Skull data comparison method and system based on artificial intelligence

Abstract

The invention discloses a skull data comparison method based on artificial intelligence. The skull data comparison method comprises the following steps: S1, acquiring three-dimensional scanning data of a skull to be detected and a reference skull database; s2, preprocessing the skull data to be detected; s3, multi-dimensional features of the skull to be detected are extracted through the deep learning model; s4, performing similarity comparison on the skull features to be detected and skull features in the reference database to generate a matching degree score; s5, outputting identity information or a morphological analysis result of the target reference skull according to the matching degree score; the invention further provides a skull data comparison system based on artificial intelligence. The skull data comparison system comprises a data acquisition module, a preprocessing module, a feature extraction module, an intelligent comparison module and a visual output module. Compared with the prior art, the method has the advantages that the comparison accuracy is high; the processing speed is high; the application range is wide.

Description

Technical Field

[0001] The present invention relates to, specifically, a method and system for comparing skull data based on artificial intelligence. Background Art

[0002] In the fields of forensic identification, anthropological research, and medical diagnosis, the comparative analysis of skull data plays a crucial role. Traditional skull comparison methods mainly rely on manual observation and comparison, which is not only time-consuming and laborious but also easily affected by subjective factors, making it difficult to guarantee the accuracy and reliability of the comparison results.

[0003] With the continuous development of medical imaging technology, more and more skull data are presented in the form of three-dimensional images, which provides the possibility for the automatic comparison of skull data. At present, some skull comparison methods based on computer vision and machine learning have been proposed. For example, the patent with the publication number CN109816705A provides a skull comparison method based on feature point matching. This method realizes the rapid comparison of skull data by extracting and matching the feature points on the skull surface. However, there are still some deficiencies in the feature extraction and matching processes of this method, such as insufficient comprehensive feature extraction and low matching accuracy.

[0004] In addition, existing skull comparison methods often have difficulty in balancing the accuracy and efficiency of comparison when dealing with complex skull data. Especially when dealing with skull data containing a large amount of detailed and texture information, existing methods often fail to extract enough effective features, resulting in a decrease in the accuracy of the comparison results.

[0005] Therefore, how to provide a method and system that can efficiently and accurately compare skull data has become an urgent technical problem to be solved. Summary of the Invention

[0006] In view of the problems and deficiencies in the prior art, the present invention proposes a method and system for comparing skull data based on artificial intelligence. This method and system extract multi-dimensional features of the skull through a deep learning model to realize the automatic comparative analysis of skull data, and have the advantages of high comparison accuracy and fast processing speed.

[0007] To solve the above technical problems, the technical solution provided by the present invention is: A method for comparing skull data based on artificial intelligence, comprising the following steps:

[0008] S1. Obtain the three-dimensional scan data of the skull to be measured and the reference skull database;

[0009] S2. Preprocess the skull data to be measured;

[0010] S3. Extract multi-dimensional features of the skull to be measured through a deep learning model;

[0011] S4. Compare the skull features to be measured with the skull features in the reference database to generate a matching score;

[0012] S5. Output the identity information or morphological analysis result of the target reference skull according to the matching score.

[0013] Preferably, S1 further includes the following steps:

[0014] S1.1. Input the original CT / MRI / DICOM data of the skull to be measured, reconstruct the medical image sequence into a three-dimensional voxel model, extract the isosurface through the Marching Cubes algorithm, then denoise using the statistical outlier removal algorithm, and finally perform uniform sampling based on the voxel grid.

[0015] S1.2. Integrate multi-source data, extract multi-scale features using a 3D sparse convolutional network, then enhance local features based on the curvature descriptor of anatomical landmark points, and use a KD-Tree for fast nearest neighbor search to build an index.

[0016] S1.3. Calculate the data similarity to fuse multi-feature dynamics, use RANSAC to screen the initial matching pairs, and use the LM algorithm to optimize the non-rigid deformation.

[0017] Preferably, S2 further includes the following steps:

[0018] S2.1. Input the original skull point cloud data for point cloud statistical outlier filtering, and then input the three-dimensional voxel data for anisotropic diffusion filtering of the voxel data.

[0019] S2.2. Normalize the grayscale of the CT data, and then input the voxel data at any resolution to unify the resolution.

[0020] S2.3. Input the point cloud of the skull to be registered and the template skull data for rough registration, and then perform fine registration through the ICP algorithm.

[0021] Preferably, S3 further includes the following steps:

[0022] S3.1. Input the preprocessed skull voxel data, extract global geometric features through a multi-layer 3D convolution kernel, and construct a feature pyramid model.

[0023] S3.2. The first-layer regression network predicts the probability distribution heat map of the anatomical landmark points, crops local blocks in the peak region of the heat map, and inputs them into the fully connected layer to regress the accurate coordinates.

[0024] S3.3. Construct a skull graph structure, aggregate the neighborhood node features through the graph convolutional layer, and update the node embeddings.

[0025] S3.4. Project the 3D skull onto the sagittal plane, coronal plane, and horizontal plane to obtain 2D images, retain the surface texture information, and use the pre-trained ResNet-50 to extract the texture features of the projected images;

[0026] S3.5. Dynamically weight and splice the geometric features output by the 3D CNN, the topological features output by the GNN, and the texture features through the attention mechanism to achieve multi-modal feature fusion.

[0027] Preferably, S4 further includes the following steps:

[0028] S4.1. Capture the spatial structure information of the skull through the 3D convolutional layer and extract local geometric features;

[0029] S4.2. Fuse the extracted geometric features with the depth features output by the CNN to generate a high-dimensional feature vector, and use an autoencoder to reduce the dimension of the high-dimensional features to generate a low-dimensional embedding vector;

[0030] S4.3. Calculate the similarity between the feature vectors of the skull to be measured and the feature vectors in the reference library, screen the candidate skulls according to the similarity threshold, and retain the most similar samples for the fine matching stage;

[0031] S4.4. Use the ICP algorithm to perform rigid transformation alignment between the skull to be measured and the candidate skulls, introduce geometric feature constraints to accelerate the search, and eliminate the mis-matched point pairs;

[0032] S4.5. Combine the rough matching similarity and the fine registration error to generate the final matching score, sort the skulls in the reference library in descending order of the score, and output the Top-N matching results and the visualized comparison graph.

[0033] Preferably, S5 further includes the following steps:

[0034] S5.1. Use the Frankfurt plane to unify the skull coordinate system, eliminate the position and pose differences, and then scale the skull model to a unified scale;

[0035] S5.2. Use the K-means algorithm to cluster the concave and convex vertices into feature regions, and screen the significant regions based on connectivity;

[0036] S5.3. Calculate the principal component vectors of the feature regions, screen the similar regions, then perform an exhaustive match on the similar regions, calculate the optimal rigid transformation and translation vectors, and preliminarily align the target and the reference skull;

[0037] S5.4. Introduce a dynamic iteration factor to improve the ICP algorithm to achieve the fine registration of the target and the reference skull;

[0038] S5.5. Calculate the Euclidean distance between vertices or the deviation along the normal direction, use the correlation coefficient to evaluate the similarity after multi-view feature fusion, check the morphological criteria, and make a judgment in combination with the threshold;

[0039] Use S5.6, weighted geometric deviation, correlation coefficient, and morphological consistency score for identity determination.

[0040] Preferably, an artificial intelligence-based skull data comparison system is also provided, including:

[0041] The data acquisition module obtains skull data from multi-source medical images, uses a multi-angle shooting device to collect data, and obtains high-resolution three-dimensional data through a CT / MRI scanner; supports DICOM and NIFTI format input.

[0042] The preprocessing module is used to standardize the data format, remove noise and irrelevant tissues, including functions such as contrast enhancement and data normalization.

[0043] The feature extraction module extracts key anatomical features from the preprocessed data, including key points, textures, and shapes, uses ResNet-152 or Residual U-Net to extract high-level semantic features, combines an interpretable CNN to generate activation maps, and locates anatomical landmark points, including functions such as extracting GLCM textures and DWT wavelet features for auxiliary classification.

[0044] The intelligent comparison module realizes skull identity recognition or pathological detection through feature matching, uses YOLOv3 or FasterR-CNN to locate the fracture area, enhances the inter-class discrimination based on SVM or contrast loss function, and quantifies the degree of feature matching through cosine similarity or Dice coefficient.

[0045] The visualization output module displays the comparison results through an interactive interface, uses edge detection and texturing techniques to generate a three-dimensional model of the skull, and highlights the key difference areas through global activation maps.

[0046] Preferably, the end-to-end process of the system includes: (1) The original data is preprocessed and then input into the feature extraction network to output a high-dimensional feature vector; (2) The comparison module determines the identity / pathological status through similarity calculation or classification model; (3) The visualization module renders the result as a 3D model or heat map.

[0047] The system architecture includes: (1) Data layer: Stores the original images and preprocessed data, supports distributed storage; (2) Algorithm layer: Integrates ResNet and U-Net models, supports parallel GPU computing; (3) Application layer: Provides a Web or desktop interactive interface, supports result export and report generation.

[0048] The interaction methods of the system include: (1) User input: Upload medical image files and select the comparison mode; (2) Result feedback: Display the stage status, allow adjustment of the model, view the marked key points, highlight the difference areas in the heat map report, and generate a formatted report.

[0049] Preferably, a computer-readable storage medium is further provided, storing a computer program, and when the program is executed by a processor, the above method steps are performed.

[0050] Preferably, an electronic device is further provided, including a memory, a processor, and a computer program stored on the memory. When the processor executes the program, the steps of the above method are implemented.

[0051] The advantages of the present invention compared with the prior art are as follows: (1) High comparison accuracy: The present invention extracts multi-dimensional features of the skull through a deep learning model, including global geometric features, topological features, texture features, etc., realizing a comprehensive description and analysis of skull data, thereby improving the comparison accuracy; (2) Fast processing speed: The present invention adopts an efficient deep learning algorithm and parallel computing technology, which can realize the rapid processing and comparison analysis of skull data, greatly improving the work efficiency; (3) Wide application range: The present invention is not only applicable to the fields of forensic identification and anthropological research, but also can be widely applied to fields such as medical diagnosis and craniofacial reconstruction, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flowchart of a method for comparing skull data based on artificial intelligence.

[0053] Figure 2 is an architecture diagram of a system for comparing skull data based on artificial intelligence. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The present invention will be further described in detail below with reference to the drawings.

[0055] Embodiment 1

[0056] As Figure 1 shown, this embodiment provides a specific implementation of step S1 of a method for comparing skull data based on artificial intelligence. The purpose of S1 is to obtain the three-dimensional scan data of the skull to be measured and the reference skull database, as follows:

[0057] (1) Input the CT / MRI / DICOM original data of the skull to be measured, and reconstruct the medical image sequence into a three-dimensional voxel model. The formula:

[0058]

[0059] where, I i is the gray value of the i-th layer slice, w i is the interpolation weight, and N is the total number of slices.

[0060] Then extract the isosurface through the Marching Cubes algorithm. The formula:

[0061] S = {(x, y, z) | f(x, y, z) = T}

[0062] Among them, f(x, y, z) is a voxel field function, and T is the bone tissue density threshold (200 - 300 HU).

[0063] Then, the statistical outlier removal algorithm is used for denoising, and the formula is:

[0064]

[0065] Among them, p is the current point, p i is the neighborhood point, μ is the average distance, σ is the standard deviation, and α = 3 is the truncation coefficient.

[0066] Finally, uniform sampling is performed based on the voxel grid, and the formula is:

[0067]

[0068] Among them, the voxel side length Δ is adaptively adjusted to retain key geometric features.

[0069] (2) Integrate multi-source data. The data sources include historical medical image databases (CT / MRI), 3D scan data of archaeological excavated skulls, and synthetic data generation (GAN enhancement). A 3D sparse convolutional network is used to extract multi-scale features, and the formula is:

[0070] F g = Conv3D(V; W global )

[0071] Among them, W global is the pre-trained weight, and a 128-dimensional global feature vector is output.

[0072] Then, local feature enhancement is performed based on the curvature descriptor of anatomical landmark points, and the formula is:

[0073]

[0074] Among them, λ 1 ≤ λ 2 ≤ λ 3 are the eigenvalues of the covariance matrix of the neighborhood of point p, representing local curvature.

[0075] Finally, KD-Tree is used for fast nearest neighbor search to build an index, and the formula is:

[0076]

[0077] Among them, ⊕ represents the concatenation of the global feature F g and the local feature F l .

[0078] (3) Calculate the data similarity to fuse multi-features dynamically, formula:

[0079] S = α· + S 几何 + β·S 解剖 + γ·S 纹理

[0080] Among them, the weights α, β, γ are dynamically adjusted through the attention mechanism.

[0081] Then use RANSAC to screen the initial matching pairs, formula:

[0082] Inliers = {(p, q)||p - T(q)|| < ∈}

[0083] Among them, T is the rigid body transformation, and ∈ = 2mm is the tolerance error.

[0084] Finally, adopt the LM algorithm to optimize the non-rigid deformation, formula:

[0085]

[0086] Among them, f θ is the thin plate spline transformation function, and λ is the regularization coefficient.

[0087] Example Two

[0088] As Figure 1 shown, this example provides a specific implementation of step S2 of the skull data comparison method based on artificial intelligence. The purpose of S2 is to preprocess the skull data to be measured, specifically as follows:

[0089] (1) Input the original skull point cloud data P = {p i ∣p i ∈R 3} to perform point cloud statistical outlier filtering. For each point p i , calculate the average distance μ i and the standard deviation σ i of its k nearest neighbors (k = 30), formula:

[0090]

[0091] Then perform dynamic threshold elimination, formula:

[0092] Retain the point if μ i <μ global + 3σ global

[0093] Among them, μ global is the global average distance, which adapts to the density differences in different regions through dynamic thresholds.

[0094] Then input the three-dimensional voxel data V(x, y, z) for anisotropic diffusion filtering of the voxel data. The formula is:

[0095]

[0096] Among them, g(d) is the edge-preserving function, and K is the control smoothing intensity (K = 50HU).

[0097] (2) Normalize the grayscale of the CT data, that is, unify the CT values of different scanning devices to the standard bone density range to eliminate device differences. The formula is:

[0098]

[0099] Then input the voxel data with any resolution to unify the resolution, set the target resolution Δ = 1mm, and perform trilinear interpolation. The formula is:

[0100]

[0101] Among them, w is the linear interpolation weight function to ensure sub-voxel accuracy.

[0102] (3) Input the skull point cloud P to be registered and the template skull data P template Coarse registration: Select at least 10 anatomical landmark points such as the automatically detected Nasion, Porion, and Orbitale, and solve the rigid body transformation for coarse registration. The formula is:

[0103]

[0104] Among them, calculate the centroid Through SVD decomposition Get

[0105] R = VU T ,

[0106] Then perform fine registration through the ICP algorithm. The formula is:

[0107]

[0108] Among them, w i is the point pair weight, and λ = 0.1 is the regularization term coefficient to prevent excessive deformation.

[0109] Finally, the standards after data processing are: (1) Denoising: The signal-to-noise ratio is increased by ≥15dB; (2) Standardization: The cross-dataset comparison error is reduced by 42%; (3) Spatial alignment: The registration accuracy reaches 0.2mm.

[0110] Example 3

[0111] As shown Figure 1 in the figure, this embodiment provides a specific implementation manner of step S3 of a skull data comparison method based on artificial intelligence. The purpose of S3 is to extract multi-dimensional features of the skull to be measured through a deep learning model, which is specifically as follows:

[0112] (1) Input the preprocessed skull voxel data,

[0113] Input the preprocessed skull voxel data: Convert the three-dimensional skull data after standardization and spatial alignment into a voxel grid V ∈ R D*H*W , where D, H, and W are the depth, height, and width respectively. Extract global geometric features through a multi-layer 3D convolutional kernel. The formula is:

[0114]

[0115] where W is the weight of the 3×3×3 convolutional kernel and b is the bias term.

[0116] Then, gradually downsample through the pooling layer to generate a multi-scale feature map, capture the overall shape and local details of the skull, and construct a feature pyramid model.

[0117] (2) The first-layer regression network predicts the probability distribution heat map H ∈ R D*H*W of the anatomical landmark points. The formula is:

[0118]

[0119] where y p is the true Gaussian distribution, is the predicted value.

[0120] Then, crop the local block in the peak region of the heat map, input it into the fully connected layer to regress the accurate coordinates (x, y, z), and refine the coordinate regression through the loss function, so as to gradually optimize the positioning accuracy through multi-stage regression. The formula is:

[0121]

[0122] (3) Construct a skull graph structure, use the anatomical landmark points as graph nodes and the bone connection relationship as edges ε to generate a graph Then, aggregate the neighborhood node features through the graph convolutional layer to update the node embedding. The formula is:

[0123]

[0124] where is the neighbor of node v, and W (l) is the learnable weight to capture the topological association of structures such as the temporomandibular joint.

[0125] (4) Project the 3D skull onto the sagittal plane, coronal plane, and horizontal plane to obtain 2D images, retaining the surface texture information. Use the pre-trained ResNet-50 to extract the texture features T ∈ R of the projected images. d Formula:

[0126] T = ResNet(I proj )

[0127] where I proj is the projected image.

[0128] (5) Dynamically weight and splice the geometric features F 3D output by the 3D CNN, the topological features F GNN output by the GNN, and the texture features T through the attention mechanism to achieve multi-modal feature fusion. Formula:

[0129] α = σ(W α [F 3D ; F GNN ; T]), F fused = α 1 F 3D + α 2 F GNN + α 3 T

[0130] where σ is the sigmoid function and W α are learnable parameters.

[0131] Example 4

[0132] As Figure 1 shown, this example provides a specific implementation of step S4 of an artificial intelligence-based skull data comparison method. The purpose of S4 is to compare the similarity between the features of the skull to be measured and the features of the skull in the reference database and generate a matching score, as follows:

[0133] (1) Capture the spatial structure information of the skull through the 3D convolutional layer and extract local geometric features;

[0134] (2) Fuse the extracted geometric features with the deep features output by the CNN to generate a high-dimensional feature vector. Formula:

[0135] F skull = α·F geometric + β·F deep

[0136] where α and β are weighting coefficients.

[0137] Then use the autoencoder to reduce the dimension of the high-dimensional features and generate a low-dimensional embedding vector;

[0138] (3) Calculate the eigenvector F of the skull to be measured query with the eigenvector F in the reference library ref for similarity. The formula is:

[0139]

[0140] Take the absolute value as the matching score. The closer the value is to 1, the more similar it is. Then, according to the similarity threshold τ (τ = 0.1), screen the candidate skulls and retain the top K most similar samples for the fine matching stage.

[0141] (4) Use the ICP algorithm to perform rigid transformation (translation, rotation) alignment on the skull to be measured and the candidate skulls, and minimize the point cloud distance formula:

[0142]

[0143] where R is the rotation matrix and t is the translation vector.

[0144] Then introduce geometric feature constraints to accelerate the search and eliminate mis-matched point pairs:

[0145] Mis-matching judgment:

[0146] (5) Combine the rough matching similarity and the fine registration error to generate the final matching score. The formula is:

[0147] Score = γ·r + (1 - γ)·exp(-λ·E ICP )

[0148] where E ICP is the mean square error after ICP registration, and γ and λ are weight coefficients.

[0149] Then sort the skulls in the reference library in descending order of scores, and output the Top-N matching results and the visualization comparison diagram.

[0150] Example 5

[0151] As Figure 1 shown, this example provides a specific implementation of step S5 of an artificial intelligence-based skull data comparison method. The purpose of S5 is to output the identity information or morphological analysis results of the target reference skull according to the matching degree score, as follows:

[0152] (1) Use the Frankfurt plane to unify the skull coordinate system and eliminate the position and attitude differences. The formula is:

[0153] P F = M p L p × M p R p

[0154] Among them, M p L p and M p R p are respectively the planes formed by the left and right auricular points and the suborbital point, which are used to establish a standard coordinate system.

[0155] Then, scale the skull model to a unified scale. The formula:

[0156]

[0157] Among them, L p and R p are the coordinates of the left and right auricular points, eliminating individual size differences.

[0158] (2) Use the K-means algorithm to cluster the concave and convex vertices into feature regions, and screen the significant regions based on connectivity;

[0159] (3) Calculate the principal component vectors of the feature regions, and screen the similar regions (same type, close principal components, similar area). The formula:

[0160]

[0161] Among them, is the first principal component of region R i , and the similarity threshold is usually set to 0.1.

[0162] Then, perform an exhaustive match on the similar regions, calculate the optimal rigid body transformation and translation vector, and preliminarily align the target with the reference skull;

[0163] (4) Introduce a dynamic iteration factor to improve the ICP algorithm and achieve the fine registration of the target and the reference skull. The formula:

[0164]

[0165] Among them, p i and q i are the matching point pairs, and α is dynamically adjusted according to the error change rate (the initial value of α is 0.5).

[0166] (5) Calculate the Euclidean distance between vertices or the deviation along the normal direction. The formula:

[0167]

[0168] Among them, the unstable regions (such as boundary vertices) need to be excluded from the normal direction deviation to improve the accuracy.

[0169] Then, use the correlation coefficient to evaluate the similarity after multi-view feature fusion. The formula:

[0170]

[0171] Among them, F is the fused feature vector (such as curvature, contour angle, etc.), and ρ represents a high matching degree.

[0172] Finally, check the morphological criteria and make a determination in combination with the threshold.

[0173] (6) Weighted geometric deviation, correlation coefficient, and morphological consistency score, formula:

[0174] S total = w 1 ·FSTD + w 2 ·ρ + w 3 ·MorphScore. Among them, the weights w 1 、w 2 、w 3 are adjusted according to the task requirements.

[0175] Then perform identity matching. If S total ≤ θ match , output the reference skull identity information; then perform morphological analysis. If θ match <S total ≤ θ morph , output a morphological difference report (such as gender, racial classification).

[0176] Example Six

[0177] As Figure 1 and Figure 2 shown, this example provides a specific implementation manner of the skull data preprocessing and feature extraction method based on a deep learning model:

[0178] (1) Data acquisition and preprocessing

[0179] Use a high-precision CT scanner to obtain the three-dimensional image data of the skull to be measured, and the data format is DICOM. Extract the isosurface through the Marching Cubes algorithm to generate a three-dimensional mesh model of the skull. Use the statistical outlier removal algorithm to denoise the three-dimensional mesh model and remove noise points and abnormal points. Voxelize the processed three-dimensional mesh model to generate uniformly sampled voxel data.

[0180] (2) Feature extraction

[0181] Input the preprocessed cranial voxel data into the deep learning model, which adopts a multi-layer 3D convolutional kernel structure. Extract the global geometric features of the skull through the convolutional layer, construct a feature pyramid model to capture feature information at different scales. Use the graph convolutional layer to aggregate features of the skull graph structure, update node embeddings, and extract topological features. Project the skull onto 2D images along the sagittal plane, coronal plane, and horizontal plane, and use the pre-trained ResNet-50 to extract the texture features of the projected images. Dynamically weight and splice the geometric features, topological features, and texture features through the attention mechanism to achieve multi-modal feature fusion.

[0182] Example Seven

[0183] As Figure 1 and Figure 2 shown, this embodiment provides a specific implementation of the skull data similarity comparison and matching method:

[0184] (1) Feature vector generation

[0185] Perform dimensionality reduction on the extracted cranial features, and use an autoencoder to reduce the high-dimensional feature vector to a low-dimensional embedding vector. The low-dimensional embedding vector retains the key feature information of the skull while reducing the computational complexity.

[0186] (2) Similarity calculation and candidate screening

[0187] Calculate the cosine similarity between the feature vector of the skull to be measured and the feature vectors in the reference skull database. Screen candidates according to the similarity threshold, and retain the N most similar samples for the fine matching stage.

[0188] (3) Fine matching and alignment

[0189] Use the ICP algorithm to perform rigid transformation alignment on the skull to be measured and the candidate skulls to ensure the same position and posture of the skulls in space. Introduce geometric feature constraints to accelerate the search process, eliminate mis-matched point pairs, and improve the matching accuracy.

[0190] (4) Matching score generation and result output

[0191] Combine the rough matching similarity and the fine registration error to generate the final matching score. Arrange the skulls in the reference library in descending order of scores, and output the Top-N matching results and the visualization comparison graph.

[0192] Example Eight

[0193] As Figure 1 and Figure 2As shown, this embodiment provides a specific implementation of the construction and application of a cranial data comparison system. The data layer of the system stores the original medical image data and the preprocessed data, supporting formats such as DICOM and NIFTI. The algorithm layer integrates deep learning models (such as ResNet, U-Net, etc.) and supports parallel GPU computing to improve the processing speed. The application layer provides a Web or desktop interactive interface, supporting users to upload medical image files, select comparison modes, view comparison results, etc.

[0194] In terms of system functions, the data acquisition module obtains cranial data from multi-source medical images, supporting multi-angle shooting and high-resolution scanning. The preprocessing module standardizes the data format, removes noise and irrelevant tissues, including contrast enhancement, normalization and other processes. The feature extraction module extracts the key anatomical features of the skull, including key points, textures, shapes, etc. The intelligent comparison module realizes cranial identity recognition or pathological detection through feature matching and outputs the comparison results. The visualization output module displays the comparison results in an interactive interface, including visualization forms such as 3D models and heat maps.

[0195] In terms of system applications, this system can assist forensic identification by comparing cranial data to determine the identity information of unknown persons. It can also assist anthropological research by analyzing the morphological characteristics of the skulls of different populations and revealing the laws and trends of human evolution. It can also assist in the medical diagnosis of cranial diseases, such as cranial defects and fractures.

[0196] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A skull data comparison method based on artificial intelligence, characterized in that The following steps are involved: S1, obtaining the three-dimensional scanning data of the skull to be tested and the reference skull database; S2, preprocessing the skull data to be tested; S3, extracting multi-dimensional features of the skull to be tested through a deep learning model; S4, comparing the similarity of the skull features to be tested with the skull features in the reference database to generate a matching score; S5. Output the identity information or morphological analysis results of the target reference skull according to the matching score.

2. The skull data comparison method based on artificial intelligence according to claim 1, characterized in that The S1 further comprises the following steps: S1.

1. Input the original CT / MRI / DICOM data of the skull to be tested, reconstruct the medical image sequence into a 3D voxel model, extract the isosurface using the MarchingCubes algorithm, then use the statistical outlier removal algorithm to remove noise, and finally uniformly sample based on the voxel grid; S1.2, integrate multi-source data, use 3D sparse convolutional network to extract multi-scale features, then enhance local features based on the curvature descriptor of anatomical landmarks, and use KD-Tree for fast nearest neighbor search to build an index; S1.

3. Calculate data similarity to fuse multi-feature dynamics, use RANSAC to screen initial matching pairs, and use LM algorithm to optimize non-rigid deformation.

3. The skull data comparison method based on artificial intelligence according to claim 1, characterized in that The S2 further comprises the following steps: S2.1, input the original skull point cloud data to perform point cloud statistical outlier filtering, and then input the three-dimensional voxel data to perform voxel data anisotropic diffusion filtering; S2.2, normalize the grayscale of the CT data, and then input the voxel data of any resolution to unify the resolution; S2.

3. Input the skull point cloud to be registered and the template skull data for rough registration, and then use the ICP algorithm for fine registration.

4. The skull data comparison method based on artificial intelligence according to claim 1, characterized in that The S3 further comprises the following steps: S3.

1. Input the preprocessed skull voxel data, extract the global geometric features through multi-layer 3D convolution kernels, and construct a feature pyramid model. S3.2, the first-layer regression network predicts the probability distribution heat map of anatomical landmarks, crops local blocks in the peak area of ​​the heat map, and inputs the fully connected layer to regress the precise coordinates; S3.3, construct the skull graph structure, aggregate the neighborhood node features through the graph convolution layer, and update the node embedding; S3.4, project the 3D skull into a 2D image along the sagittal, coronal and horizontal planes, retain the surface texture information, and use the pre-trained ResNet-50 to extract the texture features of the projected image; S3.

5. Dynamically weight and concatenate the geometric features output by 3DCNN, the topological features output by GNN, and the texture features through the attention mechanism to achieve multimodal feature fusion.

5. The skull data comparison method based on artificial intelligence according to claim 1, characterized in that The S4 further comprises the following steps: S4.1, capture the spatial structure information of the skull through the 3D convolution layer and extract local geometric features; S4.2, fuse the extracted geometric features with the deep features output by CNN to generate a high-dimensional feature vector, use the autoencoder to reduce the dimension of the high-dimensional features to generate a low-dimensional embedding vector; S4.3, calculate the similarity between the feature vector of the skull to be tested and the feature vector in the reference library, screen the candidate skulls according to the similarity threshold, and retain the most similar samples to enter the fine matching stage; S4.4, use the ICP algorithm to perform rigid transformation alignment between the skull to be tested and the candidate skull, introduce geometric feature constraints to accelerate the search, and eliminate mismatched point pairs; S4.

5. Combine the coarse matching similarity and the fine registration error to generate the final matching score, sort the skulls in the reference library in descending order of the score, and output the Top-N matching results and visual comparison diagram.

6. The method for comparing skull data based on artificial intelligence according to claim 1, characterized in that The S5 further comprises the following steps: S5.

1. Use the Frankfurt plane to unify the skull coordinate system, eliminate position and posture differences, and then scale the skull model to a uniform scale; S5.

2. Use K-means algorithm to cluster the concave and convex vertices into feature regions, and select significant regions based on connectivity; S5.3, calculate the principal component vector of the feature area, screen similar areas, perform exhaustive matching on similar areas, calculate the optimal rigid body transformation and translation vector, and preliminarily align the target with the reference skull; S5.4, introduce dynamic iteration factor to improve ICP algorithm and achieve accurate registration between target and reference skull; S5.5, calculate the Euclidean distance between vertices or the deviation along the normal direction, use the correlation coefficient to evaluate the similarity after multi-view feature fusion, check the morphological criteria, and combine the threshold judgment; S5.

6. Weighted geometric deviation, correlation coefficient and morphological consistency score for identity determination.

7. A skull data comparison system based on artificial intelligence, comprising a data acquisition module, a preprocessing module, a feature extraction module, an intelligent comparison module, and a visual output module, characterized in that: The data acquisition module acquires skull data from multi-source medical images, uses multi-angle shooting equipment to collect data, and obtains high-resolution three-dimensional data through CT / MRI scanners; supports DICOM and NIFTI format input, The preprocessing module is used to standardize the data format, remove noise and irrelevant tissue, including contrast enhancement and normalized data functions; The feature extraction module extracts key anatomical features from the preprocessed data, including key points, textures, and shapes, uses ResNet-152 or Residual U-Net to extract high-level semantic features, combines interpretable CNN to generate activation maps, and locates anatomical landmarks, including extracting GLCM textures and DWT wavelet features to assist classification functions; The intelligent comparison module realizes skull identification or pathological detection through feature matching, locates the fracture area using YOLOv3 or FasterR-CNN, enhances the inter-class distinction based on SVM or contrast loss function, and quantifies the degree of feature matching through cosine similarity or Dice coefficient; The visualization output module displays the comparison results in an interactive interface, generates a three-dimensional skull model using boundary detection and texturing techniques, and highlights key difference areas through global activation mapping.

8. The artificial intelligence-based skull data comparison system according to claim 7, characterized in that: The end-to-end process of the system includes: (1) the raw data is preprocessed and then input into the feature extraction network, which outputs a high-dimensional feature vector; (2) the comparison module determines the identity / pathological status through similarity calculation or classification model; (3) the visualization module renders the results into a 3D model or heat map; The architecture of the system includes: (1) data layer: storing raw images and pre-processed data, supporting distributed storage; (2) algorithm layer: integrating ResNet and U-Net models, supporting parallel GPU computing; (3) application layer: providing a Web or desktop interactive interface, supporting result export and report generation; The interactive mode of the system includes: (1) user input: uploading medical image files and selecting comparison mode; (2) result feedback: displaying stage status, allowing adjustment of the model, viewing annotated key points, highlighting difference areas in heat map reports, and generating formatted reports.

9. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 6 are implemented.

Images