A method and system for analyzing and classifying based on CT images

Abstract

The application provides a CT image-based analysis and classification method and system, which comprises the following steps: S1, acquiring three-dimensional voxel data matrix of CT scanning image of a patient and synchronously loading lesion anatomical boundary data labeled by a radiologist, establishing a CT cubic topology network and acquiring CT composite data; S2, continuously calculating and acquiring a CT multi-modal topology feature tensor; S3, mapping the CT multi-modal topology feature tensor to a Riemann manifold space, establishing a generative adversarial network model, acquiring a 32-dimensional manifold perturbation vector, generating a perturbation CT data set; S4, establishing a training classification network model, training the training classification network model, acquiring a CT classification network model, stopping the execution of steps S3 to S4, finally executing steps S1 to S2, and generating a CT conclusion result based on the CT classification network model; and the method can better analyze and classify the CT image of the patient.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a CT image analysis and classification method and system. BACKGROUND

[0002] Under the background of the deep integration of medical image accurate diagnosis demand and artificial intelligence technology, CT image analysis and classification technology is facing problems such as multi-modal feature mining, adversarial robustness improvement and clinical interpretability enhancement. The existing benign and malignant differentiation methods mainly rely on manual analysis and classification or shallow semantic features based on convolutional neural networks, but have limitations in complex pathological scenarios.

[0003] On the one hand, manual analysis and classification of CT images have limitations. The consistency and repeatability of manual diagnosis and classification need to be improved. For example, individual differences exist when manually evaluating the topological features of CT images, which can lead to bias.

[0004] On the other hand, deep learning models face serious adversarial vulnerability problems in clinical applications. Clinical CT scans often face interference such as respiratory motion artifacts, and traditional adversarial training often does not conform to the anatomical deformation law, resulting in limited model generalization ability.

[0005] Therefore, the existing CT image analysis and classification methods often cannot meet the needs of existing clinical diagnosis. How to enhance the robustness and generalization of the model while preserving the anatomical rationality and improve the accuracy of CT image analysis and classification results has become a problem to be solved. SUMMARY

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a CT image analysis and classification method, which comprises:

[0007] S1: Obtain the three-dimensional voxel data matrix of the patient's CT scan image and simultaneously load the lesion anatomical boundary data labeled by radiologists, establish a CT cube topology network based on the three-dimensional voxel data matrix and the lesion anatomical boundary data, and obtain CT composite data based on the CT cube topology network;

[0008] S2: Perform persistent homology calculation based on the CT composite data, and obtain CT multi-modal topological feature tensors based on CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology in the CT composite data;

[0009] S3: Map the CT multi-modal topological feature tensors to the Riemannian manifold space, establish a generative adversarial network model, obtain a 32-dimensional manifold perturbation vector based on the generative adversarial network model, superimpose the 32-dimensional manifold perturbation vector and the CT multi-modal topological feature tensors, and generate a perturbed CT data set;

[0010] S4: Establishing a training classification network model, training the training classification network model based on the perturbed CT data set, obtaining a CT classification network model, and stopping the execution of steps S3 to S4;

[0011] S5: Executing steps S1 to S2, importing the CT multi-modal topological feature tensor into the CT classification network model, and generating a CT conclusion result based on the CT classification network model, wherein the CT conclusion result includes a classification result and an analysis probability result.

[0012] As a further scheme of the present application, a CT cube unit is created for each CT voxel contained in the three-dimensional voxel data matrix, and the vertex coordinates of the CT cube unit are mapped to the corresponding voxel midpoint;

[0013] A face-sharing adjacency relationship between the CT cube units is established, each CT cube unit is only adjacent to the front, rear, left, right, upper, and lower six directions, and the adjacent directions beyond the three-dimensional voxel data matrix are shielded, and an initial CT cube topological network is constructed based on the face-sharing adjacency relationship;

[0014] The lesion region in the initial CT cube topological network is mask-marked based on the lesion anatomical boundary data, and a lesion three-dimensional region is obtained;

[0015] The CT voxels in the lesion three-dimensional region that are mask-marked are subjected to three-dimensional corrosion, the three-dimensional corrosion is dynamically iterated based on the lesion anatomical boundary data, and after the three-dimensional corrosion is completed, the edge smoothing optimization is performed at the lesion-normal tissue interface;

[0016] The CT cube topological network is obtained based on the initial CT cube topological network enhanced based on the lesion three-dimensional region.

[0017] As a further scheme of the present application, the three-dimensional corrosion of the CT voxels in the lesion three-dimensional region that are mask-marked, the three-dimensional corrosion being dynamically iterated based on the lesion anatomical boundary data, comprises:

[0018] The lesion volume is obtained based on the lesion anatomical boundary data, the theoretical corrosion iteration number is obtained based on the lesion volume, the theoretical corrosion iteration number is obtained based on a volume-iteration formula, and the volume-iteration formula can be expressed as: ;

[0019] Wherein, represents the theoretical corrosion iteration number, represents the lesion volume, if the calculated based on the volume-iteration formula is negative, then =0 is forcibly set.

[0020] Based on the theoretical corrosion iteration number, three-dimensional corrosion is continuously carried out, and each round of three-dimensional corrosion can be represented as layer-by-layer traversal of the CT voxels marked by the mask in the three-dimensional region of the lesion. For each of the CT voxels marked by the mask, if there is at least one non-lesion CT voxel in the 3x3x3 neighborhood, the voxel is removed from the mask.

[0021] After each round of three-dimensional corrosion, the current lesion volume is obtained, the lesion volume ratio before and after is calculated based on the lesion volume and the current lesion volume, and the iteration threshold is set, which is 0.65. If the lesion volume ratio before and after is greater than or equal to the iteration threshold, the three-dimensional corrosion is stopped, and the last round of three-dimensional corrosion is rolled back.

[0022] After each round of three-dimensional corrosion, six-neighborhood connectivity detection is synchronously carried out based on the face-sharing adjacency relationship. If it is detected that the face-sharing adjacency relationship of the CT cubic unit in the three-dimensional region of the lesion is destroyed and split into at least one isolated region, the three-dimensional corrosion is stopped, and the last round of three-dimensional corrosion is rolled back.

[0023] As a further scheme of the present application, the edge smoothing optimization at the lesion-normal tissue interface comprises:

[0024] Based on the local CT voxels at the lesion-normal tissue interface, the HU value gradients in X / Y / Z directions are obtained, the gradient amplitudes are calculated based on the HU value gradients in X / Y / Z directions, the diffusion coefficient is obtained according to the gradient amplitudes, and the diffusion coefficient is inversely proportional to the gradient amplitude. Diffusion filtering based on the diffusion coefficient is applied along the normal direction to perform;

[0025] Based on the local surface sheet constructed at the lesion-normal tissue interface, a quadratic surface equation is fitted according to the least squares method, and the extreme curvature of the quadratic surface equation is calculated. A curvature threshold is set.

[0026] If the extreme curvature is greater than the curvature threshold, it is determined as a high curvature region, and the diffusion filtering radius is limited to be less than 2 voxels.

[0027] If the extreme curvature is less than or equal to the curvature threshold, it is determined as a low curvature region, and the diffusion filtering radius is limited to be greater than or equal to 2 voxels.

[0028] As a further scheme of the present application, the CT multi-modal topological feature tensor is mapped to the Riemannian manifold space, a generative adversarial network model is established, a 32-dimensional manifold perturbation vector is obtained based on the generative adversarial network model, and the 32-dimensional manifold perturbation vector is superimposed with the CT multi-modal topological feature tensor to generate a perturbed CT data set, comprising:

[0029] Based on the Wasserstein distance, the feature distance between different persistent graphs in the CT multi-modal topological feature tensor is calculated, and the feature distance is converted into a similarity score.

[0030] Performing multidimensional scaling based on the feature distance and the similarity score, mapping the CT multi-modal topological feature tensor to the Riemannian manifold space, and obtaining a manifold multi-modal topological feature tensor;

[0031] Injecting 16-dimensional Gaussian noise into the manifold multi-modal topological feature tensor based on the generative adversarial network model, and generating a 32-dimensional original perturbation vector;

[0032] Verifying the 32-dimensional original perturbation vector based on the Jacobian determinant, and obtaining a 32-dimensional manifold perturbation vector.

[0033] As a further scheme of the application, the method further comprises:

[0034] Establishing a feature encoding rule, which represents the correspondence between each dimension in the Riemannian manifold space and the dimensions in the CT multi-modal topological feature tensor, and when performing multidimensional scaling, matching the corresponding dimensional relationship between the manifold multi-modal topological feature tensor and the CT multi-modal topological feature tensor based on the feature encoding rule.

[0035] As a further scheme of the application, the persistent homology calculation based on the CT composite data obtains the CT multi-modal topological feature tensor based on CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology in the CT composite data, comprising:

[0036] The CT zero-dimensional persistent homology is used for analyzing connected components, the CT one-dimensional persistent homology is used for detecting ring structures, and the CT two-dimensional persistent homology is used for detecting cavity structures.

[0037] Mapping the birth threshold and the death threshold of each dimension to a two-dimensional plane to generate a persistence diagram of each dimension, and simultaneously generating a Betti number dynamic curve of each dimension according to the persistence diagram of each dimension, obtaining the CT multi-modal topological feature tensor based on the Betti number dynamic curve, the persistence diagram and the persistent homology of each dimension, wherein the CT multi-modal topological feature tensor is a 32-dimensional floating-point tensor.

[0038] As a further scheme of the application, the CT composite data is obtained based on the CT cubic topological network, comprising:

[0039] The CT composite data contains three-dimensional space coordinates of the CT cubic topological network, corrosion state markers, lesion volumes, lesion volumes after corrosion, HU values of CT voxels in the CT cubic topological network, and adjacency relationship indexes based on face-sharing adjacency relationship.

[0040] As a further scheme of the application, the method further comprises:

[0041] The lesion anatomical boundary data includes spatial coordinate data of a lesion core area, an edge infiltration area, an anatomical marker, and a lesion volume.

[0042] The size of the three-dimensional voxel data matrix is 512*512*N, N represents the number of axial slices, each CT voxel in the three-dimensional voxel matrix corresponds to a HU value, and the HU value ranges from -1000 to 2000.

[0043] In still another aspect, the embodiment of the present application also provides an analysis and classification system based on CT images, comprising:

[0044] The acquisition module is configured to acquire CT scan images of a patient and lesion anatomical boundary data labeled by a radiologist, and acquire a three-dimensional voxel data matrix based on the CT scan images of the patient.

[0045] The construction module is configured to construct a CT cubic topology network and acquire CT composite data.

[0046] The homology calculation module is configured to perform persistent homology calculation based on the CT composite data, and acquire a CT multi-modal topology feature tensor based on CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology.

[0047] The perturbation module is configured to map the CT multi-modal topology feature tensor to a Riemannian manifold space, acquire a 32-dimensional manifold perturbation vector based on a generative adversarial network model, and acquire a perturbed CT data set based on the 32-dimensional manifold perturbation vector and the CT multi-modal topology feature tensor.

[0048] The classification analysis module is configured to train a training classification network model, acquire a CT classification network model, and generate a CT conclusion result based on the CT classification network model.

[0049] Based on the above aspects, the embodiment of the present application can quantify the topology features of the cavity structure, vascular infiltration and other traditional morphological features that cannot be captured by traditional morphological methods, and can acquire HU value fluctuations through Betti number dynamic curves, verify and constrain the manifold perturbation in the Riemannian manifold space based on the Jacobian determinant, ensure that the manifold adversarial perturbation simulates real anatomical deformation such as lesion displacement caused by respiratory motion, avoid non-physiological artifacts generated by traditional perturbation, and improve the robustness and generalization of the subsequent training model, thereby better analyzing and classifying the CT images, improving the accuracy, and completing the subsequent model. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is an execution flow diagram of an analysis and classification method based on CT images provided by the embodiment of the present application.

[0051] Figure 2It is an execution flow diagram of establishing a CT cube topology network in a CT image-based analysis classification system provided by an embodiment of the application.

[0052] Figure 3 It is a schematic diagram of a CT image-based analysis classification system provided by an embodiment of the application. DETAILED DESCRIPTION

[0053] The application will be described in detail below with reference to the accompanying drawings, Figure 1 It is an execution flow diagram of a CT image-based analysis classification method provided by an embodiment of the application, Figure 2 It is an execution flow diagram of establishing a CT cube topology network in a CT image-based analysis classification system provided by an embodiment of the application, and the CT image-based analysis classification method will be described in detail below.

[0054] In step S1, a three-dimensional voxel data matrix of a CT scan image of a patient is acquired and the lesion anatomical boundary data labeled by a radiologist is synchronously loaded, a CT cube topology network is established based on the three-dimensional voxel data matrix and the lesion anatomical boundary data, and CT composite data is acquired based on the CT cube topology network.

[0055] Specifically, the size of the three-dimensional voxel data matrix of the CT scan image of the patient can be 512*512*N, N represents the number of axial slices, each CT voxel in the three-dimensional voxel matrix corresponds to a HU value, and the HU value can be in the range of (-1000, 2000).

[0056] Specifically, the CT composite data contains the three-dimensional space coordinates of the CT cube topology network, the corrosion state label, the lesion volume, the lesion volume after corrosion, the HU value of the CT voxel in the CT cube topology network, and the adjacency relationship index based on the face-sharing adjacency relationship.

[0057] Further, 512*512 is the standard matrix size of clinical CT scanning, which ensures that the single-layer image resolution is sufficient to capture the morphological details of the micro-nodules, wherein the single-layer image resolution is about 0.5-1mm, and at the same time, redundant calculation is avoided. If a high resolution such as 1024*1024 is used, the amount of calculation will increase exponentially, but the gain effect on analysis and classification is limited. The N-dimensional slice number is dynamically adjusted, which can adapt to different scanning protocols, and the representation of lesion continuity is ensured through three-dimensional reconstruction.

[0058] Further, the value range of HU value (-1000, 2000) covers the typical density range of human tissues, the calibration difference between scanning devices is eliminated through global unification, the cross-center consistency of topological feature extraction is ensured, and the HU value is kept between (-1000, 2000) to filter out extreme HU values such as metal implants, thereby avoiding the interference of some artifacts on the morphological three-dimensional corrosion operation.

[0059] In the embodiment, the lesion anatomical boundary data labeled by the radiologist includes spatial coordinate data of a lesion core area, an edge infiltration area, an anatomical marker, and a lesion volume.

[0060] Further, the spatial coordinates are represented as contour point coordinates of the lesion boundary, and the DICOM coordinate system of the CT original image is used as the reference to label the polygon vertex coordinates of the lesion area of each slice. For the lesion core area, the three-dimensional coordinates of the lesion main area are manually drawn by the physician, which usually corresponds to the solid component area with abnormally high HU value in the CT image. The spatial prior constraint can be provided for establishing the CT cubic topological network, and the interference of healthy tissues is reduced. For the edge infiltration area, the sub-area coordinates of the infiltration features such as the ground glass shadow around the lesion and the blood vessel cluster sign are labeled. For the anatomical marker, the coordinates of the anatomical structure closely related to the spatial relationship of the lesion, such as the adjacent bronchial bifurcation point, pulmonary artery branch, and pleural attachment point, are labeled. The spatial reference framework can be provided to assist in establishing the topological connection relationship between the lesion and the anatomical structure, and the generated manifold perturbation can be ensured to be complex and real anatomical constraint in subsequent adversarial training, such as the deformation range not exceeding the pleural boundary. When the lesion volume is obtained, the overall volume is automatically calculated based on the coordinates of the core area and the infiltration area.

[0061] It can be understood that based on the lesion anatomical boundary data labeled by the physician, the annotation data can be labeled through space-semantic hierarchical labeling, and the qualitative diagnostic experience of the physician can be converted into quantifiable topological features and geometric constraints for subsequent algorithms.

[0062] The CT cubic topological network is established based on the three-dimensional voxel data matrix and the lesion anatomical boundary data.

[0063] In step S11, a CT cubic unit is created for each CT voxel included in the three-dimensional voxel data matrix, and the vertex coordinates of the CT cubic unit are mapped to the midpoint of the corresponding voxel.

[0064] In the embodiment, the mapping of the vertex coordinates of the CT cubic unit to the center of the corresponding voxel can inherit the physical coordinate system of the CT scan, avoid the distortion of the topological network caused by coordinate conversion, and simplify the coordinate conversion calculation by mapping the center point. Only the linear mapping relationship between the CT voxel and the physical coordinates needs to be calculated.

[0065] Step S12, establish the face-sharing adjacency relationship between the CT cube units, each CT cube unit is only adjacent to the front, back, left, right, up and down six directions, and the adjacency directions beyond the three-dimensional voxel data matrix are shielded, and the initial CT cube topology network is constructed based on the face-sharing adjacency relationship.

[0066] In the embodiment, the effective adjacency relationship of each CT voxel in the three-dimensional space is defined, only the CT voxels in the directly adjacent front, back, left, right, up and down six directions are included, and the diagonal adjacency relationship is excluded when connecting, so as to avoid the topological noise introduced by the diagonal connection, and the CT voxels located at the edge of the scanning volume are automatically shielded from the adjacency directions beyond the matrix range, so as to ensure the physical authenticity of the topological structure.

[0067] Step S13, mask mark the lesion area in the initial CT cube topology network based on the lesion anatomical boundary data, and obtain the lesion three-dimensional area.

[0068] Step S14, three-dimensional corrosion is performed on the CT voxels in the lesion three-dimensional area which are mask marked, the three-dimensional corrosion is dynamically iterated based on the lesion anatomical boundary data, and edge smoothing optimization is performed on the lesion-normal tissue interface after the three-dimensional corrosion is completed.

[0069] The three-dimensional corrosion is performed on the CT voxels in the lesion three-dimensional area which are mask marked, and the three-dimensional corrosion is dynamically iterated based on the lesion anatomical boundary data, and the edge smoothing optimization is performed on the lesion-normal tissue interface after the three-dimensional corrosion is completed.

[0070] Step S14-11, obtain the lesion volume based on the lesion anatomical boundary data, obtain the theoretical corrosion iteration number based on the lesion volume, the theoretical corrosion iteration number is obtained based on the volume-iteration formula, and the volume-iteration formula can be expressed as: ;

[0071] Wherein, is the theoretical corrosion iteration number, is the lesion volume, if the theoretical corrosion iteration number calculated based on the volume-iteration formula is negative, , then the theoretical corrosion iteration number is forcibly set to 0.

[0072] It can be understood that the logarithmic function with base 2 is used to establish the quantitative relationship between the volume and the iteration number, the mathematical characteristics conform to the exponential decay law of the lesion volume change sensitivity to corrosion, and the use of the logarithmic function with base 2 makes the iteration number double every time, which matches the common lesion growth mode, and when the lesion volume is less than 1, , the function outputs a negative value, at this time, the theoretical corrosion iteration number is forcibly set to 0 to avoid damage to the small lesion. ​​​When exponentially increasing, the number of iterations increases linearly, ensuring that the erosion intensity is sub-linearly related to the lesion size, preventing under-processing of large-volume lesions.

[0073] It can be understood that, by setting the threshold value of , the reason is that the artifact volume generated by the voxel-level noise in the binary mask is generally smaller than , while the volume of the core area of the malignant lesion is usually , , thus ensuring that the threshold value will not erode the real pathological tissue.

[0074] In this example, the threshold value of is dynamically corrected when calculating the theoretical number of erosion iterations, which can be expressed as: if the input CT image resolution deviates from the standard value, such as voxels, the threshold value is adjusted by a proportional formula, which can be expressed as:

[0075] ;

[0076] wherein is a constant.

[0077] Step S14-12, continue three-dimensional erosion based on the theoretical number of erosion iterations, and each round of three-dimensional erosion can be expressed as traversing the CT voxels marked by the mask in the three-dimensional area of the lesion layer by layer. For each CT voxel marked by the mask, if there is at least one non-lesion CT voxel in the 3x3x3 neighborhood, the voxel is removed from the mask.

[0078] In this embodiment, an isotropic three-dimensional cubic structure kernel (3x3x3 voxels) is used, and all elements in the kernel are assigned a value of 1, ensuring that three-dimensional erosion can be uniformly performed in three-dimensional space, eliminating isolated or semi-connected voxel noise in the lesion mask. The execution logic of three-dimensional erosion is to traverse the lesion mask voxels layer by layer. For each marked voxel, if there is at least one non-lesion voxel in its 3x3x3 neighborhood, the voxel is removed from the mask. After the three-dimensional erosion iteration is completed, only the voxel set that is strongly connected to the main body of the lesion is retained, eliminating fine branches and surface burrs.

[0079] Step S14-13-1, obtain the current lesion volume after each round of three-dimensional erosion, calculate the lesion volume ratio before and after based on the lesion volume and the current lesion volume, and set the iteration threshold value to 0.65. If the lesion volume ratio before and after is greater than or equal to the iteration threshold value, stop the three-dimensional erosion and return to the last round of three-dimensional erosion.

[0080] In this embodiment, the lesion volume ratio before and after (current volume retention rate) is calculated in real time after each round of erosion iteration. The calculation formula is:

[0081] ;

[0082] Based on the lesion volume and the current lesion volume (i.e. the eroded lesion volume), the following is calculated: If 0.65, the iteration is terminated in advance to avoid over-erosion caused by reaching the theoretical number of iterations, the three-dimensional erosion is stopped, and the state of the previous three-dimensional erosion is automatically returned to, and the number of iterations of the previous round is used as the final number of iterations, to ensure that the result is always within the effective constraint range.

[0083] Step S14-13-2: After each three-dimensional erosion, six-neighbor connectivity detection is performed based on the face-sharing adjacency relationship. If it is detected that the face-sharing adjacency relationship of the CT cubic unit in the three-dimensional region of the lesion is destroyed and split into at least one isolated region, the three-dimensional erosion is stopped, and the state of the previous three-dimensional erosion is returned to.

[0084] In this embodiment, after each iteration of erosion, it is detected whether the lesion region remains a single connected component based on the face-sharing adjacency relationship, i.e. all voxels are connected through face-sharing adjacency paths. If the lesion region is divided into multiple isolated sub-regions, such as due to excessive erosion leading to the breakage of burrs, it is determined that the topological structure is damaged. Once the damage is detected, the state of the previous iteration is automatically returned to, and the number of iterations of the previous round is used as the final number of iterations, to ensure that the result of the three-dimensional erosion meets the requirement of connectivity.

[0085] Further, the connected component is represented as a set of voxels connected through face-sharing relationship adjacency relationship, which constitutes an independent connected region, i.e. there is a continuous chain of adjacent voxels between any two voxels. When detecting whether the lesion region remains a single connected component, a marker voxel can be randomly selected from the lesion region after erosion as a search starting point, and based on BFS (breadth-first search) or DFS (depth-first search), all voxels connected in the six-neighborhood are recursively accessed and marked as visited. After the traversal is completed, if there is a lesion voxel that has not been marked, it is determined that there are multiple connected components, i.e. the structure is classified. Then, the volume of each connected component is calculated. If the volume of the largest component accounts for 95% or more, it is considered to be a slight split, which may be caused by the shedding of surface burrs. If the volume of the largest component accounts for less than 95%, it is considered to be a serious split. Based on the slight split and the serious split, it can be determined whether it is caused by noise residue or real anatomical structure tearing.

[0086] As can be understood, a malignant lesion usually appears as a single, boundary-infiltrated continuous mass, and multiple connected components may cause the erroneous segmentation of satellite lesions or vascular infiltration regions, leading to distortion of subsequent feature extraction. In subsequent persistent homology calculation, the required input structure is a closed manifold, and multiple connected components will generate additional Betti numbers, interfering with the calculation of topological features.

[0087] It can be understood that, by topological structure constraint, loss of key case features can be prevented, for example, the burr and lobulation of malignant nodules are often manifested as weak connection structure, which can be avoided from being misjudged as noise and removed by topological structure constraint, and secondly, the reliability of subsequent persistent homology calculation can be guaranteed. Persistent homology relies on closed connected domain calculation Betti number, and if the structure is split, it will lead to false mutation of ring features or cavity features.

[0088] In this embodiment, adaptive enhancement mechanism can also be added in the process of three-dimensional erosion. For specific pathological categories (such as burr type malignant tumor), erosion compensation factor can be started , The iteration number is corrected by a correction formula, and the correction formula is:

[0089] ;

[0090] The iteration number is increased by the erosion compensation factor , which strengthens the ability to clear the burr-like artifacts outside the lesion. The value of the compensation factor is obtained by learning from the training set. Based on 200-500 labeled data, the erosion sufficiency evaluated by the pathologist is used as the optimization target, and the optimal compensation factor is solved by gradient descent method .

[0091] In this example, through the double constraints of the iteration formula dependent on the lesion volume and the current lesion volume, the problems of under-erosion or over-erosion caused by traditional fixed iteration number are avoided.

[0092] Wherein, the edge smoothing optimization at the lesion-normal tissue interface comprises:

[0093] Step S14-21, local CT voxels are obtained based on the lesion-normal tissue interface, X / Y / Z direction HU value gradients are obtained based on the local CT voxels, gradient amplitudes are obtained based on the X / Y / Z direction HU value gradients, diffusion coefficients are obtained according to the gradient amplitudes, the diffusion coefficients are inversely proportional to the gradient amplitudes, and diffusion filtering based on the diffusion coefficients is applied along the normal direction.

[0094] In this embodiment, the lesion-normal tissue interface refers to the three-dimensional boundary region between the lesion region (such as tumor, inflammation and other abnormal tissues) and the surrounding normal anatomical structure (such as healthy blood vessels, bronchus) in the CT image.

[0095] In this embodiment, the local CT voxel is represented as a 3x3x3 cubic neighborhood where the current target voxel is located, containing 27 voxels including itself, and the field range is consistent with the size of the combined kernel in the three-dimensional erosion operation, ensuring the uniformity of the algorithm parameters. When calculating the HU value gradient, the HU value gradient of the target voxel in the X / Y / Z direction is calculated based on the central difference method, and then the gradient amplitude is calculated based on the HU value gradient in the X / Y / Z direction:

[0096] ;

[0097] wherein, is represented as the gradient amplitude, is represented as the HU value gradient in the X direction, is represented as the HU value gradient in the Y direction, is represented as the HU value gradient in the Z direction, and the gradient amplitude is used to drive the diffusion coefficient in each direction.

[0098] Further, based on the gradient amplitude, the diffusion coefficient is defined. When the gradient amplitude is high, such as the edge of calcification or the interface of malignant infiltration, the diffusion coefficient tends to 0, which inhibits the smoothing operation and preserves the steep density jump feature. When the gradient amplitude is low, such as the inflammatory exudation area, the diffusion coefficient tends to 1, which performs strong smoothing and eliminates the step-like status voxel artifact.

[0099] Further, the filtering operation is performed along the normal direction of the interface to avoid structure blurring caused by cross-interface diffusion. The normal direction is determined by calculating the local HU gradient direction, ensuring that the smoothing only acts on the homogeneous region inside or outside the lesion.

[0100] Step S14-22, based on the local surface sheet constructed at the lesion-normal tissue interface, a quadratic surface equation is fitted according to the least squares method, and the extreme curvature of the quadratic surface equation is calculated, and a curvature threshold is set.

[0101] Step S14-23-1, if the extreme curvature is greater than the curvature threshold, it is determined as a high curvature region, and the diffusion filter radius is limited to be less than 2 voxels.

[0102] Specifically, the high curvature region corresponds to the microstructure of the burr sign vertex and the lobulation sign depression, which needs to be protected, so a diffusion filter smaller than 2 voxels is used.

[0103] Step S14-24-2, if the extreme curvature is less than or equal to the curvature threshold, it is determined as a low curvature region, and the diffusion filter radius is limited to be greater than or equal to 2 voxels.

[0104] Specifically, the low curvature region represents the main body of the lesion or the inflammatory exudation area, which can accept a larger range of smoothing, so a diffusion filter greater than or equal to 2 voxels is used.

[0105] It can be understood that, based on edge smoothing optimization, key pathological features such as evil invasion and burr sign can be preserved while noise is eliminated, and the rough lesion mask after three-dimensional corrosion is converted into a smooth and continuous three-dimensional structure, providing input conforming to differential manifold for subsequent extraction of topological features.

[0106] Step S15, obtaining a CT cubic topological network based on the initial CT cubic topological network of the enhanced three-dimensional region of the lesion.

[0107] Step S2: performing persistent homology calculation based on the CT composite data, and obtaining a CT multi-modal topological feature tensor based on CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology in the CT composite data.

[0108] Specifically, the birth threshold and the death threshold of each dimension are obtained based on the CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology, the birth threshold and the death threshold of each dimension are mapped to a two-dimensional plane to generate a persistent diagram of each dimension, and a Betti number dynamic curve of each dimension is generated according to the persistent diagram of each dimension, and a CT multi-modal topological feature tensor is obtained based on the Betti number dynamic curve of each dimension, the persistent diagram and the persistent homology, the CT multi-modal topological feature tensor being a 32-dimensional floating-point tensor.

[0109] In this example, CT zero-dimensional persistent homology is used to analyze connected components, and the CT zero-dimensional persistent homology dynamically adjusts the HU threshold from -1000HU to 2000HU with a step size of 10HU. At each HU threshold, the voxels with HU values higher than the HU threshold are marked as active, and the number of connected components in the active region is counted (value) to record the threshold point of value mutation. Mutation is defined as a change in value between adjacent HU thresholds greater than or equal to 50%, and the HU threshold point is taken as a candidate point for lobulation. If the value of a certain lasts for less than or equal to 3 HU threshold steps, it is determined to be noise disturbance and is not recorded. If the value lasts for more than 3 HU threshold steps, it is identified as a mutation point and is recorded. Further, when the HU threshold is reduced to a certain critical value and a new mutation point appears, the birth threshold is recorded. When the HU threshold continues to decrease and the originally independent connected components merge, the mutation point decreases, and the death threshold is recorded.

[0110]

[0111] ​​​In this embodiment, CT one-dimensional persistent homology is used to detect ring structures. CT one-dimensional persistent homology is a three-dimensional simplicial complex connected by face-sharing adjacency relationship. In the three-dimensional simplicial complex, a ring structure is formed by a closed path of at least four cubic cells formed by sharing faces, such as a ring-shaped gap formed by a tumor wrapped by a blood vessel, to generate a closed loop (1-dimensional hole). The birth threshold of each loop, i.e., the HU value when the loop is first formed, is calculated. For example, when the HU threshold is lowered to a certain critical value, the low-density voxels of the necrotic area are activated to form a closed loop. The death threshold of each loop, i.e., the HU value when the loop is filled, is calculated. For example, as the HU threshold increases, the voxels of the surrounding solid components are activated, causing the ring-shaped gap to close. Stable ring structures with a life cycle (death threshold-birth threshold) of ≥80HU are screened, and short-term periodic noise is filtered.

[0112] In this embodiment, CT two-dimensional persistent homology is used to detect cavity structures. A cavity structure is represented as a closed three-dimensional region surrounded by tetrahedrons, and the inside is not occupied by any activated voxel, i.e., the HU value is lower than the current threshold. For example, the necrotic area inside a malignant tumor forms a cavity due to cell liquefaction. The HU value when the cavity is first formed is obtained. When the HU threshold is lowered to a certain critical value, low-density voxels are inhibited, and activated voxels surround a closed cavity, which is recorded as a birth threshold. The HU value when the cavity is filled with high-density voxels and dies is obtained. As the HU threshold increases, the voxels of the surrounding solid components are activated, causing the cavity to close. The death threshold is recorded. The cumulative survival time of the cavity on the HU scale during the period from birth to death is calculated. Cavity structures with a life cycle (death threshold-birth threshold) of <20HU are screened. The life cycle is integrated to obtain a life cycle integral. The life cycle integral reflects the density stability of the cavity. Malignant cavities usually have higher integral values.

[0113] Further, the birth-death threshold pair (birth threshold, death threshold) of each dimension is mapped to a two-dimensional plane. The persistent diagram of each dimension is established based on the birth-death threshold pair. The Betti number dynamic curve of each dimension is generated based on the persistent diagram. The sampling interval is 10HU. At the same time, the persistent entropy is calculated to quantify the complexity of the topological structure.

[0114] Specifically, the CT multi-modal topological feature tensor is a 32-dimensional floating-point tensor, which is specifically allocated as follows:

[0115] Dimensions 0-9: sampling values of key HU thresholds (such as -800, -600,..., 1000HU) of zero-dimensional Betti number dynamic curve, reflecting the dynamic process of lesion connectivity with density change;

[0116] Dimension 10-19: The top 10 life cycles of one-dimensional Betti number dynamic curves and coordinates, where the coordinates represent the dimension (birth threshold, death threshold), reflecting the annular structure characteristics of blood vessel infiltration and malignant cavities;

[0117] Dimension 20-29: Two-dimensional Betti number dynamic curve life cycle integral, maximum continuous life cycle and spatial distribution entropy, reflecting the topological stability of necrotic areas;

[0118] Dimensions 30-31: Persistence entropy and total life cycle of zero-dimensional, one-dimensional and two-dimensional addition, which comprehensively quantifies the overall topological complexity of the lesion.

[0119] Step S3, map the CT multi-modal topological feature tensor to the Riemannian manifold space, establish a generative adversarial network model, obtain a 32-dimensional manifold perturbation vector based on the generative adversarial network model, superimpose the 32-dimensional manifold perturbation vector and the CT multi-modal topological feature tensor, and generate a perturbed CT data set.

[0120] It can be understood that processing in the Riemannian manifold space can more naturally capture the geometric and topological relationships in medical images, while for traditional methods such as Euclidean space, data is treated as a flat table, which is difficult to accurately describe the dynamic changes of lesion morphology, such as tumor infiltration growth or respiratory motion deformation, while the manifold space simulates the real anatomical deformation law through the curved geometric structure, ensuring that the generated adversarial perturbation, such as simulating lesion displacement, is physiologically reasonable, such as blood vessels not suddenly breaking and cavities not appearing out of thin air, avoiding non-physiological artifacts generated by traditional perturbation. By adding perturbation in the Riemannian manifold space, and using the sample with added perturbation as a training sample to train the subsequent CT classification network model, the robustness of the CT classification network model to adversarial interference can be improved. When the CT classification network model is analyzed, such as in the case of a small shift in the scanning angle, the CT image can still be accurately analyzed and classified.

[0121] Among them, step S3 includes:

[0122] Step S31, based on the Wasserstein distance calculation, obtain the feature distance between different persistence diagrams in the CT multi-modal topological feature tensor, and convert the feature distance to a similarity score.

[0123] Step S32, based on the feature distance and the similarity score, perform multidimensional scaling transformation, map the CT multi-modal topological feature tensor to the Riemannian manifold space, and obtain the manifold multi-modal topological feature tensor.

[0124] In this embodiment, the feature points of the zero-dimensional, one-dimensional and two-dimensional persistent diagrams are matched in pairs, the minimum cost of "transporting" all feature points, i.e., the feature distance, is calculated, the feature distance is converted into a similarity score based on Gaussian kernel conversion, the similarity score has a value range of (0-1), the closer the distance, the higher the score, an N*N distance matrix is constructed based on the feature distance and the similarity score, the 32-dimensional coordinates are obtained through eigenvalue decomposition of the N*N distance matrix, the Euclidean distance in the manifold space is approximated to the feature distance, thereby obtaining the 32-dimensional manifold coordinates, and the manifold multi-modal topological feature tensor is obtained based on the 32-dimensional manifold coordinates.

[0125] In step S33, 16-dimensional Gaussian noise is injected into the manifold multi-modal topological feature tensor based on the generative adversarial network model, and a 32-dimensional original perturbation vector is generated.

[0126] In step S34, the 32-dimensional original perturbation vector is verified based on the Jacobian determinant, and a 32-dimensional manifold perturbation vector is obtained.

[0127] In this embodiment, a lung nodule with a diameter of 12 mm is taken as an example, at this time, the CT multi-modal topological feature tensor has been generated in step S2, the CT multi-modal topological feature tensor includes persistent diagram features, Betti number dynamic curves and cavity distribution entropy, wherein the persistent diagram features are represented by a two-dimensional birth threshold of 200 and a two-dimensional death threshold of 350, indicating that a hollow structure appears when HU=200, the cavity distribution entropy is 0.68, 16-dimensional Gaussian noise is injected based on the generative adversarial network model to provide relevant randomness, a 16-dimensional perturbation vector is generated based on the generator in the generative adversarial network model, the 16-dimensional perturbation vector is injected into the CT multi-modal topological feature tensor to generate perturbation data, a perturbation of 0.08 is injected into the Betti number dynamic curve of dimensions 16 to 23, and a perturbation of 0.05 is injected into the cavity distribution entropy of dimensions 24 to 31, the perturbation amount is constrained based on the Riemann norm to keep the perturbation amount below 0.15, a 32-dimensional original perturbation vector is generated, after injecting the perturbation, differential homeomorphism verification is performed, the perturbation amount is projected from the tangent space to the manifold surface to ensure smooth deformation, the perturbation amount is verified based on the Jacobian determinant, if the Jacobian determinant verification does not affect the topological structure, the 32-dimensional original perturbation vector is regarded as a 32-dimensional manifold perturbation vector, and if the Jacobian determinant verification affects the topological structure, the perturbation is re-injected based on the generative adversarial network model.

[0128] Further, a feature coding rule is established, which represents the correspondence between each dimension in the Riemann manifold space and the dimension in the CT multi-modal topological feature tensor. When performing multi-dimensional scaling transformation, the feature coding rule is used to match the corresponding dimension relationship between the manifold multi-modal topological feature tensor and the CT multi-modal topological feature tensor.

[0129] Step S4, a training classification network model is established, the training classification network model is trained based on the perturbed CT data set, a CT classification network model is obtained, and the execution of steps S3 to S4 is stopped.

[0130] Step S5, the CT multi-modal topological feature tensor is imported into the CT classification network model, and a CT conclusion result is generated based on the CT classification network model, the CT conclusion result including a classification result and an analysis probability result.

[0131] In this embodiment, steps S4 and S5 are explained by way of example. It is assumed that a 8mm nodule is detected in a lung CT image of a patient, steps S1 to S4 have completed processing to generate a CT classification network model, steps S1 and S2 are executed to generate a CT multi-modal topological feature tensor of the CT image, the CT multi-modal topological feature tensor is imported into the CT classification network model, and a multi-modal manifold classifier (MMC-Net) in the CT classification network model operates. The multi-modal manifold classifier (MMC-Net) includes two channels for analysis, i.e., a geometric feature analysis channel and a topological feature analysis channel. The multi-modal manifold classifier analyzes the CT multi-modal topological feature tensor based on geometric features and topological features. The geometric features are input into a fully connected layer after being weighted to generate high-order geometric features reflecting malignant morphological features such as “spiculation (weight 0.45)” and “lobulation (weight 0.32)”. The topological features are fused with Betti number dynamic curve changes by a GRU to output enhanced topological feature vectors. Then, a joint feature matrix is generated based on the high-order geometric features and the enhanced topological feature vectors. The joint feature matrix is identified for a perturbation, e.g., a perturbation amount of 0.1 in the case, which belongs to a moderate perturbation, triggers the classifier, calculates the distance from the sample to the malignant / benign class center, and outputs an analysis result of a malignant probability of 92.6%. A CT conclusion result is generated, which includes a classification result of malignancy and an analysis probability result of 92.6%. The report generated above can only be used to assist a physician in diagnosis and cannot be directly used as a diagnosis result.

[0132] Figure 3 A schematic diagram of a CT image-based analysis classification system is shown, which can implement the idea of the present application.

[0133] Specifically, a CT image-based analysis classification system includes:

[0134] The acquisition module 100 is configured to acquire a CT scan image of a patient and lesion anatomical boundary data 101 labeled by a radiologist, and acquire a three-dimensional voxel data matrix based on the CT scan image of the patient.

[0135] The constructing module 102 is configured to construct a CT cubic topology network and acquire CT composite data;

[0136] The homology calculating module is configured to perform persistent homology calculation based on the CT composite data and acquire CT multi-modal topology feature tensors based on CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homologies.

[0137] The perturbation module 103 is configured to map the CT multi-modal topology feature tensors to a Riemannian manifold space, acquire a 32-dimensional manifold perturbation vector based on a generative adversarial network model, and acquire a perturbed CT data set based on the 32-dimensional manifold perturbation vector and the CT multi-modal topology feature tensors.

[0138] The classification analysis module 104 is configured to train a CT classification network model, acquire the CT classification network model, and generate a CT conclusion result based on the CT classification network model.

[0139] The specific use mode and role of the embodiment are described as follows:

[0140] Firstly, the three-dimensional voxel data matrix of the CT scan image of a patient is acquired through step S1 and the lesion anatomical boundary data labeled by a radiologist is simultaneously loaded, the CT cubic topology network is established based on the three-dimensional voxel data matrix and the lesion anatomical boundary data, and the CT composite data is acquired, then the persistent homology calculation is performed according to step S2, the CT multi-modal topology feature tensors are acquired based on the CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homologies in the CT composite data, so as to quantify the related topology features, then the CT multi-modal topology feature tensors are mapped to the Riemannian manifold space according to step S3, the 32-dimensional manifold perturbation vector is acquired based on the generative adversarial network model, the perturbation is verified and constrained based on the Jacobian determinant, so as to ensure that the adversarial perturbation simulates the real anatomical deformation, the perturbation and the topology feature tensors are superimposed to generate the perturbed CT data set, then the classification network model is established according to step S4, and the perturbed CT data set is taken as a training sample to train the classification network model, so as to acquire the CT classification network model, at this time, the training is completed, and the steps S3 to S4 can be stopped, finally, the steps S1 to S2 are executed to acquire the related data of the CT image which needs to be analyzed and classified, and the above data is processed based on the CT classification network model, so as to generate the CT conclusion result, and the classification result and the probability result can be viewed based on the conclusion result.

[0141] In addition, the embodiment of the present application further provides an electronic device, which comprises:

[0142] at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to the first embodiment of the present application.

[0143] The various components of the electronic device are described in detail as follows:

[0144] The processor is the control center of the electronic device, and can be one processor or a plurality of processing elements. For example, the processor is one or more central processing units (CPU), or is an application specific integrated circuit (ASIC), or is one or more integrated circuits configured to implement the first embodiment of the present application, such as one or more digital signal processors (DSP), or one or more field programmable gate arrays (FPGA).

[0145] The processor can perform various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory.

[0146] The memory is used to store software programs for implementing the scheme of the present application, and is controlled by the processor to perform the implementation. The specific implementation manner can refer to the above method embodiments, and will not be described here again.

[0147] The memory can be a real-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disk storage, a magneto-optical disk storage (including a compact flash, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, and the like), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and be coupled to the processor through an interface circuit of the electronic device, and the embodiments of the present application do not make a specific limitation thereto.

[0148] The above-described embodiments can be implemented, in whole or in part, by software, hardware (such as a circuit), firmware or any combination thereof. When implemented by software, the above-described embodiments can be implemented, in whole or in part, in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the flow or function described in the embodiments of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center by limited (for example, infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing a set of one or more available media. The available media can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0149] It should be understood that the term "and / or" in this document is merely used to describe associated objects, and can represent three relationships, for example, A and / or B can represent three cases of A alone, A and B, and B alone, where A and B can be singular or plural. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects before and after it, but it can also represent an "and / or" relationship. The specific meaning can be understood according to the context before and after it.

[0150] It should be understood that the size of the sequence number of each process described above in the embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0151] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for analyzing and classifying CT images, characterized in that, The method includes: S1: Obtain the three-dimensional voxel data matrix of the patient's CT scan image and simultaneously load the lesion anatomical boundary data annotated by the radiologist. Establish a CT cube topology network based on the three-dimensional voxel data matrix and the lesion anatomical boundary data. Obtain CT composite data based on the CT cube topology network. S2: Perform continuous cohomology calculation based on CT composite data, and obtain CT multimodal topological feature tensors based on the continuous cohomology of CT zero-dimensional, CT one-dimensional and CT two-dimensional data. Specifically, the zero-dimensional continuous cohomology of CT is used to analyze connected components, the one-dimensional continuous cohomology of CT is used to detect ring structures, and the two-dimensional continuous cohomology of CT is used to detect cavity structures. Birth and death thresholds for each dimension are obtained based on the zero-dimensional, one-dimensional, and two-dimensional continuous cohomology of CT. These thresholds are then mapped to a two-dimensional plane to generate persistent graphs for each dimension. Simultaneously, dynamic curves of the Betti number for each dimension are generated based on the persistent graphs. Based on the dynamic curves of the Betti number, the persistent graphs, and the continuous cohomology, a CT multimodal topological feature tensor is obtained. This CT multimodal topological feature tensor is a 32-dimensional floating-point tensor. S3: Map the CT multimodal topological feature tensor to the Riemannian manifold space, establish a generative adversarial network model, obtain a 32-dimensional manifold perturbation vector based on the generative adversarial network model, and superimpose the 32-dimensional manifold perturbation vector with the CT multimodal topological feature tensor to generate a perturbed CT dataset. Specifically, the feature distances between different persistent graphs in the CT multimodal topological feature tensor are obtained based on Wasserstein distance calculation, and the feature distances are converted into similarity scores. Multidimensional scaling transformation is performed based on the feature distances and similarity scores to map the CT multimodal topological feature tensor to the Riemannian manifold space, obtaining the manifold multimodal topological feature tensor. 16-dimensional Gaussian noise is injected into the manifold multimodal topological feature tensor based on a generative adversarial network model, generating a 32-dimensional original perturbation vector. The 32-dimensional original perturbation vector is validated based on the Jacobian determinant, obtaining the 32-dimensional manifold perturbation vector. S4: Establish a training classification network model. Train the training classification network model based on the perturbed CT dataset to obtain the CT classification network model. Stop executing steps S3 to S4. S5: Execute steps S1 to S2 to import the CT multimodal topological feature tensor into the CT classification network model, and generate CT conclusion results based on the CT classification network model. The CT conclusion results include classification results and analysis probability results.

2. The CT image-based analysis and classification method according to claim 1, characterized in that, The establishment of a CT cube topology network based on a three-dimensional voxel data matrix and lesion anatomical boundary data includes: For each CT voxel contained in the three-dimensional voxel data matrix, a CT cube cell is created and the vertex coordinates of the CT cube cell are mapped to the midpoint of its corresponding voxel. Establish face-sharing adjacency relationships between CT cube units. Each CT cube unit is only adjacent to six directions: front, back, left, right, top, and bottom. Adjacency directions beyond the three-dimensional voxel data matrix are shielded. Construct an initial CT cube topology network based on face-sharing adjacency relationships. Based on the lesion anatomical boundary data, the lesion region in the initial CT cube topology network is masked to obtain the three-dimensional region of the lesion. Three-dimensional erosion is performed on the CT voxels marked by the mask in the three-dimensional area of ​​the lesion. The three-dimensional erosion is dynamically iterated based on the lesion anatomical boundary data. After the three-dimensional erosion is completed, the edge smoothing optimization is performed at the lesion-normal tissue interface. The CT cube topology network is obtained based on the initial CT cube topology network after enhancement of the three-dimensional region of the lesion.

3. The CT image-based analysis and classification method according to claim 2, characterized in that, The process of performing three-dimensional erosion on CT voxels marked with a mask within the three-dimensional region of the lesion, with the three-dimensional erosion being dynamically iterated based on the lesion's anatomical boundary data, includes: The lesion volume is obtained based on the lesion anatomical boundary data, and the theoretical number of corrosion iterations is obtained based on the lesion volume. The theoretical number of corrosion iterations is obtained based on the volume-iteration formula, which can be expressed as: ; in, This represents the theoretical number of corrosion iterations. Represented as lesion volume, if calculated based on the volume-iteration formula. If it is negative, then a forced setting will be applied. =0; Based on the theoretical number of erosion iterations, three-dimensional erosion is continuously performed. Each round of three-dimensional erosion can be represented as traversing the CT voxels marked by the mask layer by layer in the three-dimensional region of the lesion. For each CT voxel marked by the mask, if there is at least one non-lesion CT voxel in the 3×3×3 neighborhood, then the voxel is removed from the mask. After each round of 3D erosion, the current lesion volume is obtained. The ratio of the lesion volume before and after the current lesion volume is calculated based on the lesion volume before and after the current lesion volume. An iteration threshold is set, which is 0.

65. If the ratio of the lesion volume before and after the current lesion volume is greater than or equal to the iteration threshold, the 3D erosion is stopped and the process reverts to the previous round of 3D erosion. After each round of 3D erosion, six-neighbor connectivity detection is performed synchronously based on the face-shared adjacency relationship. If the face-shared adjacency relationship of the CT cube unit in the 3D region of the lesion is detected to be broken and split into at least one isolated region, the 3D erosion is stopped and the process reverts to the previous round of 3D erosion.

4. The CT image-based analysis and classification method according to claim 2, characterized in that, The edge smoothing optimization at the lesion-normal tissue interface includes: Based on the acquisition of local CT voxels at the lesion-normal tissue interface, the HU value gradient in the X / Y / Z directions is obtained based on the local CT voxels, the gradient amplitude is calculated based on the HU value gradient in the X / Y / Z directions, the diffusion coefficient is obtained based on the gradient amplitude, and the diffusion coefficient is inversely proportional to the gradient amplitude. A diffusion filter based on the diffusion coefficient is applied along the normal direction to perform the process. Based on the local surface patch constructed at the lesion-normal tissue interface, the quadratic surface equation is fitted according to the least squares method, and the extreme curvature of the quadratic surface equation is calculated, and the curvature critical value is set. If the extreme curvature is greater than the curvature critical value, it is determined to be a high curvature region, and the diffusion filter radius is restricted to be less than 2 voxels; If the extreme curvature is less than or equal to the curvature critical value, it is determined to be a low curvature region, and the diffusion filter radius is restricted to be greater than or equal to 2 voxels.

5. The CT image-based analysis and classification method according to claim 1, characterized in that, The method further includes: A feature encoding rule is established, which represents the correspondence between each dimension in the Riemannian manifold space and the dimension in the CT multimodal topological feature tensor. When performing multidimensional scaling transformation, the corresponding dimensional relationship between the manifold multimodal topological feature tensor and the CT multimodal topological feature tensor is matched based on the feature encoding rule.

6. The CT image-based analysis and classification method according to claim 1, characterized in that, The acquisition of CT composite data based on the CT cube topology network includes: The CT composite data includes the three-dimensional spatial coordinates of the CT cube topology network, corrosion state markers, lesion volume, post-corrosion lesion volume, HU values ​​of CT voxels in the CT cube topology network, and adjacency index based on face-shared adjacency relationships.

7. The CT image-based analysis and classification method according to claim 1, characterized in that, The method further includes: The anatomical boundary data of the lesion includes the spatial coordinate data of the lesion core area, the marginal infiltration area, anatomical landmarks, and the lesion volume; The size of the three-dimensional voxel data matrix is ​​512×512×N, where N represents the number of axial slices. Each CT voxel in the three-dimensional voxel matrix corresponds to a HU value, and the HU value ranges from (-1000 to 2000).

8. A CT image-based analysis and classification system for implementing the method of any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire the patient's CT scan images and the anatomical boundary data of the lesions marked by the radiologist, and to acquire a three-dimensional voxel data matrix based on the patient's CT scan images. The building module is used to construct the CT cube topology network and acquire CT composite data; The coherence calculation module can perform continuous coherence calculation based on CT composite data, and obtain CT multimodal topological feature tensors based on CT zero-dimensional, CT one-dimensional and CT two-dimensional continuous coherence. The perturbation module is used to map the CT multimodal topological feature tensor to the Riemannian manifold space. It can obtain a 32-dimensional manifold perturbation vector based on the generative adversarial network model, and can obtain a perturbed CT dataset based on the 32-dimensional manifold perturbation vector and the CT multimodal topological feature tensor. The classification analysis module is used to train the training classification network model, obtain the CT classification network model, and generate CT conclusion results based on the CT classification network model.

Images