Multi-scale osteosarcoma CT image lesion area detection method based on characteristic distillation

Abstract

The invention relates to the technical field of image recognition, in particular to a multi-scale osteosarcoma CT image lesion area detection method based on feature distillation. The method comprises the following steps: acquiring a multi-scale osteosarcoma CT image and an attachment area image, performing forged image difference reduction to generate a unified CT image, reconstructing a three-dimensional tumor body form, extracting a bone attachment structure, performing relocation registration to obtain tumor body relocation data, and then generating a simulation attachment area framework based on the data, the method comprises the following steps: performing static feature distillation on a three-dimensional tumor body form to obtain an edge contour, defining an initial attachment area frame, performing feature comparison through a historical contour image to extract form change, evaluating the erosion degree of the attachment frame, and finally screening candidate lesion areas and positioning osteosarcoma lesion areas according to change areas. According to the invention, more efficient multi-source heterogeneous image data integration, more accurate three-dimensional tumor simulation reconstruction and clearer lesion area detection are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and particularly relates to a multi-scale osteosarcoma CT image lesion region detection method based on feature distillation. BACKGROUND

[0002] The existing CT image processing technology often cannot effectively integrate multi-source heterogeneous image data, resulting in insufficient richness of image input, and artifacts interference is widespread in multi-angle CT images, affecting the image quality and the stability of subsequent processing. The traditional method lacks high-resolution and high-consistency image basis in three-dimensional tumor reconstruction and attachment region analysis, and it is difficult to meet the strict requirements of accuracy in clinical practice. The existing technology has significant deficiencies in the extraction and repositioning registration of bone attachment structures, and cannot realize high-precision dynamic correspondence between tumors and attachment structures. There is a lack of effective feature distillation technology, resulting in low efficiency of static feature extraction, and the temporal continuity of tumor morphology changes cannot be fully reflected. The erosion degree evaluation of the attachment framework also lacks fine analysis, which limits the accurate positioning of complex osteosarcoma lesion regions. The existing region screening method is not up to the task when dealing with temporal differences and spatial gradients, affecting the screening specificity and adaptability of the lesion region. SUMMARY

[0003] Therefore, it is necessary to provide a multi-scale osteosarcoma CT image lesion region detection method based on feature distillation to solve at least one of the above technical problems.

[0004] To achieve the above-mentioned purpose, the multi-scale osteosarcoma CT image lesion region detection method based on feature distillation comprises the following steps: Step S1: Collecting multi-scale osteosarcoma CT images and osteosarcoma attachment region images; reducing the artifact difference of each angle of the multi-scale osteosarcoma CT images to generate a unified osteosarcoma CT image; and reconstructing a simulated three-dimensional tumor morphology based on the unified osteosarcoma CT image; Step S2: Extracting the bone attachment structure in the osteosarcoma attachment region image, and performing repositioning registration based on the bone attachment structure and the reconstructed simulated three-dimensional tumor morphology to obtain tumor repositioning data; and performing attachment framework simulation based on the tumor repositioning data and the osteosarcoma attachment region image to generate a simulated attachment region framework; Step S3: Distilling the static features of the simulated three-dimensional tumor morphology to obtain the static edge contour of the tumor; and defining an initial attachment region framework according to the simulated attachment region framework; Step S4: Comparing features based on a preset historical tumor contour image and the static edge contour of the tumor to extract the tumor morphology change process; and evaluating the erosion degree of the initial attachment region framework according to a preset historical tumor attachment framework image to generate an attachment framework erosion degree; Step S5: According to the change area screening of the tumor morphology change process, the candidate tumor lesion area is obtained; based on the candidate tumor lesion area and the erosion degree of the attachment framework, the osteosarcoma lesion area in the multi-scale osteosarcoma CT image is located.

[0005] The present application realizes the structured integration of multi-source heterogeneous image data through the cooperative collection of multi-scale osteosarcoma CT images and osteosarcoma attachment area images, improves the richness and integrity of image input, and reduces the difference between the forged images. A unified image expression is constructed among the multi-view CT images, effectively suppressing the structural artifacts introduced by different angles due to the forged images, ensuring the basic stability of subsequent image processing, and generating a unified osteosarcoma CT image. It provides a high consistency and high resolution image basis for three-dimensional tumor simulation reconstruction. The reconstruction link of the simulated three-dimensional tumor morphology introduces a spatial information fusion and structure level remodeling mechanism, significantly enhancing the three-dimensional expression ability of the tumor structure. The accurate extraction of the bone attachment structure combined with the spatial repositioning registration of the reconstructed simulated three-dimensional tumor morphology makes the tumor and the attachment structure realize high-precision dynamic correspondence in the spatial coordinate system, ensuring the structural reliability of subsequent registration and framework simulation. The simulation process of the attachment framework introduces a structure mapping and spatial interaction mechanism, improving the spatial relationship expression depth between the bone attachment structure and the tumor. The generation of the simulated attachment area framework enhances the integrity and boundary clarity of the attachment area structure. The tumor static feature distillation uses multi-dimensional feature extraction and redundancy reduction to output a structure that is significant and has a clear boundary. The initial attachment area framework is defined based on the simulated attachment area framework, which realizes the dynamic import of spatial structure priori, improves the accuracy of regional division, and extracts the tumor morphology change process using multi-source feature embedding and change trajectory fitting. The simulated three-dimensional tumor morphology significantly enhances the temporal continuity and spatial contrast of the morphological change expression. The erosion degree is generated in combination with feature distribution, spatial stratification and structure response, which finely depicts the local erosion features of the attachment framework. The change area screening process is based on the tumor morphology change process, which integrates temporal differences, spatial gradients and heterogeneous feature aggregation, and realizes efficient screening of structure-sensitive regions. The output of the candidate tumor lesion area ensures the specificity and multi-scale adaptability of the region screening. The joint positioning of the candidate area and the erosion degree of the attachment framework uses multi-domain weight allocation, response coupling and boundary correction processes to strengthen the structural consistency and resolution of the spatial positioning of the lesion area. The final positioning result of the osteosarcoma lesion area in the multi-scale CT image has better spatial accuracy, structural continuity and clear boundary. BRIEF DESCRIPTION OF DRAWINGS

[0006] Fig. 1 The step flowchart of the multi-scale osteosarcoma CT image lesion area detection method based on feature distillation is shown in the figure. Fig. 2 for the detailed implementation procedure of step S1; The objectives, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0007] The technical method of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0008] In addition, the accompanying drawings are only schematic illustrations of the present application, and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus repeated description thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0009] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be referred to as a second element, and similarly a second element can be referred to as a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0010] To achieve the above-mentioned object, please refer to Figs. 1-2 The multi-scale osteosarcoma CT image lesion region detection method based on feature distillation comprises the following steps: Step S1: Collecting multi-scale osteosarcoma CT images and osteosarcoma attachment region images; reducing the pseudo-contrast difference of each angle of the multi-scale osteosarcoma CT images to generate uniform osteosarcoma CT images; and reconstructing a simulated three-dimensional tumor morphology based on the uniform osteosarcoma CT images; Step S2: Extracting the bone attachment structure in the osteosarcoma attachment region image, and performing repositioning registration based on the bone attachment structure and the reconstructed simulated three-dimensional tumor morphology to obtain tumor repositioning data; and performing attachment frame simulation based on the tumor repositioning data and the osteosarcoma attachment region image to generate a simulated attachment region frame; Step S3: distilling tumor static features on the simulated three-dimensional tumor morphology, so as to obtain a tumor static edge contour; and defining an initial attachment area framework according to a simulated attachment area framework; Step S4: performing feature comparison based on a preset historical tumor contour image and the tumor static edge contour, extracting a tumor morphology change process; and performing erosion degree evaluation on the initial attachment area framework according to a preset historical tumor attachment framework image, to generate an attachment framework erosion degree; Step S5: performing change area screening according to the tumor morphology change process, so as to obtain a candidate tumor lesion area; and positioning a bone sarcoma lesion area in a multi-scale bone sarcoma CT image based on the candidate tumor lesion area and the attachment framework erosion degree.

[0011] The application realizes the structured integration of multi-source heterogeneous image data by the cooperative collection of multi-scale osteosarcoma CT images and osteosarcoma attachment area images, improves the richness and integrity of image input, constructs a unified image expression among multi-view CT images by reducing the difference degree of pseudo shadows, effectively suppresses the structural artifacts introduced by pseudo shadows at different angles, guarantees the basic stability of subsequent image processing, generates a unified osteosarcoma CT image, and provides a high consistency and high resolution image basis for three-dimensional tumor simulation reconstruction, the reconstruction link of the simulated three-dimensional tumor morphology introduces a spatial information fusion and structure level remodeling mechanism, which significantly enhances the three-dimensional expression ability of the tumor structure, the accurate extraction of the bone attachment structure combined with the spatial repositioning registration of the reconstructed simulated three-dimensional tumor morphology enables the tumor and the attachment structure to realize high-precision dynamic correspondence in the spatial coordinate system, guarantees the structural reliability of subsequent registration and framework simulation, the simulation process of the attachment framework introduces a structure mapping and spatial interaction mechanism, which improves the spatial relationship expression depth between the bone attachment structure and the tumor, the generation of the simulated attachment area framework enhances the integrity and boundary clarity of the attachment area structure, the tumor static feature distillation uses multi-dimensional feature extraction and redundancy reduction to output a static edge contour with significant structure and clear boundary, which optimizes the basic features of subsequent contour comparison and region screening, the definition of the initial attachment area framework relies on the simulated attachment area framework to realize the dynamic import of spatial structure prior, which improves the accuracy of region division, when extracting the tumor morphology change process, multi-source feature embedding and change trajectory fitting are used, which significantly enhances the temporal continuity and spatial contrast of the morphology change expression, the generation of the erosion degree combines feature distribution, spatial stratification and structure response to finely depict the local erosion features of the attachment framework, the change region screening link is based on the tumor morphology change process, and integrates the temporal difference, spatial gradient and heterogeneous feature aggregation to realize efficient screening of the structure sensitive region, the output of the candidate tumor lesion region ensures the specificity and multi-scale adaptability of region screening, the joint positioning of the candidate region and the erosion degree of the attachment framework uses processes such as multi-domain weight distribution, response coupling and boundary correction to strengthen the structural consistency and resolution ability of the spatial positioning of the lesion region, and the final positioning result of the osteosarcoma lesion region in the multi-scale CT image has better spatial accuracy, structural continuity and clear boundary.

[0012] In the embodiment of the application, the multi-scale osteosarcoma CT image lesion area detection method based on feature distillation comprises the following steps: Step S1: collect multi-scale osteosarcoma CT images and osteosarcoma attachment area images; reduce the difference degree of pseudo shadows in each angle of the multi-scale osteosarcoma CT images to generate a unified osteosarcoma CT image; and reconstruct a simulated three-dimensional tumor morphology based on the unified osteosarcoma CT image; In this embodiment, multi-scale osteosarcoma CT images and osteosarcoma attachment area images are collected. A multi-slice spiral CT scanner is used to scan the patient to obtain CT image data of different slice thicknesses (such as 1 mm, 3 mm, and 5 mm) and different angles (such as 0°, 45°, and 90°). The images of each angle in the multi-scale osteosarcoma CT images are subjected to pseudo-contrast difference reduction. An image registration algorithm is used to align the images of different angles to reduce the image differences caused by angle differences. Then, an image fusion technology is used to fuse the aligned images to generate a unified osteosarcoma CT image. Based on the unified osteosarcoma CT image, a three-dimensional reconstruction software (such as Mimics or 3D Slicer) is used to simulate the reconstruction of the three-dimensional tumor morphology. A threshold range (such as HU values between 50 and 300) is set to extract the tumor area to generate a simulated three-dimensional tumor model.

[0013] Step S2: Extract the bone attachment structure in the osteosarcoma attachment area image, and perform repositioning registration based on the bone attachment structure and the reconstructed simulated three-dimensional tumor morphology to obtain tumor repositioning data; perform attachment frame simulation based on the tumor repositioning data and the osteosarcoma attachment area image to generate a simulated attachment area frame; In this embodiment, the bone attachment structure in the osteosarcoma attachment area image is extracted. An edge detection algorithm (such as Canny Edge Detection) is used to identify the edges of the bone attachment structure, and a region growing algorithm (Region Growing Algorithm) is used to extract the complete bone attachment structure. Then, repositioning registration is performed based on the bone attachment structure and the reconstructed simulated three-dimensional tumor morphology. A rigid registration algorithm (Rigid Registration Algorithm) is used to register the bone attachment structure and the tumor model to obtain tumor repositioning data. Based on the tumor repositioning data and the osteosarcoma attachment area image, a finite element analysis software (such as ANSYS) is used to perform attachment frame simulation. Material properties (such as elastic modulus and Poisson's ratio) and boundary conditions are set to simulate the interaction between the tumor and the bone attachment structure to generate a simulated attachment area frame.

[0014] Step S3: Perform tumor static feature distillation on the simulated three-dimensional tumor morphology to obtain a tumor static edge contour; and define an initial attachment area frame based on the simulated attachment area frame; In this embodiment, the static features of the simulated three-dimensional tumor morphology are distilled, the edge features of the tumor are extracted using a feature extraction algorithm (such as Histogram of Oriented Gradients, HOG), and the static edge contour of the tumor is generated. According to the simulated attachment area framework, an initial attachment area framework is defined, and the contact area of the tumor and the bone attachment structure is determined as the basis for subsequent analysis.

[0015] Step S4: Based on the preset historical tumor contour image and the static edge contour of the tumor, feature comparison is performed to extract the tumor morphology change process; according to the preset historical tumor attachment framework image, the initial attachment area framework is evaluated for erosion degree, and the attachment framework erosion degree is generated. In this embodiment, based on the preset historical tumor contour image and the static edge contour of the tumor, feature comparison is performed, the image difference analysis algorithm (Image Difference Analysis Algorithm) is used to compare the difference between the current tumor edge contour and the historical contour, and the tumor morphology change process is extracted. According to the preset historical tumor attachment framework image, the initial attachment area framework is evaluated for erosion degree, and the morphological analysis method (Morphological Analysis Method) is used to calculate the change of the attachment framework, and the attachment framework erosion degree is generated.

[0016] Step S5: According to the tumor morphology change process, the change area is screened to obtain the candidate tumor lesion area; based on the candidate tumor lesion area and the attachment framework erosion degree, the osteosarcoma lesion area in the multi-scale osteosarcoma CT image is located.

[0017] In this embodiment, according to the tumor morphology change process, the change area is screened, the image segmentation algorithm (Image Segmentation Algorithm) is used to identify the area with significant tumor morphology change, and the candidate tumor lesion area is obtained. Based on the candidate tumor lesion area and the attachment framework erosion degree, the osteosarcoma lesion area in the multi-scale osteosarcoma CT image is located, and the specific position and range of the lesion area are determined in combination with the multi-scale image data.

[0018] Preferably, step S1 comprises the following steps: Step S11: Collecting multi-scale osteosarcoma CT images and osteosarcoma attachment area images; performing pixel-level normalization processing on the multi-scale osteosarcoma CT images to generate normalized osteosarcoma CT images; Step S12: Multi-angle recombination of the normalized osteosarcoma CT images, wherein the angle step interval is 5°, and the total rotation angle is not less than 180°, thereby obtaining the recombined osteosarcoma CT images; Step S13: uniformity processing of the pseudo-contrast of the recombined osteosarcoma CT image is performed, and heterogeneous shadow lines are weakened, so as to generate a uniform osteosarcoma CT image, wherein the window size of the shadow line weakening processing is limited to 7x7 pixels; Step S14: spatial layer fusion is performed on the uniform osteosarcoma CT image, so as to obtain a spatially fused osteosarcoma CT image. Step S15: three-dimensional morphological simulation reconstruction is performed based on the spatially fused osteosarcoma CT image, so as to generate a three-dimensional simulated tumor morphology.

[0019] In this embodiment, the stage of collecting multi-scale osteosarcoma CT images and osteosarcoma attachment area images uses a 16-layer spiral CT scanning device with a resolution of 512x512 pixels, and through different layer thickness configurations (1mm, 2mm, 5mm), cross-sectional scanning is performed, and through the control of the lifting of the scanning bed, image acquisition in the longitudinal direction is realized, the angle coverage range is 0° to 180°, and every 10° a group of scanning image data is obtained. The acquisition of the osteosarcoma attachment area image is realized by setting the scanning center in the central part of the suspected attachment area, and using a high-resolution reconstruction algorithm to collect and reconstruct the data in this area. The pixel-level normalization processing is performed on all the collected original DICOM image data, and the normalization algorithm adopts the maximum and minimum value normalization method. The pixel value in each image is subtracted by the minimum pixel value in the image and then divided by the difference between the maximum pixel value and the minimum pixel value. The pixel value range of the normalized image is limited to 0 to 1. In the process of multi-angle reconstruction of the normalized osteosarcoma CT image, the normalized CT image is imported into the image rotation module based on OpenCV, and a two-dimensional affine transformation matrix is used to rotate each image. The angle step interval is set to 5°, and the rotation center is the image center point coordinate (256, 256). Each group of images is rotated from 0° to 180°, a total of 37 images with different rotation angles are generated, all angle images are processed using the same image size and interpolation algorithm (using bilinear interpolation method), and black edge padding is used for image boundary in the image rotation process. The rotation results are sequentially saved after being named by angle. In the process of performing pseudo-contrast uniform processing on the multi-angle reconstructed osteosarcoma CT image, first, the CLAHE (Contrast Limited Adaptive Histogram Equalization) method is used to enhance the local contrast of each image, the contrast factor is set to 2.0, and the sliding window size is 7x7 pixels. High-pass filtering is performed on the enhanced image to weaken the heterogeneous shadow, and a Gaussian filter with a kernel size of 7x7 pixels is used for image smoothing, and the filter standard deviation is set to 1.5, the high-frequency residual signal is obtained by subtracting the Gaussian smoothing result, thereby weakening the Moiré effect caused by the bone structure in the CT image acquisition, the processed image reaches a balanced state in the pseudo shadow intensity, and the key ROI region is selected in each angle image for pixel distribution consistency detection, and finally the processed image is saved as a unified osteosarcoma CT image sequence, in the step of spatial layer fusion of the unified osteosarcoma CT image, a wavelet decomposition-based image fusion algorithm is used to fuse all the pseudo shadow unified images, Haar wavelet function is used to decompose each group of images to the third layer, the approximate components and detail components of each layer are extracted, then the pixel layer fusion is performed on the images of the corresponding scale, the fusion rule is to take the maximum response value in the high-frequency region and take the mean value in the low-frequency region, then the fused wavelet components are inversely transformed to reconstruct the spatial fusion image, in the fusion process, the weighted average method is used for volume fusion of the image stack in the Z-axis direction, and the weight is set according to the local gradient change degree in the image, the voxel spacing of the spatial fusion osteosarcoma CT image is maintained consistent with the original image, and the output three-dimensional data body has a dimension of 512*512*N, wherein N is the number of reconstructed images, in the three-dimensional morphological simulation reconstruction stage based on the spatial fusion osteosarcoma CT image, the VTK (Visualization Toolkit) three-dimensional reconstruction tool is used in combination with the Marching Cubes algorithm to perform surface modeling processing on the fused image, first, the region with a voxel value greater than 0.45 is set as the target region of the osteosarcoma structure, a cube is formed by using eight adjacent points between voxels, the Marching Cubes algorithm is used to judge the generation mode of the triangular facet, and an OBJ format three-dimensional model file is output, then smoothing processing operation is performed on the model, the Laplacian smoothing algorithm is used for iteration 5 times, and the smoothing factor of each iteration is set to 0.25, and finally the generated three-dimensional simulation tumor morphology structure retains complete spatial topological information and boundary contour characteristics.

[0020] Preferably, the extracting the bone attachment structure in the osteosarcoma attachment region image and the repositioning registration based on the bone attachment structure and the reconstructed simulated three-dimensional tumor morphology in step S2 comprises: extracting the bone attachment structure in the osteosarcoma attachment region image; performing multi-modal spatial registration on the bone attachment structure and the reconstructed simulated three-dimensional tumor morphology to obtain registered bone attachment structure data; projecting the registered bone attachment structure data into the osteosarcoma attachment region image to generate an osteosarcoma projection positioning; repositioning the reconstructed simulated three-dimensional tumor morphology according to the osteosarcoma projection positioning to obtain tumor repositioning data.

[0021] In this embodiment, the osteosarcoma attachment area image data obtained by spiral CT scanning is imported, the image size is 512x512 pixels, the gray level is 12 bits, the image format is DICOM, the segmentation network based on U-Net architecture is used to segment the bone attachment structure, the segmentation network is optimized by combining Dice loss function and binary cross entropy loss function in the training stage, the training data comes from the public bone tumor CT annotation data set, and the bone attachment area is manually outlined by three experienced radiologists for label verification, the image input size is adjusted to 256x256 pixels, the batch size is set to 4, the iteration round number is set to 100 rounds, each round contains 300 steps, the learning rate is initialized to 1e-4 and reduced by half every 20 rounds, after training, the network is applied to the osteosarcoma attachment area image, and the binary mask image of the bone attachment structure is output, the pixel value of 1 in the mask image represents the bone attachment structure area, and the pixel value of 0 represents the non-bone attachment area, the image extraction result of the bone attachment structure is obtained by superimposing the mask image and the original image, and the bone attachment structure image data is constructed into a voxel data set in a three-dimensional stacking form, the voxel size maintains the original spatial resolution of 0.5mmx0.5mmx1mm, in the stage of multi-modal space registration of bone attachment structure and reconstructed simulated three-dimensional tumor morphology, the multi-resolution rigid registration algorithm in ITK (Insight Segmentation and Registration Toolkit, Image Segmentation and Registration Toolkit) is selected, the initial registration transformation is set to rigid transformation, including two parameter spaces of rotation and translation, Mattes Mutual Information is used as the registration similarity measurement index, the optimization algorithm is Regular Step Gradient Descent, the maximum step is set to 1.0, the minimum step is 0.01, the iteration number is set to 300 times, and the pyramid level is 3, and the scaling factors of each layer are 1, 0.5 and 0.25, using trilinear interpolation method to resample the reconstructed simulated three-dimensional tumor morphology, set the bone attachment structure voxel data as the fixed image and the simulated three-dimensional tumor morphology as the moving image in the registration process, and finally output the affine matrix to align the simulated three-dimensional tumor morphology to the bone attachment structure space, generate the registration result of the registered bone attachment structure data and the reconstructed tumor data, in the process of projecting the registered bone attachment structure data into the osteosarcoma attachment area image, first obtain the Z-axis layer thickness and image origin coordinates of the image stack according to the spatial direction information of the DICOM image, perform coordinate conversion on the registered bone attachment structure data using the registration affine matrix, perform layer-by-layer resampling on the three-dimensional bone attachment structure data using the vtkImageReslice module in the VTK tool, the resampling method is bilinear interpolation, the step interval is consistent with the original image, which is 1mm, the image size is aligned to 512*512 pixels, after obtaining the two-dimensional slice sequence consistent with the attachment area image, the bone attachment structure image is superimposed on the attachment area image in the form of alpha channel, the transparency factor is set to 0.3 to visualize the boundary relationship between the bone attachment structure and the original image, generate the osteosarcoma projection positioning image sequence and save it according to the sequence naming rule, in the process of osteoma interface repositioning of the reconstructed simulated three-dimensional tumor morphology according to the osteosarcoma projection positioning, based on the bone attachment structure and the tumor boundary contact area in the osteosarcoma projection positioning image sequence, the three-dimensional fitting is performed by extracting the interface contour point cloud coordinates, the RANSAC (Random Sample Consensus) algorithm is used to fit the osteoma interface plane, the number of sampling points is set to 100 per round, the fitting error threshold is 0.5mm, and the iteration number is 1000 rounds, after obtaining the fitted plane, the corresponding interface coordinate plane is generated by using the vtkPlane module in VTK, then the coordinate clipping method is used to perform geometric transformation on the simulated three-dimensional tumor model, the repositioned osteoma interface is aligned to the bone attachment structure direction, and secondary space fine tuning is performed, the three-dimensional rigid transformation matrix is used to adjust the rotation and translation of the tumor model, the translation step is 0.1mm, and the rotation accuracy is 0.5 degrees, and finally the tumor repositioning data is output.

[0022] Preferably, the step S2 includes: Simulating the attachment frame based on the tumor repositioning data and the osteosarcoma attachment area image, including: Performing structure fusion simulation according to the attachment point calibration contour and the reconstructed simulated three-dimensional tumor morphology, and cutting the structure fusion attachment area image in the osteosarcoma attachment area image based on the simulated fusion structure; Modeling the frame boundary through the structure fusion attachment area image, wherein the modeling grid length is limited to 2-4mm, and the attachment frame boundary is generated. Spatial skeleton simulation is performed based on the attachment frame boundary, thereby obtaining a simulated attachment area frame.

[0023] In this embodiment, in the process of simulating the attachment point calibration of the tumor relocation data, the input data is the three-dimensional model of the tumor relocation output in the previous stage, in the STL format, the point set generation operation is performed by using the vtkPolyDataPointPlacer module in the VTK (Visualization Toolkit), first, the tumor model is mapped into a three-dimensional Cartesian coordinate system, the boundary of the attachment area is set as the main projection range, a three-dimensional grid search space is constructed with an interval of 0.1 mm, the barycentric position of each surface triangle is sampled, an attachment point is extracted every 5 mm, the effective attachment area is screened by using the volume curvature threshold, the curvature threshold is limited to 0.02, the number of point calibration is controlled between 30 and 50, and finally the attachment point calibration contour is saved in the PLY format in the form of a point cloud, including XYZ three-dimensional coordinates and normal vector information, in the operation of structure fusion simulation according to the attachment point calibration contour and the reconstructed simulated three-dimensional tumor morphology, the tumor three-dimensional model and the attachment point calibration point cloud are imported into MeshLab for fusion reconstruction processing at the same time, first, the point cloud is reconstructed by using the Poisson Surface Reconstruction algorithm, the reconstruction depth parameter is set to 10, the sample density limit is at least 20 points per unit surface area, and the smoothing factor is set to 0.4, after obtaining the attachment contour surface model based on the calibration points, the surface model is intersected with the original tumor STL model through grid Boolean operation, and the structure fusion model is output. The structure fusion model is sliced along the cross section of the osteosarcoma attachment area image through the vtkCutter module of VTK, the slicing direction is set to the axial direction, the thickness is set to 1mm, the step is 1mm, the image size is 512x512 pixels, the generated structure fusion slice sequence is aligned with the original DICOM image sequence through indexing, and the structure fusion attachment area image in the osteosarcoma attachment area image is cropped. When the framework boundary modeling is carried out through the structure fusion attachment area image, the structure fusion attachment area image is imported into the 3D Slicer platform, the "Threshold" threshold tool in the Segment Editor module is used to preliminarily extract the structure boundary, and the gray threshold range is set to [100, 600]. After preliminary segmentation, the "MarchingCubes" algorithm is called to construct the isosurface model, and the "Surface Toolbox" module is used to perform triangular mesh reconstruction on the model. The grid edge length control parameter is set to between 2mm and 4mm, all triangles exceeding the edge length threshold are cropped through the maximum edge length constraint algorithm (Maximum Edge Length Constraint), and bidirectional topological repair operation is performed to generate the attachment framework boundary. The boundary model is exported as an OBJ file and the three-dimensional space coordinates and grid topology structure corresponding to each boundary point are recorded. The total number of boundary models is controlled within 50000 faces. In the operation of spatial skeleton simulation based on the attachment framework boundary, the Geometry Nodes system integrated with Blender is used to construct the spatial skeleton simulation process. First, the attachment framework boundary OBJ file is imported, the "Surface to Curve" operation is applied to convert the boundary surface into an equidistant curve contour, and then the "Curve to Points" operation is used to generate a uniform skeleton point set on the curve contour. The interval parameter is set to generate a skeleton point every 5mm. The skeleton point set is divided into a three-dimensional skeleton grid structure through Delaunay three-dimensional triangulation. The shortest path tree algorithm is used to connect the key nodes of the skeleton to establish the center skeleton line. The connection weight is constructed according to the weighted Euclidean distance between nodes. Finally, the simulation attachment area framework skeleton data structure is output in JSON format, recording all the positions and connection relationships of the skeleton nodes. The node coordinate precision is set to four decimal places. The generated skeleton structure is combined with the three-dimensional tumor model and the attachment point calibration structure for joint visualization verification to complete the spatial skeleton simulation output.

[0024] Especially important is that the simulation attachment point calibration of the tumor relocation data includes: The tumor relocation data is sampled and rearranged by three-dimensional grid to obtain the distribution of candidate attachment points. measure a spatial gradient of each point to the center of the tumor surface boundary based on the candidate attachment point distribution, and perform a boundary centricity measurement screening based on the spatial gradient to obtain a screened attachment point set, wherein the number of attachment points is not less than 30; topological neighborhood constraint fitting is performed according to the screened attachment point set to generate a preliminary attachment point fitting structure; a local surface profile is reconstructed through the preliminary attachment point fitting structure, so as to generate a point position profile; global interpolation reconstruction is performed based on the point position profile, and boundary completion is performed based on the point position profile, so as to obtain an attachment point calibration profile.

[0025] In this embodiment, the repositioned tumor data is extracted using a dense voxel segmentation method based on CT images to obtain three-dimensional binary tumor volume data, and the Marching Cubes algorithm is used to convert voxel information into a triangular mesh model to generate a triangular mesh model. All triangular facet point coordinates form a tumor surface point set. The grid model is divided into three-dimensional space, and the entire tumor bounding box is divided into equidistant voxel blocks with a 2mm cube size. In each voxel block, all grid vertex coordinates falling within it are collected, and the geometric center is calculated using these coordinates, i.e. the average value of each coordinate as the barycenter. Obtain the geometric barycenter of all non-empty voxel blocks to form the first set of candidate attachment point set, in order to ensure the uniformity of the spatial distribution of these points on the tumor structure surface, K-means clustering is performed on all barycenter coordinates, the clustering number is set to 40, and the three-dimensional Euclidean distance is used as the distance measure in the clustering process. Finally, the nearest point to the barycenter is extracted from each cluster as the representative point to form a candidate attachment point set with uniform distribution, and the three-dimensional coordinates of each point and its position index on the tumor surface are saved. The surface boundary is extracted from the tumor three-dimensional grid structure, and the alpha-shape method is used to construct the tumor boundary contour. The shape control parameter a is set to 1.2 to remove the internal recessed area and only keep the external boundary facets. Subsequently, the long axis direction is extracted from the boundary facets. By calculating the covariance matrix of all boundary points and extracting its principal eigenvector, the spatial principal axis direction of the tumor is determined, and 20% of the boundary point set is extracted at both ends of the principal axis direction as the edge point set. Calculate the distance between each candidate attachment point and all points in the edge point set, extract the nearest edge point, and calculate the direction vector of the attachment point pointing to the edge point.Further, the direction vector is calculated with the tumor geometric center direction, and the angle less than 30 degrees is considered to have good boundary centripetal characteristics and is reserved. Finally, not less than 35 attachment points are screened, and each element of the reserved point set records its three-dimensional coordinates, corresponding boundary center point coordinates, direction vector and angle value as attribute information. For the screened attachment point set, three-dimensional Delaunay triangulation is used to construct its topological adjacency structure. In the splitting process, the attachment points are regarded as a point set, and the connection rule is defined as a non-overlapping tetrahedral set formed by all points in three-dimensional space. Each tetrahedron is determined by four attachment points to form an edge connection relationship. Then, all the connected edges are processed by neighborhood merging. The neighborhood of each point is defined as the set of all points connected to it, and further geometric adjustment is performed on each neighborhood. The position of the current point is updated to the weighted average of all points in its neighborhood using Laplace smoothing. The weight is set according to the inverse of the edge connection distance. The smaller the distance, the higher the weight. Three rounds of iteration are performed to ensure geometric continuity. After completion, the final coordinates of each point, the connection edge relationship, and the neighborhood point index are organized into a preliminary fitting structure diagram. The moving least squares surface reconstruction method is used to construct a local neighborhood window with each attachment point in the preliminary fitting structure as the center. The neighborhood radius is set to 5mm. The position coordinates of all points in each neighborhood are collected, and local quadratic surface fitting is performed on them. The fitting weight function is a Gaussian function. The smaller the distance between the point and the center point, the higher the weight. The fitting result is used to determine the local normal direction and local curvature information of the point. All local fitting points are spliced into a continuous curved surface structure. Further, the coordinates, normal vector and spatial curvature of all surface points are calculated to form a point profile dataset, which is stored in a three-dimensional point cloud format. Each point records its position, local fitting center point index, normal vector and fitting residual. On the basis of the point profile, a global interpolation reconstruction operation is performed. Radial basis function interpolation is used, and a thin plate spline function is used as the interpolation kernel. The coordinates and normal vector information of all point profile data points are fitted. The fitting range is limited to the three-dimensional bounding box range of the tumor surface grid. The interpolation density is set to generate a new point every 2mm. The interpolation points uniformly cover the entire profile area. The boundary points are extracted from the generated fitting surface. The boundary is identified by detecting the spatial curvature change amplitude. Any point with a normal direction change greater than 20 degrees within a neighborhood radius of 1mm is marked as a boundary point. All boundary points form a boundary curve in the order of spatial connection. At least 30 points are uniformly sampled in the boundary curve segment as the final attachment point marker. Each marker records its three-dimensional coordinates in the surface, tangent direction, normal direction and position index in the boundary curve. The final output of the profile is a structure array, which is used in subsequent tumor structure attachment simulation and response modeling.

[0026] Preferably, the tumor static feature distillation on the simulated three-dimensional tumor morphology in step S3 comprises: Multi-scale static feature profiling is performed on the three-dimensional simulated tumor morphology data to obtain multi-scale tumor features; Region boundary fusion is performed based on the multi-scale tumor features to generate boundary-fused tumor features; The boundary-fused tumor features are subjected to feature sparse distillation, and static morphology mapping is performed based on the sparse tumor features to obtain tumor static morphology; The tumor static morphology is subjected to contour segmentation to generate a tumor static edge contour.

[0027] In the operation of multi-scale static feature analysis of three-dimensional simulated tumor morphology data in this embodiment, the input data is the simulated three-dimensional tumor STL grid model generated in the previous step. The multi-scale static feature extraction is realized by constructing a voxel analysis structure based on a scale space pyramid. The STL model is converted into a three-dimensional voxel data format, and the voxel size is set to 0.5 mm, 1 mm and 2 mm in sequence. The surface average curvature, principal curvature direction distribution, local normal vector change rate and voxel density distribution of the voxel model at each scale are calculated respectively. A local statistical window based on Gaussian weighting is used to perform weighted filtering processing on the above parameters, and the window size is set to 3x3x3, 5x5x5 and 7x7x7 respectively. The VSRM (Voxel Scale Response Model) based on the voxel space multi-scale response model is used to calculate the static response intensity at different scales, and the composite feature vector set containing the curvature atlas, normal distribution tensor field and density image at different scales is output. All multi-scale tumor feature vectors are uniformly encoded in a four-dimensional tensor structure, and each dimension represents the coordinate position, scale level, feature channel and response intensity. In the processing of regional boundary fusion based on multi-scale tumor features, the four-dimensional feature tensor is first normalized by scale maximum response to generate a scale response atlas. A regional aggregation algorithm based on local boundary saliency is used to fuse the boundary response. The regional fusion convolution module based on heteroscedastic extension is used in the specific algorithm, the convolution kernel size is set to 5x5x5, the step is 1, the weight is initialized using He normal distribution, and the boundary stability criterion is used to control the regional growth direction, and the criterion value is that the boundary point gray gradient and the normal angle are less than 15 degrees. The fused boundary response map is converted into a three-dimensional surface curve graph structure, and then the Laplacian smoothing algorithm is used to smooth the fused boundary. The iteration number is set to 20 times, and the Laplace weight coefficient is set to 0.6. The tumor features after boundary fusion are composed of 32 channel boundary feature maps and scale response maps. Finally, the fusion feature set containing the spatial position, boundary response and scale label is saved in HDF5 format for subsequent distillation processing. In the process of feature sparse distillation of the boundary fused tumor features, a sparse feature distillation network SFDN (Sparse Feature Distillation Network) based on feature redundancy reduction is constructed. The input channel number is set to 64, a three-layer convolution network structure is used, the convolution kernel size is 3x3x3, 1x1x1 and 3x3x3 respectively, the activation function is ReLU, and the output channel is a 16-dimensional sparse feature map. The mean square error loss function is used in the network training stage, and a sparse regularization term is added. The regularization term uses L1 norm penalty, and the regularization coefficient is set to 0.001, The distillation network outputs the spatially sparse tumor feature map after the sparse graph generation is completed, and then the sparse feature map is remapped to the three-dimensional spatial structure through the static morphology mapping module. The static morphology mapping process is realized by density interpolation of sparse feature points according to the spatial position. The interpolation method adopts Trilinear Interpolation (three linear interpolation), and generates a static morphology map containing tumor surface density features, local principal direction and local symmetry. In the contour segmentation operation of the tumor static morphology, the static morphology map is first re-encoded into a three-dimensional tensor graph structure, which is input into the improved 3D U-Net structure. The network contains four layers of encoder and decoder, and uses Skip Connection (skip connection) connection method to ensure feature alignment. The number of convolution channels is 16, 32, 64 and 128 in turn, the convolution kernel is set to 3x3x3, the activation function uses LeakyReLU, the negative half-axis slope is set to 0.01, and the loss function uses the mixed loss of Dice coefficient and cross entropy, in which the Dice weight is set to 0.7. Batch Normalization is used in the training process to prevent gradient disappearance. The training rounds are 100, the data batch size is 8, the image scaling size is 128x128x64, the model outputs the static edge contour Mask graph, and the Marching Cubes algorithm is used to convert the segmentation Mask graph into an isosurface contour grid. The grid is exported in STL format, and the three-dimensional boundary point normal information is attached for the next stage of boundary refinement modeling and tumor contour expansion processing.

[0028] Preferably, defining the initial adhesion area framework according to the simulation adhesion area framework in step S3 comprises: Conducting spatial constraint mapping of the simulation adhesion area framework according to the static edge contour of the tumor to obtain a contour spatial constraint; Conducting boundary optimization of the simulation adhesion area framework based on the contour spatial constraint to generate an optimized adhesion area framework; Conducting local texture correction on the optimized adhesion area framework and re-projecting it into the initial adhesion area framework.

[0029] In the operation of spatial constraint mapping of the simulation adhesion area framework according to the static edge contour of the tumor, the input data is the static edge contour grid of the tumor obtained by three-dimensional segmentation and the adhesion area framework STL model generated by the previous simulation. To construct an accurate spatial constraint, the static edge contour of the tumor is first resampled by three-dimensional point cloud, with a point cloud spacing of 1.5 mm. A fast nearest neighbor index structure is constructed using the KD tree algorithm, and the resampled point cloud is used as the static edge constraint point set. Then, the simulation adhesion area framework STL grid is converted into a point cloud structure. The distance between point pairs and surfaces is used as the error evaluation criterion to construct a constraint energy function. The position of each point of the simulation framework is adjusted by minimizing the function. The edge attraction term, the rigidity preservation term and the local consistency term are included in the constraint energy function. The weight of the edge attraction term is set to 0.6, the weight of the rigidity term is set to 0.3, and the weight of the consistency term is set to 0.1. The gradient descent method is used for iteration 100 times to obtain the displacement mapping vector set of the framework point set. The vector set is applied to the vertex position of the simulation adhesion area framework grid, and the adhesion area framework structure after the contour spatial constraint mapping is finally output. In the process of boundary optimization of the simulation adhesion area framework based on contour spatial constraint, the adhesion area framework grid after contour spatial constraint is first loaded as a vertex-face structure graph, and each boundary vertex is marked as an edge feature point. A local fitting optimization algorithm based on multi-direction boundary curvature evaluation is used for boundary optimization. In the optimization step, the main direction curvature value of each boundary point in the three-dimensional space and the average curvature value of its neighborhood are calculated. If the curvature difference is greater than 0.3 is classified as a mutation point, and local quadratic B-spline fitting is used to replace the neighborhood control point set of the mutation point to generate a new frame structure. The number of control points is limited to 5 to 7. The original topological structure is preserved during the fitting process. The overall structure consistency is maintained through the centroid alignment method. After fitting, the boundary area is reconstructed into a triangular mesh. The Delaunay triangulation method is used to fill the local edge holes to ensure the continuity of the boundary. The optimized frame is smoothed by the normal consistency filter with a filter radius of 3mm. The final output is the optimized simulation adhesion area frame structure. In the operation of local texture correction and re-projection of the optimized adhesion area frame to the initial adhesion area frame, first, the optimized adhesion area frame is registered with the voxel level of the original CT image. The registration method is based on mutual information registration. The reference image is the original CT slice sequence, and the reconstructed image is the voxel volume corresponding to the adhesion area frame. The mutual information window size is set to 16x16x16. After registration, the three-dimensional coordinates of each vertex of the frame are extracted and projected to the original CT image index space. The bicubic interpolation method is used to assign a gray value to each vertex as the local texture response value. Then, the local texture correction model based on color consistency adjustment is used for texture compensation. The model calculates the local gray deviation by comparing the gray value of each vertex with the average gray value of its 6-neighborhood. If the deviation exceeds 8 gray levels, local re-calibration is performed. The re-calibration strategy is to retain the local maximum response gray value and reassign the remaining neighborhood gray values to the linear interpolation between the maximum value and the average value. The interpolation coefficient is 0.5. The corrected frame is re-projected with the gray value of each vertex as the texture coordinate value. The OpenGL graphics rendering pipeline is used for GPU accelerated processing of the re-projection. The re-projected frame is output as the initial adhesion area frame.

[0030] Preferably, the feature comparison based on the preset historical tumor contour image and the tumor static edge contour in step S4 includes: Time series registration is performed on the preset historical tumor contour image to obtain a time series registration historical tumor contour image; The time series registration historical tumor contour image is projected and compared with the tumor static edge contour to generate contour projection comparison data; Gradual morphological tracking is performed based on the contour projection comparison data to obtain morphological tracking data; The morphological change trend in the morphological tracking data is extracted; The local deformation region in the tumor static edge contour is detected according to the morphological change trend; Integrate all local deformation regions and conduct time sequence shape migration modeling to obtain the shape change process of the tumor body.

[0031] In this embodiment, for historical tumor contour images, first, extract the tumor contours at different time points from the CT image dataset. These contour images are usually labeled by medical imaging experts, and the labeling process uses edge detection algorithms such as Canny edge detection to ensure the accuracy of the extracted contours. Then, use multi-modal image registration techniques to time-series register these historical images, using a rigid transformation model based on anatomical structure (such as affine transformation or rigid registration), optimized by mutual information method, to ensure spatial alignment of all historical images. After each registration, save the registered image data and check the registration accuracy to ensure that the geometric shapes of the historical images remain consistent. Project the time-series registered historical tumor contour images against the static edge contour of the tumor. First, extract the static edge contour using the Sobel operator for edge detection to ensure that each edge point is clearly and accurately positioned. Then, perform contour projection comparison between the static edge contour image and each time-series registered historical image using a point-by-point distance calculation method to calculate the Euclidean distance and angular difference between each historical contour and the static edge contour points, ensuring accurate projection of each historical contour against the static contour. These data record the changes of each contour feature point at different time points, and the generated projection comparison data contain the spatial positions of each contour feature point and their corresponding change trends. Based on the contour projection comparison data, use the progressive morphological tracking method to first construct a continuous tracking model based on distance and angle for each contour feature point at each time point. Use curve fitting methods (such as B-spline curves) to model the contour, and fit the contour points in the time series to generate the contour morphology curve at the corresponding time point. Then, compare these contour curves with the contour curve at the previous time point to calculate the quantitative data of morphological changes, including curvature changes, shape expansion or contraction, etc., to ensure that the morphological tracking results can dynamically show the trends of tumor morphology in time evolution. The output data of this process includes contour curve changes, curvature changes, and morphological evolution trajectories at each time point. After obtaining the morphological tracking data, use the principal component analysis (PCA) method to analyze the morphological change data. First, input all the contour curve data at different time points into the PCA algorithm for eigenvalue decomposition to extract the main components of morphological changes, ensuring that the extracted change trends can cover most of the features of morphological changes. By analyzing the contribution rate of the principal components, find the feature dimension that has the greatest impact on the tumor morphology change, and then determine the dominant trend of the morphological change. The specific data includes the rate of morphological change and the direction of the trend.The morphological change trend of the extracted morphology will reflect the morphological change rule of the tumor in the whole time period. According to the data of the morphological change trend, the local deformation area in the static edge contour image is detected by using image segmentation and region growing algorithm. First, the significant deformation area is determined by setting a threshold (for example, the area with a change amplitude greater than a certain critical value). Then, the region growing algorithm is used to expand the surrounding area according to the change direction and rate of the contour, and finally the contour image of the local deformation area is formed. This process identifies and demarcates the range of the local deformation area by calculating the similarity (such as shape difference, edge curvature, etc.) between the contour of each local area and other areas, and generates the changed area in the static edge contour image. After integrating the data of all local deformation areas, a hidden Markov model (HMM) is used to model the morphological migration. First, the state transition matrix of each region is established according to the change rule of the local deformation area in the time series. By analyzing the dynamic characteristics of these changes, the trend of the tumor morphological change is inferred. Then, according to the morphological features of each local deformation area, a time series morphological migration model is constructed, and the morphological change prediction data at each time point is output. By modeling the historical data and combining the existing morphological change trend, the complete tumor morphological change process data is generated, ensuring that the process can fully describe the time evolution of the tumor morphology.

[0032] Preferably, the erosion degree evaluation of the initial attachment area framework in step S4 according to the preset historical tumor attachment framework image includes: Obtaining a historical tumor attachment framework image; Extracting a historical attachment framework structure in the historical tumor attachment framework image; Locking the bony anchor point of the historical attachment framework structure to obtain an anchor point-locked historical attachment framework; Constructing a skeleton vector field according to the anchor point-locked historical attachment framework; Performing framework bending tensor deduction on the anchor point-locked historical attachment framework based on the skeleton vector field, to obtain tensor deduction framework evolution data; Performing normal growth curve prediction through the tensor deduction framework evolution data, and inferring an ideal evolution attachment area framework based on the predicted normal growth curve; Performing framework projection comparison on the ideal evolution attachment area framework and the initial attachment area framework, and screening out a difference area framework; Evaluating the erosion degree of the initial attachment area framework according to the difference area framework, to generate an attachment framework erosion degree.

[0033] In this embodiment, CT image data of a specific patient is extracted from a medical image library. These data are obtained by scanning and labeled by an imaging expert to determine the attachment area of the historical tumor. Multi-modal image registration techniques are used to integrate multi-period CT images to ensure that images at different time points can be spatially aligned. The registration method used is rigid registration and affine registration. In the registration process, an optimization algorithm (such as mutual information method) is used to minimize the spatial error between images. The historical tumor attachment framework image is obtained. After obtaining the historical tumor attachment framework image, edge detection algorithm is used to extract the attachment area in the image. Sobel operator is used to extract the edge of the attachment framework to ensure that the tumor attachment area in the image is clear and visible. Morphological operations (such as dilation and erosion) are used to further optimize the edge extraction effect, so as to obtain the geometric structure of the historical attachment framework. In the extraction process, the threshold value of edge intensity needs to be set to ensure that only the true attachment structure is retained, but not the noise or pseudo-edge. The extracted attachment framework structure is analyzed, and an anchor point recognition algorithm based on morphological features is used. First, the feature of the local region of the attachment framework structure is extracted to identify the bone anchor point. The determination of the anchor point depends on the structural features of the attachment framework image, such as regions with large curvature values or irregular shapes, which usually represent the attachment points. Adaptive threshold method is used to determine the accurate position of the anchor point. The obtained anchor point position data will be used as the basis for subsequent framework analysis. After locking these anchor points, a historical attachment framework image with anchor point markers is generated. After obtaining the locked historical attachment framework image, the skeleton vector field is constructed. First, the local region of the attachment framework is segmented, and the Voronoi diagram segmentation method is used to grid the attachment region. The center of each grid is regarded as the bone branch point of the skeleton. Then, the gradient vector field method is used to calculate the vector field of each grid unit. The obtained skeleton vector field reflects the mechanical changes in the attachment region. Each point of the vector field represents the direction and strength of the skeleton structure at that point. Using the skeleton vector field, the tensor deduction method is used to deduce the framework bending of the anchor-locked historical attachment framework. First, the bending tensor model of the attachment region is established, and the bending tensor is defined as a mathematical expression describing the local region deformation. In the deduction process, according to the local direction and intensity change of the skeleton vector field, the bending degree of each skeleton point is calculated step by step. The calculation formula of the bending tensor is: tensor = (vector difference / local distance) x local deformation. The local deformation is derived by this formula. The obtained tensor deduction data will present the morphological change process of the attachment framework, and provide a basis for the subsequent growth curve prediction.The framework evolution data obtained from tensor deduction is used for prediction of the normal growth curve. First, the stable change trend in the framework evolution process is extracted, and the evolution data is smoothed using a time series analysis-based method (such as weighted average method) to obtain the normal growth curve of the attachment area. The normal growth curve reflects the expected change rule of the attachment area framework under normal growth conditions. To ensure the accuracy of the prediction, regression analysis is performed using data from multiple time points, and the growth trend of the attachment area is fitted by the least squares method (Least Squares). Finally, the normal growth curve is obtained. According to the normal growth curve, the ideal evolution of the attachment area framework is inferred. First, compare the normal growth curve with the actual historical data to ensure that the prediction result conforms to the historical growth rule. Then, optimize the ideal evolution of the attachment area framework using a numerical optimization algorithm (such as gradient descent method) to ensure that it is consistent with the morphology of the historical attachment area framework. Finally, an ideal evolution framework is derived, which represents the most ideal morphology of the attachment area under the condition of no external interference. Compare the ideal evolution of the attachment area framework with the initial attachment area framework by framework projection. First, map the two frameworks to the same two-dimensional plane for comparison. Calculate the morphological difference between the two frameworks using the least squares method to ensure that the geometric changes such as rotation, scaling and translation are considered in the comparison process. The comparison result outputs the difference of each region in the two frameworks, forming the difference region framework data. According to the difference region framework selected in the framework projection comparison, calculate the erosion degree of the initial attachment area framework. Use morphological analysis methods (such as erosion and dilation operations) in image processing to quantify the difference region and evaluate the magnitude of its morphological change. Calculate the area ratio of the difference region as an indicator of the erosion degree to obtain the erosion index of each region. Finally, generate an attachment framework erosion degree data reflecting the erosion degree of the initial attachment area framework.

[0034] Especially important is that the framework bending tensor deduction of the anchor-locked historical attachment framework based on the skeleton vector field includes: Perform direction consistency normalization on the skeleton vector field to obtain a normalized skeleton vector field; Perform structure node layering on the anchor-locked historical attachment framework to generate layered structure nodes; Perform node vector coupling on the normalized skeleton vector field and the layered structure nodes to obtain a coupled node vector field; Perform multi-scale tensor mapping according to the coupled node vector field to generate multi-scale tensor mapping data; Perform bending response deduction based on the multi-scale tensor mapping data to obtain a bending response tensor; Integrate the time sequence structure evolution through the bending response tensor to obtain the tensor deduction framework evolution data.

[0035] In this embodiment, when processing the skeleton vector field generated after the segmentation of the osteosarcoma CT image, first, the direction vectors of all skeleton points are subjected to a normalization operation, and the direction vectors of each skeleton point are normalized by the Euclidean norm by using a unit vector standardization processing method, that is, the vector v=(x,y,z) is divided by its own norm √(x 2 +y 2 +z 2) operation, ensure that the length of each vector is uniform for 1, in order to ensure the consistency of the direction, the main direction of the method of discrimination, calculation of each skeleton point around the radius of 2 mm within the neighborhood vector average direction as the main direction, using the dot product (dot product) to determine whether the current vector and the main direction of the same direction, if the dot product is less than zero, the current vector will be negative, so that it is reversed to the main direction direction, after the above processing, all the skeleton point vector has a consistent direction and uniform scale, form the normalized skeleton vector field, when the anchor point locking history of the analysis of the framework structure, first of all, according to the position of the anchor point in the framework of global connection path scanning, using the traversal method based on the construction of the graph structure to extract all the connection path from the anchor point, and according to the structure of the topological depth of the layer operation, using breadth first search (Breadth-First Search) algorithm in the three-dimensional structure of the path extraction, and in each layer record and anchor point distance in the Euclidean distance range 1.0 mm to 5.0 mm between the nodes, the construction of the first layer structure node, then continue to expand the non overlapping basis to obtain more outer layer node, in turn to build the second layer, the third layer, until cover the whole attachment framework, the result of the structure of the hierarchical index way of recording, and mark each layer node of the spatial coordinates, connection edge information and local curvature value, all nodes in the form of three-dimensional structure tree (3D structural tree) expression, the normalized skeleton vector field and the hierarchical structure of the coupling operation of the nodes of the attachment framework, first of all, according to the three-dimensional coordinate position of each structure node, in the skeleton vector field search radius of 1 mm neighborhood vector set, if there are multiple skeleton vector in the range, then calculate the distance weight between the node and each vector point, the closer the distance, the greater the weight, the weight calculation formula is weight = 1 / (1+d), wherein d is the Euclidean distance, through the weighted average way of all matching vector superposition to form the coupling vector of the structure node, if no vector point is matched, the structure node does not participate in the coupling generation, through this way, each structure node and its spatial distribution of the skeleton vector field data coupling, get a structure node based on the coupling node vector field, the data format of the coupling field is each node contains position triplet (x, y, z), hierarchical level, coupling vector value (x', y',z') and weight values, the coupling node vector field data is input into a tensor generation module, first, the node set of different hierarchical levels is grouped, a local tensor representation model is constructed according to the distribution of the node set of each layer, each node is regarded as a vector source point, and the coupling vector is used for constructing the main axis direction of the tensor. The strain characteristics in the region are represented in the form of a symmetric second-order tensor. The tensor calculation adopts a covariance matrix construction method. The covariance matrix is constructed for the coupling vector set of all nodes in each layer, and the eigenvalues and eigenvectors are calculated. The eigenvectors are used to construct the main axis direction of the tensor. The eigenvalues represent the extension of the tensor in the corresponding direction. The operation is repeated to generate the tensor for all hierarchical nodes. Finally, a tensor set composed of multiple scale layers is obtained. Each tensor set records the corresponding hierarchical level, tensor matrix, main axis direction and structure position. All data are combined to generate a multi-scale tensor mapping data structure. The structure is organized by three-dimensional space index. Each voxel corresponds to one or more tensor data entries. The multi-scale tensor mapping data is input into a calculation module based on structural response simulation. The response stress field simulation is performed for each group of tensor data. The response gradient of each main axis direction in the tensor field is calculated by the structural finite difference method. The initial bending perturbation vector is set to act on the anchor point level. The bending strain is iteratively calculated at each structure layer node position along the tensor main axis direction. The response value of each dimension in the tensor field is calculated by the partial derivative. The response value is recorded in the form of a three-dimensional curvature tensor. The tensor value reflects the curvature bending trend of the structure in the corresponding space to the initial perturbation. The response tensors generated by all structure levels are combined into a complete three-dimensional response tensor atlas. The data format is position coordinates plus three-dimensional curvature response value (x_r, y_r, z_r). The response tensor is fused with the historical evolution data in the time dimension. First, the attached framework structure at the corresponding time point in the multi-period CT image is called. The attached structures at different time points are superimposed into a unified three-dimensional space coordinate by image registration method. The three-dimensional curvature response value of each position in the response tensor is used as an evolution vector. The response trend is continuously fitted by the time interpolation method. The response trend of each anchor point region is converted into a time sequence change trajectory. The evolution trajectory is fitted by using a B-spline curve to generate a continuous evolution path. Each path records the structure position, time step and corresponding tensor response value change. Finally, the evolution paths of all regions are combined to generate the tensor deduction framework evolution data. The data structure is a four-dimensional array, which includes spatial three-dimensional coordinates and time dimension t, and records the structure change state and response tensor at the corresponding time point. It is used as an important data basis for structure evolution modeling for subsequent structure prediction and lesion analysis tasks.

[0036] Preferably, the change region screening according to the change process of the tumor body morphology in step S5 comprises: fitting the tumor morphology change process as morphology trajectory aggregated data; performing multi-dimensional morphology response stratification according to the morphology trajectory aggregated data to obtain stratified morphology response data; performing local window segmentation on the reconstructed simulated three-dimensional tumor morphology based on the stratified morphology response data to generate a local window segmentation region; performing candidate region label assignment based on the local window segmentation region to obtain a candidate tumor lesion region.

[0037] In this embodiment, the images of each time period are reconstructed in three dimensions, the contours and morphological information of the tumor are extracted by using edge detection algorithms for osteosarcoma such as Canny edge detection, and then the edges are optimized through morphological operations to eliminate noise and irregular details, obtaining clear tumor contours. Then the morphological changes of the images at multiple time points are fitted, the morphological features of the tumor at each time point are converted into vector data, and finally these vector data are combined into morphological trajectory aggregation data through weighted averaging method. The morphological trajectory aggregation data represents the change trend and spatial distribution of the tumor morphology at different time points, and the data can be represented as a series of coordinate points and deformation metrics, ensuring that the dynamic changes of the tumor over time can be fully reflected. Based on the morphological trajectory aggregation data, a multi-dimensional data analysis method is used to analyze the changes in tumor morphology. First, the PCA (Principal Component Analysis) method is used to reduce the dimension of the morphological trajectory data, and the main components affecting the morphological changes of the tumor are extracted. Then, the K-means clustering algorithm is used to divide different morphological changes into multiple levels, forming multiple morphological response levels, ensuring that each level reflects the characteristic interval of the morphological changes of the tumor. The data of each level represents the characteristic information of the morphological response of the tumor at this level. These hierarchical morphological response data provide a strong basis for subsequent regional screening, and the hierarchical data can be represented by the mean and standard deviation of each cluster center, helping further morphological analysis and lesion region positioning. Using hierarchical morphological response data, the three-dimensional tumor morphology image is cut into local windows. First, according to the different response regions in the hierarchical data, the three-dimensional tumor morphology image is divided into multiple small regions by algorithm, and 3D image cutting algorithm (such as Octree decomposition) is used for uniform segmentation, dividing the morphology of the tumor into different small windows. Each small window represents a part of the tumor, ensuring that each region contains relatively uniform morphological information. These local window cutting regions reflect the spatial distribution and morphological characteristics of the tumor. The size of the cutting region is usually set to 5x5x5mm, ensuring that high-precision local information can be obtained at different resolutions. The generated local window segmentation region provides accurate spatial cutting basis for subsequent lesion region screening.The candidate region label assignment is performed on the local window pane segmentation area obtained by cutting. First, the morphological characteristics of each local window pane area are calculated, such as area, perimeter, volume and other geometric indicators. Whether the area meets the characteristics of the lesion is determined through these indicators. A support vector machine (SVM) is used for classification. The support vector machine determines whether each window pane is a lesion area by using the classifier obtained by training. The training process uses labeled lesion area data for supervised learning. Each local window pane is assigned a label. The label value is 0 (normal area) or 1 (lesion area). The final candidate tumor lesion area is the set of all window panes labeled as lesion areas. These candidate areas will be further analyzed in subsequent steps. The threshold setting in the labeling process is adjusted according to the actual labeling of historical data to ensure high classification accuracy.

[0038] Preferably, the positioning of the osteosarcoma lesion area in the multi-scale osteosarcoma CT image based on the candidate tumor lesion area and the attachment frame erosion degree in step S5 comprises: Abnormality evaluation of the attachment frame erosion degree according to the preset erosion abnormality assignment, thereby obtaining an erosion abnormality score; Screening of the abnormal frame erosion area according to the erosion abnormality score; Extraction of the three-dimensional spatial positioning of the abnormal frame erosion area and the candidate tumor lesion area; Projection of the abnormal frame erosion area and the candidate tumor lesion area onto the same coordinate system based on the three-dimensional spatial positioning, thereby obtaining abnormal erosion area coordinates and candidate lesion area coordinates; Determination of the overlapping coordinates in the abnormal erosion area coordinates and the candidate lesion area coordinates; Positioning of the same coordinate area in the multi-scale osteosarcoma CT image based on the overlapping coordinates; Definition of the same coordinate area as the osteosarcoma lesion area.

[0039] In this embodiment, the tumor attachment framework and its corresponding erosion data in the historical data set are used for calibration. A standard erosion score based on osteosarcoma morphological changes is set. Multidimensional scaling analysis method is used to model the morphological characteristics of the attachment framework. Feature extraction algorithm is used to quantitatively analyze the morphological changes of the tumor. Eigenvalue calculation is performed according to the boundary morphology of the erosion area. Image gradient calculation method is used to obtain the gradient distribution of the erosion degree. Finally, the framework is assigned an abnormal value according to the set erosion score standard. The abnormal area is screened out by setting a threshold. The abnormal evaluation result obtained is compared with the actual data of the erosion situation. The erosion abnormal score of each region is obtained. The region with a score higher than the threshold is considered as an abnormal region. The score is quantified according to the morphological changes in the osteosarcoma CT image and the difference with the normal region. Finally, the erosion abnormal score is obtained. According to the erosion abnormal score, the abnormal framework erosion area is screened out by setting a preset threshold. First, the hierarchical screening method is used to sort all regions. Based on the erosion abnormal score, the regions with scores exceeding the threshold are screened out. The threshold is set as the boundary point of the region with a score exceeding 95%. This threshold is determined after a large amount of experimental data is verified. The abnormal region obtained contains the part of the tumor attachment framework with more serious erosion. Through the way of layer-by-layer screening, all regions with high scores are extracted from the original image. Image segmentation algorithm (such as region growing method) is used to further extract the complete erosion area, so as to ensure that the abnormal area screened out can reflect the true situation of the lesion. All screened regions are labeled. The abnormal framework erosion area and the candidate tumor lesion area are spatially positioned by three-dimensional reconstruction algorithm. First, CT image data is used for reconstruction. Surface reconstruction based method (such as Marching Cubes algorithm) is used to reconstruct each slice in the CT image into a three-dimensional model. Then, the three-dimensional coordinates of the abnormal area and the candidate lesion area are extracted to obtain the spatial positioning information of each region. In this process, the centroid coordinates of each region are calculated to ensure the accuracy of the region positioning. The three-dimensional boundary of these regions is further extracted to form an accurate spatial description. These spatial coordinates are used for subsequent region projection and comparison processing. The extracted spatial positioning data is in millimeters to ensure that the accuracy of each region positioning can meet the needs of subsequent calculation and comparison.The three-dimensional spatial coordinates of the abnormal framework erosion area and the candidate tumor lesion area are uniformly mapped. First, a uniform coordinate system is set, and the three-dimensional coordinates of the two areas are aligned using the Least Squares Method. The mapping relationship is adjusted by calculating the minimum error of the coordinates to ensure that the projected coordinates are as close as possible to the actual position, ensuring that the mapping of the two types of areas to the same three-dimensional space will not be distorted. The accurate positions of the abnormal erosion area coordinates and the candidate lesion area coordinates in the uniform coordinate system are obtained. This projection process is crucial to accurately express the spatial relationship between the regions. Each coordinate point in the projection result will be used for the next coordinate coincidence analysis. The abnormal erosion area coordinates and the candidate lesion area coordinates projected onto the same coordinate system are matched. First, the Centroid Method is used to compare the coordinates of the two areas, and the Euclidean Distance between each coordinate point is calculated. By setting a distance threshold, coordinate points with a distance less than the threshold are selected as coincident coordinates. The threshold is usually set to 1-2 mm to ensure the accuracy of the coincident coordinates. During the screening process, spatial clustering analysis is also performed to further optimize the matching results and ensure the accuracy of the coincident coordinate region. Finally, through this analysis step, the coincident coordinate point set of the abnormal erosion area and the candidate lesion area is obtained. Using the coincident coordinate point set, based on the aforementioned spatial positioning algorithm, these coordinate points are mapped to the multi-scale image hierarchy of the CT image. First, through Multiresolution Analysis, the coincident coordinates are searched at different scale levels to ensure that each coordinate point can be accurately positioned at different image resolutions. Then, Image Registration is used to accurately align the images at each scale, and the coincident coordinate region is located on the aligned image. Finally, the corresponding regions of the coincident coordinates at different scales are marked as the same coordinate region. This process further improves the positioning accuracy through multi-layer image fusion technology to ensure the accurate extraction of the same coordinate region. The same coordinate region obtained through the above steps is defined as the osteosarcoma lesion area. In this process, the rationality of the defined region is further verified through the morphological features, size, and differences with normal regions of the lesion area. Morphological Closing is used to smooth the boundary to ensure that the lesion area boundary is clear and distinct from the surrounding tissue. Finally, these same coordinate regions are marked as osteosarcoma lesion areas, and according to the clinical standard, the spatially accurate lesion area data is generated.

[0040] Thus, the embodiments should be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the description given above, and all changes which come within the meaning and range of equivalency of the claims are intended to be embraced therein.

[0041] The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and changes will readily occur to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-scale osteosarcoma CT image lesion region detection method based on feature distillation, characterized by, The method comprises the following steps: Step S1: Collecting multi-scale osteosarcoma CT images and osteosarcoma attachment area images; performing pseudo-contrast difference reduction on each angle of the multi-scale osteosarcoma CT images to generate uniform osteosarcoma CT images; and reconstructing a simulated three-dimensional tumor morphology based on the uniform osteosarcoma CT images; Step S2: Extracting bone attachment structures in the osteosarcoma attachment area images, and performing repositioning registration based on the bone attachment structures and the reconstructed simulated three-dimensional tumor morphology to obtain tumor repositioning data; and performing attachment framework simulation based on the tumor repositioning data and the osteosarcoma attachment area images to generate a simulated attachment area framework; Step S3: Performing tumor static feature distillation on the simulated three-dimensional tumor morphology to obtain a tumor static edge contour; and defining an initial attachment area framework according to the simulated attachment area framework; Step S4: Comparing features based on preset historical tumor contour images and the tumor static edge contour to extract a tumor morphology change process; and performing erosion degree evaluation on the initial attachment area framework according to preset historical tumor attachment framework images to generate an attachment framework erosion degree; Step S5: Screening a change area according to the tumor morphology change process to obtain a candidate tumor lesion area; Based on the candidate tumor lesion area and the attachment framework erosion degree, an osteosarcoma lesion area in the multi-scale osteosarcoma CT images is located.

2. The multi-scale bone sarcoma CT image lesion region detection method based on feature distillation according to claim 1, characterized in that, Step S1 comprises the following steps: Step S11: Collecting multi-scale osteosarcoma CT images and osteosarcoma attachment area images; and performing pixel-level normalization processing on the multi-scale osteosarcoma CT images to generate normalized osteosarcoma CT images; Step S12: Reconstructing the normalized osteosarcoma CT images at multiple angles, wherein the angle step interval is 5°, and the total rotation angle is not less than 180°, to obtain reconstructed osteosarcoma CT images; Step S13: Performing pseudo-contrast uniform processing on the reconstructed osteosarcoma CT images, and weakening heterogeneous shadows to generate uniform osteosarcoma CT images, wherein the shadow weakening processing window size is limited to 7x7 pixels; Step S14: Performing spatial layer fusion on the uniform osteosarcoma CT images to obtain spatially fused osteosarcoma CT images; Step S15: Performing three-dimensional morphology simulation reconstruction based on the spatially fused osteosarcoma CT images to generate a three-dimensional simulated tumor morphology.

3. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The extraction of the bone attachment structures in the osteosarcoma attachment area images and the repositioning registration based on the bone attachment structures and the reconstructed simulated three-dimensional tumor morphology in step S2 comprise: Extracting bone attachment structures in the osteosarcoma attachment area images; Performing multi-modal spatial registration on the bone attachment structures and the reconstructed simulated three-dimensional tumor morphology to obtain registered bone attachment structure data; Projecting the registered bone attachment structure data into the osteosarcoma attachment area images to generate osteosarcoma projection positioning; Performing bone tumor interface repositioning on the reconstructed simulated three-dimensional tumor morphology based on the osteosarcoma projection positioning to obtain tumor repositioning data.

4. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The attachment framework simulation based on the tumor repositioning data and the osteosarcoma attachment area images in step S2 comprises: Performing simulated attachment point calibration on the tumor repositioning data, wherein the number of calibration points is not less than 30 to obtain an attachment point calibration contour; According to the attachment point calibration contour and the reconstructed simulation three-dimensional tumor morphology, a structure fusion simulation is performed, and based on the simulated fusion structure, a structure fusion attachment area image in the osteosarcoma attachment area image is segmented; A frame boundary model is established through the structure fusion attachment area image, wherein the grid length of the model is limited to 2-4 mm, and an attachment frame boundary is generated; Based on the attachment frame boundary, a spatial skeleton simulation is performed, so as to obtain a simulated attachment area frame.

5. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The tumor static feature distillation of the simulated three-dimensional tumor morphology in step S3 comprises: Multi-scale static feature analysis is performed on the three-dimensional simulated tumor morphology data, so as to obtain multi-scale tumor features; Based on the multi-scale tumor features, a region boundary fusion is performed, so as to generate a boundary fusion tumor feature; The boundary fusion tumor feature is subjected to feature sparse distillation, and based on the sparse tumor feature, a static morphology is mapped, so as to obtain a tumor static morphology; The tumor static morphology is subjected to contour segmentation, so as to generate a tumor static edge contour.

6. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The initial attachment area frame defined according to the simulated attachment area frame in step S3 comprises: According to the tumor static edge contour, a spatial constraint of the simulated attachment area frame is obtained; Based on the contour spatial constraint, a boundary optimization of the simulated attachment area frame is performed, so as to generate an optimized attachment area frame; The optimized attachment area frame is subjected to local texture correction, and is re-projected into an initial attachment area frame.

7. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The feature comparison based on the preset historical tumor contour image and the tumor static edge contour in step S4 comprises: The preset historical tumor contour image is subjected to time sequence registration, so as to obtain a time sequence registered historical tumor contour image; The time sequence registered historical tumor contour image is projected and compared with the tumor static edge contour, so as to generate contour projection comparison data; Based on the contour projection comparison data, a gradual morphology tracking is performed, so as to obtain morphology tracking data; A morphology change trend in the morphology tracking data is extracted; According to the morphology change trend, a local deformation region in the tumor static edge contour is detected; All local deformation regions are integrated, and a time sequence morphology migration model is established, so as to obtain a tumor morphology change process.

8. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The erosion degree evaluation of the initial attachment area frame according to the preset historical tumor attachment frame image in step S4 comprises: The historical tumor attachment frame image is obtained; A historical attachment frame structure in the historical tumor attachment frame image is extracted; An anchor point of the historical attachment frame structure is locked, so as to obtain an anchor point locked historical attachment frame; A skeleton vector field is constructed according to the anchor point locked historical attachment frame; Based on the skeleton vector field, a frame bending tensor of the anchor point locked historical attachment frame is deduced, so as to obtain tensor deduced frame evolution data; Through the tensor deduced frame evolution data, a normal growth curve is predicted, and based on the predicted normal growth curve, an ideal evolved attachment area frame is inferred; The ideal evolved attachment area frame and the initial attachment area frame are subjected to frame projection comparison, so as to screen out a difference area frame; According to the difference area frame, an erosion degree of the initial attachment area frame is evaluated, so as to generate an attachment frame erosion degree.

9. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The change area screening according to the tumor morphology change process in step S5 comprises: Fitting the tumor morphology change process as morphology trajectory aggregation data; According to the morphology trajectory aggregation data, multi-dimensional morphology response layering is carried out to obtain layered morphology response data; Based on the layered morphology response data, local window segmentation is carried out on the reconstructed simulation three-dimensional tumor morphology to generate local window segmentation region; Based on the local window segmentation region, candidate region label assignment is carried out to obtain the candidate tumor lesion region.

10. The feature distillation-based multi-scale osteosarcoma CT image lesion region detection method according to claim 1, characterized in that, The step S5 based on the candidate tumor lesion region and the attachment framework erosion degree to locate the osteosarcoma lesion region in the multi-scale osteosarcoma CT image includes: According to the preset erosion anomaly assignment, the abnormality of the attachment framework erosion degree is evaluated, so as to obtain the erosion anomaly score; According to the erosion anomaly score, the abnormal framework erosion region is screened out; Extracting the three-dimensional spatial positioning of the abnormal framework erosion region and the candidate tumor lesion region; Based on the three-dimensional spatial positioning, the abnormal framework erosion region and the candidate tumor lesion region are projected onto the same coordinate system to obtain the abnormal erosion region coordinates and the candidate lesion region coordinates; Determine the overlapping coordinates in the abnormal erosion region coordinates and the candidate lesion region coordinates; Based on the overlapping coordinates, the same coordinate region in the multi-scale osteosarcoma CT image is located; The same coordinate region is defined as the osteosarcoma lesion region.

Images