CT image-based analysis and classification method and system

Abstract

The invention provides an analysis and classification method and system based on a CT image, and the method comprises the steps: S1, obtaining a three-dimensional voxel data matrix of a CT scanning image of a patient, synchronously loading lesion anatomy boundary data marked by a radiologist, building a CT cubic topological network, obtaining CT composite data, carrying out the continuous coherence calculation according to S2, and obtaining a CT data matrix; and S3, mapping the CT multi-modal topological feature tensor to a Riemannian manifold space based on the step S3, establishing a generative adversarial network model, obtaining a 32-dimensional manifold disturbance vector, generating a disturbance CT data set, establishing a training classification network model according to the step S4, training the training classification network model, and obtaining a CT classification network model. Stopping executing the steps S3 to S4, finally executing the steps S1 to S2, and generating a CT conclusion result based on the CT classification network model; through the method, the CT image of the patient can be better analyzed and classified.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method and system for analyzing and classifying CT images. Background Technology

[0002] Against the backdrop of the deep integration between the demand for precise medical imaging diagnosis and artificial intelligence technology, CT image analysis and classification technology is facing challenges such as multimodal feature mining, improving adversarial robustness, and enhancing clinical interpretability. Existing methods for differentiating between benign and malignant tumors mainly rely on manual analysis and classification or are based on shallow semantic features from convolutional neural networks, but these methods reveal limitations in complex pathological scenarios.

[0003] On the one hand, there are limitations to the analysis and classification of CT images by humans. The consistency and repeatability of manual diagnosis and classification need to be improved. For example, there are individual differences when visually evaluating the topological features of CT images, which can easily lead to deviations.

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

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

[0006] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an analysis and classification method based on CT images, the method comprising:

[0007] 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.

[0008] 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.

[0009] 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.

[0010] 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.

[0011] 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.

[0012] As a further aspect of the present invention, 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 midpoint of its corresponding voxel.

[0013] 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.

[0014] 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.

[0015] 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.

[0016] 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.

[0017] As a further aspect of the present invention, the three-dimensional erosion of CT voxels marked by a mask within the three-dimensional region of the lesion, wherein the three-dimensional erosion is dynamically iterated based on the lesion's anatomical boundary data, includes:

[0018] 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: ;

[0019] 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;

[0020] 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.

[0021] 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.

[0022] 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.

[0023] As a further aspect of the present invention, the edge smoothing optimization at the lesion-normal tissue interface includes:

[0024] 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.

[0025] 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.

[0026] 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;

[0027] 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.

[0028] As a further aspect of the present invention, the step of mapping the CT multimodal topological feature tensor to the Riemannian manifold space, establishing a generative adversarial network (GAN) model, obtaining a 32-dimensional manifold perturbation vector based on the GAN model, and superimposing the 32-dimensional manifold perturbation vector with the CT multimodal topological feature tensor to generate a perturbed CT dataset includes:

[0029] The feature distance between different persistent graphs in the CT multimodal topological feature tensor is obtained based on Wasserstein distance calculation, and the feature distance is converted into a similarity score;

[0030] Multidimensional scaling transformation is performed based on feature distance and similarity score to map the CT multimodal topological feature tensor to the Riemannian manifold space, thereby obtaining the manifold multimodal topological feature tensor.

[0031] Based on the generative adversarial network model, 16-dimensional Gaussian noise is injected into the multimodal topological feature tensor of the manifold, and a 32-dimensional original perturbation vector is generated.

[0032] The 32-dimensional original perturbation vector is verified based on the Jacobian determinant to obtain the 32-dimensional manifold perturbation vector.

[0033] As a further aspect of the present invention, the method further includes:

[0034] 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.

[0035] As a further aspect of the present invention, the continuous cohomology calculation based on CT composite data, which obtains the CT multimodal topological feature tensor based on the continuous cohomology of CT zero-dimensional, CT one-dimensional, and CT two-dimensional data, includes:

[0036] 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.

[0037] The birth and death thresholds of each dimension are obtained based on the continuous cohomology of CT zero-dimensional, CT one-dimensional and CT two-dimensional dimensions. The birth and death thresholds of each dimension are mapped to the two-dimensional plane to generate persistent graphs of each dimension. At the same time, the dynamic curves of Betti number of each dimension are generated based on the persistent graphs of each dimension. Based on the dynamic curves of Betti number of each dimension, persistent graphs and continuous cohomology, the CT multimodal topological feature tensor is obtained. The CT multimodal topological feature tensor is a 32-dimensional floating-point tensor.

[0038] As a further aspect of the present invention, the acquisition of CT composite data based on a CT cube topology network includes:

[0039] 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.

[0040] As a further aspect of the present invention, the method further includes:

[0041] 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;

[0042] 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).

[0043] Furthermore, embodiments of the present invention also provide an analysis and classification system based on CT images, comprising:

[0044] 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.

[0045] The building module is used to construct the CT cube topology network and acquire CT composite data;

[0046] 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.

[0047] 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.

[0048] 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.

[0049] Based on the above, the embodiments of this application can quantify topological features that traditional morphology often cannot capture, such as cavity structures and vascular infiltration, through three-dimensional continuous cohomology. Furthermore, the fluctuation of HU value can be obtained through the dynamic curve of Betti number. In the Riemannian manifold space, the manifold perturbation is verified and constrained based on the Jacobian determinant, ensuring that the manifold adversarial perturbation simulates real anatomical deformations, such as lesion displacement caused by respiratory motion. This avoids non-physiological artifacts generated by traditional perturbations, improves the robustness and generalization of the subsequent training model, and thus better analyzes and classifies CT images, improves accuracy, and further enhances the model after its completion. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the execution flow of an analysis and classification method based on CT images provided in an embodiment of the present invention.

[0051] Figure 2This is a schematic diagram of the execution flow for establishing a CT cube topology network in a CT image-based analysis and classification system provided by an embodiment of the present invention.

[0052] Figure 3 This is a schematic diagram of an analysis and classification system based on CT images provided in an embodiment of the present invention. Detailed Implementation

[0053] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a schematic diagram illustrating the execution flow of a CT image-based analysis and classification method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the execution flow for establishing a CT cube topology network in a CT image-based analysis and classification system provided by an embodiment of the present invention. The following is a detailed description of this CT image-based analysis and classification method.

[0054] Step 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, and obtain CT composite data based on the CT cube topology network.

[0055] Specifically, the size of the three-dimensional voxel data matrix of the patient's CT scan image can be 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 can range from (-1000 to 2000).

[0056] Specifically, 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.

[0057] Furthermore, 512×512 is the standard matrix size for clinical CT scans, ensuring that the resolution of a single-layer image is sufficient to capture the morphological details of tiny nodules. The resolution of a single-layer image is approximately 0.5-1mm, while avoiding redundant calculations. If a high resolution such as 1024×1024 is used, the computational load will increase exponentially, but the gain effect on analysis and classification will be limited. The number of N-dimensional slices is dynamically adjusted to adapt to different scanning protocols, and the continuity of lesion characterization is ensured through three-dimensional reconstruction.

[0058] Furthermore, the HU value range (-1000, 2000) covers the typical density range of human tissues. By eliminating calibration differences between scanning devices through global unification, cross-center consistency of topological feature extraction is ensured. Moreover, the HU value is kept between (-1000, 2000) to filter out extreme HU values, such as metal implants, thus avoiding some artifacts from interfering with morphological three-dimensional erosion operations.

[0059] In this embodiment, the anatomical boundary data of the lesion marked by the radiologist includes the spatial coordinate data of the lesion core area, the marginal infiltration area, anatomical landmarks, and the lesion volume.

[0060] Furthermore, spatial coordinates are represented as the coordinates of the layer-by-layer contour points of the lesion boundary. Based on the DICOM coordinate system of the original CT image, the coordinates of the polygon vertices of the lesion region in each slice are marked. For the core area of ​​the lesion, the three-dimensional coordinates of the main lesion area are manually delineated by the physician. This usually corresponds to the solid component area with abnormally high HU values ​​in the CT image. This can provide spatial prior constraints for establishing the CT cube topology network and reduce interference from healthy tissue. For the marginal infiltrative area, the coordinates of the sub-regions of infiltrative features such as ground-glass opacities and vascular cluster signs around the lesion are marked. For anatomical landmarks, the coordinates of anatomical structures closely related to the lesion space are marked, such as adjacent bronchial bifurcation points, pulmonary artery branches, and pleural attachment points. This can provide a spatial reference frame to help establish the topological connection relationship between the lesion and anatomical structures. In subsequent adversarial training, it can ensure that the generated manifold perturbation is combined with real anatomical constraints, such as the deformation range not exceeding the pleural boundary. When obtaining the lesion volume, the total volume is automatically calculated based on the coordinates of the core area and the infiltrative area.

[0061] Understandably, lesion anatomical boundary data based on physician annotations can transform physicians' qualitative diagnostic experience into quantifiable topological features and geometric constraints for subsequent algorithms through spatial-semantic hierarchical annotation of such data.

[0062] The establishment of a CT cube topology network based on a three-dimensional voxel data matrix and lesion anatomical boundary data includes:

[0063] Step S11: Create a CT cube unit for each CT voxel contained in the three-dimensional voxel data matrix and map the vertex coordinates of the CT cube unit to the midpoint of its corresponding voxel.

[0064] In this embodiment, mapping the vertex coordinates of the CT cube unit to the corresponding voxel center allows the physical coordinate system inherited from the CT scan to be inherited, avoiding topological distortion caused by coordinate inconsistencies. Furthermore, center point mapping simplifies coordinate transformation calculations, requiring only the calculation of the linear mapping relationship between the CT voxel and the physical coordinates.

[0065] Step S12: 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. An initial CT cube topology network is constructed based on the face-sharing adjacency relationships.

[0066] In this embodiment, the effective adjacency relationship of each CT voxel in three-dimensional space is defined, including only CT voxels that are directly adjacent in the six directions of front, back, left, right, top, and bottom. When making connections, diagonal adjacency relationships are excluded to avoid topological noise introduced by oblique connections. For CT voxels located at the edge of the scan volume, adjacency directions that exceed the matrix range are automatically shielded to ensure the physical authenticity of the topological structure.

[0067] Step S13: Based on the lesion anatomical boundary data, mask the lesion region in the initial CT cube topology network to obtain the three-dimensional region of the lesion.

[0068] Step S14: Perform three-dimensional erosion 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, perform edge smoothing optimization at the lesion-normal tissue interface.

[0069] This includes three-dimensional erosion of CT voxels marked by a mask within the three-dimensional region of the lesion, with dynamic iteration based on the lesion's anatomical boundary data.

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

[0071] 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.

[0072] Understandably, using a base-2 logarithmic function to establish a quantitative relationship between volume and iteration number aligns with the exponential decay law of corrosion sensitivity to changes in lesion volume. Using a base-2 logarithmic function doubles the volume processing order with each iteration, matching common lesion growth patterns. hour, The function outputs a negative value, at which point a forced setting is required. =0, to avoid accidental damage to small lesions, when When the growth is exponential, the number of iterations increases linearly, ensuring that the corrosion intensity and lesion size have a sublinear relationship, and preventing undertreatment of large-volume lesions.

[0073] Understandable, As The threshold is because the volume of artifacts generated by voxel-level noise in a binary mask is generally smaller than that of the target image. The volume of the core area of ​​a malignant lesion is usually This ensures that the threshold does not corrode real pathological tissue.

[0074] In this example, when calculating the theoretical corrosion iteration count, for The threshold is dynamically corrected. Dynamic correction can be expressed as follows: if the resolution of the input CT image deviates from the standard value, such as... For voxels, the threshold is adjusted according to a proportional formula, which can be expressed as:

[0075] ;

[0076] in, It is represented as a constant.

[0077] Steps S14-12: Three-dimensional erosion is continuously performed based on the theoretical number of erosion iterations. Each round of three-dimensional erosion can be represented as traversing the CT voxels marked by the mask layer by layer within 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, the voxel is removed from the mask.

[0078] In this embodiment, an isotropic three-dimensional cubic core (3×3×3 voxels) is used, with all elements in the core assigned a value of 1. This ensures that the three-dimensional erosion can be performed uniformly in three-dimensional space, eliminating noise from isolated or semi-connected voxels in the lesion mask. The execution logic of the 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 3×3×3 neighborhood, the voxel is removed from the mask. After the three-dimensional erosion iteration is completed, only the set of voxels strongly connected to the lesion body is retained, eliminating fine branches and surface burrs.

[0079] Step S14-13-1: After each round of three-dimensional erosion, obtain the current lesion volume, calculate the ratio of the lesion volume before and after the current lesion volume based on the lesion volume, and set an iteration threshold of 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, stop the three-dimensional erosion and revert to the previous round of three-dimensional erosion.

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

[0081] ;

[0082] Calculated based on the lesion volume and the current lesion volume (i.e., the volume of the lesion after corrosion). ,like If the value is 0.65, the iteration is terminated early to avoid over-erosion caused by reaching the theoretical number of iterations. The three-dimensional erosion is stopped, and the system automatically reverts to the state of the previous three-dimensional erosion round. The number of the previous iteration is used as the final iteration number to ensure that the result is always within the effective constraint range.

[0083] Step S14-13-2: After each round of three-dimensional 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 three-dimensional region of the lesion is detected to be broken and split into at least one isolated region, then the three-dimensional erosion is stopped and the process reverts to the previous round of three-dimensional erosion.

[0084] In this embodiment, after each erosion iteration, the lesion region is checked to see if it remains a single connected component based on the surface shared adjacency relationship, that is, all voxels are connected through the surface shared adjacency path. If it is divided into multiple isolated sub-regions, or if the burrs break due to excessive erosion, it is determined to be a topological structure damage. Once damage is detected, it automatically reverts to the state of the previous iteration and uses the number of the previous iteration as the final iteration number to ensure that the result of the three-dimensional erosion meets the connectivity requirements.

[0085] Furthermore, a connected component is represented as an independent connected region consisting of all voxels connected by face-sharing adjacency relationships. That is, there is a continuous chain of adjacent voxels between any two voxels. When detecting whether a lesion region remains a single connected component, a marked voxel can be randomly selected from the eroded lesion region as the search starting point. Based on BFS (Breadth-First Search) or DFS (Depth-First Search), all voxels connected by the six neighborhoods are recursively visited and marked as visited. After traversal, if there are unmarked lesion voxels, they are determined to be multi-connected components, i.e., structural classification. Then, the volume of each connected component is calculated. If the volume of the largest component is greater than or equal to 95%, it is considered a slight split, which may be caused by the shedding of surface burrs. If the volume of the largest component is less than 95%, it is considered a severe classification. Based on the classification of slight and severe splits, it can be determined whether it is caused by residual noise or actual anatomical tearing.

[0086] Understandably, malignant lesions typically present as single, continuous masses with infiltrative borders. Multi-connected components may missegment satellite lesions or vascular infiltrative regions, leading to distortion in subsequent feature extraction. Furthermore, during subsequent continuous homology calculations, a closed manifold is required as the input structure, and multi-connected components may incorrectly generate additional Betti numbers, interfering with the calculation of topological features.

[0087] Understandably, topological constraints can prevent the loss of key case features. For example, the spiculation and lobulation features of malignant nodules often exhibit weak connectivity structures. Topological constraints can prevent these features from being misjudged as noise and removed. Secondly, they can ensure the reliability of subsequent continuous homology calculations. Continuous homology relies on the calculation of Betti numbers in closed connected domains. If the structure splits, it can lead to false mutations in ring features or cavity features.

[0088] In this embodiment, an adaptive enhancement mechanism can also be incorporated into the three-dimensional corrosion process, which can activate corrosion compensation factors for specific pathological categories (such as spiculated malignant tumors). , The number of iterations is corrected using a modified formula, which is as follows:

[0089] ;

[0090] The number of iterations is adjusted using the corrosion compensation factor. Enlargement enhances the ability to clearly see spiculated artifacts around lesions, compensating factors. The value of is learned from the training set. Based on 200-500 labeled data cases, the optimal compensation factor is obtained by using the adequacy of corrosion as assessed by pathologists as the optimization objective and the gradient descent method. .

[0091] In this example, the problem of under-corrosion or over-corrosion caused by the traditional fixed number of iterations is avoided by using the dual constraints of the iterative formula dependent on lesion volume and the current lesion volume.

[0092] Among these, edge smoothing optimization at the lesion-normal tissue interface includes:

[0093] Steps S14-21: Obtain local CT voxels at the lesion-normal tissue interface; obtain HU value gradients in the X / Y / Z directions based on the local CT voxels; calculate the gradient amplitude based on the HU value gradients in the X / Y / Z directions; obtain the diffusion coefficient based on the gradient amplitude; the diffusion coefficient is inversely proportional to the gradient amplitude; and apply diffusion filtering based on the diffusion coefficient along the normal direction.

[0094] In this embodiment, the lesion-normal tissue interface refers to the three-dimensional boundary area between the lesion area (such as abnormal tissues like tumors and inflammation) and the surrounding normal anatomical structures (such as healthy blood vessels and bronchi) in a CT image.

[0095] In this embodiment, a local CT voxel is represented as a 3×3×3 cubic neighborhood of the current target voxel, containing 27 voxels and including itself. This neighborhood size is consistent with the binding kernel size in the 3D erosion operation, ensuring the uniformity of algorithm parameters. When calculating the HU value gradient, the HU value gradient of the target voxel in the X / Y / Z directions is calculated based on the central difference method. Subsequently, the gradient magnitude is calculated based on the HU value gradients in the X / Y / Z directions.

[0096] ;

[0097] in, Represented as gradient magnitude, This is represented as the HU gradient in the X direction. This is represented as the gradient of the HU value in the Y direction. It is represented as the HU value gradient in the Z direction, and the gradient magnitude is used to drive the diffusion coefficients in all directions.

[0098] Furthermore, the diffusion coefficient is defined based on the gradient magnitude. When the gradient magnitude is high, such as at the edge of calcification or the interface of malignant infiltration, the diffusion coefficient approaches 0, which suppresses the smoothing operation and preserves the steep density jump characteristics. When the gradient magnitude is low, such as in the inflammatory exudative area, the diffusion coefficient approaches 1, which performs strong smoothing and eliminates the step-like state artifacts.

[0099] Furthermore, the filtering operation is performed along the interface normal direction to avoid cross-interface diffusion that could lead to structural ambiguity. The normal direction is determined by calculating the local HU gradient direction to ensure that smoothing only applies to homogeneous regions inside or outside the lesion.

[0100] Steps S14-22: Construct a local surface patch at the lesion-normal tissue interface, fit the quadratic surface equation using the least squares method, calculate the extreme curvature of the quadratic surface equation, and set the curvature critical value.

[0101] Step S14-23-1: 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 limited to less than 2 voxels.

[0102] Specifically, the high curvature region corresponds to microstructures such as burr apex and lobed depression, which need to be protected. Therefore, a diffusion filter with less than 2 voxels is used.

[0103] Step S14-24-2: 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.

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

[0105] Understandably, edge smoothing optimization can eliminate noise while preserving key pathological features such as malignant infiltration and spiculation, transforming the rough lesion mask after three-dimensional erosion into a smooth and continuous three-dimensional structure, providing input conforming to the differential manifold for subsequent extraction of topological features.

[0106] Step S15: Obtain the CT cube topology network based on the initial CT cube topology network after enhancement of the three-dimensional region of the lesion.

[0107] Step S2: Perform continuous cohomology calculation based on CT composite data, and obtain the CT multimodal topological feature tensor based on the continuous cohomology of CT zero-dimensional, CT one-dimensional and CT two-dimensional data.

[0108] Specifically, the birth and death thresholds of each dimension are obtained based on the continuous cohomology of CT zero-dimensional, CT one-dimensional, and CT two-dimensional dimensions. The birth and death thresholds of each dimension are mapped to the two-dimensional plane to generate persistent graphs of each dimension. At the same time, the dynamic curves of the Betti number of each dimension are generated based on the persistent graphs of each dimension. Based on the dynamic curves of the Betti number of each dimension, the persistent graphs, and the continuous cohomology, the CT multimodal topological feature tensor is obtained. The CT multimodal topological feature tensor is a 32-dimensional floating-point tensor.

[0109] In this example, CT zero-dimensional continuous cohomology is used to analyze connected components. CT zero-dimensional continuous cohomology dynamically adjusts the HU threshold from -1000HU to 2000HU in a step of 10HU. At each HU threshold, voxels with HU values ​​higher than the HU threshold are marked as active, and the number of connected components in the active regions is counted. Value), record The threshold point for value mutation, where mutation is defined as the value between adjacent HU thresholds. If the value changes by more than or equal to 50%, the HU threshold point is used as a candidate point for the segmented sign. The candidate points for the segmented sign are then evaluated. If a certain... If a value remains less than or equal to 3 HU threshold steps, it is considered noise disturbance and is not recorded. If the value is consistently greater than 3 HU threshold steps, it is identified as a mutation point and recorded.

[0110] Furthermore, when the HU threshold decreases to a certain critical value and a new mutation point appears, its birth threshold is recorded. When the HU threshold continues to decrease, the originally independent connected components merge and the number of mutation points decreases, and the extinction threshold is recorded.

[0111] In this embodiment, CT one-dimensional continuous cohomology is used to detect ring structures. CT one-dimensional continuous cohomology is used to construct a three-dimensional simple complex. Cubic units are connected through face-sharing adjacency relationships. In the three-dimensional simple complex, the ring structure is composed of at least four cubic units forming a closed path through shared faces. For example, a ring cavity formed by a blood vessel surrounding a tumor generates a closed loop (1-dimensional hole). The birth threshold of each loop is calculated, i.e., the HU value when the loop is first formed. For example, when the HU threshold decreases to a certain critical value, the low-density voxels in the necrotic area are activated, forming a closed loop. The extinction threshold of each loop is calculated, i.e., the HU value when the loop is filled. For example, as the HU threshold increases, the voxels of the surrounding solid components are activated, causing the ring cavity to close. Stable ring structures with a lifespan (extinction threshold - birth threshold) ≥ 80 HU are screened, filtering out short-term periodic noise.

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

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

[0114] Specifically, the CT multimodal topological feature tensor is a 32-dimensional floating-point tensor, with the following specific allocation:

[0115] Dimensions 0-9: Sampled values ​​of key HU thresholds (e.g., -800, -600, ..., 1000 HU) in the zero-dimensional Betti number dynamic curve, reflecting the dynamic process of lesion connectivity changing with density;

[0116] Dimensions 10-19: The first 10 life cycles and coordinates of the one-dimensional Betti number dynamic curve, where the coordinates represent dimensions (birth threshold, death threshold), reflecting the ring structure characteristics of vascular infiltration and malignant cavities;

[0117] Dimensions 20-29: Two-dimensional Betti number dynamic curves, life cycle integrals, maximum persistent life cycle, and spatial distribution entropy, reflecting the topological stability of the necrotic region;

[0118] Dimensions 30-31: Persistent entropy and the total lifespan of zero-dimensional, one-dimensional and two-dimensional combined, comprehensively quantifying the overall topological complexity of lesions.

[0119] Step 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.

[0120] Understandably, processing within a Riemannian manifold space allows for a more natural capture of the geometric and topological relationships in medical images. Traditional methods, such as those using Euclidean space, treat data as flat tables, making it difficult to accurately describe the dynamic changes in lesion morphology, such as tumor infiltration and growth or respiratory deformation. Manifold space, through its curved geometric structure, simulates the real anatomical deformation patterns, ensuring that the algorithm-generated perturbations, such as lesion displacement simulations, are physiologically plausible. For example, blood vessels will not suddenly rupture, and cavities will not appear out of nowhere, avoiding non-physiological artifacts caused by traditional perturbations. By adding perturbations within a Riemannian manifold space and using the perturbated samples as training samples for subsequent CT classification network model training, the robustness of the CT classification network model against interference can be improved. When the CT classification network model is performing analysis, even with slight shifts in the scanning angle, it can still perform relatively accurate analysis and classification of CT images.

[0121] Step S3 includes:

[0122] Step S31: Based on Wasserstein distance calculation, obtain the feature distance between different persistent graphs in the CT multimodal topological feature tensor, and convert the feature distance into a similarity score.

[0123] Step S32: Perform multi-dimensional scaling transformation based on feature distance and similarity score to map the CT multimodal topological feature tensor to the Riemannian manifold space and obtain the manifold multimodal topological feature tensor.

[0124] In this embodiment, feature points are matched pairwise in the zero-dimensional, one-dimensional, and two-dimensional persistent graphs. The minimum cost of "transferring" all feature points, i.e., the feature distance, is calculated. Based on Gaussian kernel transformation, the feature distance is converted into a similarity score. The similarity score ranges from 0 to 1, with higher scores for closer distances. An N×N distance matrix is ​​constructed based on the feature distance and similarity score. By decomposing the eigenvalues ​​of the N×N distance matrix, 32-dimensional coordinates are obtained, making the Euclidean distance in the manifold space approximate the feature distance, thus obtaining 32-dimensional manifold coordinates. Based on the 32-dimensional manifold coordinates, the manifold multimodal topological feature tensor is obtained.

[0125] Step S33: Based on the generative adversarial network model, inject 16-dimensional Gaussian noise into the manifold multimodal topological feature tensor and generate a 32-dimensional original perturbation vector.

[0126] Step S34: Verify the 32-dimensional original perturbation vector based on the Jacobian determinant to obtain the 32-dimensional manifold perturbation vector.

[0127] In this embodiment, a lung nodule with a diameter of 12mm is used as an example. At this time, step S2 has generated a CT multimodal topological feature tensor. The CT multimodal topological feature tensor includes persistent graph features, Betti number dynamic curves, and cavity distribution entropy. The persistent graph features are represented by a two-dimensional birth threshold dimension of 200 and a two-dimensional disappearance threshold dimension of 350, indicating that a cavity structure appears when HU=200. The cavity distribution entropy is 0.68. Based on the generative adversarial network model, 16-dimensional Gaussian noise is injected to provide relevant randomness. Based on the generator in the generative adversarial network model, a 16-dimensional perturbation vector is generated and injected into the CT multimodal topological feature tensor to generate perturbation data. For the dimension For dynamic curves with Betti numbers of degrees 16 to 23, a perturbation of 0.08 is injected. For cavity distribution entropy of dimensions 24 to 31, a perturbation of 0.05 is injected. The perturbation is constrained based on the Riemann norm to keep it below 0.15, generating a 32-dimensional original perturbation vector. After injecting the perturbation, a homeomorphism verification is performed. The perturbation is projected from the tangent space onto the manifold surface to ensure smooth deformation. The perturbation is verified based on the Jacobian determinant. If the Jacobian determinant verification does not affect the topology, the 32-dimensional original perturbation vector is considered as a 32-dimensional manifold perturbation vector. If the Jacobian determinant verification affects the topology, the perturbation is re-injected based on the generative adversarial network model.

[0128] Furthermore, feature encoding rules are established, which represent 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 rules.

[0129] Step 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.

[0130] Step S5: Perform steps S1 to S2 to import the CT multimodal topological feature tensor into the CT classification network model. Generate CT conclusion results based on the CT classification network model. The CT conclusion results include classification results and analysis probability results.

[0131] In this embodiment, steps S4 and S5 are explained using examples. Assume an 8mm nodule is detected in a patient's lung CT image. Steps S1 to S4 have completed the processing and generated a CT classification network model. Steps S1 and S2 are then executed to generate the CT multimodal topological feature tensor of the CT image. This tensor is imported into the CT classification network model, where the multimodal manifold classifier (MMC-Net) operates. The MMC-Net includes two channels for analysis: a geometric feature analysis channel and a topological feature analysis channel. The MMC-Net analyzes the CT multimodal topological feature tensor based on geometric and topological features. The geometric features are weighted and then input into the fully connected layer. The system generates high-order geometric features to reflect malignant morphological features such as "spiculation sign (weight 0.45)" and "lobulation sign (weight 0.32)". For topological features, it uses GRU to fuse the dynamic curve changes of the Betti number and outputs an enhanced topological feature vector. Then, it generates a joint feature matrix based on the high-order geometric features and the enhanced topological feature vector. It identifies perturbations in the joint feature matrix. For example, in this case, the perturbation amount is 0.1, which is a moderate perturbation. The classifier is triggered to calculate the distance from the sample to the center of the malignant / benign class. The output analysis result is a malignancy probability of 92.6%. The CT conclusion result is generated. The classification result included in the CT conclusion result is malignant, and the analysis probability result is 92.6%. The above-generated report can only be used to assist doctors in diagnosis and cannot be directly used as a diagnostic result.

[0132] Figure 3 The diagram illustrates a schematic of a CT image-based analysis and classification system that can realize the ideas of this application, according to some embodiments of this application.

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

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

[0135] Module 102 is used to construct a CT cube topology network and acquire CT composite data;

[0136] 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.

[0137] The perturbation module 103 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.

[0138] The classification analysis module 104 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.

[0139] The specific usage and function of this embodiment are explained below:

[0140] First, in step S1, a three-dimensional voxel data matrix of the patient's CT scan image is acquired, and lesion anatomical boundary data annotated by the radiologist is simultaneously 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 obtained. Next, in step S2, continuous cohomology calculation is performed. Based on the CT zero-dimensional, CT one-dimensional, and CT two-dimensional continuous cohomology in the CT composite data, CT multimodal topological feature tensors are obtained, thereby quantifying the relevant topological features. Then, in step S3, the CT multimodal topological feature tensors are mapped to the Riemannian manifold space. A 32-dimensional manifold perturbation vector is obtained based on a generative adversarial network model, and based on the Jacobian model… The determinant is used to verify and constrain the perturbation, thereby ensuring that the adversarial perturbation simulates real anatomical deformation. The perturbation is superimposed with the topological feature tensor to generate a perturbation CT dataset. Then, a classification network model is built according to step S4, and the perturbation CT dataset is used as training samples to train the classification network model to obtain the CT classification network model. At this point, the training is complete, and steps S3 to S4 can be stopped. Finally, steps S1 to S2 are executed to obtain the relevant data of the CT images that need to be analyzed and classified, and the above data is processed based on the CT classification network model to generate CT conclusion results. Based on the conclusion results, the classification results and probability results can be viewed.

[0141] Furthermore, embodiments of the present invention also provide an electronic device, comprising:

[0142] At least one processor; and at least one memory communicatively connected to the processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform the method proposed in Embodiment 1 of the present invention.

[0143] The following is a detailed introduction to the various components of the electronic device:

[0144] In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement Embodiment 1 of this invention, such as one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs).

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

[0146] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. For specific implementation methods, please refer to the above method embodiments, which will not be repeated here.

[0147] The memory can be a real-only memory (ROM) or other type of static storage device capable of storing static information and instructions, a random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; this embodiment of the invention does not specifically limit this.

[0148] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as 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, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via limited means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0149] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0150] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0151] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method of analyzing and classifying based on CT images, characterized by, The method comprises: S1: obtaining a three-dimensional voxel data matrix of a CT scan image of a patient and synchronously loading lesion anatomical boundary data labeled by a radiologist, establishing a CT cubic topology network based on the three-dimensional voxel data matrix and the lesion anatomical boundary data, and obtaining CT composite data based on the CT cubic topology network; S2: performing persistent homology calculation based on the CT composite data, obtaining a CT multi-modal topology feature tensor based on CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology in the CT composite data; S3: mapping the CT multi-modal topology feature tensor to a Riemannian manifold space, establishing a generative adversarial network model, obtaining a 32-dimensional manifold perturbation vector based on the generative adversarial network model, superimposing the 32-dimensional manifold perturbation vector and the CT multi-modal topology feature tensor, and generating a perturbed CT data set; 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; S5: executing steps S1 to S2, importing the CT multi-modal topology 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 comprises a classification result and an analysis probability result.

2. The method of claim 1, wherein, The method comprises: creating a CT cubic unit for each CT voxel contained in the three-dimensional voxel data matrix and mapping the CT cubic unit vertex coordinates to the corresponding voxel midpoint; establishing a face-sharing adjacency relationship between the CT cubic units, wherein each CT cubic unit is only adjacent to the front, rear, left, right, upper and lower six directions, and the adjacency directions beyond the three-dimensional voxel data matrix are shielded, and an initial CT cubic topology network is constructed based on the face-sharing adjacency relationship; masking the lesion regions in the initial CT cubic topology network based on the lesion anatomical boundary data to obtain a lesion three-dimensional region; performing three-dimensional corrosion on the mask-marked CT voxels in the lesion three-dimensional region, wherein the three-dimensional corrosion is dynamically iterated based on the lesion anatomical boundary data, and after the three-dimensional corrosion is completed, edge smoothing optimization is performed on the lesion-normal tissue interface; obtaining the CT cubic topology network based on the enhanced initial CT cubic topology network of the lesion three-dimensional region.

3. The method of claim 2, wherein, The method comprises: Based on the lesion anatomical boundary data, a lesion volume is obtained, and based on the lesion volume, a theoretical corrosion iteration number is obtained, the theoretical corrosion iteration number being obtained based on a volume-iteration formula, which can be expressed as: ; wherein, is expressed as the number of theoretical corrosion iterations, is expressed as the lesion volume, if calculated based on the volume-iteration formula is negative, then set =0; continuously performing three-dimensional corrosion based on the theoretical corrosion iteration number, wherein each round of three-dimensional corrosion can be represented as layer-by-layer traversal of the mask-marked CT voxels in the lesion three-dimensional region, and for each mask-marked CT voxel, if there is at least one non-lesion CT voxel in the 3*3*3 neighborhood, the voxel is removed from the mask; after each round of three-dimensional corrosion is completed, obtaining a current lesion volume, calculating a lesion volume ratio based on the lesion volume and the current lesion volume, setting an iteration threshold, wherein the iteration threshold is 0.65, and if the lesion volume ratio 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; After each round of three-dimensional corrosion, the six-neighbor connectivity detection is synchronously performed based on the face-sharing adjacency relationship. If the face-sharing adjacency relationship of the CT cubic unit in the three-dimensional region of the lesion is detected to be 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.

4. The method of claim 2, wherein, The edge smoothing optimization at the lesion-normal tissue interface comprises: Based on the local CT voxels obtained at the lesion-normal tissue interface, the HU value gradients in X / Y / Z directions are obtained based on the local CT voxels, 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, the diffusion coefficient is inversely proportional to the gradient amplitude, and the diffusion filtering based on the diffusion coefficient is applied along the normal direction to perform; Based on the local surface sheet constructed at the lesion-normal tissue interface, the quadratic surface equation is fitted according to the least square method, and the extreme curvature of the quadratic surface equation is calculated, and the curvature threshold is set; 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; 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.

5. The method of claim 1, wherein, 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, the 32-dimensional manifold perturbation vector is superimposed with the CT multi-modal topological feature tensor, and a perturbed CT data set is generated, comprising: The feature distance between different persistent graphs in the CT multi-modal topological feature tensor is calculated based on the Wasserstein distance, and the feature distance is converted into a similarity score; The CT multi-modal topological feature tensor is mapped to the Riemannian manifold space based on the feature distance and the similarity score for multidimensional scaling transformation, and a manifold multi-modal topological feature tensor is obtained; Based on the generative adversarial network model, 16-dimensional Gaussian noise is injected into the manifold multi-modal topological feature tensor, and a 32-dimensional original perturbation vector is generated; The 32-dimensional original perturbation vector is verified based on the Jacobian determinant, and a 32-dimensional manifold perturbation vector is obtained.

6. The method of claim 5, wherein, The method further comprises: A feature coding rule is established, which represents the corresponding relationship between each dimension in the Riemannian manifold space and the dimension in the CT multi-modal topological feature tensor. When multidimensional scaling transformation is performed, the corresponding dimensional relationship between the manifold multi-modal topological feature tensor and the CT multi-modal topological feature tensor is matched based on the feature coding rule.

7. The method of claim 1, wherein, The CT multi-modal topological feature tensor is obtained based on the CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology in the CT composite data, comprising: 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; The birth threshold and the death threshold of each dimension are obtained based on 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 persistence diagram of each dimension, and a Betti number dynamic curve of each dimension is generated according to the persistence diagram of each dimension, and a CT multi-modal topological feature tensor is obtained based on the Betti number dynamic curve, the persistence diagram and the persistent homology of each dimension, and the CT multi-modal topological feature tensor is a 32-dimensional floating-point tensor.

8. The method of claim 1, wherein, The CT composite data is obtained based on the CT cubic topological network, and the CT composite data comprises: The CT cubic topological network comprises three-dimensional space coordinates, an erosion state label, a lesion volume, an erosion lesion volume, a CT voxel HU value in the CT cubic topological network and an adjacent relationship index based on a face-sharing adjacent relationship.

9. The method of claim 1, wherein, The method further comprises: The lesion anatomical boundary data comprises spatial coordinate data of a lesion core area, an edge infiltration area and an anatomical marker, and a lesion volume; 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.

10. A CT image-based analysis classification system for implementing any of the methods of claims 1 to 9, characterized by, The method comprises: An acquisition module is configured to acquire a CT scan image 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 image of the patient; A construction module is configured to construct a CT cubic topological network and acquire CT composite data; A homology calculation module is configured to perform persistent homology calculation based on the CT composite data, and obtain a CT multi-modal topological feature tensor based on CT zero-dimensional, CT one-dimensional and CT two-dimensional persistent homology; A perturbation module is configured to map the CT multi-modal topological feature tensor to a Riemannian manifold space, obtain a 32-dimensional manifold perturbation vector based on a generative adversarial network model, and obtain perturbed CT data based on the 32-dimensional manifold perturbation vector and the CT multi-modal topological feature tensor; A classification analysis module is configured to train a CT classification network model, and generate a CT conclusion result based on the CT classification network model.

Images