Method for diagnosing head deformity of children

Abstract

The invention relates to a method for diagnosing head deformity of children. The method comprises the following steps: S1; generating a three-dimensional model of the child head based on a priori auxiliary child head multi-view three-dimensional reconstruction algorithm; s2; s2, performing attitude adjustment and slicing processing on the three-dimensional model generated in the step S1 to obtain a two-dimensional slice image; (S3); calculating an exponential characteristic value of the two-dimensional slice image obtained in the step S2; (S4); and S3, inputting the exponential characteristic value in S3 into a malformation classification model, and outputting a malformation diagnosis result. Compared with the prior art, the method has the advantages that radiation-free safety detection is realized; performing high-precision three-dimensional reconstruction; carrying out high-precision teratosis judgment and classification; the operation is convenient and the cost is low.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain neurosurgery, in particular to a method for diagnosing head deformity of children. BACKGROUND

[0002] Head deformity refers to a series of abnormalities and deviations of the head of children, which is manifested as asymmetry of the skull or inconsistency of the facial shape. The types of head deformity mainly include plagiocephaly, brachycephaly and craniosynostosis, among which craniosynostosis is a common cause of skull deformation. It is a congenital defect. Head deformity is a common disease in pediatric neurosurgery. If not treated in time, it not only affects the appearance of the head, but also may be accompanied by visual, hearing impairment, balance, epilepsy, brain dysplasia, etc., which affects the health and growth and development of children. Therefore, how to quickly and accurately identify the head deformity of children is of great significance to the health of the head of children.

[0003] Patent CN202110267241.2 discloses a measurement method for customizing an orthopedic helmet and an obtained orthopedic helmet, and the steps are as follows: S1: head deformity patient data acquisition: that is, a plurality of camera arrays are arranged around the deformity patient to take 360° no dead angle photos of the head at the same time; to obtain the 2D picture of the head of the deformity patient;

[0004] S2: deformity head model acquisition: that is, the data of each picture is input as the point of the initial photo, the computer image recognition neural network system is used, the encoder-decoder architecture is used, the points of each picture are converted into the neural network of the feature map or vector, which is used to predict the specific information of the shape of the head of the patient, and then the real three-dimensional head model of the patient is generated; S3: corresponding parameter measurement of deformity head: that is, the three-dimensional head model is measured by index to obtain the data of the customized helmet; the parameter measurement includes head circumference index measurement, head index measurement and head asymmetry index measurement, S4: customization of orthopedic helmet through the above measurement indexes. However, the data acquisition has limitations, and the neural network reconstruction has the risk of distortion, resulting in inaccurate judgment structure.

[0005] The existing diagnosis method needs to use expensive medical equipment to assist in diagnosis. For families with relatively poor conditions, the diagnosis cost is high, the time is long, the diagnosis method is not very friendly, and the judgment accuracy is low. Therefore, how to make the diagnosis method more convenient, lower in cost, safer and more efficient is an important direction for the development of the diagnosis of head deformity of children. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art and provide a method for diagnosing head deformity of children, which is safe and radiation-free, has high-precision three-dimensional reconstruction, high-precision judgment and classification, and is convenient and low in cost.

[0007] The object of the present application can be achieved by the following technical solutions:

[0008] The construction of a three-dimensional model of a child's head for deformity analysis is an effective way to overcome existing problems. The present application first constructs a high-precision three-dimensional model of a child's head based on prior-assisted multi-view reconstruction algorithm and a special mesh cap, then processes the constructed three-dimensional model to calculate the corresponding characteristic values, and finally imports the characteristic values into a deformity classification model for deformity diagnosis. Figure Three

[0009] The present application realizes rapid and accurate screening and diagnosis of child head deformity through multi-view reconstruction technology and machine learning algorithm. The system can efficiently generate a three-dimensional model of a child's head under non-radiation conditions, and make deformity judgment through a high-precision deformity classification model, thereby providing technical support for early screening, diagnosis and treatment of child head deformity. Figure Three

[0010] The present application provides a method for diagnosing child head deformity, comprising the following steps:

[0011] S1: generating a three-dimensional model of a child's head based on prior-assisted multi-view reconstruction algorithm of a child's head; Figure Three

[0012] S2: adjusting the posture and slicing the three-dimensional model generated in S1 to obtain a two-dimensional slice image;

[0013] S3: calculating the exponential characteristic value of the two-dimensional slice image obtained in S2;

[0014] S4: inputting the exponential characteristic value of S3 into a deformity classification model to output a deformity diagnosis result.

[0015] Further, in S1, the prior-assisted multi-view reconstruction algorithm of a child's head comprises the following steps: Figure Three

[0016] A1: pre-processing the input two-dimensional child head multi-view image, including processing tangential distortion and radial distortion; since the camera may have distortion, the distortion includes tangential distortion and radial distortion, so the captured image needs to be processed for distortion. The processed picture is input into the prior-assisted multi-view reconstruction algorithm of a child's head to generate a three-dimensional model of a child's head. Figure Three

[0017] A2: extracting feature points in the two-dimensional child head multi-view image and performing feature point matching;

[0018] A3: calculating the fundamental matrix and the essential matrix according to epipolar geometry, and decomposing the camera internal and external parameters; ​​​​​

[0019] A4: According to the matched feature points and the camera internal and external parameters, the coordinates of the feature points in the three-dimensional space are calculated by triangulation, and a sparse point cloud is generated;

[0020] A5: Start dense reconstruction, and select the best neighbor image for each multi-view image to calculate a depth map;

[0021] A6: By propagation strategy, the image blocks with high similarity are matched to generate an initial depth map and filter; the filtering involves projecting the points in the image to the world coordinate system, and then comparing with the depth information on the neighbor frames to determine which points are valid;

[0022] A7: Utilize the plane priori to assist in reconstructing weak texture areas, and generate a priori depth map by triangulation algorithm; when processing weak texture areas, the plane priori can be used to assist three-dimensional reconstruction; the plane priori assumes that the surface in these areas can be well approximated by a plane; by triangulation algorithm, the points in the filtered initial depth map are connected into a triangular mesh, thereby generating a priori depth map;

[0023] A8: Construct a fusion method considering both photometric consistency measurement and plane priori information, introduce multi-view geometric consistency to delete matching errors, and generate a dense point cloud as a three-dimensional model.

[0024] The fusion method is to find and judge the low texture area in the initial depth map, and adjust the depth value and normal vector of the area to the depth value and normal vector of the priori depth map.

[0025] Further, in A1, to ensure the efficiency and quality of reconstruction, the photographed two-dimensional child skull multi-view images are full-view images of the skull, and the shooting conditions of the two-dimensional child skull multi-view images include:

[0026] The included angle between the current frame and the image plane of the previous frame is 15 degrees;

[0027] The number of images is 24-30 frames.

[0028] Further, in A6, the propagation strategy is a red and black checkerboard propagation strategy, which is suitable for GPU processors.

[0029] Further, in A7, the plane priori discrimination formula (i.e. the discrimination formula of the image block being a low texture area) is: assuming that the image block size is m x n,

[0030]

[0031] where ω i is the weight of each pixel in the image block, which considers the influence of color and spatial distance; Δg i = (gi– g r ) 2 is the square of the color intensity difference between pixel i and reference pixel r in the image block, Δx i =(x i– x r ) 2 is the square of the pixel distance difference between pixel i and reference pixel r in the image block, σ g , σ x are Gaussian distribution parameters for controlling the influence of color distance and spatial distance respectively; W(i,r) is the color intensity value of pixel point (i,r), W is the average color intensity of pixels in the image block, and l is the texture threshold. The smaller the value, the weaker the texture of the region.

[0032] Further, the three-dimensional reconstruction algorithm is combined with a grid cap composed of a black grid and an ArUco code to solve the problem that the presence of hair cover and a large number of weak texture areas on the surface of the child's head will cause the reconstruction effect to weaken.

[0033] Further, in S2, the posture adjustment includes calculating the convex hull of the point cloud, making the Z-axis direction consistent with the convex hull direction, and making the horizontal plane of the point cloud parallel to the ground.

[0034] Further, in S2, the slice processing includes selecting the most complete slice of the point cloud and projecting it onto a two-dimensional plane to generate a two-dimensional image.

[0035] Further, in S3, the index characteristic values of the two-dimensional slice image include:

[0036] Cranial Radius (CR) index; Fourier coefficients with a centroid distance function of the oriented cranial boundary points are used to describe the location and extent of protruding areas (e.g. restrictive or compensatory growth) in the skull;

[0037] Cranial Partial Slope (CPS) index; the sum of the differences in frontal bone centroid distance is used to describe the degree of frontal bone tilting due to compensatory growth.

[0038] Cranial Extreme Spot (CES) index;

[0039] Near Cranial Extreme Spot (NCES) index; the distance distribution between boundary points is used to describe the location and distribution of protruding areas in the skull, CES is calculated as the variance of extreme points and the variance of extreme point areas, and NCES index is calculated as the variance of near extreme points.

[0040] Cranial Index (CI); is the ratio of the maximum width and the maximum length of the skull.

[0041] Cranial Vault Asymmetry Index (CVAI); is the percentage difference of the diagonal line pointing of the head.

[0042] Anterior Midline Width Ratio (AMWR); calculates the ratio of the 1 / 4 quantile line of the skull and the midline.

[0043] Anterior Posterior Width Ratio (APWR); calculates the ratio of the 3 / 4 quantile line of the skull and the midline.

[0044] Left-Right Height Ratio (LRHR); calculates the ratio of the minimum and maximum of the left AP (maximum length) and the right AP of the skull.

[0045] Further, in S4, the malformation classification model is trained using child skull malformation data, and the training constructs a primary classifier as SVM, KNN, RandomForest, XGBoost, and a secondary classifier as a Stacking heterogeneous ensemble learning classification model of RandomForest, and the malformation classification model is deployed to the algorithm processing platform.

[0046] Compared with the prior art, the present application has the following advantages:

[0047] (1) Non-radiation safe detection. Compared with traditional CT scanning, this method does not need to use radioactive equipment, avoiding potential harm to children by radiation. Multi-view optical imaging is used instead of X-ray imaging.

[0048] (2) High-precision three-dimensional reconstruction. Through the child skull multi-view Figure Three reconstruction algorithm based on priori assistance and the grid cap, a high-quality three-dimensional model can be quickly generated. The grid cap (black grid + ArUco code) solves the reconstruction problem of hair occlusion and weak texture area, and introduces a plane prior to improve the reconstruction quality.

[0049] (3) High-precision malformation classification. The Stacking heterogeneous ensemble model can accurately judge the skull malformation.

[0050] (4) Convenient operation and low cost. The present application only needs to shoot multi-view images of the child's skull. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 Flowchart of the method for child skull malformation diagnosis;

[0052] Figure 2 For this prior aid-based pediatric cranioscopic multiple vision Figure Three Flowchart of the 3D reconstruction algorithm;

[0053] Figure 3 This is a schematic diagram of a mesh cap;

[0054] Figure 4 A schematic diagram illustrating the processing of a 3D model of a child's skull; Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0056] Example 1

[0057] This embodiment provides a method for diagnosing cranial malformations in children, such as... Figure 1 As shown, it includes the following steps:

[0058] S1; Prior-assisted cephalometrics in children Figure Three A 3D reconstruction algorithm generates a 3D model of a child's skull;

[0059] S2; The 3D model generated in S1 is subjected to pose adjustment and slicing to obtain a 2D slice image;

[0060] S3; Calculate the 9 exponential feature values ​​of the two-dimensional slice image obtained in S2;

[0061] S4; Input the exponential feature value of S3 into the deformity classification model and output the deformity diagnosis result.

[0062] In a specific implementation, S1, in the case of children's cranial multiple vision based on prior aids... Figure Three In the 3D reconstruction algorithm, a pre-selection process is performed before choosing the best neighboring frame. This pre-selection considers three main conditions: the angle between common viewpoints, resolution similarity, and minimum coverage area. During the pre-selection stage, several neighboring frames are selected for each image. Next, an energy optimization problem using a Markov random field is used to select the best neighboring frame, minimizing the overall energy function. The algorithm matches image patches with high similarity, using a weighted normalized cross-correlation method to calculate the photometric consistency measure of the image patches, and uses the result as an index to measure the similarity of the image patches.

[0063] like Figure 2 , 4 As shown, pediatric cephalometrics based on prior aids Figure Three The 3D reconstruction algorithm includes the following steps:

[0064] A1: Preprocess the input two-dimensional child head multi-view image, including processing tangential distortion and radial distortion; since there may be distortion in the shooting camera, the distortion includes tangential distortion and radial distortion, so the shot image needs to be processed for distortion. The processed picture is input to the priori auxiliary based child head multi-view reconstruction algorithm to generate a three-dimensional model of the child head. Figure Three

[0065] A2: Extract feature points in the two-dimensional child head multi-view image and perform feature point matching;

[0066] A3: Calculate the fundamental matrix and the essential matrix according to the epipolar geometry, and decompose the camera internal and external parameters;

[0067] A4: According to the matched feature points and the camera internal and external parameters, the coordinates of the feature points in the three-dimensional space are calculated by triangulation, and a sparse point cloud is generated;

[0068] A5: Start dense reconstruction, and select the best neighbor image for each multi-view image to calculate the depth map;

[0069] A6: Through the propagation strategy, the image blocks with high similarity are matched to generate an initial depth map and filter; the filtering involves projecting the points in the image to the world coordinate system, and then comparing with the depth information on the neighbor frame to determine which points are valid;

[0070] A7: Reconstruct weak texture areas with plane priori assistance, and generate a priori depth map through a triangulation algorithm; when processing weak texture areas, plane priori can be used to assist three-dimensional reconstruction; the plane priori assumes that the surface in these areas can be well approximated by a plane; through the triangulation algorithm, the points in the filtered initial depth map are connected into a triangular mesh, thereby generating a priori depth map;

[0071] A8: Construct a fusion method considering photometric consistency measurement and plane priori information, introduce multi-view geometric consistency to delete matching errors, and generate a dense point cloud as a three-dimensional model.

[0072] The fusion method is to find and judge the low texture area in the initial depth map, and adjust the depth value and normal vector of the area to the depth value and normal vector of the priori depth map.

[0073] In the specific embodiment, in A1, to ensure the efficiency and quality of reconstruction, the shot two-dimensional child head multi-view image is a full-view image of the head, and the shooting conditions of the two-dimensional child head multi-view image include:

[0074] The included angle between the current frame and the plane of the previous frame image is 15 degrees;

[0075] ​The number of images is 24-30 frames.

[0076] In the specific embodiment, in A6, the propagation strategy is a red and black checkerboard propagation strategy, which is suitable for a GPU processor.

[0077] In the specific embodiment, in A7, the plane prior discriminant formula (i.e., the discriminant formula of the image block being a low-texture region) is: assuming that the image block size is m x n,

[0078]

[0079] where ω i is the weight of each pixel in the image block, which takes into account the influence of color and spatial distance; Δg i =(g i– g r ) 2 is the square of the color gray value difference between pixel i and reference pixel r in the image block, Δx i =(x i– x r ) 2 is the square of the pixel distance difference between pixel i and reference pixel r in the image block, σ g , σ x are Gaussian distribution parameters for controlling the influence of color distance and spatial distance, respectively; W(i,r) is the color gray value of pixel point (i,r), W is the average color gray value of the pixels in the image block, and l is a texture threshold value. The smaller the value, the weaker the texture of the region.

[0080] As shown in Figure 3 , in the specific embodiment, the three-dimensional reconstruction algorithm is used in combination with a grid cap composed of a black grid and an ArUco code. It is used to solve the problem that the presence of hair occlusion and a large number of weak texture regions on the surface of a child's head can cause the reconstruction effect to weaken.

[0081] As shown in Figure 4 , in the specific embodiment, in S2, the pose adjustment includes calculating the convex hull of the point cloud, making the Z-axis direction consistent with the convex hull direction, and making the horizontal plane of the point cloud parallel to the ground. Because the child's head can be regarded as a convex hull. After completing the slicing, the sliced point cloud is projected onto a two-dimensional plane for processing. The nine index eigenvalues are calculated by performing index eigenvalue calculation on the two-dimensional image of the slice. Finally, the obtained index eigenvalues are imported into the malformation classification model for malformation judgment.

[0082] In the specific embodiment, in S2, the slice processing includes selecting the most complete slice of the point cloud and projecting it onto a two-dimensional plane to generate a two-dimensional image.

[0083] In the specific embodiment, in S3, the index eigenvalues of the two-dimensional slice image include:

[0084] Cranial Radius (CR) Index; Fourier coefficients with centroid distance function with directional cranial boundary points are used to describe the location and extent of protruding areas in the skull (e.g. restrictive or compensatory growth);

[0085] Cranial Partial Slope (CPS) Index; the sum of the frontal bone centroid distance differences is used to describe the extent of frontal bone tilting due to compensatory growth.

[0086] Cranial Extreme Spot (CES) Index;

[0087] Near Cranial Extreme Spot (NCES) Index; the distance distribution between boundary points is used to describe the location and distribution of protruding areas in the skull, the CES is calculated as the variance of the extreme points and the variance of the extreme point area, the NCES index is calculated as the variance of the near extreme points.

[0088] Cranial Index (CI); is the ratio of the maximum width and maximum length of the skull.

[0089] Cranial Vault Asymmetry Index (CVAI); is the percentage difference of the head diagonal pointing.

[0090] Anterior Midline Width Ratio (AMWR) Index; calculates the ratio of the cranial 1 / 4 quantile line to the midline.

[0091] Anterior Posterior Width Ratio (APWR) Index; calculates the ratio of the cranial 3 / 4 quantile line to the midline.

[0092] Left-Right Height Ratio (LRHR) Index; calculates the ratio of the minimum and maximum of the left and right AP (maximum length) of the skull.

[0093] In the specific embodiment, in S4, the malformation classification model is trained using child skull malformation data, the training constructs a primary classifier as SVM, KNN, RandomForest, XGBoost, and a secondary classifier as RandomForest Stacking Heterogeneous Ensemble Learning Classification Model, and the malformation classification model is deployed to the algorithm processing platform.

[0094] Components not specifically described in this example are existing components that can be purchased in the open market.

[0095] The foregoing description of the embodiments has been presented for the purpose of enabling those skilled in the art to practice the application. It is understood that various modifications can be made to the embodiments disclosed herein without departing from the spirit and scope of the application. Therefore, the above description should not be taken as limiting the scope of the application and the application is defined only by the appended claims.

Claims

1. A method of diagnosing a cranial deformity in a child, characterized in that, The method comprises the following steps: S1: generating a three-dimensional model of a child's head based on a priori auxiliary child head multi-view three-dimensional reconstruction algorithm; S2: adjusting the posture and slicing the three-dimensional model generated in S1 to obtain a two-dimensional slice image; S3: calculating the exponential characteristic value of the two-dimensional slice image obtained in S2; S4: inputting the exponential characteristic value of S3 into a deformity classification model to output a deformity diagnosis result.

2. The method of diagnosing cranial deformities in children according to claim 1, characterized in that, In S1, the priori auxiliary child head multi-view three-dimensional reconstruction algorithm comprises the following steps: A1: preprocessing the input two-dimensional child head multi-view image, including processing tangential distortion and radial distortion; A2: extracting feature points in the two-dimensional child head multi-view image and performing feature point matching; A3: calculating the fundamental matrix and the essential matrix, and decomposing the camera internal and external parameters; A4: calculating the coordinates of the feature points in the three-dimensional space according to the matched feature points and the camera internal and external parameters, and generating a sparse point cloud through triangulation; A5: starting dense reconstruction, and selecting the best neighbor image for each multi-view image to calculate a depth map; A6: matching the image blocks with high similarity through a propagation strategy, generating an initial depth map and filtering; A7: using a plane priori to assist in reconstructing a weak texture area, and generating a priori depth map through a triangulation algorithm; A8: constructing a fusion method considering photometric consistency measurement and plane priori information, introducing multi-view geometric consistency to delete matching errors, and generating a dense point cloud as a three-dimensional model.

3. The method of diagnosing cranial deformities in children according to claim 2, characterized in that, In A1, the photographed two-dimensional child head multi-view image is a full-view image of the head, and the shooting conditions of the two-dimensional child head multi-view image include: The included angle between the current frame and the plane of the last frame image is 15 degrees; The number of images is 24-30 frames.

4. The method of diagnosing cranial deformities in children according to claim 2, wherein In A6, the propagation strategy is a red and black checkerboard propagation strategy, which is suitable for GPU processors.

5. The method of diagnosing cranial deformities in children according to claim 2, wherein In A7, the plane priori discriminant formula (i.e. the discriminant formula of the image block in the low texture area) is: assuming that the image block size is m×n, where ω i is the weight of each pixel in the image block; Δg i = (g i– g r ) 2 is the square of the color intensity difference between pixel i and reference pixel r in the image block, Δx i = (x i– x r ) 2 is the square of the pixel distance difference between pixel i and reference pixel r in the image block, σ g , σ x are Gaussian distribution parameters for controlling the influence of color distance and spatial distance respectively; W(i, r) is the color intensity value of pixel point (i, r), W is the average value of pixel color intensity in the image block, and l is the texture threshold.

6. The method of diagnosing cranial deformities in children according to claim 1, wherein The three-dimensional reconstruction algorithm is combined with a grid cap, and the grid cap is composed of a black grid and an ArUco code.

7. The method of diagnosing cranial deformities in children according to claim 1, wherein In S2, the posture adjustment includes calculating the convex hull of the point cloud, making the Z-axis direction consistent with the convex hull direction, and making the point cloud horizontal plane parallel to the ground.

8. The method of diagnosing cranial deformities in children according to claim 1, wherein In S2, the slicing processing includes selecting the most complete slice of the point cloud and projecting it onto a two-dimensional plane to generate a two-dimensional image.

9. The method of diagnosing cranial deformities in children according to claim 1, wherein In S3, the exponential characteristic value of the two-dimensional slice image includes: skull radius index; skull part slope; skull pole index; near skull pole index; skull index; skull vault asymmetry index; front width ratio index; back width ratio index; left and right height ratio index; calculating the ratio of the minimum value and the maximum value of the left AP and the right AP of the head.

10. The method of diagnosing cranial deformities in children according to claim 1, wherein In S4, the deformity classification model is trained using child head deformity data, and the training constructs a primary classifier as SVM, KNN, RandomForest, XGBoost, and a secondary classifier as a RandomForest Stacking heterogeneous ensemble learning classification model. The deformity classification model is deployed to an algorithm processing platform.

Images