General surgery department abdominal trauma grading evaluation method based on machine learning

Abstract

The invention relates to the technical field of intelligent surgery, in particular to a general surgery department abdominal trauma grading evaluation method based on machine learning, which comprises the following steps: constructing an anatomical model based on a medical image, analyzing vascular topological affinity, quantifying liquid dispersion amount, and calculating an artery compensation ratio after portal vein occlusion, according to the method, the two-dimensional image is analyzed into the three-dimensional voxel matrix, the voxels of the vascular skeleton and the lesion are extracted, the topological homology characteristics between the blood vessel and the lesion are captured, the spatial affinity of the damaged tissue to the blood vessel network is quantified, and the damage assessment accuracy is improved. The method comprises the following steps: screening a liquid dispersion region conforming to physical permeation characteristics, evaluating tissue survival potential according to a compensatory perfusion proportion of a disjunction region, constructing a theoretical extremely-dangerous critical state vector containing multi-dimensional physiological characteristics, converting a trauma severity into a measurable spatial deviation value, and calculating the trauma severity. And precise quantitative grading from anatomical morphology to functional compensation dimension is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent surgical technology, and in particular to a machine learning-based method for grading and assessing abdominal trauma in general surgery. Background Technology

[0002] The field of intelligent surgical technology encompasses a medical practice system that deeply integrates artificial intelligence, computer vision, and robotics with modern surgery. It involves using computer-aided digital reconstruction and analysis of multimodal medical images, assisting doctors in performing precise operations through surgical robots or optical navigation, using algorithmic models to quantitatively analyze surgical risks, lesion characteristics, and prognoses, and constructing a full-process digital diagnosis and treatment environment covering preoperative planning, real-time intraoperative assistance, and postoperative rehabilitation assessment. The traditional general surgery abdominal trauma grading assessment method refers to the clinical grading of the severity of damage to solid or hollow organs in the abdomen after external impact. Clinicians mainly rely on the patient's enhanced CT scan images of the abdomen to visually observe and identify the depth of organ lacerations, the surface area ratio of subcapsular hematoma, the diameter of intrastantial hematoma, and whether there is active bleeding and contrast agent leakage. Combined with the trauma organ injury grading standards, the anatomical and geometric parameters of the injured area are manually measured on the imaging workstation using calipers. Then, the grading table is checked item by item to determine the specific injury level. At the same time, the patient's systolic blood pressure, pulse rate, and abdominal physical examination data are combined for comprehensive judgment to complete the determination of the severity of the condition.

[0003] Traditional manual assessment mainly relies on two-dimensional tomographic images for visual judgment, which cannot accurately reconstruct the three-dimensional non-Euclidean spatial morphology of organ damage. Measuring the depth of lacerations or the diameter of hematomas on a single level with calipers alone is insufficient to quantify the volume of irregular wounds. Manual subjective interpretation of images is easily affected by differences in experience, leading to the omission of key anatomical details. Simple geometric morphological measurements ignore the topological connectivity between the vascular skeleton and damaged tissues. Static anatomical parameters cannot dynamically reflect the compensatory perfusion function of blood vessels and the microscopic diffusion behavior of fluid media in the interstitial space. Rigid lookup-based grading cannot accurately define the pathological critical state of complex complications, resulting in a lag bias in the quantitative rating of the severity of the condition. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a machine learning-based method for grading and assessing abdominal trauma in general surgery, comprising the following steps: S1: Acquire abdominal enhanced tomographic scan images, parse them into a three-dimensional voxel matrix, segment them into a high-density voxel set, a low-density voxel set and a background voxel set according to density distribution characteristics, and construct the portal vein skeleton and arterial skeleton based on the background voxel set, and output anatomical structure feature data. S2: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton, transform the distance between the portal vein skeleton and the arterial skeleton and map it to the low-density voxel set, establish a simple complex filter flow according to the distance, count the number of existing features, and generate a vascular affinity topology vector. S3: Call the anatomical structure feature data, calculate the edge gradient vector of the high-density voxel set and the decay rate of the magnitude sequence along the gradient vector, filter the voxel set with linear and gradual decay and no abrupt changes and calculate the volume, and output the liquid diffusion volume parameters. S4: Call the anatomical structure feature data, identify the portal vein interruption node and downstream fragmented voxel set truncated by the low-density voxel set, count the number of fragmented voxels located within the perfusion radius of the arterial skeleton and calculate the proportion, and output the compensatory perfusion ratio. S5: Normalize the vascular affinity topology vector, liquid diffusion volume parameter, and compensatory perfusion ratio, combine them into a multidimensional feature vector, construct a critical state vector and calculate the weighted distance with the multidimensional feature vector, match the grading interval and output the abdominal trauma grading assessment result.

[0005] As a further aspect of the present invention, the vascular affinity topology vector includes the number of near-vascular connectivity components, the number of near-vascular topology loops, and the length of the topological feature persistence interval; the liquid diffusion volume parameter includes the volume value of the liquid permeation region, the gradient continuity index of the permeation region edge, and the average density value of the permeation region; the compensatory perfusion ratio includes the total volume of the portal vein injury region, the volume of the arterial compensation coverage region, and the percentage of functional preservation; and the abdominal trauma grading assessment results include trauma severity grading labels, multidimensional feature weighted comprehensive scores, and anatomical risk and functional preservation map data.

[0006] As a further aspect of the present invention, the step of obtaining the anatomical structure feature data specifically includes: S101: Acquire abdominal enhanced tomographic scan images and parse them into a three-dimensional voxel matrix. Analyze the gray intensity distribution characteristics of voxel units within the matrix. Classify and filter various tissues based on density differences under contrast conditions. Divide the three-dimensional voxel matrix into a high-density voxel set, a low-density voxel set, and a background voxel set to generate a basic voxel segmentation set. S102: Call the basic voxel segmentation set and extract the background voxel set, perform morphological thinning and shrinking processing on the background voxel set, strip non-core tissue pixels and retain single-pixel width connected paths, calculate the geometric center line coordinates of the residual paths, track and classify the center lines according to the vascular anatomy connectivity rules, construct the portal vein connectivity network and arterial connectivity network, and generate vascular skeleton data. S103: Call the high-density voxel set and low-density voxel set in the basic voxel segmentation set, combine them with the vascular skeleton data, establish a three-dimensional spatial mapping coordinate system, map the spatial distribution coordinates of each voxel set to the coordinate system, perform spatial position registration with the vascular skeleton data as a reference, fuse the physical density attributes and skeleton topology information of each group of voxels, and generate anatomical structure feature data.

[0007] As a further aspect of the present invention, the process of classifying and screening various tissues based on density differences under imaging conditions, and dividing the three-dimensional voxel matrix into a high-density voxel set, a low-density voxel set, and a background voxel set, specifically involves: The gray intensity values ​​of all voxel units in the three-dimensional voxel matrix are statistically analyzed to establish a global gray-level statistical histogram reflecting the overall distribution pattern of abdominal tissue density. The probability density function of the global gray-level statistical histogram is fitted using a Gaussian mixture model to extract three independent Gaussian distribution components, and the gray-level mean of each Gaussian distribution component is calculated. The three Gaussian distribution components are arranged in order of increasing gray value, and are defined as low attenuation distribution component, background distribution component and high attenuation distribution component respectively. Calculate the gray value of the intersection point of the probability density curves between the low-attenuation distribution component and the background distribution component, and set it as the low-density segmentation threshold value. Calculate the gray value of the intersection point of the probability density curves between the background distribution component and the high attenuation distribution component, and set it as the high density segmentation threshold value; Traverse the voxel units in the three-dimensional voxel matrix and extract the voxel units whose gray intensity values ​​are greater than the high-density segmentation threshold value to the high-density voxel set. Voxel units with gray intensity values ​​less than the low-density segmentation threshold are extracted into a low-density voxel set. Voxel units whose grayscale intensity values ​​fall within the closed interval between the low-density segmentation threshold and the high-density segmentation threshold are extracted into the background voxel set.

[0008] As a further aspect of the present invention, the step of obtaining the vascular affinity topology vector specifically comprises: S201: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton. Establish a three-dimensional spatial distance field with the vascular skeleton as a reference. Calculate the Euclidean distance parameter from all voxels in the field to the nearest vascular skeleton point. Traverse the voxel units in the low-density voxel set, retrieve the corresponding position coordinates of each voxel unit in the distance field, assign the Euclidean distance parameter to the corresponding low-density voxel, and generate a vascular distance attribute mapping set. S202: Call the blood vessel distance attribute mapping set, sort the voxels in the mapping set in ascending order of distance parameters, construct a simple complex evolution sequence, analyze the morphological changes of connected components and topological holes during the sequence evolution, record the generation and extinction times of each topological structure in the simple complex, obtain the persistence characteristics of the laceration structure as the blood vessel distance changes, and generate topological feature persistence interval data. S203: Call the continuous interval data of the topological features, filter the topological features whose generation time is within the range of the starting parameters, classify and count the number of coherent classes in different dimensions within the target range, analyze the distribution density of the topological features on the distance parameter axis, arrange and combine the statistically obtained quantity indicators and distribution density indicators in a preset order, construct a digital vector, and generate a vascular affinity topological vector.

[0009] As a further aspect of the present invention, the step of obtaining the liquid dispersion volume parameter specifically includes: S301: Call the anatomical structure feature data and extract the high-density voxel set, filter the edge voxels in the set that are at the junction of high-density tissue and background, calculate the gray-level change gradient of the edge voxels in the local three-dimensional environment, construct vector data containing the direction of the maximum rate of change and gradient intensity information, and generate edge gradient vector field data. S302: Call the edge gradient vector field data, analyze the pointing distribution of the gradient vector in three-dimensional space, collect the gradient intensity of continuous neighborhood voxels along the extension direction of the gradient vector, construct the intensity change sequence distributed along the spatial distance, calculate the differential decay rate of the gradient intensity relative to the spatial position in the sequence, and generate a spatial decay rate sequence. S303: Call the spatial decay rate sequence, analyze the evolution of gradient intensity along the spatial extension path, select a set of voxels whose gradient magnitude decreases linearly and gradually with distance and without abrupt changes as liquid dispersion characteristic regions, calculate the cumulative space occupancy, and generate liquid dispersion volume parameters.

[0010] As a further aspect of the present invention, the process of selecting a set of voxels whose gradient magnitude decreases linearly and smoothly with distance and without abrupt changes as liquid dispersion feature regions specifically involves: The discrete gradient magnitude data contained in the spatial decay rate sequence are fitted by least squares linear regression to calculate the linear correlation coefficient and the slope of the fitted line, which characterize the overall trend of the sequence. Calculate the absolute difference between the gradient magnitudes of adjacent spatial locations in the spatial decay rate sequence point by point to construct a first-order difference sequence that reflects the local numerical jump characteristics; The arithmetic mean of the gradient magnitudes of all voxel units in the background voxel set is calculated and set as the reference benchmark for gradual decay. The preset lower limit of the linear correlation coefficient is set as the linear correlation threshold. Based on the numerical distribution characteristics of the first-order difference sequence, the average value and standard deviation of the absolute values ​​of the differences within the sequence are calculated, and the sum of the average value and three times the standard deviation is set as the dynamic step judgment threshold. Sequences whose linear correlation coefficient is greater than the linear correlation threshold, whose slope value is negative and whose absolute value is less than the smooth decay reference are selected and determined to meet the linear smooth decay condition. Traverse the first-order difference sequence corresponding to the sequence that satisfies the linear gradual decay condition, check whether there are any numerical points in the sequence that are greater than the dynamic step judgment threshold, and remove the sequences that have such numerical points. The set of voxel units corresponding to sequences that are filtered by linear and gradual decay conditions and do not contain the numerical points are extracted and marked as liquid dispersion feature regions.

[0011] As a further aspect of the present invention, the step of obtaining the compensatory perfusion ratio specifically includes: S401: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton, establish the three-dimensional spatial superposition mapping relationship between the low-density voxel set and the portal vein vascular skeleton, analyze the path connectivity status of the vascular skeleton in the low-density area, determine the spatial geometric position of the physical truncation of the skeleton path, mark the breakpoints of missing upstream connectivity and use them as the starting point of blood flow transmission interruption, and generate a set of portal vein interruption nodes. S402: Call the set of portal vein interruption nodes, perform downstream path traversal operation according to the topological extension direction of the vascular anatomy, search for the distribution area of ​​the vascular branch network downstream of the interruption node, collect all voxel units within the extension range of the branch network, construct the spatial distribution data of the damaged tissue that has lost portal vein connectivity, and generate the downstream interrupted voxel set. S403: Call the downstream fragmented voxel set and associate it with the arterial skeleton in the anatomical structure feature data, calculate the spatial proximity between the fragmented voxel and the nearest arterial skeleton, screen voxel units located within the coverage area of ​​arterial blood permeation, statistically analyze the proportional relationship between the size of the screened voxels and the total size of the fragmented voxel set, and generate the compensatory perfusion ratio.

[0012] As a further aspect of the present invention, the steps for obtaining the abdominal trauma grading assessment results are specifically as follows: S501: Obtain the vascular affinity topology vector, liquid diffusion volume parameters and compensatory perfusion ratio, analyze the differences in physical dimensions and the characteristics of data distribution range, and perform normalization mapping processing. After fusing the processed multi-source feature data, establish a unified high-dimensional feature space coordinate system and generate multi-dimensional feature vectors. S502: Call the multidimensional feature vector to construct a critical reference benchmark composed of the theoretical limit states of each evaluation dimension, establish a graded evaluation interval sequence containing multiple severity levels, analyze the contribution weight difference of each feature dimension in the trauma severity assessment system, and calculate the weighted Euclidean distance between the current feature vector and the critical reference benchmark in the multidimensional feature space to generate distance difference calculation data. S503: Based on the graded evaluation interval sequence, retrieve the corresponding interval of the distance difference calculation data, determine the severity level of the current trauma status, and generate an abdominal trauma grading assessment result.

[0013] As a further aspect of the present invention, the process of constructing a critical reference benchmark composed of theoretical limit states of each evaluation dimension and establishing a graded evaluation interval sequence containing multiple severity levels specifically involves: Determine the maximum theoretical value in the vascular affinity topology vector that represents the complete overlap between the laceration and the vascular skeleton space, and define it as the topological erosion limit value; The value of the liquid diffusion volume parameter that represents the maximum physical capacity of the abdominal anatomical space is determined and defined as the diffusion capacity limit value. The zero value in the compensatory perfusion ratio, which represents the complete lack of compensation in the artery, is determined and defined as the perfusion loss limit. Normalization is performed on the topological erosion limit value, the diffusion capacity limit value, and the infusion loss limit value, and the critical reference benchmark in the multidimensional feature space is constructed by combining them. Retrieve classified historical abdominal trauma case data, calculate the weighted Euclidean distance between the multidimensional feature vector of each historical case and the critical reference benchmark, and construct a sample distance set; The K-means clustering algorithm is used to iteratively calculate the distance set of the samples to obtain the cluster center values ​​corresponding to different trauma severity, and then sort them in ascending order of value; Calculate the arithmetic mean of the values ​​of two adjacent cluster centers, set it as the hierarchical boundary threshold, and divide the continuous numerical range based on the hierarchical boundary threshold to establish a hierarchical evaluation interval sequence.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by resolving two-dimensional images into a three-dimensional voxel matrix and extracting vascular skeleton and lesion voxels, the topological coherence features between blood vessels and lesions are captured, the spatial affinity of damaged tissue to the vascular network is quantified, liquid diffusion areas that conform to physical permeability characteristics are screened, the tissue survival potential is assessed based on the compensatory perfusion ratio of the detached area, and a theoretical critical state vector containing multidimensional physiological features is constructed. The severity of trauma is transformed into a measurable spatial deviation value, achieving precise quantitative grading from anatomical morphology to functional compensation dimensions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides a machine learning-based method for grading and assessing abdominal trauma in general surgery, comprising the following steps: S1: Acquire abdominal enhanced tomographic scan images, parse them into a three-dimensional voxel matrix, segment them into a high-density voxel set, a low-density voxel set and a background voxel set according to density distribution characteristics, and construct the portal vein skeleton and arterial skeleton based on the background voxel set, and output anatomical structure feature data. S2: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton, transform the distance between the portal vein skeleton and the arterial skeleton and map it to the low-density voxel set, establish a simple complex filter flow according to the distance, count the number of existing features, and generate a vascular affinity topology vector. S3: Call the anatomical structure feature data, calculate the edge gradient vector of the high-density voxel set and the decay rate of the magnitude sequence along the gradient vector, filter the voxel set with linear and gradual decay and no abrupt changes and calculate the volume, and output the liquid diffusion volume parameters. S4: Call anatomical structure feature data, identify the portal vein interruption node and downstream fragmented voxel set truncated by low-density voxel set, count the number of fragmented voxel sets located within the perfusion radius of the arterial skeleton and calculate the proportion, and output the compensatory perfusion ratio. S5: Normalize the vascular affinity topology vector, liquid diffusion volume parameter, and compensatory perfusion ratio, combine them into a multidimensional feature vector, construct the critical state vector and calculate the weighted distance with the multidimensional feature vector, match the grading interval and output the abdominal trauma grading assessment result.

[0023] The vascular affinity topology vector includes the number of near-vascular connectivity components, the number of near-vascular topological loops, and the length of the topological feature persistence interval. The fluid diffusion volume parameters include the volume of the fluid permeation region, the gradient continuity index at the edge of the permeation region, and the average density of the permeation region. The compensatory perfusion ratio includes the total volume of the portal vein injury region, the volume of the arterial compensation coverage region, and the percentage of functional preservation. The abdominal trauma grading assessment results include trauma severity grading labels, multidimensional feature weighted comprehensive scores, and anatomical risk and functional preservation map data.

[0024] Please see Figure 2 The specific steps for obtaining anatomical structural feature data are as follows: S101: Acquire abdominal enhanced tomographic scan images and parse them into a three-dimensional voxel matrix. Analyze the gray intensity distribution characteristics of voxel units within the matrix. Classify and filter various tissues based on density differences under contrast conditions. Divide the three-dimensional voxel matrix into a high-density voxel set, a low-density voxel set, and a background voxel set to generate a basic voxel segmentation set. First, the data interface for medical image storage and transmission is connected to read the DICOM format abdominal enhanced tomographic image sequence. During the reading process, the header information is parsed to obtain pixel spacing, slice thickness, intercept, and slope parameters. Using the intercept and slope parameters, the original stored grayscale values ​​are converted into standard Henlein unit CT values, thereby constructing a three-dimensional voxel matrix covering the entire abdominal region. The dimensions of this matrix are set to 512 x 512 pixels, and the number of layers is set between 200 and 500 layers depending on the scanning range. Next, statistical analysis is performed on the voxel units within the matrix, traversing the grayscale intensity values ​​of approximately 50 million voxel units in the matrix. The statistical results are mapped onto a histogram coordinate system with CT value on the horizontal axis and frequency on the vertical axis. Subsequently, a Gaussian mixture model was introduced to fit the global grayscale statistical histogram. This model was initialized with three Gaussian components, corresponding to low-attenuation substances, soft tissue, and high-attenuation substances, respectively. Iterative calculations were performed using the expectation-maximization algorithm. In the E-step, the posterior probability of each grayscale value belonging to each Gaussian component was calculated. In the M-step, the mean, variance, and mixing weight of each Gaussian component were updated based on the posterior probability. The iterative process continued until the increment of the log-likelihood function was less than a preset convergence threshold of 0.001. For example, after iterative calculations, the mean of the first Gaussian component was -50 HU, representing the fat-fluid mixture region; the mean of the second Gaussian component was 45 HU, representing normal solid organs; and the mean of the third Gaussian component was 120 HU, representing contrast-filled blood vessels and bone. Based on the above mean sorting results, the intersection point of the probability density curves of the first and second components is calculated. Assuming the calculated result of this intersection point is 10 HU, it is set as the low-density segmentation threshold. The intersection point of the probability density curves of the second and third components is calculated. Assuming the calculated result of this intersection point is 90 HU, it is set as the high-density segmentation threshold. Finally, a voxel classification operation is performed, traversing the 3D voxel matrix. Voxels with CT values ​​greater than 90 HU are marked as high-density voxel sets, voxels with CT values ​​less than 10 HU are marked as low-density voxel sets, and voxels with CT values ​​between 10 HU and 90 HU are marked as background voxel sets, thus generating the basic voxel segmentation set.

[0025] S102: Call the basic voxel segmentation set and extract the background voxel set, perform morphological thinning and shrinking processing on the background voxel set, strip non-core tissue pixels and retain connected paths with a single pixel width, calculate the geometric center line coordinates of the residual paths, track and classify the center lines according to the vascular anatomy connectivity rules, construct the portal vein connectivity network and arterial connectivity network, and generate vascular skeleton data. First, a background voxel set is retrieved from the basic voxel segmentation set. This set mainly consists of unreinforced solid organ tissue. A parallel thinning algorithm that preserves the 3D Euler number is applied to it. This algorithm iteratively peels away the boundary voxels on the surface of the voxel set through multiple rounds. In each iteration, the configuration of the 26 neighboring voxels in a 3x3x3 neighborhood centered on the current voxel is checked. If the removal of the current voxel does not change the local connectivity and is not an endpoint, it is removed. This peeling process is repeated until the structure formed by all remaining voxels in the set can no longer be thinned, thus obtaining a 3D connected path with a width of only one pixel. Next, the residual path is geometrically analyzed to extract the 3D spatial coordinates of each voxel on the path, and these coordinates are connected in an orderly manner to form a geometric centerline. Subsequently, the central line is traced according to the tree-like bifurcation rules of vascular anatomy. The starting point of the tracing is set as the entrance of the main portal vein and the entrance of the abdominal aorta. The tracing extends distally along the central line. At each bifurcation node, the bifurcation angle, branches, and their parent-child relationships are recorded. If the bifurcation angle is less than 90 degrees and the branch path length is greater than 5 mm, it is considered a valid vascular branch. This process clearly divides the complex central line network into a portal vein connectivity network and an arterial connectivity network. For example, during the tracing process, a main path with a length of 120 mm and an average diameter of 8 mm is identified as the main portal vein, along with several secondary branches connected to it, each with a length between 20 mm and 50 mm. Finally, the determined coordinate sets of the portal vein network and the arterial network are structured and stored to generate vascular skeleton data containing node coordinates, edge connectivity, and radius information.

[0026] S103: Call the high-density voxel set and low-density voxel set in the basic voxel segmentation set, combine with the vascular skeleton data, establish a three-dimensional spatial mapping coordinate system, map the spatial distribution coordinates of each voxel set to the coordinate system, perform spatial position registration with the vascular skeleton data as a reference, and fuse the physical density attributes and skeleton topology information of each group of voxels to generate anatomical structure feature data. A unified three-dimensional Cartesian coordinate system was established, with the origin set at the center of the top-left voxel of the first layer of the scanned image sequence. The X, Y, and Z axes corresponded to the row, column, and depth directions of the image, respectively, with a resolution of 0.625 mm. Next, the high-density and low-density voxel sets were accessed, and the row, column, and layer numbers of each voxel in the original matrix were read. An affine transformation matrix was used to convert these voxels into spatial physical coordinates within this unified coordinate system. The transformation involved translation and scaling operations to ensure that the spatial position of the voxels strictly corresponded to the actual anatomical dimensions. Simultaneously, the vascular skeleton data was imported into this coordinate system. Since the skeleton data was generated based on the same set of original images, coordinate alignment was directly performed. Based on this, spatial registration and attribute fusion operations were performed. For each spatial position within the coordinate system, its corresponding physical density attribute (CT value) and its topological position information relative to the vascular skeleton were also associated. For example, for a spatial point with coordinates x=150, y=200, z=50, it is recorded that it belongs to a low-density voxel set, has a CT value of -20HU, and its Euclidean distance to the nearest portal vein skeletal node is recorded as 5.5 mm. In this way, the originally discrete voxel data and the abstract skeletal structure are integrated into a composite data structure containing multi-dimensional attributes, namely anatomical structural feature data, providing a rigorously registered digital foundation for subsequent distance field calculations and topology analysis.

[0027] Please see Figure 3 The specific steps for obtaining the vascular affinity topology vector are as follows: S201: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton. Establish a three-dimensional spatial distance field with the vascular skeleton as the reference benchmark. Calculate the Euclidean distance parameter from all voxels in the field to the nearest vascular skeleton point. Traverse the voxel units in the low-density voxel set, retrieve the corresponding position coordinates of each voxel unit in the distance field, assign the Euclidean distance parameter to the corresponding low-density voxel, and generate a vascular distance attribute mapping set. Using all discrete points on the extracted vascular skeleton as zero-distance seed points, a three-dimensional distance field matrix with the same size as the original image is initialized. The values ​​corresponding to skeleton points in the matrix are set to 0, and the values ​​at other locations are set to infinity. Then, an Euclidean distance transformation algorithm, such as the Mett algorithm or the linear time distance transformation algorithm, is applied. This process is divided into two stages. The first stage involves a forward scan, traversing along the positive Z-axis, Y-axis, and X-axis from the origin. For the current voxel, the distance to its neighboring voxels with known distance values ​​within the mask coverage area is calculated, and the minimum distance value is updated. The second stage involves a backward scan, traversing in the reverse direction and correcting the distance values ​​to ensure that all voxels record the accurate Euclidean distance to the nearest vascular skeleton point. For example, after calculation, the straight-line distance of a low-density voxel to the nearest blood vessel point is 3.2 mm. Subsequently, each voxel unit in the low-density voxel set is traversed, and its spatial coordinates are indexed in the three-dimensional distance field matrix. The corresponding distance value is directly read and assigned as a new attribute channel to the low-density voxel. This process transforms a simple geometric location into a biologically meaningful "vascular affinity" attribute, generating a vascular distance attribute mapping set containing tens of thousands of key-value pairs. The key is the voxel ID, and the value is the Euclidean distance from the voxel to the vascular skeleton. For example, voxel ID1024 corresponds to a distance of 4.5 mm, and voxel ID1025 corresponds to a distance of 4.6 mm.

[0028] S202: Call the blood vessel distance attribute mapping set, sort the voxels in the mapping set in ascending order of distance parameters, construct the simple complex evolution sequence, analyze the morphological changes of connected components and topological holes during the sequence evolution, record the generation and extinction times of each topological structure in the simple complex, obtain the persistence characteristics of the laceration structure as the blood vessel distance changes, and generate the persistence interval data of the topological features. All low-density voxels in the vessel distance attribute mapping set are sorted in ascending order according to their distance parameters, i.e., voxels closer to the vessel are listed first. A simple complex evolution sequence is constructed by gradually accumulating voxels as the distance threshold increases. Initially, the distance threshold is 0, and the complex is empty. As the threshold increases slightly, for example, by 0.1 mm each time, voxels that meet the distance condition are added to the complex. During the voxel addition process, the Vitoris-Lipps complex construction rule is used. When the distance between voxels is less than the current threshold, edges are connected, and triangles or tetrahedrons are constructed. At the same time, the continuous homology algorithm is used to monitor the topological changes of the complex in real time, recording the generation and merging of the 0th dimension homology class (connected components), the formation and filling of the 1st dimension homology class (holes), and the enclosure and filling of the 2nd dimension homology class (cavities). For each topological feature, the distance threshold at which it first appears is recorded as the generation time, and the distance threshold at which it merges or disappears is recorded as the disappearance time. For example, a connected component representing a laceration region is generated at a distance parameter of 1.2 mm. As the distance increases, this region expands and merges with another large connected region at a distance parameter of 5.8 mm. Therefore, the survival interval for recording this feature is from 1.2 mm to 5.8 mm. This process transforms static anatomical structures into dynamic topological evolution history, generating persistent topological feature data containing a large amount of binary data, accurately reflecting the distribution pattern of laceration morphology with vascular distance.

[0029] S203: Call the continuous interval data of topological features, filter the topological features that are within the range of the starting parameters at the time of generation, classify and count the number of coherent classes in different dimensions within the target range, analyze the distribution density of topological features on the distance parameter axis, arrange and combine the statistically obtained quantity indicators and distribution density indicators in a preset order, construct a digital vector, and generate a vascular affinity topological vector. First, the range of initial parameters of interest is defined. Considering the close connection between abdominal traumatic bleeding and blood vessels, the distance parameter of interest is set to a range of 0 mm to 20 mm. Next, the data of the topological feature persistence interval is traversed, and all homology features whose generation time falls within this interval are selected. Then, this interval is further subdivided into several sub-intervals, for example, each 5 mm is a statistical surface element. The number of homology features of different dimensions and the sum of their persistence lengths are counted in each sub-interval. For example, in the 0-5 mm interval, the number of 0th dimension features (i.e., connected components) is 150, with an average persistence length of 3.5 mm; in the 5-10 mm interval, the number is 80, with an average persistence length of 2.1 mm. Simultaneously, the distribution density of features on the distance axis is analyzed, and the rate of change of the number of features generated per unit distance is calculated. Finally, these statistical indicators are combined into a one-dimensional numerical vector according to a preset logical order. The structure of this vector is shown in Table 1, containing components such as the total number of features within the interval, the average lifetime, and the maximum lifetime. For example, the final generated vascular affinity topology vector is [150, 3.5, 8.2, 80, 2.1, 4.5, ...]. This vector digitally represents the density and spatial complexity of low-density lesions (such as hematomas or lacerations) distributed around the vascular skeleton.

[0030] Table 1. Example of eigencomponents of vascular affinity topological vectors Feature number Feature Name Numerical Examples Explanation of physical meaning F1 Near-field connectivity 150 Number of independent laceration areas within the 0-5mm range F2 Near-field durability 3.5 Average length (mm) of a laceration extending along a blood vessel within the 0-5 mm range. F3 Midfield connectivity 80 Number of independent laceration areas within the range of 5-10mm F4 Midfield endurance 2.1 Average length (mm) of a laceration extending along a blood vessel within the 5-10 mm range. F5 Topological entropy 0.85 The degree of disorder in the distribution of laceration morphology As shown in Table 1, this vector quantifies the spatial morphological characteristics of the trauma lesion through multi-dimensional topological parameters.

[0031] Please see Figure 4 The specific steps for obtaining the liquid dispersion volume parameters are as follows: S301: Call anatomical structure feature data and extract high-density voxel set, filter edge voxels in the set at the junction of high-density tissue and background, calculate the gray-level change gradient of edge voxels in local three-dimensional environment, construct vector data containing the direction of maximum change rate and gradient intensity information, and generate edge gradient vector field data. The high-density voxel set is traversed to identify voxels with at least one neighbor belonging to either the background or low-density voxel set in a 26-neighborhood in 3D space, and these are marked as edge voxels. For each edge voxel, a local 3x3x3 computation window is established, and its gray-level partial derivatives in the X, Y, and Z directions are calculated using either the 3D Sobel operator or the Pruitt operator. Specifically, the operator template is convolved with the gray-level values ​​within the local window to obtain the three directional components Gx, Gy, and Gz. Subsequently, a gradient vector is synthesized based on these three components. The gradient direction points towards the direction with the largest gray-level change rate, typically from the interior of a high-density organ to a low-density region or background, and its magnitude, i.e., the gradient intensity, is equal to the square root of the sum of the squares of the three components. For example, if the partial derivatives of an edge voxel in the three directions are 30, 40, and 0, its gradient magnitude is 50 HU per voxel, and its gradient direction vector is 0.6, 0.8, and 0. Finally, all edge voxels and their calculated gradient vector information are structured and organized to generate edge gradient vector field data, which accurately describes the gray-scale variation characteristics and potential fluid diffusion direction at the organ parenchyma boundary.

[0032] S302: Call the edge gradient vector field data, analyze the pointing distribution of the gradient vector in three-dimensional space, collect the gradient intensity of continuous neighborhood voxels along the extension direction of the gradient vector, construct the intensity change sequence along the spatial distance distribution, calculate the differential decay rate of the gradient intensity relative to the spatial position in the sequence, and generate the spatial decay rate sequence. The edge gradient vector field data is retrieved. Starting from each edge voxel, ray projection sampling is performed outward along the spatial direction indicated by its gradient vector, with a step size of 1 pixel unit (approximately 0.625 mm). Along each ray, the gradient intensity values ​​of all voxel units passed within a preset length, such as 10 mm, are continuously collected. This forms a series of one-dimensional numerical sequences, for example, [120, 110, 95, 80, 65, ...]. This sequence is then differentiated, calculating the ratio of the gradient intensity difference between adjacent sampling points to the spatial distance, thus obtaining the differential attenuation rate of the gradient magnitude relative to the spatial location. For example, if the gradient magnitude of the first point is 120, the second point is 115, and the distance is 0.625 mm, then the attenuation rate at that point is -8 HU per millimeter. By performing this operation on the edge voxels of the entire field, thousands of data streams reflecting the trend of grayscale changes extending outward from the surface of the organ are constructed, namely spatial decay rate sequences. These sequences record how quickly the concentration of contrast agents or blood decreases as they diffuse in the interstitial space or free peritoneum.

[0033] S303: Call the spatial decay rate sequence to analyze the evolution of gradient intensity along the spatial extension path, select a set of voxels whose gradient magnitude decreases linearly and gradually with distance and without abrupt changes as liquid dispersion feature regions, calculate the cumulative space occupancy, and generate liquid dispersion volume parameters. First, least squares linear regression is performed on the original gradient magnitude data corresponding to each spatial decay rate sequence. Distance is the independent variable x, and gradient magnitude is the dependent variable y. The slope k and linear correlation coefficient r of the fitted line are calculated. Next, a first-order difference sequence is constructed, and the absolute value of the difference between the gradient magnitudes of adjacent points in the sequence is calculated. Simultaneously, the average gradient magnitude of all voxels in the background voxel set is calculated, assumed to be 5 HU per millimeter, as a reference for smooth decay; the linear correlation threshold is set to 0.9. The mean μ and standard deviation σ of the difference sequence are calculated, and the dynamic step threshold is set to μ plus 3σ. The screening process is as follows: First, sequences with a linear correlation coefficient r greater than 0.9 and a negative slope k with an absolute value less than 5 HU per millimeter are found. This indicates that the gradient is uniform and slowly decreasing, conforming to the physical laws of free diffusion in liquids, rather than the abrupt boundary of solid tissue. Then, the corresponding first-order difference sequence is checked. If the difference value of all points is less than the dynamic step threshold, it indicates that no structures causing abrupt changes, such as bone or organ capsules, are encountered along the path. Sequences that pass the double screening are identified as fluid diffusion pathways. Finally, the total number of voxel units covered by all sequences meeting the criteria is counted, assumed to be N. Combined with the physical volume of a single voxel, Vvoxel (e.g., 0.244 cubic millimeters), a multiplication operation is performed, N multiplied by Vvoxel, to obtain the final fluid diffusion volume parameter. For example, screening 100,000 voxels yields a diffusion volume of 24,400 cubic millimeters, or 24.4 milliliters. This parameter accurately quantifies the scale of ascites or active bleeding.

[0034] Please see Figure 5 The specific steps for obtaining the compensatory perfusion ratio are as follows: S401: Call anatomical structure feature data and extract low-density voxel set and vascular skeleton, establish a three-dimensional spatial superposition mapping relationship between low-density voxel set and portal vein vascular skeleton, analyze the path connectivity status of vascular skeleton in low-density area, determine the spatial geometric position of physical truncation of skeleton path, mark the breakpoint of missing upstream connectivity and use it as the starting point of blood flow transmission interruption, and generate a set of portal vein interruption nodes. First, the 3D coordinates of the low-density voxel set are superimposed with the coordinate system of the portal vein vascular skeleton. Since they are in the same registration space, spatial position comparison can be performed directly. By calculating the minimum Euclidean distance between each low-density voxel and each point on the vascular skeleton, if a segment of the vascular skeleton is completely surrounded by low-density voxels, and the surrounding radius exceeds 1.5 times the vessel radius, and the connectivity detection at both ends of the segment shows a path interruption (i.e., it is impossible to reach the other end of the segment through the skeleton network without passing through the low-density region), then the segment of the vessel is determined to have been physically truncated. In practice, a connected component labeling algorithm from graph theory is used. If a node loses connection to the main portal vein entrance after the edges covered by the low-density region are removed, the node is marked as a breakpoint. The entire skeleton network is traversed to identify all nodes that meet the above characteristics, and their 3D coordinates and the ID of the vessel branch to which they belong are recorded, generating a set of portal vein interruption nodes. For example, a breakpoint was found in the posterior segment of the right lobe of the liver, with coordinates x = 100, y = 120, and z = 60. This point connects to the main trunk upstream and several small branches downstream, with the continuity interrupted due to hematoma compression or tearing.

[0035] S402: Call the set of portal vein interruption nodes, perform downstream path traversal operation according to the topological extension direction of the vascular anatomy, search for the distribution area of ​​the vascular branch network downstream of the interruption node, collect all voxel units within the extension range of the branch network, construct the spatial distribution data of the damaged tissue that has lost portal vein connectivity, and generate the downstream interrupted voxel set. Starting from each node in the generated set of portal vein interruption nodes, a depth-first search or breadth-first search is performed along the blood flow direction (away from the main trunk) based on the original tree-like topology of the vascular skeleton. During the search, all child nodes and edges belonging to the downstream of the interruption node are marked until the end of the vascular tree is reached. These marked vascular branches were originally supplied by the interruption node, but now lose their direct portal vein blood source due to the interruption. Next, based on the Voronoi diagram or a distance-based region growing algorithm, the tissue regions in the organ parenchyma that are responsible for supplying blood to these downstream vascular branches are delineated. Specifically, each voxel in the organ is assigned to the nearest vascular skeleton point. If the nearest skeleton point of a voxel belongs to the marked downstream branch network, then the voxel is determined to be part of the damaged area. All voxel units that meet this condition are traversed and collected to construct a three-dimensional point set, which spatially depicts the extent of potential ischemia in the organ parenchyma caused by portal vein injury, generating a set of downstream interrupted voxels. For example, through search and region mapping, approximately 500,000 voxel units were identified as belonging to the disjointed region, accounting for about 15% of the total liver volume.

[0036] S403: Call the downstream fragmented voxel set and associate it with the arterial skeleton in the anatomical structure feature data, calculate the spatial proximity between the fragmented voxel and the nearest arterial skeleton, screen voxel units located within the coverage of arterial blood permeation, statistically analyze the ratio of the screened voxel size to the total size of the fragmented voxel set, and generate the compensatory perfusion ratio. First, the arterial skeleton information from the anatomical structural feature data is accessed, and the effective perfusion radius of the artery is set. This radius is usually set based on physiological parameters, such as the effective permeation distance of a small artery being 5 mm. Next, each voxel in the downstream fragmented voxel set is traversed, and its Euclidean distance to the nearest arterial skeleton node is calculated. If this distance is less than or equal to the set effective perfusion radius of 5 mm, the voxel is determined to be within the compensatory blood supply range of the artery, even though it has lost portal vein blood supply, and belongs to the viable or functionally compensated area; if the distance is greater than 5 mm, it is determined to be a completely ischemic area. The number of voxels in the fragmented voxel set that meet the arterial perfusion conditions is counted and denoted as N_comp; at the same time, the total number of fragmented voxels is obtained and denoted as N_total. A division operation is performed to calculate the ratio of N_comp to N_total, which is the compensatory perfusion ratio. As shown in Table 2, if the fragmented region has a total of 500,000 voxels, of which 150,000 voxels are located within a 5 mm radius of the arterial skeleton, then the compensatory perfusion ratio is 150,000 divided by 500,000, which equals 0.30. This ratio is a value between 0 and 1; a higher value indicates stronger arterial compensatory capacity and a lower risk of tissue necrosis. This ratio is ultimately output as a key quantitative indicator for assessing organ function prognosis.

[0037] Table 2 Example of Compensated Infusion Ratio Calculation Data Data item name Numerical Examples unit Notes Total number of fragmented voxels (N_total) 500,000 indivual Total amount of blood lost from portal vein Artery coverage voxel count (N_comp) 150,000 indivual Voxel quantity ≤ 5 mm from the arterial skeleton Infusion radius setting 5.0 mm Thresholds set based on physiological osmotic pressure Compensated perfusion ratio 0.30 - N_comp / N_total Percentage of areas at risk of ischemia 0.70 - 1 - Compensated perfusion ratio As shown in Table 2, the calculation results directly reflect the hemodynamic compensation status of the damaged area.

[0038] Please see Figure 6 The specific steps for obtaining the abdominal trauma grading assessment results are as follows: S501: Obtain vascular affinity topology vectors, liquid diffusion volume parameters and compensatory perfusion ratios, analyze the differences in physical dimensions and the characteristics of data distribution range, perform normalization mapping processing, fuse the processed multi-source feature data to establish a unified high-dimensional feature space coordinate system, and generate multi-dimensional feature vectors. The system receives three core parameters generated in the preceding steps: a vascular affinity topology vector (containing multiple components), a fluid diffusion volume parameter (unit: cubic millimeters), and a compensatory perfusion ratio (dimensionless ratio). Due to the inconsistent physical dimensions and vastly different numerical ranges of these parameters (e.g., volumes can reach tens of thousands while ratios are only 0.x), normalization is required. A min-max normalization method is employed. For each parameter x, its theoretical physiological minimum value (min) and maximum value (max) are determined. The operation (x minus min) is then performed divided by (max min), mapping it to a closed interval between 0 and 1. For example, for fluid diffusion volume, min is set to 0, and max to 500 ml (500,000 cubic millimeters); for compensatory perfusion ratio, min is 0, and max is 1. Specifically, for the multidimensional vascular affinity topology vector, principal component analysis is used to extract its first principal component as a representative value before normalization. The three normalized values ​​after processing are combined in a predetermined order (topology, diffusion, compensation) to construct a three-dimensional feature vector, for example, V = [0.65, 0.42, 0.30]. This establishes a unit cube feature space with a side length of 1, i.e., a unified high-dimensional feature space coordinate system, where each point represents a specific trauma pathological state. This generates a standardized multidimensional feature vector, eliminating the interference of different dimensions on subsequent distance calculations.

[0039] S502: Call the multidimensional feature vector to construct a critical reference benchmark composed of the theoretical limit states of each evaluation dimension, establish a graded evaluation interval sequence containing multiple severity levels, analyze the contribution weight difference of each feature dimension in the trauma severity assessment system, and calculate the weighted Euclidean distance between the current feature vector and the critical reference benchmark in the multidimensional feature space to generate distance difference calculation data. A critical reference baseline is constructed, which means defining a vector representing the "theoretically most severe trauma state". According to pathological logic, the most severe trauma corresponds to: lacerations completely overlapping with blood vessels (the normalized topological erosion limit is 1), ascites reaching a lethal dose (the normalized diffusion capacity limit is 1), and complete arterial uncompensation (the normalized perfusion loss limit is 0, which is the worst case, so it needs to be subtracted from the original value by 1 or the direction needs to be adjusted; here, "the smaller the distance, the more severe" is set, so the value corresponding to the extremely dangerous state is taken. Assuming that the normalization direction is uniformly that the larger the value, the more severe it is, the compensation ratio is taken as its complement, i.e., 1 minus the ratio, and the extremely dangerous value is 1). Therefore, the constructed critical reference baseline vector is W_critical = [1, 1, 1]. Next, the weights of each feature dimension are determined, which is usually determined based on expert experience or the analytic hierarchy process. For example, the topological weight w1 is set to 0.4, the diffusion weight w2 to 0.3, and the compensation weight w3 to 0.3. The weighted Euclidean distance between the current multidimensional feature vector V = [v1, v2, v3] and the critical baseline W_critical = [1, 1, 1] is then calculated. The calculation logic is as follows: First, the squares of the differences between each component are calculated, i.e., the squares of (v1 - 1), (v2 - 1), and (v3 - 1); then, these are multiplied by their corresponding weights w1, w2, and w3, respectively; finally, the three products are added together and the square root is taken. For example, if V = [0.65, 0.42, 0.30] (note that the third term has been converted to a non-compensation rate of 0.7), and the weights are as above, the distance value can be obtained by substituting these values. The smaller this distance value, the closer the current state is to a critical state, and the more severe the injury. This specific distance difference calculation data is then generated.

[0040] S503: Based on the graded evaluation interval sequence, retrieve the corresponding interval of the distance difference calculation data, determine the severity level of the current trauma status, and generate the abdominal trauma grading assessment result; The established tiered assessment interval sequence is invoked. This sequence is based on K-means cluster analysis of large-scale historical case data. Assume that after cluster analysis, the historical data forms four cluster centers on the distance axis, corresponding to critical, severe, moderate, and mild cases, with center values ​​of 0.15, 0.40, 0.65, and 0.90, respectively. The mean of adjacent centers is calculated as the tiered boundary thresholds: the boundary between critical and severe cases is (0.15 + 0.40) divided by 2 equals 0.275; the boundary between severe and moderate cases is 0.525; and the boundary between moderate and mild cases is 0.775. Based on this, the intervals are divided as follows: 0 to 0.275 is critical (Level IV), 0.275 to 0.525 is severe (Level III), 0.525 to 0.775 is moderate (Level II), and 0.775 to 1.0 is mild (Level I). As shown in Table 3, the distance difference calculation data generated in step S502 (assuming a calculation result of 0.35) is compared and retrieved with the above intervals. Since 0.35 falls between 0.275 and 0.525, the current trauma status is determined to fall into the Grade III severe injury interval. Finally, the abdominal trauma grading assessment result, including the specific grading label "Grade III" and comprehensive score data, is output, providing an intuitive basis for clinical decision-making.

[0041] Table 3 Reference Table for Abdominal Trauma Grading Assessment Intervals Classification and grading Severity Distance range Clinical Corresponding Description Level IV endangered [0.000, 0.275] Extremely critical condition, requires immediate damage control surgery Level III severe (0.275, 0.525] Severe injury requiring close monitoring or emergency surgery Level II Moderate disease (0.525, 0.775] Moderate injury, vital signs relatively stable Level I Mild cases (0.775, 1.000] Minor injury, conservative treatment and observation As shown in Table 3, by mapping the calculation results to standard intervals, the automated and accurate grading of complex abdominal trauma was achieved.

[0042] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine learning-based method for grading and assessing abdominal trauma in general surgery, characterized in that, Includes the following steps: S1: Acquire abdominal enhanced tomographic scan images, parse them into a three-dimensional voxel matrix, segment them into a high-density voxel set, a low-density voxel set and a background voxel set according to density distribution characteristics, and construct the portal vein skeleton and arterial skeleton based on the background voxel set, and output anatomical structure feature data. S2: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton, transform the distance between the portal vein skeleton and the arterial skeleton and map it to the low-density voxel set, establish a simple complex filter flow according to the distance, count the number of existing features, and generate a vascular affinity topology vector. S3: Call the anatomical structure feature data, calculate the edge gradient vector of the high-density voxel set and the decay rate of the magnitude sequence along the gradient vector, filter the voxel set with linear and gradual decay and no abrupt changes and calculate the volume, and output the liquid diffusion volume parameters. S4: Call the anatomical structure feature data, identify the portal vein interruption node and downstream fragmented voxel set truncated by the low-density voxel set, count the number of fragmented voxels located within the perfusion radius of the arterial skeleton and calculate the proportion, and output the compensatory perfusion ratio. S5: Normalize the vascular affinity topology vector, liquid diffusion volume parameter, and compensatory perfusion ratio, combine them into a multidimensional feature vector, construct a critical state vector and calculate the weighted distance with the multidimensional feature vector, match the grading interval and output the abdominal trauma grading assessment result.

2. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 1, characterized in that, The anatomical structural feature data includes a three-dimensional voxel matrix, voxel set partitioning results, and skeleton construction results. The vascular affinity topology vector includes the number of near-vascular connectivity components, the number of near-vascular topological loops, and the length of the topological feature persistence interval. The fluid diffusion volume parameters include the volume value of the fluid infiltration region, the gradient continuity index of the infiltration region edge, and the average density value of the infiltration region. The compensatory perfusion ratio includes the total voxel volume of the portal vein injury region, the voxel volume of the arterial compensation coverage region, and the percentage of functional preservation. The abdominal trauma grading assessment results include trauma severity grading labels, multidimensional feature weighted comprehensive scores, and anatomical risk and functional preservation map data.

3. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 1, characterized in that, The specific steps for obtaining the anatomical structural feature data are as follows: S101: Acquire abdominal enhanced tomographic scan images and parse them into a three-dimensional voxel matrix. Analyze the gray intensity distribution characteristics of voxel units within the matrix. Classify and filter various tissues based on density differences under contrast conditions. Divide the three-dimensional voxel matrix into a high-density voxel set, a low-density voxel set, and a background voxel set to generate a basic voxel segmentation set. S102: Call the basic voxel segmentation set and extract the background voxel set, perform morphological thinning and shrinking processing on the background voxel set, strip non-core tissue pixels and retain single-pixel width connected paths, calculate the geometric center line coordinates of the residual paths, track and classify the center lines according to the vascular anatomy connectivity rules, construct the portal vein connectivity network and arterial connectivity network, and generate vascular skeleton data. S103: Call the high-density voxel set and low-density voxel set in the basic voxel segmentation set, combine them with the vascular skeleton data, establish a three-dimensional spatial mapping coordinate system, map the spatial distribution coordinates of each voxel set to the coordinate system, perform spatial position registration with the vascular skeleton data as a reference, fuse the physical density attributes and skeleton topology information of each group of voxels, and generate anatomical structure feature data.

4. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 3, characterized in that, The process of classifying and screening various tissues based on density differences under imaging conditions, and dividing the three-dimensional voxel matrix into high-density voxel sets, low-density voxel sets, and background voxel sets, is specifically as follows: The gray intensity values ​​of all voxel units in the three-dimensional voxel matrix are statistically analyzed to establish a global gray-level statistical histogram reflecting the overall distribution pattern of abdominal tissue density. The probability density function of the global gray-level statistical histogram is fitted using a Gaussian mixture model to extract three independent Gaussian distribution components, and the gray-level mean of each Gaussian distribution component is calculated. The three Gaussian distribution components are arranged in order of increasing gray value, and are defined as low attenuation distribution component, background distribution component and high attenuation distribution component respectively. Calculate the gray value of the intersection point of the probability density curves between the low-attenuation distribution component and the background distribution component, and set it as the low-density segmentation threshold value. Calculate the gray value of the intersection point of the probability density curves between the background distribution component and the high attenuation distribution component, and set it as the high density segmentation threshold value; Traverse the voxel units in the three-dimensional voxel matrix and extract the voxel units whose gray intensity values ​​are greater than the high-density segmentation threshold value to the high-density voxel set. Voxel units with gray intensity values ​​less than the low-density segmentation threshold are extracted into a low-density voxel set. Voxel units whose grayscale intensity values ​​fall within the closed interval between the low-density segmentation threshold and the high-density segmentation threshold are extracted into the background voxel set.

5. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 4, characterized in that, The specific steps for obtaining the vascular affinity topology vector are as follows: S201: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton. Establish a three-dimensional spatial distance field with the vascular skeleton as a reference. Calculate the Euclidean distance parameter from all voxels in the field to the nearest vascular skeleton point. Traverse the voxel units in the low-density voxel set, retrieve the corresponding position coordinates of each voxel unit in the distance field, assign the Euclidean distance parameter to the corresponding low-density voxel, and generate a vascular distance attribute mapping set. S202: Call the blood vessel distance attribute mapping set, sort the voxels in the mapping set in ascending order of distance parameters, construct a simple complex evolution sequence, analyze the morphological changes of connected components and topological holes during the sequence evolution, record the generation and extinction times of each topological structure in the simple complex, obtain the persistence characteristics of the laceration structure as the blood vessel distance changes, and generate topological feature persistence interval data. S203: Call the continuous interval data of the topological features, filter the topological features whose generation time is within the range of the starting parameters, classify and count the number of coherent classes in different dimensions within the target range, analyze the distribution density of the topological features on the distance parameter axis, arrange and combine the statistically obtained quantity indicators and distribution density indicators in a preset order, construct a digital vector, and generate a vascular affinity topological vector.

6. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 5, characterized in that, The specific steps for obtaining the liquid dispersion volume parameter are as follows: S301: Call the anatomical structure feature data and extract the high-density voxel set, filter the edge voxels in the set that are at the junction of high-density tissue and background, calculate the gray-level change gradient of the edge voxels in the local three-dimensional environment, construct vector data containing the direction of the maximum rate of change and gradient intensity information, and generate edge gradient vector field data. S302: Call the edge gradient vector field data, analyze the pointing distribution of the gradient vector in three-dimensional space, collect the gradient intensity of continuous neighborhood voxels along the extension direction of the gradient vector, construct the intensity change sequence distributed along the spatial distance, calculate the differential decay rate of the gradient intensity relative to the spatial position in the sequence, and generate a spatial decay rate sequence. S303: Call the spatial decay rate sequence, analyze the evolution of gradient intensity along the spatial extension path, select a set of voxels whose gradient magnitude decreases linearly and gradually with distance and without abrupt changes as liquid dispersion characteristic regions, calculate the cumulative space occupancy, and generate liquid dispersion volume parameters.

7. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 6, characterized in that, The process of selecting a set of voxels whose gradient magnitude decreases linearly and smoothly with distance and without abrupt changes as the liquid dispersion feature region is as follows: The discrete gradient magnitude data contained in the spatial decay rate sequence are fitted by least squares linear regression to calculate the linear correlation coefficient and the slope of the fitted line, which characterize the overall trend of the sequence. Calculate the absolute difference between the gradient magnitudes of adjacent spatial locations in the spatial decay rate sequence point by point to construct a first-order difference sequence that reflects the local numerical jump characteristics; The arithmetic mean of the gradient magnitudes of all voxel units in the background voxel set is calculated and set as the reference benchmark for gradual decay. The preset lower limit of the linear correlation coefficient is set as the linear correlation threshold. Based on the numerical distribution characteristics of the first-order difference sequence, the average value and standard deviation of the absolute values ​​of the differences within the sequence are calculated, and the sum of the average value and three times the standard deviation is set as the dynamic step judgment threshold. Sequences whose linear correlation coefficient is greater than the linear correlation threshold, whose slope value is negative and whose absolute value is less than the smooth decay reference are selected and determined to meet the linear smooth decay condition. Traverse the first-order difference sequence corresponding to the sequence that satisfies the linear gradual decay condition, check whether there are any numerical points in the sequence that are greater than the dynamic step judgment threshold, and remove the sequences that have such numerical points. The set of voxel units corresponding to sequences that are filtered by linear and gradual decay conditions and do not contain the numerical points are extracted and marked as liquid dispersion feature regions.

8. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 7, characterized in that, The steps for obtaining the compensatory perfusion ratio are as follows: S401: Call the anatomical structure feature data and extract the low-density voxel set and vascular skeleton, establish the three-dimensional spatial superposition mapping relationship between the low-density voxel set and the portal vein vascular skeleton, analyze the path connectivity status of the vascular skeleton in the low-density area, determine the spatial geometric position of the physical truncation of the skeleton path, mark the breakpoints of missing upstream connectivity and use them as the starting point of blood flow transmission interruption, and generate a set of portal vein interruption nodes. S402: Call the set of portal vein interruption nodes, perform downstream path traversal operation according to the topological extension direction of the vascular anatomy, search for the distribution area of ​​the vascular branch network downstream of the interruption node, collect all voxel units within the extension range of the branch network, construct the spatial distribution data of the damaged tissue that has lost portal vein connectivity, and generate the downstream interrupted voxel set. S403: Call the downstream fragmented voxel set and associate it with the arterial skeleton in the anatomical structure feature data, calculate the spatial proximity between the fragmented voxel and the nearest arterial skeleton, screen voxel units located within the coverage area of ​​arterial blood permeation, statistically analyze the proportional relationship between the size of the screened voxels and the total size of the fragmented voxel set, and generate the compensatory perfusion ratio.

9. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 8, characterized in that, The specific steps for obtaining the abdominal trauma grading assessment results are as follows: S501: Obtain the vascular affinity topology vector, liquid diffusion volume parameters and compensatory perfusion ratio, analyze the differences in physical dimensions and the characteristics of data distribution range, and perform normalization mapping processing. After fusing the processed multi-source feature data, establish a unified high-dimensional feature space coordinate system and generate multi-dimensional feature vectors. S502: Call the multidimensional feature vector to construct a critical reference benchmark composed of the theoretical limit states of each evaluation dimension, establish a graded evaluation interval sequence containing multiple severity levels, analyze the contribution weight difference of each feature dimension in the trauma severity assessment system, and calculate the weighted Euclidean distance between the current feature vector and the critical reference benchmark in the multidimensional feature space to generate distance difference calculation data. S503: Based on the graded evaluation interval sequence, retrieve the corresponding interval of the distance difference calculation data, determine the severity level of the current trauma status, and generate an abdominal trauma grading assessment result.

10. The machine learning-based grading assessment method for abdominal trauma in general surgery according to claim 9, characterized in that, The process of constructing a critical reference benchmark composed of the theoretical limit states of each evaluation dimension and establishing a graded evaluation interval sequence containing multiple severity levels is as follows: Determine the maximum theoretical value in the vascular affinity topology vector that represents the complete overlap between the laceration and the vascular skeleton space, and define it as the topological erosion limit value; The value of the liquid diffusion volume parameter that represents the maximum physical capacity of the abdominal anatomical space is determined and defined as the diffusion capacity limit value. The zero value in the compensatory perfusion ratio, which represents the complete lack of compensation in the artery, is determined and defined as the perfusion loss limit. Normalization is performed on the topological erosion limit value, the diffusion capacity limit value, and the infusion loss limit value, and the critical reference benchmark in the multidimensional feature space is constructed by combining them. Retrieve classified historical abdominal trauma case data, calculate the weighted Euclidean distance between the multidimensional feature vector of each historical case and the critical reference benchmark, and construct a sample distance set; The K-means clustering algorithm is used to iteratively calculate the distance set of the samples to obtain the cluster center values ​​corresponding to different trauma severity, and then sort them in ascending order of value; Calculate the arithmetic mean of the values ​​of two adjacent cluster centers, set it as the hierarchical boundary threshold, and divide the continuous numerical range based on the hierarchical boundary threshold to establish a hierarchical evaluation interval sequence.

Images