MVI-positive hepatocellular carcinoma high-risk low-risk classification method and application

Abstract

This invention relates to the field of medical image processing and discloses a method and application for high-risk / low-risk classification of MVI-positive hepatocellular carcinoma. The method includes: acquiring and registering three-dimensional enhanced CT image sequences to generate a tumor target area and a background liver parenchyma mask; calculating the temporal gradient to construct a two-dimensional temporal perfusion vector and extracting a subset of target blood supply for screening; determining a baseline vector based on the background liver parenchyma, projecting the screened vector to obtain orthogonal components and generating an intensity scalar field; combining the tumor boundary normal vector field with the spatial gradient for weighted summation to generate a target feature field; extracting image features from the target feature field and concatenating clinical variables, inputting the data into a classifier model to output high-risk / low-risk classification results. By screening active blood supply areas to eliminate interference from necrotic tissue, using baseline vector projection to eliminate measurement bias caused by individual circulatory differences, and combining spatial gradients to quantify the diffusion direction of abnormal blood supply, the accuracy of the model's high-risk / low-risk classification prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically to a method and application for classifying high-risk and low-risk hepatocellular carcinoma with MVI positivity. Background Technology

[0002] Hepatocellular carcinoma (HCC) is a common malignant tumor in clinical practice. Microvascular invasion (MVI) is a significant risk factor for postoperative recurrence. Preoperative identification of intraoperative heterogeneity and high-risk phenotypes in MVI-positive HCC is crucial for clinical treatment planning. Currently, preoperative assessment based on medical imaging-assisted diagnostic techniques, especially using three-dimensional enhanced CT image sequences, has become routine in the field of medical imaging research. Existing auxiliary diagnostic methods typically involve segmenting lesions in enhanced CT images and then extracting morphological, textural, and absolute grayscale features of the tumor region, or directly extracting high-dimensional features from the image region using neural network models and inputting the extracted high-dimensional features into a classifier to obtain prediction results.

[0003] Existing feature extraction methods, when processing enhanced CT sequences, largely rely on the absolute grayscale values ​​of the tumor region for quantitative evaluation. This approach is easily affected by individual patient differences in systemic blood circulation, such as cardiac output, and fluctuations in the contrast agent injection rate of the CT equipment, leading to systematic measurement biases in cross-sample evaluation. Furthermore, liver cancer lesions exhibit internal heterogeneity, often containing liquefied necrotic tissue or delayed-enhancing fibrous tissue. Conventional methods mostly treat the entire tumor target area as a whole for feature extraction, making it difficult to separate necrotic areas lacking blood supply from areas with abnormal proliferation and active blood supply. This makes the extracted imaging features susceptible to interference from invalid tissue, affecting the feature quality of the input data.

[0004] In terms of spatial feature computation, existing image feature extraction methods mainly focus on the static morphological boundaries of lesions or scalar grayscale distribution information. Microvascular invasion is essentially a process of tumor neovascularization expanding into the peripheral normal liver parenchyma, exhibiting a clear directionality in three-dimensional space. Conventional feature extraction and spatial pooling methods fail to effectively calculate and retain the spatial direction information of abnormal blood supply invading outward along the anatomical morphology, making it difficult for the final generated features to simultaneously account for the intensity of metabolic abnormalities and spatial diffusion trends. This lack of spatial direction information limits the accuracy of classification models in identifying intra-arterial heterogeneity and high-risk phenotypes in MVI-positive HCC. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and application for classifying high-risk and low-risk hepatocellular carcinoma (HCC) with MVI-positive tumors. This method solves the problem that existing auxiliary diagnostic methods for microvascular invasion in HCC mainly rely on static spatial morphological characteristics or absolute grayscale values ​​for assessment. The assessment results are easily affected by individual patient systemic circulatory differences and fluctuations in scanning equipment parameters. Furthermore, existing methods struggle to eliminate interference from necrotic and fibrotic tissues within the tumor and lack quantitative calculation methods for the spatial diffusion trend of blood perfusion in the lesion. This makes it difficult for existing methods to stratify the internal structure of MVI-positive HCC and identify truly high-risk individuals.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a method for classifying MVI-positive hepatocellular carcinoma into high-risk and low-risk types, comprising the following steps: S10. Obtain the three-dimensional enhanced CT image sequence of the target object, and perform registration on the arterial phase image and portal venous phase image using the plain scan image as the reference coordinate system to generate the registered image sequence. S20. Generate a tumor target area mask and a background liver parenchyma mask based on the registered image sequence; S30. Based on the tumor target area mask and the background liver parenchyma mask, extract the voxel gray values, calculate the inflow time gradient and outflow time gradient, generate the two-dimensional temporal perfusion vector corresponding to each spatial voxel, and extract the voxel spatial coordinates with positive inflow time gradient and negative outflow time gradient to construct a subset of primary active blood supply to the tumor, and use the subset of primary active blood supply to purify the two-dimensional temporal perfusion vector. S40. Extract features from the two-dimensional temporal perfusion vector within the background liver parenchyma mask to determine the physiological baseline vector, and project the two-dimensional temporal perfusion vector within the tumor target area mask onto the physiological baseline vector to obtain orthogonal components. S50. Calculate the modulus of the orthogonal components in three-dimensional space to generate an abnormal metabolic intensity scalar field. Construct a boundary normal vector field based on the tumor target area mask. Perform spatial gradient operation on the abnormal metabolic intensity scalar field. Generate an invasion direction weight field by multiplying the spatial gradient with the boundary normal vector field. Use the invasion direction weight field to perform weighting on the abnormal metabolic intensity scalar field to generate an invasion dynamics feature field. S60. Spatial pooling is performed on the invasion dynamics feature field to generate image feature vectors. The image feature vectors are concatenated with the clinical variables of the target object to form a joint feature vector. The joint feature vector is input into the classifier model to output the probability distribution of the target object category.

[0007] Furthermore, the three-dimensional enhanced CT image sequence includes plain scan images, arterial phase images, and portal venous phase images; in step S10, the registration of the arterial phase images and portal venous phase images includes: using the plain scan images as the reference coordinate system, performing affine transformations on the arterial phase images and portal venous phase images respectively for global spatial coarse alignment, and constructing an objective function containing a smoothing penalty term to perform B-spline non-rigid registration and interpolation resampling to generate a spatially aligned registered image sequence.

[0008] Further, in step S20, generating the tumor target mask and the background liver parenchyma mask based on the registered image sequence includes: extracting the three-dimensional spatial coordinate set of the lesion in the registered image sequence to generate the tumor target mask; extracting the spatial coordinates of the entire liver parenchyma of the target object to generate the initial liver mask; performing morphological dilation on the tumor target mask to generate a boundary buffer zone; extracting the main vascular region in the registered image sequence to generate a vascular region mask; and removing the tumor target mask, the boundary buffer zone, and the vascular region mask from the initial liver mask through Boolean difference algebra operations to generate the background liver parenchyma mask.

[0009] Further, in step S30, calculating the inflow time gradient and outflow time gradient includes: extracting the scanning time nodes of the plain scan image, the arterial scan image, and the portal venous scan image; calculating the grayscale difference between the arterial scan image and the plain scan image at the same spatial coordinates and dividing it by the inflow time interval to obtain the inflow time gradient; calculating the grayscale difference between the portal venous scan image and the arterial scan image at the same spatial coordinates and dividing it by the outflow time interval to obtain the outflow time gradient.

[0010] Further, in step S30, purifying the two-dimensional time-series perfusion vector using a subset of primary active blood supply to the tumor includes: constructing a two-dimensional time-series perfusion vector field with the inflow time gradient as the horizontal axis component and the outflow time gradient as the vertical axis component; setting inflow tolerance thresholds and outflow tolerance thresholds; generating an effective logical mask when the inflow time gradient is greater than the inflow tolerance threshold and the outflow time gradient is less than a negative number of the outflow tolerance threshold; and performing a point-by-point scalar multiplication operation on the two-dimensional time-series perfusion vector field and the effective logical mask to output the purified two-dimensional time-series perfusion vector.

[0011] In a preferred embodiment of the present invention, in step S40, feature extraction of the two-dimensional temporal perfusion vector within the background liver parenchyma mask to determine the physiological baseline vector includes: performing regional average accumulation calculation of the two-dimensional temporal perfusion vector within the global anatomical space, and determining the calculated average perfusion vector as the physiological baseline vector.

[0012] Preferably, the method of extracting features from the two-dimensional temporal perfusion vector within the background liver parenchyma mask to determine the physiological baseline vector is replaced by: performing multivariate kernel density estimation on the two-dimensional temporal perfusion vector within the background liver parenchyma mask, and determining the vector corresponding to the main peak of the probability density distribution as the physiological baseline vector.

[0013] Further, in step S40, projecting the two-dimensional temporal perfusion vector within the tumor target mask onto the physiological baseline vector to obtain orthogonal components includes: performing algebraic normalization on the physiological baseline vector to extract parallel unit vectors, and solving for orthogonal unit vectors perpendicular to the parallel unit vectors in the two-dimensional orthogonal plane to construct a two-dimensional orthogonal feature basis; performing vector inner product projection calculation on the purified two-dimensional temporal perfusion vector and the orthogonal unit vectors to separate the orthogonal components characterizing abnormal tumor blood supply.

[0014] Further, in step S50, constructing the boundary normal vector field based on the tumor target area mask includes: performing a three-dimensional Euclidean distance transformation on the binarized tumor target area mask to establish a spatial distance field that is uniformly increasing outward; applying a three-dimensional Gaussian filter to the spatial distance field for smoothing; calculating the spatial first derivative of the smoothed spatial distance field to obtain the boundary normal vector field pointing outward from the tumor.

[0015] Further, in step S50, performing spatial gradient calculation and weighted generation of invasion dynamics feature field on the abnormal metabolic intensity scalar field includes: calculating the partial derivative vectors of the abnormal metabolic intensity scalar field in the three spatial coordinate axes to extract the spatial gradient; performing a vector inner product operation between the spatial gradient and the boundary normal vector field, and removing the negative gradient components pointing into the tumor interior through a rectification function to obtain the invasion direction weight field; adjusting the invasion direction weight field using a gain control coefficient and fusing it with the original abnormal metabolic intensity scalar field to generate an invasion dynamics feature field characterizing tumor-specific microvascular invasion features.

[0016] Further, in step S60, the spatial pooling of the invasion dynamics feature field to generate image feature vectors includes: performing spatial pyramid pooling by applying multi-level scale grid division along the three spatial dimensions of the invasion dynamics feature field; performing max pooling operation in each divided local grid region to extract invasion dynamics extrema; and concatenating and stitching the extrema of each local grid region to generate a fixed-length one-dimensional image feature vector.

[0017] Further, in step S60, concatenating the image feature vector with the clinical variables of the target object to form a joint feature vector includes: obtaining continuous numerical variables such as the alpha-fetoprotein concentration, age level, and BCLC count of the target object as clinical variables; mapping the clinical variables to a distribution interval with zero mean and unit variance using a standard score normalization algorithm to generate a one-dimensional clinical feature vector; and cascading and concatenating the image feature vector and the one-dimensional clinical feature vector in the feature dimension to construct a multimodal fusion feature as a joint feature vector.

[0018] Furthermore, in step S60, the classifier model is a deep feedforward neural network model, and its network architecture is configured as follows: the model contains an input layer, multiple hidden layers and an output layer in sequence. A linear rectified function activation layer and a regularization layer are configured between adjacent fully connected layers. The output layer uses an activation function to map the features calculated by the network to a specified probability range for output.

[0019] A second aspect of the present invention provides a high-risk and low-risk classification system for MVI-positive hepatocellular carcinoma, comprising: The data acquisition module is used to acquire three-dimensional enhanced CT image sequences of the target object; The image registration module is used to perform registration on arterial phase images and portal venous phase images using plain scan images as a reference coordinate system, and generate a registered image sequence. A mask construction module is used to generate tumor target area masks and background liver parenchyma masks based on the registered image sequence; The vector computation module is used to calculate the inflow and outflow time gradients based on the tumor target area mask and the background liver parenchyma mask to generate a two-dimensional time-series perfusion vector, and to perform filtration and purification using a subset of primary active blood supply to the tumor. The vector decomposition module is used to determine the physiological baseline vector and project the purified two-dimensional time-series perfusion vector onto the physiological baseline vector to obtain orthogonal components. The feature reconstruction module is used to calculate the modulus to generate an abnormal metabolic intensity scalar field, construct a boundary normal vector field and perform a dot product operation between the spatial gradient and the normal vector field to generate an invasion direction weight field, and use the invasion direction weight field to generate an invasion dynamics feature field. The classification prediction module is used to generate image feature vectors by spatial pooling of the invasion dynamics feature field. The joint feature vector after concatenating clinical variables is input into the classifier model to output the probability distribution of the target object category.

[0020] A third aspect of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the steps of the MVI-positive hepatocellular carcinoma high-risk and low-risk classification method provided in the first aspect.

[0021] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the method for classifying high-risk and low-risk MVI-positive hepatocellular carcinoma provided in the first aspect.

[0022] The fifth aspect of this invention provides an application of the MVI-positive hepatocellular carcinoma high-risk and low-risk classification method provided in the first aspect above in the preparation of microvascular invasion risk assessment or tumor auxiliary diagnostic software.

[0023] This invention provides a method and application for classifying high-risk and low-risk hepatocellular carcinoma with MVI positivity. It has the following beneficial effects: 1. This invention constructs a two-dimensional temporal perfusion vector by calculating the inflow and outflow time gradients, and extracts spatial coordinates with positive inflow and negative outflow time gradients to construct a target blood supply subset, thereby filtering the perfusion vector. This step, based on the rapid inflow and outflow perfusion characteristics of contrast agents, locks the active blood supply area within the target region, effectively eliminating data interference from liquefied necrosis or delayed enhancement tissue within solid tumors, and increasing the proportion of microvascular proliferation region features in the model input data.

[0024] 2. This invention establishes a baseline vector based on perfusion data within the background liver parenchyma region, and obtains the orthogonal components of the tumor target area data by performing inner product projection calculations onto this baseline vector. This calculation process reduces the dependence of feature extraction on absolute grayscale values, weakens the systematic measurement bias caused by individual differences in blood circulation among different patients and fluctuations in the contrast agent injection rate, thereby accurately extracting abnormal perfusion features relative to the blood flow state of normal liver tissue.

[0025] 3. This invention solves for the spatial gradient of the calculated intensity scalar field and combines it with the boundary normal vector field constructed based on tumor morphology to generate a directional weight field for spatial weighting. This step quantifies the spatial direction of the expansion of abnormal blood supply into peripheral normal tissue, so that the final output feature field simultaneously possesses both metabolic intensity and spatial direction information. This compensates for the deficiency of conventional scalar feature extraction, which easily misses the spatial diffusion direction, and improves the classification prediction accuracy of subsequent classification models. Attached Figure Description

[0026] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a flowchart of the image sequence acquisition and deformation field constrained registration process of the present invention; Figure 4 This is a flowchart of the dual-space mask isolation and background extraction process of the present invention; Figure 5This is a flowchart of the temporal dynamics vector mapping and active blood supply purification process of the present invention; Figure 6 This is a flowchart of the physiological baseline extraction and vector orthogonal decomposition process of the present invention; Figure 7 This is a flowchart of the fluid dynamics feature reconstruction and invasion direction weighting of the present invention; Figure 8 This is a flowchart of the feature tensor pooling and multimodal collaborative prediction process of the present invention; Figure 9 This is a scatter plot of the dimensionality reduction distribution of the multidimensional feature manifold of the present invention; Figure 10 This is a heatmap showing the contribution of multimodal features in this invention.

[0027] The module includes: 10. Data acquisition module; 20. Image registration module; 30. Mask construction module; 40. Vector calculation module; 50. Vector decomposition module; 60. Feature reconstruction module; and 70. Classification prediction module. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] See attached document Figure 1 , Figure 1 This is an architectural block diagram of a high-risk and low-risk classification system for MVI-positive hepatocellular carcinoma according to an embodiment of the present invention. The present invention provides a high-risk and low-risk classification system for MVI-positive hepatocellular carcinoma, comprising: a data acquisition module 10, an image registration module 20, a mask construction module 30, a vector calculation module 40, a vector decomposition module 50, a feature reconstruction module 60, and a classification prediction module 70.

[0030] The data acquisition module 10 acquires a sequence of three-dimensional enhanced CT images of the target object. The three-dimensional enhanced CT image sequence includes plain phase images, arterial phase images, and portal venous phase images.

[0031] The image registration module 20 receives the three-dimensional enhanced CT image sequence output by the data acquisition module 10. Using the plain scan image as the reference coordinate system, the image registration module 20 performs affine transformation and B-spline non-rigid registration on the arterial phase image and the portal venous phase image, respectively, and generates a spatially aligned registered image sequence through interpolation and resampling.

[0032] The mask construction module 30 generates a tumor target mask and a background liver parenchyma mask based on the registered image sequence. The tumor target mask corresponds to the set of three-dimensional spatial coordinates of the lesion. The background liver parenchyma mask is generated by removing the location of the tumor target mask, the boundary buffer zone, and the vascular region from the extracted whole liver parenchyma region.

[0033] The vector calculation module 40 extracts the corresponding voxel gray values ​​based on the tumor target area mask and the background liver parenchyma mask. The module obtains the inflow temporal gradient by calculating the difference between the arterial phase image and the plain scan image, and the outflow temporal gradient by calculating the difference between the portal venous phase image and the arterial phase image. Combining the inflow and outflow temporal gradients, the module generates a two-dimensional temporal perfusion vector corresponding to each spatial voxel. Based on the quadrant distribution attributes of the two-dimensional temporal perfusion vector, the module extracts the spatial coordinates of voxels with positive inflow temporal gradients and negative outflow temporal gradients, constructing a subset of primary active blood supply to the tumor. The module performs a Boolean masking operation on the two-dimensional temporal perfusion vector using this subset, filtering out voxels within the tumor that exhibit low perfusion necrosis or delayed enhancement fibrosis, and outputs a purified two-dimensional temporal perfusion vector.

[0034] The vector decomposition module 50 acquires the two-dimensional temporal perfusion vector within the background liver parenchyma mask, and determines the physiological baseline vector based on the average value of the two-dimensional temporal perfusion vector of each voxel within the background liver parenchyma region. The vector decomposition module 50 projects the two-dimensional temporal perfusion vector within the tumor target area mask onto the reference direction corresponding to the physiological baseline vector, and separates the abnormal metabolic orthogonal components relative to the physiological baseline direction.

[0035] As an optional implementation, the vector decomposition module 50 can also perform multivariate kernel density estimation on the two-dimensional temporal perfusion vector within the background liver parenchyma mask, and use the vector corresponding to the main peak of the probability density distribution as the physiological baseline vector.

[0036] The feature reconstruction module 60 receives the orthogonal component dataset, calculates the Euclidean modulus of the abnormal metabolic orthogonal components corresponding to each spatial voxel, and generates a scalar field of abnormal metabolic intensity. The feature reconstruction module 60 constructs a spatial distance field based on the tumor target area mask and further calculates the boundary normal vector field pointing outwards from the tumor. The feature reconstruction module 60 performs a three-dimensional spatial gradient operation on the abnormal metabolic intensity scalar field and multiplies this spatial gradient by the boundary normal vector field to obtain an invasion direction weight field characterizing the trend of abnormal metabolism spreading to the tumor periphery. The feature reconstruction module 60 uses the invasion direction weight field to perform voxel-by-voxel weighting on the abnormal metabolic intensity scalar field to generate an invasion dynamics feature field.

[0037] In this invention, the aforementioned invasion dynamics feature field is used to characterize the intensity of abnormal blood supply to the tumor site and its spatial tendency to expand outward along the normal direction of the tumor boundary.

[0038] The classification prediction module 70 performs spatial pooling on the invasion dynamics feature field to generate a fixed-length image feature vector. The classification prediction module 70 concatenates the image feature vector with the clinical variables of the target object to form a joint feature vector. The classification prediction module 70 inputs the joint feature vector into a preset classifier model for calculation and outputs the probability distribution of the target object's category.

[0039] As a preferred implementation, the spatial pooling process is spatial pyramid pooling performed along the three spatial dimensions of the invasion dynamics feature field; the classifier model is a deep feedforward neural network model. As an alternative implementation, the spatial pooling process can also form an image statistical feature vector by calculating at least one of the statistics of mean, variance, skewness, and kurtosis of the invasion dynamics feature field.

[0040] See attached document Figure 2 , Figure 2 This is a flowchart of a high-risk / low-risk classification method for MVI-positive hepatocellular carcinoma according to an embodiment of the present invention. The present invention provides a high-risk / low-risk classification method for MVI-positive hepatocellular carcinoma, comprising the following steps: S10: Obtain the three-dimensional enhanced CT image sequence of the target object, and perform registration on the arterial phase image and portal venous phase image using the plain scan image as the reference coordinate system to generate the registered image sequence. S20, Generate a tumor target area mask and a background liver parenchyma mask based on the registered image sequence; S30: Based on the tumor target area mask and the background liver parenchyma mask, extract voxel gray values, calculate the inflow time gradient and outflow time gradient, and generate a two-dimensional temporal perfusion vector corresponding to each spatial voxel. The horizontal axis of the two-dimensional temporal perfusion vector is defined as the inflow time gradient, and the vertical axis is defined as the outflow time gradient. Extract the spatial coordinates of voxels located in the fourth quadrant, i.e., the inflow time gradient is positive and the outflow time gradient is negative, to construct a subset of primary active blood supply to the tumor. Use the subset of primary active blood supply to filter out voxels that show necrosis or fibrosis metabolic characteristics, and output the purified two-dimensional temporal perfusion vector. S40, perform regional averaging calculation on the two-dimensional temporal perfusion vector within the background liver parenchyma mask to determine the corresponding physiological baseline vector, and perform projection on the two-dimensional temporal perfusion vector within the tumor target mask onto the physiological baseline vector to obtain orthogonal components; as an optional implementation, multivariate kernel density estimation can also be performed on the two-dimensional temporal perfusion vector within the background liver parenchyma mask, and the vector corresponding to the main peak can be used as the physiological baseline vector. S50 calculates the Euclidean modulus of orthogonal components in three-dimensional space to generate an abnormal metabolic intensity scalar field; constructs a spatial distance field based on the tumor target mask and calculates the boundary normal vector field; performs spatial gradient operation on the abnormal metabolic intensity scalar field, and generates an invasion direction weight field by multiplying the spatial gradient with the boundary normal vector field; and uses the invasion direction weight field to perform weighting on the abnormal metabolic intensity scalar field to generate an invasion dynamics feature field. S60 performs spatial pooling on the invasion dynamics feature field to generate an image feature vector. The image feature vector is then concatenated with the clinical variables of the target object to form a joint feature vector. This joint feature vector is then input into the classifier model to output the probability distribution of the target object category.

[0041] As a preferred implementation, spatial pooling is a multi-level spatial pyramid pooling.

[0042] As an optional embodiment of the present invention, the high-risk or low-risk classification method or system provided by the present invention can also be combined with a conventional preliminary screening model for microvascular invasion status to construct a two-step clinical cascade assessment strategy. In specific implementation, the medical system can first use basic imaging and clinical laboratory data to assess whether the target subject has microvascular invasion, i.e., determine whether it is positive or negative; subsequently, only for the patient cohort initially screened or pathologically diagnosed as having positive microvascular invasion, the invasion dynamics feature extraction and high-risk or low-risk classification process of the present invention is triggered and executed. This two-step strategy can be flexibly integrated into existing clinical workflows, avoiding over-calculation for low-risk patients with negative microvascular invasion while achieving precise targeted risk stratification for patients with positive microvascular invasion.

[0043] The following section will provide a detailed explanation of each step in the above method, taking into account the specific implementation mechanism.

[0044] See attached document Figure 3 , Figure 3 This is a flowchart of image sequence acquisition and deformation field constrained registration according to an embodiment of the present invention. In this embodiment, step S10 of the present invention, through the interactive cooperation between the data acquisition module 10 and the image registration module 20, may further include the following specific execution steps: S110, the data acquisition module 10 receives the three-dimensional enhanced CT image sequence of the target object and constructs the original three-dimensional spatial tensor matrix.

[0045] Specifically, the data acquisition module 10 acquires a sequence of three-dimensional enhanced CT images, including plain scan images, arterial phase images, and portal venous phase images. Subsequently, the system converts the image data, conforming to medical digital imaging and communication standard formats, into a three-dimensional spatial tensor matrix. For ease of mathematical expression and subsequent calculations, the three-dimensional physical space voxel coordinates are set as follows: The set of phase scan times is defined as follows: .in, Corresponding to the time points of the plain scan period, Corresponding to the arterial phase time point, This corresponds to the portal venous phase time point. Based on this, the data acquisition module 10 generates the original three-dimensional spatial tensor matrix. The original three-dimensional spatial tensor matrix completely records the grayscale attenuation values ​​of the corresponding spatial voxel positions at different time points, thus forming the basic data input for subsequent spatial registration calculations.

[0046] In a preferred embodiment, after completing subsequent image registration, the system uniformly resamples the multi-phase images to a preset spatial resolution, so that subsequent morphological operations, distance transformations, gradient calculations, and pooling processes are performed at a unified physical scale. Preferably, the preset spatial resolution is an isotropic voxel resolution.

[0047] S120, the image registration module 20 uses the flat scan image as the reference coordinate system to perform global spatial coarse alignment based on affine transformation.

[0048] After acquiring the basic data input, the image registration module 20 receives the original three-dimensional spatial tensor matrix output by the data acquisition module 10. To overcome positional deviations caused by changes in patient position between different scanning phases, this embodiment fixes the tensor of the plain scan image. As a reference spatial coordinate system, the image registration module 20 applies tensors to the arterial phase image respectively. portal venous phase image tensor Perform affine transformation calculations.

[0049] During the calculation process, the system constructs a linear transformation matrix containing translation, rotation, and scaling parameters, and applies this matrix to perform a global linear mapping of the three-dimensional coordinates of the target phase image. The final output is an intermediate image tensor with globally rigid spatial alignment. This step effectively reduces the rigid displacement error caused by the overall body movement of the target object during scanning. The iterative process for solving the specific translation, rotation, and scaling parameters in the affine transformation matrix can be performed using conventional gradient descent optimization algorithms. The specific mathematical derivation principles are well-known in the field and will not be elaborated upon here.

[0050] S130, the image registration module 20 constructs an objective function containing a smoothing penalty term, performs B-spline non-rigid registration and interpolation resampling, and generates a registered image sequence.

[0051] While global spatial coarse alignment addresses the overall rigid displacement problem, it struggles to handle local soft tissue deformations caused by respiratory movements or organ peristalsis. As a preferred approach, after acquiring the intermediate image tensor, the image registration module 20 applies a B-spline non-rigid registration algorithm to compensate for local deformations. The non-rigid spatial transformation mapping function is defined as follows: To prevent topological tearing and overlap of spatial voxel coordinates during local deformation matching, the image registration module 20 constructs a registration objective function that includes a spatial smoothing regularization penalty term. The registration objective function is specifically defined as follows: ; In the formula, This instructs that independent registration calculations be performed on the arterial phase image tensor and the portal venous phase image tensor, respectively. This represents the normalized mutual information similarity measure, used to measure the joint grayscale consistency between the reference image and the deformed image; This represents the target phase image tensor after non-rigid transformation mapping; This represents the range of action of the integral in the voxel space of the global 3D image; Represents a three-dimensional voxel integral element; This represents the result of the Laplacian operator acting on a non-rigid spatial transformation mapping function, and characterizes the second-order rate of change of the spatial displacement vector of adjacent voxels; The squared Euclidean norm 2 of a vector is used to quantize the energy magnitude of the second-order operatorization rate; The smoothing regularization coefficient is preferably in the range of 0.1 to 1.0.

[0052] The negative sign of the similarity term in the objective function is used to transform the problem of maximizing image similarity into the problem of minimizing the objective function.

[0053] In the specific implementation of the calculation, the image registration module 20 solves for the minimum value of the above-mentioned registration objective function through iterative calculation. Preferably, when the relative rate of change of the objective function corresponding to a preset number of consecutive iterations is less than a preset convergence threshold, the iteration stops, and the current corresponding spatial transformation mapping is determined as the optimal non-rigid deformation field.

[0054] Based on the optimal non-rigid deformation field, the image registration module 20 calls a cubic spline interpolation algorithm to resample and calculate the original voxel grayscale data, thereby outputting a registered image sequence with substantially aligned spatial coordinates. The sequence includes flat scan images in a unified coordinate reference system. Arterial phase images And portal venous phase images This establishes an accurate spatiotemporal computational foundation for subsequent steps.

[0055] When it is necessary to perform spatial mapping on the tumor mask or liver mask generated by the segmentation step, the system performs a spatial transformation on the corresponding mask that is consistent with the image registration, and uses a resampling method suitable for preserving the boundary attributes of the binary mask to complete the coordinate alignment.

[0056] See attached document Figure 4 , Figure 4 This is a flowchart of dual spatial mask isolation and background extraction according to an embodiment of the present invention. In this embodiment, step S20 of the present invention, through spatial logic operations of the mask construction module 30, can further include the following specific execution steps: S210, the mask construction module 30 defines the boundaries of tumor lesions based on the registered image sequence and generates a tumor target area mask.

[0057] The mask construction module 30 receives a sequence of registered images in a unified coordinate reference system. Based on the contrast enhancement difference between the arterial and portal venous phase images, the system extracts a set of three-dimensional spatial coordinates of the lesion. The three-dimensional physical space voxel coordinates are set as follows: The mask construction module 30 constructs a binary tumor target region mask with the same size as the original image. .

[0058] When spatial coordinates When located inside a tumor lesion, set ; Conversely, when spatial coordinates When located outside the tumor lesion, set .

[0059] The above process can effectively define the structural location boundaries of abnormally proliferating tissues within the target body.

[0060] As a preferred approach, when performing three-dimensional contour segmentation of tumor lesions, those skilled in the art can use a three-dimensional convolutional neural network segmentation model, such as a three-dimensional U-Net network, for calculation. If such a model structure is adopted, its input receives the above-mentioned registered and aligned multi-phase CT tensor sequence. After the three-dimensional convolutional layer of the internal encoder extracts multi-scale spatial features, the decoder reconstructs a voxel-level classification probability map with the same size as the original image through jump-linking. After thresholding, the above-mentioned binarized target area mask can be output.

[0061] As an alternative implementation, tumor target masks can also be generated through manual delineation, semi-automatic region growth segmentation, image segmentation, or other existing medical image segmentation methods that can obtain the spatial boundaries of the lesion.

[0062] When using an automated segmentation model, it is preferable to have the segmentation results manually reviewed or revised to improve the accuracy of the tumor target mask.

[0063] S220, Mask construction module 30 extracts the spatial coordinates of the entire liver parenchyma of the target object and generates an initial liver mask.

[0064] After defining the tumor boundary, in order to establish a normal physiological reference frame relative to the pathological tissue, the system needs to obtain the distribution range of normal liver tissue unaffected by tumor metabolism. In this embodiment, the mask construction module 30 uses plain scan images or portal venous images as the anatomical basis to extract the connected region containing the entire liver organ boundary, thereby generating a binarized initial liver mask. .

[0065] In the initial liver mask, the voxel coordinates belonging to the interior of the liver parenchyma are assigned a value of 1, while the voxel coordinates of the surrounding tissue spaces and other organs outside the liver are assigned a value of 0. For the three-dimensional extraction and internal hole filling operations of the initial liver organ region, those skilled in the art can combine morphological closing operations for region extraction. The image connected component analysis operation is a well-known technique in the art and will not be described in detail here.

[0066] As a preferred implementation, the initial liver mask can be obtained by performing threshold segmentation, connected component filtering, hole filling, and boundary smoothing on the portal venous phase image. As an alternative implementation, the initial liver mask can also be obtained using existing automatic liver segmentation models or through manual revision.

[0067] S230, the mask construction module 30 performs morphological dilation and grayscale thresholding to construct a background liver parenchyma mask.

[0068] Since there are usually high-incidence areas of microvascular invasion around liver cancer lesions, and the contrast agent perfusion pattern of the main blood vessels in the liver is significantly different from that of the capillary bed in the liver parenchyma, if these areas are calculated directly without processing, these areas will seriously interfere with the establishment of the subsequent normal physiological metabolic baseline.

[0069] Based on this, the mask construction module 30 uses a morphological dilation algorithm to mask the tumor target area. Spatial expansion is performed to construct a boundary buffer zone. From a spatial geometry perspective, the dilation operation slides through a predefined 3D structural kernel in image space, uniformly expanding the 3D morphology of the tumor outwards. Specifically, a 3D spherical structural element is defined as follows: Its radius is set to Based on the statistical distribution of the physical distance of peripheral diffusion from microvascular invasion, the radius... The value is typically set between 10 mm and 20 mm to cover high-risk areas of tumor infiltration at the tumor margin. The mask construction module 30 performs dilation calculations to generate a buffer zone mask. The morphological calculation process is defined as follows: ; In the formula, The morphological expansion operation symbol is used to topologically extend the three-dimensional morphology of a lesion outward by a specified distance. It should be understood that this is a buffer zone mask. Includes tumor target area mask The corresponding lesion body area and the boundary buffer area formed by its outward expansion.

[0070] During the synchronous or alternating execution phase of processing the tumor boundary buffer zone, the mask construction module 30 utilizes the portal venous phase image tensor generated by registration. Extract the highlighted main blood vessel regions within the liver. Set the grayscale segmentation threshold for the main blood vessels to [value missing]. Based on the distribution pattern of Huntsfield units in conventional medical contrast-enhanced CT, The value is typically set between 150 HU and 200 HU. The system generates a vascular region mask through voxel-by-voxel grayscale comparison. The specific logical conditions are as follows: When the gray value of local coordinates At that time, assign a value ; Conversely, set .

[0071] With this basic data complete, the mask construction module 30 performs Boolean difference algebra operations within the global anatomical scope defined by the initial liver mask, effectively stripping away the tumor lesion itself, the outwardly extending boundary buffer zone, and the main vascular region. Background: Liver parenchyma mask The specific spatial algebraic calculation formula is defined as follows: ; In the formula, Indicates the initial liver mask in spatial coordinates The value at; Indicates the boundary buffer zone mask in spatial coordinates The value at; Represents the spatial coordinates of the vascular region mask. The value at the specified position; • In binary state, it is equivalent to a spatial logical AND operation; Expression This represents the point-by-point inversion of the boundary buffer zone mask, expressed as: This indicates that the mask for the vascular region is inverted point by point. Since the boundary buffer zone mask includes the tumor body region and its outward buffer region, the above calculation simultaneously removes the tumor lesion itself, the boundary buffer zone, and the main vascular region.

[0072] The final output remains a binary state of 0 or 1. After the above calculations, the mask construction module 30 outputs an independent tumor target region mask. and background liver parenchyma mask This allows for the isolation and extraction of the dual spatial masks. The background liver parenchyma mask locks in the coordinates of the liver parenchyma capillary bed, which is unaffected by tumor invasion and large vessel blood flow disturbance, providing reliable spatial data support for the subsequent establishment of an individual's systemic physiological metabolic baseline vector.

[0073] Preferably, the system also verifies the number of effective voxels in the background liver parenchyma mask; when the number of effective voxels is lower than a preset threshold, the baseline construction calculation of the current sample can be stopped, or a preset alternative background region can be used to reconstruct the physiological baseline vector.

[0074] See attached document Figure 5 , Figure 5 This is a flowchart of temporal dynamic vector mapping and active blood supply purification according to an embodiment of the present invention. In this embodiment, step S30 of the present invention, through data processing and mathematical transformation by the vector calculation module 40, can further include the following specific execution steps: S310, the vector calculation module 40 obtains the registered and aligned image tensor sequence and calculates the first-order temporal gradient field of the contrast agent flow.

[0075] In contrast-enhanced CT scans, the dynamic evolution of contrast agent concentration within tissues directly reflects the local microcirculation status. Based on the physical principle that CT attenuation values, typically expressed in Huntsfield units, are approximately linearly positively correlated with contrast agent iodine concentration, the system calculates the rate of voxel grayscale evolution over time using a time-difference method to quantify blood perfusion characteristics.

[0076] In this embodiment, the vector calculation module 40 extracts the set of scan time nodes from the original medical digital imaging and communication standard data header file, wherein... , and These correspond to the scan times for the plain scan phase, arterial phase, and portal venous phase, respectively. The inflow phase time interval is defined as... The outflow period interval is .

[0077] Since multi-phase scanning inevitably involves a temporal sequence, and All values ​​are greater than zero, thus ensuring the validity of subsequent time difference calculations. The vector calculation module 40 calculates each three-dimensional spatial coordinate... Calculate the scalar gradient during the inflow period respectively. and outflow scalar gradient The specific algebraic formulas are defined as follows: ; ; In the formula, , and These represent the plain scan phase, arterial phase, and portal venous phase in the same spatial coordinates. The grayscale attenuation value at the location. The calculated first-order temporal gradient field eliminates the absolute numerical error caused by the inconsistency of scanning time intervals and can objectively reflect the inflow and outflow rates of contrast agent in the local microvascular network.

[0078] S320, the vector calculation module 40 orthogonally combines the first-order temporal gradient fields to construct a two-dimensional temporal infusion vector field.

[0079] After calculating the independent scalar gradients, the vector calculation module 40 combines the inflow and outflow scalar gradients corresponding to each voxel in an orthogonal coordinate system, aiming to construct a two-dimensional temporal perfusion vector field with a temporal evolution direction. The system uses the inflow scalar gradient... As a component of the horizontal axis, the outflow scalar gradient As a component of the vertical axis, at each spatial voxel position Constructing a two-dimensional column vector The specific algebraic expression of the two-dimensional temporal injection vector field is as follows: ; In the formula, This represents the matrix transpose operation, used to convert a row vector into a column vector. Representing spatial coordinates The two-dimensional temporal injection vector at that location. The constructed two-dimensional temporal injection vector field. This method transforms discrete multi-phase scalar images into continuous kinetic vector representations, providing reliable data support for subsequent metabolic space analysis.

[0080] S330, the vector calculation module 40 performs Boolean masking filtering based on the blood flow metabolism quadrant attributes, and outputs the purified active blood supply vector field.

[0081] Significant spatial heterogeneity often exists within malignant solid tumors such as hepatocellular carcinoma. As a preferred approach, the system further extracts voxel subsets characterizing specific high-risk pathological attributes. Typical microvascular invasion lesions typically exhibit rapid enhancement in the arterial phase due to abnormal hepatic artery supply, and rapid clearance in the portal venous phase due to arteriovenous fistula or accelerated venous return. This rapid inflow and outflow pathological metabolic characteristic, mapped onto the aforementioned two-dimensional orthogonal coordinate system, is mainly distributed in the fourth quadrant, where the inflow gradient is positive and the outflow gradient is negative. In contrast, liquefied necrotic areas within tumors typically exhibit a low-gradient state with no significant enhancement, while benign lesions such as cavernous hemangiomas often show delayed enhancement, i.e., a positive outflow gradient.

[0082] Based on the aforementioned differences in physical properties, the vector calculation module 40 constructs a logical mask. The inflow tolerance threshold for filtering background noise from the equipment is set to [value]. The outflow tolerance threshold is Considering the baseline drift of CT scans and slight fluctuations in the background perfusion of healthy tissues and organs, and The value is typically set within the real range of 0 to 5 Huntsfield units per second. Those skilled in the art can determine the specific magnitude of the above tolerance threshold by statistically analyzing the variance of grayscale fluctuations in healthy tissues from the same imaging position during the plain scan period. The system generates the mask by performing a voxel-by-voxel logical judgment: Preferably, the logical judgment is performed within the spatial range defined by the tumor target area mask. This occurs when the following conditions are met: and hour; Then assign a value: ; Otherwise, assign a value: ; In the formula, Indicates the inflow tolerance threshold. The values ​​represent the outflow tolerance threshold, and the preferred range for both is 0 to 50 Huntsfield units per second.

[0083] Furthermore, the vector computation module 40 injects the original two-dimensional temporal vector field. With logical mask The formula for point-by-point scalar multiplication is: ; In the formula, The purified temporal perfusion vector field is represented by the multiplication sign, indicating a scalar multiplication operation between a scalar and a column vector. Through this quadrant-based Boolean masking filter, the system can effectively eliminate ineffective metabolic voxels exhibiting liquefactive necrosis, cystic degeneration, or delayed enhancement, while retaining a subset of active blood supply representing high-risk abnormal angiogenesis. The purified vector field not only reduces dimensional redundancy in subsequent computations but also improves the signal-to-noise ratio of tumor invasiveness features. In a preferred embodiment, the purified temporal perfusion vector field also satisfies tumor target area mask constraints, limiting the retained voxels to a subset of active blood supply within the tumor region exhibiting rapid-entry, rapid-hemispheric characteristics.

[0084] See attached document Figure 6 , Figure 6 This is a flowchart of physiological baseline extraction and vector orthogonal decomposition according to an embodiment of the present invention. In this embodiment, step S40 of the present invention, through algebraic projection and spatial statistical calculation by the vector decomposition module 50, may further include the following specific execution steps: S410, the vector decomposition module 50 extracts the individualized physiological metabolic baseline vector based on the background liver parenchyma mask.

[0085] In contrast-enhanced CT scans, differences in cardiac output among patients and objective fluctuations in contrast agent injection rates can lead to systematic biases in the absolute grayscale changes of medical image sequences. To establish a quantitative standard unaffected by individual differences in the systemic circulatory system, normal liver tissue unaffected by tumor metabolic invasion was systematically extracted as a benchmark reference.

[0086] Vector decomposition module 50 obtains the background liver parenchyma mask generated in step S20. and the original two-dimensional temporal perfusion vector field constructed in step S30 In specific calculations, the system calculates the average perfusion vector of the background region within the global anatomical space, defining it as the physiological metabolic baseline vector. .

[0087] The specific formula for calculating spatial integrals is defined as follows: ; In the formula, Represents the integration range in the voxel space of the global 3D image; Represents a three-dimensional voxel integral element; This indicates the background liver parenchyma mask in spatial coordinates. The value at; This represents the scalar multiplication result of the binary mask and the two-dimensional perfusion vector. The numerator represents the cumulative sum of all two-dimensional temporal perfusion vectors within the background liver parenchyma region, and the denominator represents the total number of effective voxels within the background liver parenchyma region.

[0088] As an optional implementation, the system can also perform two-dimensional nuclear density estimation on the two-dimensional time-series perfusion vector in the background liver parenchyma region, and determine the two-dimensional vector corresponding to the main peak of the probability density as the physiological metabolic baseline vector.

[0089] Considering that the actual objects processed in medical image processing are discrete voxel grids, the aforementioned continuous integration operation is equivalent to a discrete summation operation in engineering implementation. Assuming the background liver parenchyma mask meets the preset effective voxel count requirement, the denominator is greater than zero, thus ensuring the validity of the baseline vector calculation.

[0090] S420, Vector decomposition module 50 constructs a two-dimensional orthogonal feature basis based on physiological metabolic baseline vectors.

[0091] After obtaining individualized physiological metabolic baselines, the vector decomposition module 50 establishes a reconstructed reference coordinate system within a two-dimensional kinetic plane. The system then converts the physiological metabolic baseline vector... Algebraic normalization is performed to extract parallel unit vectors representing the direction of normal physiological perfusion. The specific calculation method is to... Divide by its Euclidean norm 2 To avoid zero-vector division crashes caused by extreme pathological dead zones or abnormal data acquisition, the system presets a minimal constant. As a denominator protection term, its value is set to 10⁻⁶. The specific algebraic formula for parallel unit vectors is: ; As a preferred approach, the system further solves for orthogonal unit vectors perpendicular to the parallel unit vectors in a two-dimensional orthogonal plane. .set up: ; In the formula, Indicates the inflow gradient component. Let represent the outflow gradient component. Then the orthogonal unit vector can be represented as: ; In the formula, The Euclidean norm 2 of the physiological metabolic baseline vector, and the parallel unit vector. Orthogonal unit vectors Together, they constitute a two-dimensional orthogonal feature basis describing the local microvascular metabolic state. The constructed orthogonal feature basis removes the original absolute gray-level gradient coordinate axis limitation and establishes a relative observation coordinate system based on the evolution direction of normal liver blood flow in the patient.

[0092] It should be understood that the two-dimensional orthogonal feature basis corresponds to the two-dimensional infusion property representation at each three-dimensional spatial voxel location; therefore, the two-dimensional vector components at each voxel location can be further mapped to a scalar field or vector field defined on the three-dimensional voxel mesh for use in subsequent spatial gradient and orientation weighted calculations.

[0093] S430, the vector decomposition module 50 performs vector projection calculations of the tumor's active blood supply and separates abnormal metabolic feature components.

[0094] The blood supply to solid malignant tumors such as liver cancer mainly originates from arterialized neovascularization, causing their contrast agent perfusion pattern to deviate significantly from the liver parenchyma metabolic baseline dominated by the portal vein. Vector decomposition module 50 obtains the purified active blood supply vector field output from step S30. The system then projects the voxel-by-voxel onto an orthogonal feature basis. In this embodiment, the system focuses on calculating the active blood supply on orthogonal unit vectors. The orthogonal projection component physically removes the same in-direction perfusion background signal as normal liver parenchyma, quantifying the degree of abnormal blood supply deviating from the normal physiological metabolic baseline. The separated abnormal metabolic feature vector field is defined as... The specific formulas for vector inner product projection and reconstruction are as follows: ; In the formula, This represents the projection coefficient of the purified infusion vector onto the direction of the orthogonal unit vector. This represents the algebraic inner product operation of two column vectors, used to calculate the scalar coefficients of the projection; the multiplication on the right-hand side of the projection formula is the multiplication of the scalar coefficients and the column vectors. The scalar multiplication operation is used to recover the spatial vector properties in orthogonal directions.

[0095] Through orthogonal decomposition calculations, the vector decomposition module 50 effectively decouples the mixed systemic circulatory background signal from the original dynamic signal of local tumor pathological features, outputting an abnormal metabolic vector containing only tumor-specific microvascular invasion features. This decoupling operation not only improves the signal-to-noise ratio of high-risk blood supply feature expression but also provides a unified and objective data foundation for subsequent large-sample quantitative comparative analysis across individual patients.

[0096] See attached document Figure 7 , Figure 7 This is a flowchart of fluid dynamics feature reconstruction and invasion direction weighting according to an embodiment of the present invention. In this embodiment, step S50 of the present invention, through spatial gradient calculation and direction mapping of the feature reconstruction module 60, may further include the following specific execution steps: S510, Feature Reconstruction Module 60 calculates the local intensity distribution of abnormal metabolic features and generates a metabolic intensity scalar field.

[0097] In medical image analysis, to transform abstract metabolic vectors into physical quantities characterizing tumor invasiveness, the system needs to perform dimensionality reduction assessments of the degree of metabolic abnormalities by incorporating anatomical spatial structures. In this embodiment, the system extracts the spatial distribution intensity of abnormal blood supply. The feature reconstruction module 60 acquires the abnormal metabolic feature vector field output in step S40. Based on vector algebra theory, the module assigns each spatial coordinate... The Euclidean norm 2 of the abnormal metabolic feature vector is calculated, mapping it from a two-dimensional metabolic space to a scalar value in a three-dimensional physical space. The specific calculation formula is defined as follows: ; In the formula, Representing local coordinates Abnormal metabolic intensity at the site, This represents the Euclidean modulus of the abnormal metabolic feature vector. This scalar field aims to reflect the activity level of abnormal microvascular perfusion in tumors. High-intensity regions correspond to the lesion core or active periphery with dense neovascularization and vigorous metabolism.

[0098] Preferably, the abnormal metabolic intensity scalar field is defined in a three-dimensional voxel grid consistent with the registered image sequence, so that spatial gradient calculations can be performed directly thereafter.

[0099] S520, the feature reconstruction module 60 constructs a spatial distance field and a boundary normal vector field based on the tumor target area mask, and establishes the anatomical direction of tumor infiltration.

[0100] Based on the establishment of scalar intensity distribution, quantitative analysis of the physical trend of microvascular invasion relies on the accurate extraction of the anatomical direction of tumor tissue infiltration into the peripheral healthy liver. As a preferred method, the feature reconstruction module 60 extracts the binary tumor target mask generated in step S20. A three-dimensional Euclidean distance transformation is performed on it. Since the Euclidean distance transformation can establish a spatial field that uniformly increases outward from the tumor boundary, its gradient direction naturally points towards the peripheral region away from the lesion interior, highly matching the physiological characteristics of tumor cells breaking through the capsule and spreading outward. The shortest physical distance from the tumor boundary voxel to the tumor surface is defined as... Its unit is millimeters.

[0101] To obtain smooth and continuous geometric orientations, the module modulates the range field. Apply a three-dimensional Gaussian filter. Set the standard deviation of the Gaussian kernel function to 1. Combined with the conventional spatial resolution of medical images, The value is set to 1.0 to 2.0 voxel spacing.

[0102] In practical applications, those skilled in the art can determine the slice thickness and interslice spacing parameters based on the specific CT scanning equipment. Dynamic adjustments are made to balance smoothing effects with boundary feature preservation. The feature reconstruction module calculates the spatial first derivative of the smoothed distance field to obtain the normal vector field pointing outwards from the tumor. The specific gradient normalization formula is as follows: ; In the formula, Represents the spatial distance field in coordinates The value at; This represents the gradient vector of the spatial distance field in three-dimensional space. Represents the normalized spatial boundary vector; For the denominator protection term, its preferred value is 10. -5 Constructed boundary normal vector field This provides a directional reference with clear anatomical significance for subsequent quantification of invasion trends. For the Euclidean distance transformation and Gaussian smoothing filtering of 3D images, those skilled in the art can implement them using conventional image processing algorithm libraries; the underlying computational logic is well-known in the field and will not be elaborated upon here.

[0103] Preferably, the boundary normal vector field and the abnormal metabolic intensity scalar field are defined in the same three-dimensional spatial coordinate system, thereby enabling the voxel-by-voxel directional weighted calculation to be performed at the same spatial location.

[0104] S530, Feature Reconstruction Module 60 calculates the spatial diffusion gradient of the metabolic intensity scalar field and performs directional weighting to generate an invasion dynamics feature field.

[0105] Clinical pathological studies have shown that the spread of tumor cells to surrounding tissues triggers compensatory angiogenesis in the peripheral blood vessels, which manifests macroscopically as a tendency for highly abnormal metabolic zones to spread to surrounding tissues. Based on this principle, the feature reconstruction module 60 pairs metabolic intensity scalar fields. Computational space three-dimensional gradient The extracted gradient vector characterizes the spatial rate of change of abnormal metabolic intensity and the direction of its outward diffusion.

[0106] Liquefaction necrosis or tissue heterogeneity within a tumor can also generate internal spatial gradient interference. Therefore, in its implementation, the system focuses on extracting invasive behavior that spreads outward from the tumor. The feature reconstruction module 60 converts the metabolic intensity spatial gradient... With the aforementioned boundary normal vector A vector inner product operation is performed to separate the dynamic component consistent with the tumor invasion direction. The specific invasion direction weighted formula is defined as follows: ; In the formula, The scalar field representing the intensity of abnormal metabolism in spatial coordinates Spatial gradient at a given location; This represents the inner product operation of three-dimensional vectors; Indicates the weight field of the invasion direction; This indicates a rectification operation used to preserve the positive diffusion trend along the boundary normal vector pointing outwards from the tumor.

[0107] After obtaining the aforementioned directional features, the system spatially fuses the directional weighted weights with the original abnormal metabolic intensity to output an invasion dynamics feature field. The specific algebraic composition method is as follows: ; In the formula, This represents scalar multiplication. This is the gain control coefficient, with a value ranging from 0.5 to 1.5, used to adjust the influence of the hydrodynamic diffusion gradient on the overall features. In practice, technicians can retrospectively analyze historical case cohorts with clear pathological gold standards for microvascular invasion, and use conventional hyperparameter optimization algorithms such as grid search to determine the optimal feature discrimination. value.

[0108] After the aforementioned spatial reconstruction and orientation-weighted calculations, the feature field output by the feature reconstruction module 60 helps to incorporate the hydrodynamic trend of the lesion's expansion into the peripheral liver parenchyma while preserving the absolute degree of metabolic abnormality at the local level. This extraction method, which takes into account both metabolic and spatial anatomical features, provides structurally transparent and physically meaningful data support for the quantitative assessment of microvascular invasion.

[0109] In this invention, the invasion dynamics characteristic field This serves as the basis for image input features in subsequent classification and prediction steps.

[0110] See attached document Figure 8 , Figure 8 This is a flowchart of feature tensor pooling and multimodal collaborative prediction according to an embodiment of the present invention. In this embodiment, step S60 of the present invention, through the dimensionality reduction mapping and network model calculation of the classification prediction module 70, may further include the following specific execution steps: S610, the classification prediction module 70 performs spatial pyramid pooling processing on the invasion dynamics feature field to extract scale-invariant global image feature vectors.

[0111] The lesion volume of hepatocellular carcinoma varies objectively among different patients. Directly flattening the three-dimensional feature field often leads to a mismatch in the dimension of subsequent network input features. To address the issue of data dimension uniformity, this embodiment introduces a spatial pyramid pooling mechanism. The classification prediction module 70 acquires the three-dimensional invasion dynamics feature field output in step S50. The system applies grids of different scales along the three spatial dimensions of the feature field.

[0112] As a preferred approach, the grid division levels are set to 1×1×1, 2×2×2, and 4×4×4. Within each local grid region, the module performs max pooling to extract the invasion dynamics extrema within the grid region. After completing the pooling calculations for all grids, the system concatenates the obtained extrema from each region to generate a fixed-length one-dimensional image feature vector. Through the multi-scale pooling process described above, the system not only effectively compresses high-dimensional spatial data, but also preserves the local extreme features of microvascular invasion at different anatomical levels, ensuring the structural consistency of the input dimensions of subsequent network models.

[0113] As an alternative implementation, the classification prediction module 70 can also calculate at least one of the following statistics among the mean, variance, skewness and kurtosis of the invasion dynamics feature field to form an image statistical feature vector.

[0114] S620, the classification prediction module 70 extracts patient-related clinical test indicators and constructs a multimodal fusion input vector.

[0115] The mechanisms of microvascular invasion, besides manifesting as hydrodynamic changes on medical imaging, are also related to macroscopic biochemical metabolic indicators in patients. Based on the principle of multidimensional feature complementarity, the system simultaneously acquires key clinical laboratory data of the target subject, specifically including alpha-fetoprotein concentration, age level, and BCLC count. Due to the large range of physical dimensions among different clinical laboratory indicators, the module performs standardization on continuous numerical variables, using a Z-score normalization algorithm to map them to a distribution interval with zero mean and unit variance, thereby generating a one-dimensional clinical feature vector. Based on this, the classification and prediction module 70 divides the image feature vector into feature dimensions. Or the corresponding image statistical feature vector and clinical feature vector Cascaded concatenation is performed to construct a multimodal fusion feature vector for final prediction. This multimodal data fusion approach overcomes the information limitations of single imaging assessments, enabling the system to comprehensively quantify the risk of local tumor invasion by combining the patient's systemic biochemical microenvironment.

[0116] Preferably, the clinical laboratory indicators are taken from the test results most recently obtained from the CT scan and within a preset time window. For missing clinical variables, the system uses any specific strategy among mean imputation, median imputation, or multiple imputation algorithms to impute the missing indicators before participating in the fusion calculation.

[0117] S630, the classification and prediction module 70 inputs the multimodal fusion feature vector into the deep feedforward neural network model and outputs the quantified probability of microvascular invasion. In this embodiment, the deep feedforward neural network model is a preferred implementation of the aforementioned preset classifier model.

[0118] In this embodiment, the multimodal fusion feature vector is formed by concatenating the image feature vector or image statistical feature vector obtained in step S610 with the clinical feature vector obtained in step S620.

[0119] After acquiring the multimodal fusion features, the system performs collaborative prediction using a deep feedforward neural network model. In practice, the network model employs a multilayer perceptron architecture, internally containing an input layer, multiple hidden layers, and an output layer. The number of nodes in the model's input layer is related to the multimodal fusion feature vector. The dimensions remain strictly consistent. Input data flows sequentially through three fully connected hidden layers.

[0120] As a preferred approach, the number of nodes in each hidden layer is set to 128, 64, and 32, respectively. To introduce non-linear feature mapping capabilities and alleviate the gradient vanishing problem in deep networks, a linear rectified function is connected as an activation layer after each hidden layer. Furthermore, to reduce the risk of overfitting the model to a limited number of medical samples, a regularization layer with a dropout rate of 0.3 is configured between adjacent fully connected layers.

[0121] After the data stream completes forward propagation through the hidden layer, it enters the output layer, which contains only a single node. The output layer uses the Sigmoid activation function to map the continuous real-valued features calculated by the network to a probability range of 0 to 1. The specific output prediction formula is as follows: ; In the formula, This indicates the predicted probability of microvascular invasion. This represents the Sigmoid activation function; This represents the weight vector of the output layer; This represents the feature vector output by the last hidden layer; This represents the bias term of the output layer. From a specific business perspective, it refers to the output probability value. This directly indicates the risk level of microvascular invasion in the current patient's liver cancer lesions; the closer the probability is to 1, the clearer the physical state of microvascular invasion in the target object. For the forward propagation of the fully connected layers and the matrix operations of the activation functions in the neural network model, those skilled in the art can implement them using conventional deep learning frameworks. The underlying computational logic is well-known in the field and will not be elaborated upon here.

[0122] S640, the classification prediction module 70 performs backpropagation based on clinical pathology labels to complete the parameter optimization and training of the neural network model.

[0123] Neural network models that have not undergone parameter optimization cannot be directly used for clinical risk prediction. To ensure that the aforementioned models have reliable discriminative capabilities, supervised learning-based model training must be performed before the system is deployed for inference.

[0124] The samples used for model training are defined as a set of multimodal fusion feature vectors from patients with hepatocellular carcinoma who have visited the hospital in the past; the corresponding labels are defined as the binary classification results of microvascular invasion confirmed by clinical postoperative pathological examination, where invasion is marked as 1 and no invasion is marked as 0. The training samples and label data are all derived from de-identified archived records of the hospital's imaging and communication systems and electronic medical records systems, thereby ensuring the legality and objectivity of the data sources.

[0125] In the training step, the system uses binary cross-entropy as the loss function to quantify the distribution difference between the model's output probability and the true pathological label. The specific loss function is defined as follows: ; In the formula, This represents the total number of samples in the current training batch; Indicates the first The true pathological label of each sample; The model represents the first The predicted probability of each sample. The system uses an adaptive moment estimator optimizer to perform gradient descent calculations, with an initial learning rate of 0.001.

[0126] During each iteration, the optimizer updates the weight matrix and bias parameters of each fully connected layer based on the gradients calculated by the backpropagation algorithm. To avoid overtraining, an early stopping mechanism is configured in the training process, i.e., training is terminated when the validation set loss no longer decreases for 10 consecutive iterations, in order to preserve the model weight parameters with the best generalization performance. Based on the disclosed sample sources, network structure, and optimization logic, those skilled in the art can implement the model training and deployment process accordingly.

[0127] Preferably, during the model inference phase, the standardized parameters of the clinical variables used for the test samples are consistent with those used during the training phase.

[0128] To further aid in understanding the technical solutions disclosed in this invention, a specific application embodiment is provided. In this embodiment, the microvascular invasion prediction system is deployed on a hepatobiliary surgical imaging diagnostic platform. The target patient undergoes a routine preoperative multi-phase enhanced CT scan of the abdomen. The system automatically acquires the arterial and portal venous phase CT sequences of the target patient from the image archiving and communication system, and simultaneously retrieves clinical biochemical test data such as alpha-fetoprotein, age, and BCLC count from the electronic medical record system.

[0129] After acquiring the input data, the system performs spatial segmentation of the liver parenchyma and tumor target area through the preprocessing and mask extraction modules. The dynamic vector field construction module calculates the gray-level gradient between adjacent enhancement phases to generate the original two-dimensional temporal perfusion vector field. The baseline analysis and vector decomposition module orthogonally decouples the hybrid dynamic signal based on the physiological metabolic baseline of the background normal liver, separating the vector components characterizing the abnormal blood supply of the tumor.

[0130] The feature reconstruction and weighting module combines the spatial anatomical orientation of tumor invasion to generate a three-dimensional invasion dynamics feature field. The pooling and collaborative prediction module concatenates the image features reduced by spatial pyramid pooling with the clinical features normalized by standard scores, and inputs them into a trained multilayer perceptron prediction model. The system outputs a probability of microvascular invasion in the target object of 0.87.

[0131] Based on the aforementioned quantitative indicators, the medical terminal outputs a comprehensive intervention strategy that expands the scope of anatomical liver resection and combines it with postoperative adjuvant interventional therapy. Postoperative pathological section results confirm the objective existence of microvascular invasion, validating the reference value of the system's predictive output.

[0132] To verify the technical effectiveness of the present invention, the system combined retrospective clinical data for experimental verification and mechanism analysis. During the experimental phase, desensitized data were collected from 320 patients with primary hepatocellular carcinoma who possessed complete preoperative imaging, clinical laboratory indicators, and clear postoperative pathological criteria, and were confirmed to have positive microvascular invasion. In this invention, the clinical gold standard for high-risk and low-risk groups was clearly defined as the patient's early postoperative recurrence risk. Specifically, the system uniformly set the follow-up time window and grouping threshold to 12 months after radical resection. Patients who developed new intrahepatic lesions or extrahepatic metastases within 12 months of postoperative follow-up via enhanced CT or MRI were defined as high-risk; those who showed no evidence of recurrence during the 12-month follow-up period were defined as low-risk. During the training data label generation phase, the system automatically parsed follow-up records from the electronic medical record system, assigning a label of 1 to samples with a recurrence-free survival time of 12 months or less as the high-risk group, and assigning a label of 0 to samples with a recurrence-free survival time greater than 12 months as the low-risk group. Based on the aforementioned objective clinical criteria, 115 cases were assessed as high-risk and 205 cases as low-risk in this experimental dataset. All experimental data were randomly divided into a model training set and an independent test set at a fixed ratio of 7:3.

[0133] See attached document Figure 9 , Figure 9 This is a scatter plot of the dimensionality-reduced distribution of a multidimensional feature manifold according to an embodiment of the present invention. Figure 9 This invention was used to verify the inter-class separability of the multimodal fusion features constructed in this invention when characterizing the high- and low-risk subtypes of hepatocellular carcinoma with microvascular invasion. The system utilizes a t-distributed random nearest neighbor embedding manifold dimensionality reduction algorithm to project the high-dimensional multimodal fusion feature vector onto a two-dimensional plane. In the figure, each scatter point represents an independent test set sample, circular points represent low-risk group samples, and intersection points represent high-risk group samples. The data distribution pattern reflects that features extracted using traditional 3D convolution exhibit large-area inter-class aliasing and overlap within the two-dimensional projection space. After applying the orthogonal decoupling of the physiological baseline and invasion direction weighting of this invention, high-risk and low-risk group samples form two clusters with clear classification boundaries in the dimensionality-reduced space. This distribution pattern, at the data mechanism level, confirms that the system decouples the original hybrid dynamic signal, filters out background noise from normal liver parenchyma blood flow that causes feature aliasing, and improves the data fidelity of local invasion features of lesions.

[0134] See attached document Figure 10 , Figure 10 This is a heatmap of multimodal feature contribution according to an embodiment of the present invention. Figure 10 This demonstrates the weighting mechanism of various input features on the final decision result when a deep feedforward neural network outputs the probability of a high-risk subtype. (Appendix) Figure 10The vertical axis represents different samples in the test set, and the horizontal axis represents specific feature nodes of the input, including spatial pyramid pooling extrema at different scales and multiple clinical variables. The brightness represents the contribution of local features calculated by the Shapley algorithm and the interpretation algorithm. The heatmap distribution shows that when predicting high-risk samples, the invasive dynamics extrema feature regions at the 2x2x2 grid scale and the 4x4x4 grid scale are highlighted, reflecting that the high-frequency invasive features of microvessels at the local tumor boundary dominate the model inference. Clinical feature nodes such as alpha-fetoprotein, age, and BCLC stage simultaneously show secondary contribution weights. The multimodal collaborative mechanism constructed by the joint activation state verification system avoids feature suppression or information loss, achieves complementarity between local imaging invasive phenotypes and overall clinical variables, and establishes the logical completeness and implementation reliability of the technical solution from the perspective of data flow within the model.

[0135] At the technical and microscopic mechanism levels, this invention, through systematic identification and output of high-risk imaging phenotypes, not only accurately corresponds to the high risk of early postoperative recurrence in macroscopic follow-up but also exhibits a high correlation with poor prognosis and enhanced biological invasiveness in previous studies. Combined with previous molecular mechanism studies using transcriptomic sequencing, such as population transcriptomic sequencing and single-cell transcriptomic mapping analysis, it is shown that the high-frequency invasive features and multimodal fusion features of local microvessels extracted by this system have a significant biological mapping relationship with specific highly invasive molecular subtypes, such as malignant subtypes representing high proliferation and high risk of microvascular invasion. Abnormal expression of these features is often accompanied by the remodeling of the tumor local immune microenvironment and abnormal activation of angiogenesis pathways. These molecular biological mechanisms not only provide a solid theoretical foundation for the scientific validity of the orthogonal decoupling of physiological baselines and the invasion direction weighting algorithm in this invention but also further confirm the important clinical significance of the high-risk or low-risk classification results output by this invention in guiding the formulation of personalized neoadjuvant and postoperative adjuvant therapy plans.

[0136] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying a high-risk low-risk subtype of MVI-positive hepatocellular carcinoma, characterized by, Includes the following steps: S10. Obtain the three-dimensional enhanced CT image sequence of the target object and perform registration to generate a registered image sequence; S20. Generate a tumor target area mask and a background liver parenchyma mask based on the registered image sequence; S30. Calculate the inflow time gradient and outflow time gradient based on the tumor target area mask and the background liver parenchyma mask, and construct a two-dimensional time-series perfusion vector; extract the spatial coordinates where the inflow time gradient is positive and the outflow time gradient is negative to construct a target blood supply subset, and use the target blood supply subset to filter the two-dimensional time-series perfusion vector; S40. Based on the two-dimensional temporal perfusion vector within the background liver parenchyma mask, determine the baseline vector, and perform projection calculation on the selected two-dimensional temporal perfusion vector within the tumor target area mask to the baseline vector to obtain orthogonal components; S50. Calculate the modulus of the orthogonal components to generate an intensity scalar field, and construct a boundary normal vector field based on the tumor target mask; Calculate the spatial gradient of the intensity scalar field, generate a directional weight field based on the spatial gradient and the boundary normal vector field, and use the directional weight field to weight the intensity scalar field to generate the target feature field; S60. Pool the target feature field to extract image feature vectors, concatenate the image feature vectors with the clinical variables of the target object to construct a joint feature vector, input the joint feature vector into a classifier model, and output the classification result of whether the target object belongs to the high-risk group or the low-risk group.

2. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, The step of generating a tumor target mask and a background liver parenchyma mask based on the registered image sequence includes: The tumor target mask is generated by extracting the set of three-dimensional spatial coordinates of the lesion from the registered image sequence. Extract the spatial coordinates of the entire liver parenchyma of the target object to generate an initial liver mask, perform morphological dilation on the tumor target area mask to generate a boundary buffer zone, and extract the main blood vessel region in the registered image sequence to generate a blood vessel region mask; The tumor target area mask, the boundary buffer zone, and the vascular region mask are removed from the initial liver mask using Boolean difference operations to generate the background liver parenchyma mask.

3. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, The three-dimensional enhanced CT image sequence includes plain scan images, arterial phase images, and portal venous phase images; the calculation of inflow and outflow time gradients based on the tumor target mask and the background liver parenchyma mask includes: The grayscale difference between the arterial phase image and the plain scan image at the same spatial coordinates is calculated and divided by the inflow phase time interval to obtain the inflow time gradient. The grayscale difference between the portal venous phase image and the arterial phase image at the same spatial coordinates is calculated and divided by the efferent time interval to obtain the efferent time gradient.

4. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, The step of extracting spatial coordinates where the inflow time gradient is positive and the outflow time gradient is negative to construct a target blood supply subset, and using the target blood supply subset to filter the two-dimensional time-series perfusion vector, includes: A two-dimensional time-series infusion vector field is constructed by using the inflow time gradient as the horizontal axis component and the outflow time gradient as the vertical axis component. Extract the voxel space coordinates where the inflow time gradient is greater than the inflow tolerance threshold and the outflow time gradient is less than the negative number of the outflow tolerance threshold, and generate a logical mask as the target blood supply subset; The two-dimensional time-series infusion vector field is multiplied point-by-point by the logical mask to output the filtered two-dimensional time-series infusion vector.

5. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, The determination of the baseline vector based on the two-dimensional temporal perfusion vector within the background liver parenchyma mask includes: The two-dimensional temporal perfusion vector within the background liver parenchyma mask is calculated by regional averaging and summing in the global anatomical space, and the calculated average perfusion vector is determined as the baseline vector; or... Multivariate kernel density estimation is performed on the two-dimensional temporal perfusion vector within the background liver parenchyma mask, and the vector corresponding to the main peak of the probability density distribution is determined as the baseline vector.

6. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, Projecting the selected two-dimensional temporal perfusion vector within the tumor target mask onto the baseline vector yields orthogonal components, including: The baseline vector is algebraically normalized to extract parallel unit vectors. Orthogonal unit vectors perpendicular to the parallel unit vectors are solved in a two-dimensional orthogonal plane to construct a feature basis. The orthogonal components are obtained by performing a vector inner product projection calculation on the two-dimensional temporal perfusion vector selected within the tumor target mask and the orthogonal unit vector.

7. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, The process of constructing a boundary normal vector field based on the tumor target region mask, calculating the spatial gradient of the intensity scalar field, generating a direction weight field based on the spatial gradient and the boundary normal vector field, and weighting the intensity scalar field using the direction weight field to generate a target feature field includes: A spatial distance field is established by performing a three-dimensional Euclidean distance transformation and filtering smoothing on the binarized tumor target mask, and the spatial first derivative of the spatial distance field is calculated to obtain the boundary normal vector field. The spatial gradient is obtained by calculating the partial derivative vectors of the intensity scalar field along the three spatial coordinate axes; The spatial gradient and the boundary normal vector field are subjected to a vector inner product operation, and the vector components with a positive value of the inner product operation result are retained by a linear rectified function ReLU or a threshold truncation function to complete the elimination of negative components, thereby obtaining the directional weight field. The directional weight field is adjusted using a gain control coefficient, and the adjusted directional weight field is multiplied and fused with the intensity scalar field to generate the target feature field.

8. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, The step of performing pooling processing on the target feature field to extract image feature vectors includes: Spatial pyramid pooling is performed by applying a multi-level mesh partitioning along the spatial dimension of the target feature field. Max pooling is performed within each defined local grid region to extract extreme value data; The extreme value data of each local grid region are concatenated and stitched together to generate one-dimensional data as the image feature vector.

9. The method for high-risk and low-risk classification of MVI-positive hepatocellular carcinoma according to claim 1, characterized in that, In step S60, the image feature vector is concatenated with the clinical variables of the target object to construct a joint feature vector. The joint feature vector is then input into a classifier model to output the classification result indicating whether the target object belongs to the high-risk group or the low-risk group, including: The alpha-fetoprotein concentration, age level, and BCLC count of the target subjects were obtained as clinical variables. The clinical variables are mapped to a fixed distribution interval using a standard score algorithm to generate a clinical feature vector. The image feature vector and the clinical feature vector are then concatenated and spliced ​​to construct the joint feature vector. The joint feature vector is input into a deep feedforward neural network model, and classification features are extracted through the hidden layer. The activation function of the output layer is used to map and output the high-risk or low-risk classification probability or classification result.

10. The application of the MVI-positive hepatocellular carcinoma high-risk / low-risk typing method as described in any one of claims 1 to 9 in the preparation of an auxiliary assessment device / system / software for MVI-positive hepatocellular carcinoma high-risk / low-risk typing.

Images