorted by time as the dynamic component feature subset. Specifically, this includes: splitting the compensated spatial discrete set into time-series signals, traversing all units in the compensated spatial discrete set, grouping them according to the coordinates of the three-dimensional spatial grid points corresponding to each unit, and grouping units corresponding to the same three-dimensional spatial grid point but different time parameters into one group to form a time-series sequence of Doppler frequency shift characteristic parameters corresponding to each grid point. This sequence contains the compensated peak and mean data of the grid point at different time points, which can completely reflect the temporal change trend of blood flow frequency shift at this spatial location.

[0046] An adaptive wavelet thresholding denoising algorithm is employed to process the Doppler frequency shift characteristic parameter time series corresponding to each grid point, achieving separation of dynamic and static components. The wavelet basis function and the number of decomposition levels are set. The preferred wavelet basis function is the db4 wavelet (adapted to the smoothness and abrupt changes of ultrasound time series signals). The number of decomposition levels is set to 3 to 5, with 4 levels preferred in this implementation. The number of decomposition levels is determined based on the length and fluctuation frequency of the time series to ensure effective separation of components at different frequencies. The wavelet basis function is set as follows: Time series After four levels of wavelet decomposition, its wavelet decomposition expression is: ,in The low-frequency components (corresponding to the static components) are the result of 4-layer decomposition. (j=1,2,3,4) represents 4 high-frequency components (corresponding to dynamic components); each time series is substituted into the wavelet basis function for 4-level wavelet decomposition, resulting in 1 low-frequency component and 4 high-frequency components. The low-frequency component corresponds to the static components in the time series signal, mainly including stable and unchanging components such as background noise and inherent equipment interference; the high-frequency components correspond to the dynamic components in the time series signal, mainly including the dynamic changes in blood flow velocity and blood flow concentration, which are the core components related to tumor susceptibility characteristics.

[0047] Thresholding is performed on the high-frequency components. An adaptive threshold is set. The threshold is calculated by first calculating the standard deviation of each high-frequency component, and then multiplying the standard deviation by a preset coefficient (the coefficient is between 1.2 and 1.8, preferably 1.5) to obtain the adaptive threshold for that high-frequency component. The coefficients of the high-frequency components whose absolute values ​​are less than the threshold are set to 0, while the coefficients whose absolute values ​​are greater than or equal to the threshold are retained. This achieves noise suppression in the high-frequency components while preserving the core dynamic components. The processed high-frequency components and the original low-frequency components are subjected to inverse wavelet transform. The processed high-frequency components are added together and then fused with the low-frequency components to obtain the reconstructed time series sequence. This reconstructed sequence is the separated dynamic component time series sequence. The original low-frequency components are the static component time series sequence. After separation, the dynamic component time series sequence of each grid point is integrated with the spatial coordinate parameters and time parameters of that grid point to form a dynamic component feature subset.

[0048] In this embodiment, the global adaptive compensation amount generated by weighting coefficients and local heterogeneity adjustment parameters can comprehensively reflect the overall blood flow heterogeneity level of the entire monitoring area, realize the adaptive adjustment of the compensation amount, ensure that the compensation effect fits the actual situation of the monitoring area, and effectively suppress the interference caused by global blood flow heterogeneity.

[0049] In a preferred embodiment of the present invention, step 6, which involves sorting the Doppler frequency shift feature parameters in the dynamic component feature subset by time and inputting them into a pre-trained improved dense connection network model to obtain dynamic monitoring results of tumor susceptibility, may include: Step 601: Extract the compensated Doppler frequency shift feature parameters carried by all units from the dynamic component feature subset, and sort them in ascending order according to the time parameter corresponding to each compensated Doppler frequency shift feature parameter to obtain a one-dimensional time series signal. Specifically, this includes: obtaining the complete dynamic component feature subset set separated in step 503, which contains the dynamic component time series sequence, spatial coordinate parameters, and time parameters after integrating all grid points; traversing the dynamic component feature subset corresponding to each grid point in the dynamic component feature subset set one by one, and extracting the compensated Doppler frequency shift feature parameters carried by all units from each dynamic component feature subset. The compensated Doppler frequency shift feature parameters are obtained after the compensation processing in step 502, and contain the compensated peak value and compensated mean value of each unit, which reflect the dynamic changes in blood flow. The core data of the process is also a key parameter closely related to tumor susceptibility characteristics. After extraction, all extracted compensated Doppler frequency shift feature parameters are collected, and the time parameter corresponding to each compensated Doppler frequency shift feature parameter is extracted. This time parameter is the time parameter at the center of the sliding window in step 103, used to characterize the monitoring time point corresponding to each Doppler frequency shift feature parameter. All compensated Doppler frequency shift feature parameters are sorted according to the time parameter corresponding to each parameter, arranged in ascending order from earliest to latest. During the sorting process, it is ensured that each compensated Doppler frequency shift feature parameter corresponds one-to-one with its corresponding time parameter, without any misalignment. After sorting, all sorted compensated Doppler frequency shift feature parameters are concatenated sequentially to form a one-dimensional time series signal.

[0050] Step 602 involves normalizing the one-dimensional time series signal, linearly mapping its numerical range to a preset standard interval to obtain a normalized dynamic feature sequence. This includes: defining the preset standard interval; considering the adaptation requirements of feature data in tumor susceptibility monitoring, the preset standard interval is set between 0 and 1. This interval effectively eliminates interference from different numerical ranges, ensuring the accuracy of subsequent model training and inference; performing numerical statistics on the one-dimensional time series signal obtained in step 601; first, identifying the maximum and minimum values ​​of all compensated Doppler frequency shift feature parameters in the one-dimensional time series signal; comparing the value of each parameter one by one, recording the parameter with the largest value as the maximum value and the parameter with the smallest value as the minimum value; and calculating the normalization process. The denominator is calculated by subtracting the minimum value from the maximum value to obtain the numerical difference. If the numerical difference is 0, it means that all compensated Doppler frequency shift characteristic parameters have the same value. In this case, the values ​​of all parameters are uniformly set to 0.5 to ensure that the normalized data is within the preset standard range. If the numerical difference is not 0, the normalization calculation is performed on each compensated Doppler frequency shift characteristic parameter in the one-dimensional time series signal. The calculation method is to subtract the minimum value from the value of each compensated Doppler frequency shift characteristic parameter, and then divide the difference by the previously calculated numerical difference. Through this calculation process, the value of each parameter is linearly mapped to the preset standard range of 0 to 1. After all parameters have been normalized, they are concatenated in the original time order to form a normalized dynamic characteristic sequence.

[0051] Step 603: The normalized dynamic feature sequence is input into a pre-trained improved densely connected network model. The improved densely connected network model sequentially includes an initial convolutional layer, multiple densely connected blocks, multiple transition layers, a global average pooling layer, and a fully connected output layer. The normalized dynamic feature sequence is processed by the initial convolutional layer to extract shallow local temporal features, resulting in a shallow feature tensor. The shallow feature tensor is then transformed step-by-step through each densely connected block and each transition layer to obtain a deep feature tensor. The deep feature tensor is compressed into a fixed-length feature vector by the global average pooling layer. The fixed-length feature vector is then processed by the activation function of the fully connected output layer to output a two-dimensional probability vector, which includes a tumor susceptibility positive category. The probability of negative tumor susceptibility category and the probability of negative tumor susceptibility category are specifically included: clarifying the detailed construction, training and implementation process of the improved dense connection network model. This model is based on the traditional dense connection network and is adapted to the extraction of Doppler frequency shift temporal features and the judgment of tumor susceptibility category. Its detailed construction process is as follows: the initial convolutional layer is used to extract shallow local temporal features. The convolutional kernel size is set to 3, the number of convolutional kernels is 64, the stride is 1, and the same padding is used. Specifically, when the initial convolutional layer performs convolution operation on the normalized dynamic feature sequence, an appropriate number of 0s are padded at both ends of the sequence so that after the convolution operation is completed, the length of the output shallow feature tensor is completely consistent with the length of the input normalized dynamic feature sequence, ensuring that the convolution is performed correctly. The length of the subsequent feature sequence remains unchanged, and the ReLU activation function is used to alleviate the gradient vanishing problem and lay the foundation for subsequent feature extraction. Multiple densely connected blocks are used to extract deep temporal features. Four densely connected blocks are used, each containing six convolutional layers. Each convolutional layer has a kernel size of 3, with the kernel counts being 64, 64, 128, 128, 256, and 256 respectively. Each convolutional layer is followed by a ReLU activation function and a batch normalization layer. The batch normalization layer accelerates model convergence and reduces overfitting. Each densely connected block uses a dense connection approach, meaning that the input of each convolutional layer contains the outputs of all previous convolutional layers within that block. This fully utilizes shallow features and improves feature extraction efficiency. The integrity and relevance of the input are ensured; multiple transition layers correspond one-to-one with densely connected blocks, with four transition layers set up. Each transition layer contains a convolutional layer and a pooling layer. The convolutional kernel size of the convolutional layer is 1, and the number of convolutional kernels is half the number of output features of the last convolutional layer of the previous densely connected block, which is used to compress the feature dimension and reduce the amount of computation. The pooling layer uses average pooling with a pooling kernel size of 2 and a stride of 2, which is used to retain core features and reduce the feature dimension. The global average pooling layer is used to compress the deep feature tensor into a fixed-length feature vector. The global average pooling method is used to average the calculation of each feature channel of the deep feature tensor to obtain a fixed-length feature vector. The fixed length is set to 256 to ensure that the input format of the subsequent fully connected layers is consistent.The fully connected output layer outputs the probability of tumor susceptibility category. It consists of two fully connected layers: the first layer has 128 neurons and uses the ReLU activation function, while the second layer has 2 neurons and uses the softmax activation function. This maps the feature vector to a two-dimensional probability vector, corresponding to the probabilities of positive and negative tumor susceptibility categories.

[0052] The training process of the improved densely connected network model is as follows: A training dataset is constructed, which is derived from the dynamic component feature subsets of tumor-related samples and normal samples of different types and stages. Following the methods in steps 601 and 602, the compensated Doppler frequency shift feature parameters of each sample are arranged in ascending order of time and normalized to obtain the normalized dynamic feature sequence, which is used as the input data of the model. At the same time, each sample is labeled, with tumor-related samples labeled as tumor susceptibility positive and normal samples labeled as tumor susceptibility negative, thus forming a complete training dataset. The training dataset was randomly divided into training and test sets in a 7:3 ratio to ensure dataset diversity and representativeness. Model parameters were initialized using Xavier initialization for all convolutional and fully connected layer weights to avoid gradient vanishing due to excessively large or small initial weights. Bias terms were initialized to 0, and batch normalization layer parameters were initialized to default values. Core training parameters were set: the cross-entropy loss function was used to measure the deviation between the model's predictions and the true labels; the Adam optimizer was used to adaptively adjust the learning rate and accelerate model convergence; the number of iterations was set to 100 epochs, and the batch size to 32. Accuracy on both the training and test sets was calculated every 10 epochs to monitor model training performance in real time. An early stopping strategy was employed: if the test set accuracy did not improve for 10 consecutive epochs, training was stopped to prevent overfitting. The model parameters at this point were saved as a pre-trained improved densely connected network model.

[0053] After model training is complete, the normalized dynamic feature sequence obtained in step 602 is input into the pre-trained improved densely connected network model. The process is as follows: the normalized dynamic feature sequence is input into the initial convolutional layer. Through convolution operations and the ReLU activation function of the initial convolutional layer, shallow local temporal features in the one-dimensional time series are extracted. These shallow local temporal features include basic features such as peak abrupt changes in Doppler shift and mean fluctuations. After extraction, a shallow feature tensor is obtained. The shallow feature tensor is then input into the first densely connected block. Through the progressive operations of multiple convolutional layers within the densely connected block, the shallow features are fully integrated to extract more representative mid-level temporal features. The mid-level temporal features undergo convolution and pooling processing in the first transition layer to compress the feature dimension while retaining the core features, resulting in a compressed mid-level feature tensor. This feature tensor then passes through the remaining three densely connected blocks and three transition layers. Each time it passes through a densely connected block, the feature... The correlation and representativeness are further improved. After each transition layer, the feature dimension is further compressed, finally obtaining a deep feature tensor. This deep feature tensor contains deep correlation information between Doppler frequency shift temporal features and tumor susceptibility features. The deep feature tensor is input to a global average pooling layer. Through global average pooling operation, each feature channel is averaged, compressing the deep feature tensor into a feature vector of fixed length 256, eliminating the interference caused by the difference in feature dimension. The fixed-length feature vector is then input to a fully connected output layer. After the operation of the first fully connected layer and the ReLU activation function, the feature vector is dimensionality reduced to a 128-dimensional feature vector. After the operation of the second fully connected layer and the softmax activation function, the 128-dimensional feature vector is mapped to a two-dimensional probability vector. The two values ​​of this two-dimensional probability vector correspond to the probability of the tumor susceptibility positive category and the probability of the tumor susceptibility negative category, respectively. The sum of the two probabilities is 1.

[0054] Step 604: Compare the probabilities of the positive and negative tumor susceptibility categories in the two-dimensional probability vector, and output the category with the larger probability value as the result of dynamic tumor susceptibility monitoring. Specifically, this includes: obtaining the two-dimensional probability vector output in step 603; identifying the corresponding categories of the two values ​​in the vector, where the first value corresponds to the probability of the positive tumor susceptibility category and the second value corresponds to the probability of the negative tumor susceptibility category; comparing these two probability values ​​one by one; if the probability of the positive tumor susceptibility category is greater than the probability of the negative tumor susceptibility category, it indicates that the Doppler shift state of the monitored sample is positive. If the correlation between the Doppler frequency shift feature and the positive tumor susceptibility feature is stronger, then the positive tumor susceptibility feature is output as the dynamic monitoring result of tumor susceptibility for this monitoring sample. If the probability of the negative tumor susceptibility category is greater than the probability of the positive tumor susceptibility category, it indicates that the correlation between the Doppler frequency shift feature and the negative tumor susceptibility feature of this monitoring sample is stronger, then the negative tumor susceptibility feature is output as the dynamic monitoring result of tumor susceptibility for this monitoring sample. If the two probability values ​​are equal, it indicates that the Doppler frequency shift feature of this monitoring sample cannot clearly distinguish the tumor susceptibility category, then a prompt message is output, suggesting that further monitoring data needs to be supplemented and the judgment re-evaluated.

[0055] This embodiment achieves accurate and dynamic monitoring of tumor susceptibility characteristics by constructing, training, optimizing, dynamically evaluating, and adjusting the tumor susceptibility monitoring model, thus forming a closed-loop monitoring system.

[0056] like Figure 2 As shown, embodiments of the present invention also provide a machine learning-based dynamic monitoring system for tumor susceptibility, comprising: The module is used to acquire dynamic monitoring data of the subjects and transform the dynamic monitoring data into a spatial discrete set in three-dimensional space. Each unit in the spatial discrete set contains spatial coordinate parameters, time parameters, and Doppler frequency shift feature parameters. The fitting module is used to fit the spatial coordinate parameters of all units in the spatial discrete set and construct a spatial trend feature analytical hypersurface. The partitioning module is used to determine a boundary-constrained four-way domain based on the projection boundaries of each unit in the spatial discrete set on the spatial trend feature analytic hyperplane; and to partition the boundary-constrained four-way domain based on the projection clustering density distribution of the spatial coordinate parameters of each unit in the spatial discrete set on the spatial trend feature analytic hyperplane, thereby obtaining multiple grid point heterogeneous partitioning blocks. The computation module is used to map each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block to obtain discrete attribution feature subsets; extract the distribution moment features of the Doppler frequency shift feature parameter of each discrete attribution feature subset, and calculate the local heterogeneity adjustment parameter; The separation module is used to weight and aggregate all local heterogeneity adjustment parameters to generate a global adaptive compensation quantity; based on the global adaptive compensation quantity, the spatial discrete set is separated to obtain a dynamic component feature subset; The output module is used to sort the Doppler frequency shift feature parameters in the dynamic component feature subset by time and input them into a pre-trained improved dense connection network model to obtain the dynamic monitoring results of tumor susceptibility.

[0057] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0058] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0059] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0060] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring of tumor susceptibility based on machine learning, characterized in that, The method includes: Step 1: Acquire the dynamic monitoring data of the subjects and transform the dynamic monitoring data into a spatial discrete set in three-dimensional space. Each unit in the spatial discrete set contains spatial coordinate parameters, time parameters, and Doppler frequency shift characteristic parameters. Step 2: Fit the spatial coordinate parameters of all units in the spatial discrete set to construct a spatial trend feature analytical hypersurface; Step 3: Determine a boundary-constrained four-way domain based on the projection boundaries of each unit in the spatial discrete set onto the spatial trend feature analytic hyperplane; and divide the boundary-constrained four-way domain into multiple grid point heterogeneous partitioning blocks based on the projection clustering density distribution of the spatial coordinate parameters of each unit in the spatial discrete set onto the spatial trend feature analytic hyperplane. Step 4: Map each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block to obtain discrete attribution feature subsets; extract the distribution moment features of the Doppler frequency shift feature parameter of each discrete attribution feature subset, and calculate the local heterogeneity adjustment parameter; Step 5: Weighted aggregation of all local heterogeneity adjustment parameters to generate a global adaptive compensation quantity; Based on the global adaptive compensation quantity, the spatial discrete set is separated to obtain a dynamic component feature subset; Step 6: After sorting the Doppler frequency shift feature parameters in the dynamic component feature subset by time, input them into the pre-trained improved dense connection network model to obtain the dynamic monitoring results of tumor susceptibility.

2. The method for dynamic monitoring of tumor susceptibility based on machine learning according to claim 1, characterized in that, The dynamic monitoring data of the subjects is acquired and transformed into a spatial discrete set in three-dimensional space. Each cell in the spatial discrete set contains spatial coordinate parameters, time parameters, and Doppler frequency shift characteristic parameters, including: Acquire dynamic monitoring data of the subject, wherein the dynamic monitoring data is continuously acquired Doppler ultrasound time-series signals, and each acquisition time corresponds to a set of spatial location codes; Based on the spatial location encoding, the spatial sampling points at each acquisition time are resampled in a grid to obtain uniformly distributed three-dimensional spatial grid points. Each three-dimensional spatial grid point carries the interpolation result of the original Doppler frequency shift. Based on the interpolation result, sliding window energy detection is performed on the time axis to extract the peak value and mean value of the Doppler frequency shift within each sliding window. The combination of the peak value and mean value is used as the Doppler frequency shift feature parameter of the corresponding three-dimensional spatial grid point at the center time of the sliding window, resulting in a time-sorted sequence of Doppler frequency shift feature parameters. Based on the Doppler frequency shift characteristic parameter sequence, the three-dimensional spatial coordinates corresponding to each Doppler frequency shift characteristic parameter are taken as spatial coordinate parameters, the corresponding sliding window center time is taken as time parameters, and the Doppler frequency shift characteristic parameter itself is taken as the third dimension. The spatial coordinate parameters, time parameters, and Doppler frequency shift characteristic parameters are collectively encapsulated into a unit. All units are organized according to spatial position and time order to obtain a spatial discrete set in three-dimensional space.

3. The method for dynamic monitoring of tumor susceptibility based on machine learning according to claim 2, characterized in that, By fitting the spatial coordinate parameters of all cells in the spatial discrete set, a spatial trend feature analytical hypersurface is constructed, including: Spatial coordinate parameters of all units are extracted from the spatial discrete set, and each spatial coordinate parameter contains three-dimensional spatial coordinate values; all spatial coordinate parameters are centered to obtain a set of centered spatial coordinate parameters. A three-dimensional quadratic surface is fitted to the centered set of spatial coordinate parameters. During the fitting process, two horizontal dimensions of the three dimensions of the spatial coordinate parameters are used as independent variables and the third vertical dimension is used as the dependent variable to obtain the fitting coefficients. Based on the fitting coefficients, a continuous three-dimensional spatial surface function is constructed, which is the spatial trend feature analytical hyperface.

4. The method for dynamic monitoring of tumor susceptibility based on machine learning according to claim 3, characterized in that, Step 3 includes: Using the spatial trend feature analysis hyperplane as the projection plane, the spatial coordinate parameters of each unit in the spatial discrete set are vertically projected onto the spatial trend feature analysis hyperplane to obtain the projection point corresponding to each unit. Extract the minimum bounding rectangle boundary of all projection points, and use the minimum bounding rectangle boundary as the constraint boundary to determine a rectangular region on the spatial trend feature analysis hyperplane. The rectangular region is the boundary constraint four-way domain. Density-based spatial clustering is performed on all projection points to obtain multiple projection point clusters. Each projection point cluster corresponds to a high-density clustering region. The minimum bounding closure region is calculated for each projection point cluster to obtain multiple minimum bounding closure regions. Intersection calculation is performed between each minimum bounding closed region and the boundary constraint four-way domain to obtain the intersection closed region of each minimum bounding closed region within the boundary constraint four-way domain; the inclusion relationship between each intersection closed region is determined, and each intersection closed region is marked as a root closed region or a sub-closed region, wherein a region not included by any other intersection closed region is marked as a root closed region, and a region included by at least one other intersection closed region is marked as a sub-closed region. Remove all sub-closed regions directly embedded within the root closed region to obtain a set of regions with a porous structure; take each removed sub-closed region as input and recursively perform the removal operation, that is, remove the innermost sub-closed regions directly embedded within the sub-closed region from the sub-closed region until there are no more removable sub-closed regions; collect all non-overlapping regions generated in the entire recursive process as grid point heterogeneity partition blocks.

5. The method for dynamic monitoring of tumor susceptibility based on machine learning according to claim 4, characterized in that, By mapping each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block, a discrete attribution feature subset is obtained; Extract the distribution moment features of the Doppler frequency shift feature parameters of each discrete subset of the attributed features, and calculate the local heterogeneity adjustment parameters, including: Based on the heterogeneity of each grid point, the blocks are divided. All units in the spatial discrete set are traversed. It is determined whether the projection point of the spatial coordinate parameter of each unit on the spatial trend feature analysis hyperplane falls within the grid point heterogeneity division block. If it falls within the block, the unit is assigned to the set corresponding to the block. After traversing all blocks, each block corresponds to a set of units. The set of units is the discrete attribution feature subset. Extract the Doppler frequency shift feature parameters of all units in each discrete attribution feature subset, and calculate the first-order raw moment, second-order central moment and third-order central moment of the Doppler frequency shift feature parameters, which are used as mean feature, variance feature and skewness feature, respectively. The ratio of variance to mean is calculated to obtain the coefficient of variation; the product of the absolute value of skewness and the coefficient of variation is calculated to obtain the initial value of local heterogeneity; the initial value of local heterogeneity and the area of ​​the grid heterogeneity partition corresponding to the discrete subset of the feature are normalized to obtain the local heterogeneity adjustment parameter for each discrete subset of the feature.

6. The method for dynamic monitoring of tumor susceptibility based on machine learning according to claim 5, characterized in that, All local heterogeneity adjustment parameters are weighted and aggregated to generate a global adaptive compensation quantity; Based on the global adaptive compensation amount, the spatial discrete set is separated to obtain a dynamic component feature subset, including: Calculate the proportion of the area of ​​the grid heterogeneity partition block corresponding to each discrete attribution feature subset to the total area of ​​all grid heterogeneity partition blocks, and use the proportion as the weighting coefficient of each discrete attribution feature subset; calculate the weighted local heterogeneity adjustment parameter by combining the local heterogeneity adjustment parameter of each discrete attribution feature subset with the weight of each discrete attribution feature subset; sum all the weighted local heterogeneity adjustment parameters to generate the global adaptive compensation quantity; Subtract the global adaptive compensation amount from the Doppler frequency shift feature parameter of each cell in the spatial discrete set to obtain the compensated Doppler frequency shift feature parameter of each cell; repackage the compensated Doppler frequency shift feature parameter of each cell with the spatial coordinate parameter and time parameter of each cell to obtain the compensated spatial discrete set; Calculate the mean and standard deviation of the compensated Doppler frequency shift characteristic parameters of all elements in the compensated spatial discrete set, and use the sum of the mean and standard deviation by a preset multiple as the dynamic threshold; extract all elements in the compensated spatial discrete set whose absolute value of the compensated Doppler frequency shift characteristic parameters is greater than the dynamic threshold, and use the set of the extracted elements' compensated Doppler frequency shift characteristic parameters sorted by time as the dynamic component feature subset.

7. The method for dynamic monitoring of tumor susceptibility based on machine learning according to claim 6, characterized in that, Step 6 includes: The compensated Doppler frequency shift feature parameters carried by all units are extracted from the dynamic component feature subset, and then sorted in ascending order according to the time parameter corresponding to each compensated Doppler frequency shift feature parameter to obtain a one-dimensional time series signal. Normalization is performed on a one-dimensional time series signal to linearly map the numerical range of the one-dimensional time series signal to a preset standard interval, thereby obtaining a normalized dynamic feature sequence. The normalized dynamic feature sequence is input into a pre-trained improved densely connected network model. This improved densely connected network model sequentially comprises an initial convolutional layer, multiple densely connected blocks, multiple transition layers, a global average pooling layer, and a fully connected output layer. The normalized dynamic feature sequence is processed by the initial convolutional layer to extract shallow local temporal features, resulting in a shallow feature tensor. This shallow feature tensor is then transformed step-by-step through each densely connected block and each transition layer to obtain a deep feature tensor. The deep feature tensor is compressed into a fixed-length feature vector by the global average pooling layer. This fixed-length feature vector is then processed by the activation function of the fully connected output layer to output a two-dimensional probability vector, which includes the probability of a positive tumor susceptibility category and the probability of a negative tumor susceptibility category. The probability of the positive tumor susceptibility category is compared with the probability of the negative tumor susceptibility category in the two-dimensional probability vector. The category with the larger probability value is output as the result of dynamic monitoring of tumor susceptibility.

8. A machine learning-based dynamic monitoring system for tumor susceptibility, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The module is used to acquire dynamic monitoring data of the subjects and transform the dynamic monitoring data into a spatial discrete set in three-dimensional space. Each unit in the spatial discrete set contains spatial coordinate parameters, time parameters, and Doppler frequency shift feature parameters. The fitting module is used to fit the spatial coordinate parameters of all units in the spatial discrete set and construct a spatial trend feature analytical hypersurface. The partitioning module is used to determine a boundary-constrained four-dimensional domain based on the projected boundaries of each unit in the spatial discrete set on the spatial trend characteristic analytical hyperface. Based on the projection clustering density distribution of the spatial coordinate parameters of each unit in the spatial discrete set onto the spatial trend feature analysis hyperplane, the boundary constraint four-directional domain is partitioned to obtain multiple grid point heterogeneous partitioning blocks. The computation module is used to map each unit in the spatial discrete set to the corresponding grid point heterogeneity partitioning block to obtain a discrete subset of belonging features; Extract the distribution moment features of the Doppler frequency shift feature parameters of each discrete attribution feature subset, and calculate the local heterogeneity adjustment parameter; The separation module is used to weight and aggregate all local heterogeneity adjustment parameters to generate a global adaptive compensation quantity. Based on the global adaptive compensation amount, the spatial discrete set is separated to obtain a dynamic component feature subset; The output module is used to sort the Doppler frequency shift feature parameters in the dynamic component feature subset by time and input them into a pre-trained improved dense connection network model to obtain the dynamic monitoring results of tumor susceptibility.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images