An Automatic Assessment Method for Cranial Morphology Based on Pediatric CT Images

Abstract

This invention belongs to the field of computer medical technology and relates to an automatic assessment method for cranial morphology based on pediatric CT images. The method involves acquiring original pediatric cranial CT images; identifying the voxel coordinates of the nasal root point, left tragus point, and right tragus point in the original pediatric cranial CT images; redirecting the original pediatric cranial CT images to a standard anatomical coordinate system to obtain a normalized cranial image; performing three-dimensional connected component extraction and morphological closure operations on the normalized cranial image to obtain a binary mask of the complete skull; locating the pediatric cranial measurement plane and extracting the skull contour to obtain a collision point set; calculating the cranial asymmetry index and cranial proportion; and comparing the results with a predefined clinical grading scale to generate an assessment result. This invention solves the problems of weak anti-interference ability, strong subjectivity in key point localization, lack of standardized spatial alignment, and fragmented automated processes in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of computer medical technology, and in particular relates to an automatic assessment method for cranial morphology based on pediatric CT images. Background Technology

[0002] In pediatric clinical diagnosis, accurate assessment of cranial deformities (such as plagiocephaly, brachycephaly, and scaphoid cephaly) is crucial for infant and toddler development. Traditional assessment methods mainly include physical measurement, manual observation of imaging, and laser scanning, but these methods generally suffer from problems such as large errors, poor repeatability, and long processing times. Although CT images are the gold standard for diagnosing cranial deformities and can clearly show the skeletal structure, there is still a lack of CT image analysis software specifically for the automatic assessment of pediatric cranial morphology. The development of such an automated system faces the following major technical challenges: (1) Weak anti-interference ability: the hard background such as the scanning bed in CT images is similar in gray to the bones, which can easily interfere with subsequent segmentation and measurement; (2) Difficulty in locating anatomical landmarks: the locations of key points such as the root of the nose (SE), the left tragus (TR_L), and the right tragus (TR_R) are hidden. Manually locating these points in hundreds of CT slices is not only time-consuming and laborious, but also relies on subjective experience, resulting in significant differences in the cranial index (CVAI, CR) measured by different doctors; (3) Lack of standardized spatial alignment: changes in head posture (such as pitch, lateral tilt, and rotation) often occur during the scanning of infants and young children, resulting in deviations between the original axial images and the standard anatomical plane, affecting the accuracy of measurement; (4) Low degree of automation: there is a lack of end-to-end automated processing flow from the original images to the diagnostic report.

[0003] In summary, there is an urgent need to develop a new technical method to overcome the above-mentioned bottlenecks and achieve accurate, efficient, and automated assessment of pediatric skull morphology. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic assessment method for cranial morphology based on pediatric CT images, which solves the technical problems of low diagnostic accuracy, poor repeatability, and time-consuming and labor-intensive processes caused by the lack of specialized automated analysis software, the reliance on manual methods in current clinical practice, weak anti-interference ability, strong subjectivity in key point localization, lack of standardized spatial alignment, and broken automation processes.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is an automatic assessment method for cranial morphology based on pediatric CT images, comprising the following steps: S1: Acquire raw pediatric head CT images ; S2: Using a pre-trained deep learning keypoint detection model, identify raw pediatric head CT images. Voxel coordinates of the middle nasal root point, left tragus point, and right tragus point; S3: Based on the voxel coordinates of the nasal root point, left tragus point, and right tragus point, a standard anatomical coordinate system is constructed, and the original pediatric head CT images are transformed using affine transformation. Redirecting to this coordinate system yields the corrected head image; S4: Perform three-dimensional connected component extraction and morphological closure operations on the corrected skull image to obtain a binary mask of the complete skull. S5: Based on a binary mask of the complete skull, perform planar localization and skull contour extraction for pediatric skull measurement to obtain a collision point set; S6: Convert the pixel distance of each collision point in the collision point set into physical length, calculate the cranial asymmetry index and cranial proportion, and then compare it with a predefined clinical grading scale to generate the assessment results.

[0006] Furthermore, the specific steps of S3 are as follows: S3.1: Calculate the midpoint M of the two tragus points based on the coordinates of the left and right tragus points, and use it as the origin of the standard anatomical coordinate system; S3.2: Construct a standard anatomical coordinate system based on three anatomical landmarks, including the Y-axis vector. Z-axis vector X-axis vector ; S3.3: For the Y-axis vector Z-axis vector X-axis vector Normalize and construct the rotation matrix. ; S3.4: Using a rotation matrix Construct an affine transformation matrix, as shown in Equation (1), to transform the original pediatric head CT image. Reprojecting the image onto the standard anatomical coordinate system yields the uprighted skull image. ; (1) in: This is a head image after the procedure was completed. Original pediatric head CT images, Let the coordinates be the center of rotation. It is a rotation matrix.

[0007] Furthermore, the specific steps of S4 are as follows: S4.1: Using the nasal root point, left tragus point, and right tragus point detected by S2 as seed points, and setting an adaptive threshold range. ,in, The average gray level of the image. The standard deviation of the noise level is used to extract the connected components in three-dimensional space to obtain the main region of the child's skull; S4.2: Perform two-dimensional morphological closure operations layer by layer on the extracted main area of ​​the child's skull to obtain a binary mask of the complete skull.

[0008] Furthermore, the specific steps of S5 are as follows: S5.1: Determine the Level 0 plane based on the nasal root point, left tragus point, and right tragus point; S5.2: Using a complementary strategy of roundness detection and volume centroid method, identify the top direction of the child's head, determine the physical position of the top of the head along the determined top direction, construct a plane parallel to the Level 0 plane, and determine this plane as the Level 10 top plane; S5.3: Divide the vertical distance between the Level 0 plane and the Level 10 overhead plane into 10 equal parts using linear interpolation, and calculate the z-coordinate of the Level 3 plane; S5.4: Using the midpoint of both tragus Centered on the projection on the Level 3 plane, rays are emitted in multiple directions to detect collision points with the binary mask of the skull and extract the collision point set.

[0009] Furthermore, the specific steps for identifying the direction of the top of a child's head using a complementary strategy of roundness detection and volume centroid method, as described in S5.2, are as follows: S5.2.1: Extracting raw pediatric head CT images In the non-background region of the whole slice sequence, determine the Z-axis span of the brain tissue; S5.2.2: Extract slice images at 1 / 4 and 3 / 4 of the span respectively, perform edge detection and cranial ring repair on the extracted images to obtain the outermost closed contour of the skull; S5.2.3: Calculate the roundness value of the two slice images using a roundness detection algorithm; S5.2.4: Compare the roundness values ​​of the two slice images. If the slice image at 1 / 4 is close to a regular circle or ellipse, while the slice image at 3 / 4 is irregular, then the lower index direction is determined to be the top of the head; otherwise, the higher index direction is determined to be the top of the head. S5.2.5: If the judgment in S5.2.4 fails, calculate the geometric centroid of the z-axis for all slices containing brain tissue. S5.2.6: Compare the preset z-coordinate of Level 0 with the geometric centroid of the z-axis. If the z-coordinate of Level 0 is greater than the geometric centroid of the z-axis, then Level 10 is located at the low index end of the slice sequence; otherwise, Level 10 is located at the high index end of the slice sequence.

[0010] Furthermore, the specific steps in S5.4 are as follows: S5.4.1: Set the sampling angle range θ∈[0,2π), and emit rays from the center point along multiple preset angles; S5.4.2: Perform a collision point search for each ray; S5.4.3: Collect all collision points of the rays to form a collision point set.

[0011] Furthermore, the specific steps in S5.4.2 are as follows: S5.4.2.1: For each angle... The ray, setting the initial distance scalar ; S5.4.2.2: Calculate the unit direction vector As shown in equation (2), the coordinates of the current sampling point are updated. As shown in equation (3): (2) in: It is a unit direction vector. For ray index sequence number; (3) in: This represents the x-coordinate of the current sampling point in the real-time voxel of the Level 3 axis plane. This represents the ordinate of the current sampling point in the real-time voxel on the Level 3 axis plane. Midpoint of both tragus The x-coordinate of the orthographic projection onto the Level 3 plane. Midpoint of both tragus The ordinate of the orthographic projection onto the Level 3 plane. To start from the origin The cumulative distance scalar along the ray direction from the starting point; S5.4.2.3: Query the value of the binary mask of the complete skull at the current sampling point (x,y) Mask(x,y). If Mask(x,y)=1, record this point as the first collision point and terminate the extension of the current ray; if Mask(x,y)=0, continue stepping and repeat S5.4.2.2-S5.4.2.3 until each ray finds a collision point.

[0012] Furthermore, the specific steps of S6 are as follows: S6.1: Convert the pixel distance between corresponding points in the collision point set into physical length to obtain the preset skull feature value; The preset skull feature values ​​are: anteroposterior diameter, lateral diameter, long diagonal, and short diagonal; S6.2: Based on the preset skull feature values ​​obtained in S6.1, calculate the skull asymmetry index and skull proportion, and then compare them with the predefined clinical grading scale to generate the assessment results.

[0013] The beneficial effects of this invention are as follows: First, in terms of detection efficiency, traditional manual observation of a single imaging examination takes 15-30 minutes, physical measurement (head shape ruler / soft ruler) takes 10-20 minutes, and laser measurement (equipment debugging and pressure processing) takes 5-10 minutes. This invention, however, requires only one-click upload and can automatically generate a report within 10 seconds, greatly shortening the detection time. Second, in terms of measurement accuracy, using manual observation by experienced radiologists as the gold standard, the mean square error (MSE) of traditional physical measurements such as CVAI and CR is 0.1-0.2, while the MSE of the output results of this invention is only 0.03-0.06, significantly improving accuracy. Furthermore, this invention reduces the qualification requirements for operators. Traditional methods require personnel with specialized learning and training to use head shape rulers / soft rulers or laser equipment, or require manual observation by experienced radiologists. This invention, however, can be completed by general computer operators with one-click upload, automatic analysis, simple operation, and easy promotion. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of an automatic assessment method for skull morphology based on pediatric CT images according to an embodiment of the present invention. Figure 2 These are schematic diagrams of the primitive pediatric skull and the pediatric skull closed by two-dimensional morphology, where (a) is the primitive pediatric skull and (b) is the binary mask of the complete skull obtained by two-dimensional morphological closure. Figure 3 This is a schematic diagram of the method for determining the Level 0 and Level 3 planes, where (a) is a schematic diagram of the division of Level 0 with other planes, and (b) is a schematic diagram of the establishment of the corrected coordinate axes; Figure 4 These are the front-to-back diameter, left-to-right diameter, and diagonal, where (a) is a schematic diagram of the front-to-back diameter, left-to-right diameter, and diagonal, and (b) is a diagram of the system output. Figure 5 It is an assessment report output by an automatic assessment method for cranial morphology based on pediatric CT images. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, a specific embodiment of the present invention provides an automatic assessment method for cranial morphology based on pediatric CT images, which specifically includes the following steps: S1: Acquire raw pediatric head CT images ; S2: Using a pre-trained deep learning keypoint detection model, identify raw pediatric head CT images. The system outputs the voxel coordinates in the image space of three anatomical landmarks: the root of the nose (SE), the left tragus (TR_L), and the right tragus (TR_R).

[0018] In the specific implementation of this application, the pre-trained deep learning keypoint detection model is Med-YOLO. The training method of the pre-trained deep learning keypoint detection model is to use a batch of CT image data that are manually and precisely annotated by senior radiologists as the standard answer (i.e., the "gold standard") to teach the model to learn.

[0019] S3: Based on the voxel coordinates of the nasal root point, left tragus point, and right tragus point, a standard anatomical coordinate system is constructed. The original pediatric head CT images are then transformed using affine transformation. Redirecting to this coordinate system yields the corrected head image; S3.1: Based on the coordinates of the left tragus (TR_L) and the right tragus (TR_R), calculate the midpoint M of the two tragus points, and use it as the origin of the standard anatomical coordinate system, represented as: (1) in: The midpoint between the two traguss, Let's define the coordinates of the left tragus. Here are the coordinates of the right ear stigma point.

[0020] S3.2: Construct a standard anatomical coordinate system based on three anatomical landmarks, including the Y-axis vector. Z-axis vector X-axis vector ; The specific steps in S3.2 are as follows: Define the Y-axis vector , used to represent the anteroposterior axis of the skull, is represented as: (2) in: The Y-axis vector For normalization, Here are the coordinates of the nasal root point.

[0021] Define auxiliary vector , represented as: (3) in: This is an auxiliary vector.

[0022] Define the Z-axis vector , used to represent the vertical axis of the skull, is represented as: (4) in: It is the Z-axis vector.

[0023] Define the X-axis vector , used to represent the lower left axis of the skull, is represented as: (5) like Figure 3 The diagram shown illustrates the method for determining the measurement plane and coordinate axes in this invention. Figure 3 (b) shows a standard anatomical coordinate system constructed based on the root of the nose (SE), the left tragus (TR_L), and the right tragus (TR_R), where point M is the midpoint between the two tragus points and serves as the origin of the coordinate system. The Y-axis points towards the root of the nose, the Z-axis is perpendicular to the plane formed by the line connecting the Y-axis and the tragus points, the X-axis is determined by the cross product of the Y-axis and the Z-axis, A is a virtual reference point, and M is the midpoint between the two tragus points.

[0024] S3.3: For the Y-axis vector Z-axis vector X-axis vector Normalize, and use the normalized X-axis vector Normalized Y-axis vector The normalized Z-axis vector Construct rotation matrix , represented as: (6) in: This is the normalized X-axis vector. The normalized Y-axis vector. The normalized Z-axis vector. It is a rotation matrix; S3.4: Using a rotation matrix Construct an affine transformation matrix, as shown in Equation (7), to transform the original pediatric head CT image. Reprojecting the image onto the standard anatomical coordinate system yields the uprighted skull image. This is to eliminate positional deviations during the scanning of the child.

[0025] (7) in: This is a head image after the procedure was completed. Original pediatric head CT images, Original pediatric head CT images The rotation center coordinates are used to move the original image to the origin to perform the rotation transformation. It is a rotation matrix.

[0026] In a specific embodiment of the present invention, the rectified head image The patient's head is positioned in a standard midline position, eliminating postural deviations. This step resolves the problem of inaccurate selection of measurement slices caused by the child's tilted position.

[0027] S4: Perform three-dimensional connected component extraction and morphological closure operations on the corrected skull image to obtain a binary mask of the complete skull. S4.1: Using the nasal root point, left tragus point, and right tragus point detected in S2 as seed points, set a threshold range. ,in, The average gray level of the image. The standard deviation σ of the noise level is used to extract connected components in three-dimensional space to obtain the main region of the child's skull, such as... Figure 2 As shown in (a), there is a gap, which will affect the subsequent measurement results; In the specific embodiments of this application, the threshold range is set based on the following considerations: 1) Basis for setting the lower limit of -600: In CT values, air is approximately -1000 HU, fat is approximately -100 to -50 HU, and water is 0 HU. Setting the lower limit to -600 HU effectively captures the outer contour of the head skin. Due to insufficient skull calcification in infants and young children, areas of weak bone often appear fractured in CT images. Extracting only the high-value bone portion will lead to segmentation failure. Relaxing the lower threshold to -600 HU allows the skin, soft tissue, and skull to be extracted as a connected whole, effectively avoiding segmentation discontinuities caused by cranial sutures. On the other hand, the -600 HU setting filters out random noise from the air portion of the image background. Compared to starting growth directly from -1000 HU (air), it significantly reduces the computational cost of the region growing algorithm and prevents the growth process from "over-diffusion" into air areas outside the field of view.

[0028] 2) Basis for setting the upper limit value Mean+2σ: Differences in CT equipment manufacturers (such as GE, Siemens, Philips) and scanning doses can cause fluctuations in the average gray level (Mean) and noise level (standard deviation σ) of images. Using a fixed upper limit threshold (such as Mean+1000) may fail in low-dose scan images. Adaptive statistical values ​​are used instead. As an upper limit, it can be dynamically adjusted according to the actual grayscale distribution of each image, eliminating extreme high-value points caused by artifacts or noise, making the edges of the segmented head body smoother, and avoiding abnormal "burrs" caused by artifacts from affecting the accuracy of subsequent measurements.

[0029] In summary, the threshold range is set in a way that takes into account the physiological characteristics of children's skulls, image noise suppression, and adaptability to different equipment, thus ensuring the accuracy and robustness of skull segmentation.

[0030] S4.2: For the extracted main area of ​​the pediatric skull, perform a two-dimensional morphological closure operation layer by layer, as shown in Equation (8), to obtain a binary mask of the complete skull, as shown in Equation (8). Figure 2 As shown in (b).

[0031] (8) in: This is a complete binary mask of the skull obtained after the closing operation. The original two-dimensional axial slice binary image to be processed. This is a dilation operator, which expands the highlighted areas (i.e., the skull) in an image outwards, connecting adjacent fractured bone points. A structuring element is a "probe" or operator in morphological operations (such as a circular operator with radius r). Its size determines the width of the crack that the algorithm can automatically stitch / fill. The erosion operator is responsible for shrinking the edges inward after expansion, thus offsetting the increase in size and restoring the original thickness of the skull.

[0032] The reason for adopting a "3D followed by 2D" composite strategy in S4.1 and S4.2 is that, firstly, the raw CT data obtained clinically is itself a three-dimensional volumetric image, and the three key anatomical points (nasal root point, left tragus point, and right tragus point) are not located in the same axial slice in space, making it impossible to locate them simultaneously on a single two-dimensional plane. Therefore, it is necessary to rely on three-dimensional space for key point detection and the construction of an anatomical coordinate system. At the same time, the determination of the Level 3 measurement plane also needs to be completed in three-dimensional morphology for spatial localization. However, the feature parameters ultimately used for clinical diagnosis—anteroposterior diameter (AP), lateral diameter (LR), and diagonal length—are essentially two-dimensional geometric quantities defined on a specific axial plane (i.e., the Level 3 plane) under standard anatomical position. Therefore, after completing the three-dimensional spatial normalization and measurement plane localization, subsequent calculations revert to the two-dimensional level, which not only conforms to clinical measurement standards but also simplifies the complexity of feature extraction.

[0033] This step aims to address the issue of insufficient skull calcification in children (where weak bone appears fractured on CT scans) and ensure that the skull mask is a closed ring structure during subsequent Level 3 plane extraction.

[0034] S5: Based on the binary mask of the skull, the localization of the pediatric skull measurement plane and the extraction of the skull contour are performed to obtain the collision point set; S5.1: Determine the Level 0 plane based on the nasal root point, left tragus point, and right tragus point; S5.2: Using a complementary strategy of roundness detection and volume centroid method, identify the top direction of the child's head, determine the physical position of the top of the head along the determined top direction, construct a plane parallel to the Level 0 plane, and determine this plane as the Level 10 top plane; In a specific embodiment of the present invention, from the original pediatric cranial CT images Three core anatomical landmarks are identified: the root of the nose (SE), the left tragus (TR_L), and the right tragus (TR_R). The physical plane determined by these three points is the Level 0 plane. Then, a plane parallel to Level 0 and passing through the top of the head is defined as Level 10. The vertical distance from Level 0 to Level 10 is divided into 10 equal parts, and the third equal part is taken as Level 3.

[0035] In a specific embodiment of the present invention, the specific steps for identifying the direction of the top of a child's head using a complementary strategy of roundness detection and volume centroid method described in S5.2 are as follows: S5.2.1: Extracting raw pediatric head CT images In the non-background region of the whole slice sequence, determine the Z-axis span of the brain tissue; S5.2.2: Extract slice images at 1 / 4 and 3 / 4 of the span respectively, perform edge detection and cranial ring repair on the extracted images to obtain the outermost closed contour of the skull; In a specific embodiment of this application, the slice image at 1 / 4 is located above the eye socket and above the skull base, while the slice image at 3 / 4 is close to the skull base and facial bones (such as the nasal bone, eye socket, and sphenoid bone).

[0036] S5.2.3: The roundness values ​​of the two slice images are calculated using the roundness detection algorithm, as shown in Equation (9): (9) in: The area of ​​the two-dimensional region enclosed by the outermost closed contour of the skull extracted by Otsu thresholding in the target slice. This represents the perimeter of the closed contour. The perimeter is obtained by the algorithm through calculating the sum of the Euclidean chain code lengths of the contour pixel sequence. For roundness.

[0037] S5.2.4: Compare the roundness values ​​of the two slice images. If the slice image at 1 / 4 is close to a regular circle or ellipse, while the slice image at 3 / 4 is irregular, then the lower index direction is determined to be the top, and Level10 is located at the lower index end of the slice sequence; otherwise, the higher index direction is determined to be the top, and Level10 is located at the higher index end of the slice sequence.

[0038] S5.2.5: If the shape features of the slice image are not significant, the volume centroid method is used to determine the direction of the top of the child's head, specifically: S5.2.6: Calculate the z-axis geometric centroid of all brain tissue slices, compare the preset Level 0 z-coordinate with the z-axis geometric centroid. If the Level 0 z-coordinate is greater than the z-axis geometric centroid, then Level 10 is located at the low index end of the slice sequence; otherwise, Level 10 is located at the high index end of the slice sequence.

[0039] S5.3: Divide the vertical distance between Level 0 and Level 10 into 10 equal parts using linear interpolation, and calculate the z-coordinate of the Level 3 plane, as shown in equation (10): (10) in: The coordinates on the z-axis in the 3D image after Level 3 conversion. The coordinates on the z-axis in the 3D image after Level 0 is corrected. The coordinates on the z-axis in the 3D image after Level 10 conversion; like Figure 3 The diagram shown illustrates the method for determining the measurement plane and coordinate axes in this invention. Figure 3 (a) shows the equal division between the Level 0 plane and the Level 10 top plane. The Level 3 plane is located at a distance of 30% from Level 0 towards the top of the head. SE is the root of the nose point, and t is the vertical distance parameter related to the plane division.

[0040] S5.4: Using the midpoint of both tragus Centered on the projection of the Level 3 plane, rays are emitted in multiple directions. Each ray extends outward with a fixed step size until it contacts the complete binary skull mask generated by S4.2. The collision point positions are recorded and the collision point set is extracted. In a specific embodiment of the present invention, step S5.4 is as follows: S5.4.1: Set the sampling angle range θ∈[0,360°), and emit rays from the center point along multiple preset angles, specifically: 90° and 270° as front-to-back diameters, 0° and 180° as left-to-right diameters, 60° and 240° as diagonal 1, and 120° and 300° as diagonal 2; S5.4.2: Perform a collision point search for each ray; The specific steps in S5.4.2 are as follows: S5.4.2.1: For each angle... The ray, setting the initial distance scalar ; S5.4.2.2: Calculate the unit direction vector As shown in equation (11), the coordinates of the current sampling point are updated. As shown in equation (12): (11) in: is the unit direction vector, and i is the ray index sequence number; (12) in: This represents the x-coordinate of the current sampling point in the real-time voxel of the Level 3 axis plane. This represents the ordinate of the current sampling point in the real-time voxel on the Level 3 axis plane. for The x-coordinate of the orthographic projection onto the Level 3 plane. for The ordinate of the orthographic projection onto the Level 3 plane. To start from the origin The cumulative distance scalar along the ray direction from the starting point.

[0041] S5.4.2.3: Query the value of the binary mask of the complete skull at the current sampling point (x,y) Mask(x,y). If Mask(x,y)=1, record this point as the first collision point and terminate the extension of the current ray; if Mask(x,y)=0, continue stepping, the calculation method is as shown in Equation (13), and repeat S5.4.2.2-S5.4.2.3 until each ray finds a collision point; (13) in: The step distance; S5.4.3: Collect all collision points of the rays to form a collision point set.

[0042] S6: Convert the pixel distance of each collision point in the collision point set into physical length, calculate the cranial asymmetry index and cranial proportion, and then compare it with a predefined clinical grading scale to generate the assessment results.

[0043] S6.1: Set the collision points The pixel distance between any two corresponding points is converted into physical length to obtain the preset skull feature value; The preset skull feature values ​​are: anteroposterior diameter (AP), lateral diameter (LR); long diagonal, and short diagonal.

[0044] like Figure 4 (a) is a schematic diagram of the geometric definition of cranial feature values ​​in this invention. On the Level 3 axial plane, with the projection of the midpoint M of the two traguss as the center, the straight-line distance between two points obtained by emitting rays in the 0° and 180° directions and colliding with the inner table of the skull is defined as the left-right diameter (LR); the straight-line distance between two points obtained by emitting rays in the 90° and 270° directions and colliding with the skull is defined as the anteroposterior diameter (AP); the straight-line distance between two points obtained by emitting rays in the 60° and 240° directions and colliding with the skull is defined as one diagonal, and the straight-line distance between two points obtained by emitting rays in the 120° and 300° directions and colliding with the skull is defined as another diagonal, wherein the longer one is defined as the long diagonal and the shorter one is defined as the short diagonal. Figure 4 (b) is an example diagram of the measurement results automatically drawn and output by the system based on the actual collision point set.

[0045] S6.2: Based on the preset skull feature values ​​obtained in S6.1, compare them with the predefined clinical grading scale to generate the assessment results.

[0046] Clinical diagnostic indicators were calculated and graded according to a standard scale to obtain the cranial morphology assessment results.

[0047] The specific steps are as follows: S6.2.1: The cervical asymmetry index (CVAI) is calculated as shown in Equation (14), and the cervical proportion (CR) is calculated as shown in Equation (15): (14) in: The long diagonal It is the short diagonal. The index of skull asymmetry; CR = LR / AP (15) Where: LR is the left-right diameter, AP is the anteroposterior diameter, and CR is the cranial proportion.

[0048] S6.2.2: Compare the calculated CVAI value with the predefined plagiocephaly grading scale to determine the severity of the plagiocephaly abnormality; at the same time, compare the calculated CR value with the predefined brachycephaly / scaphoid grading scale to determine the abnormal type and grade of the skull length-width ratio.

[0049] Head tilt assessment: Determined according to the CVAI range shown in Table 1 (e.g., <3.5% normal, >11.00% extremely severe).

[0050] Table 1 Grading Scale for Plagiocephaly

[0051]

[0052] Short head / scaphoid head evaluation: determined based on CR value (e.g., <0.74 for scaphoid head, ≥1.0 for severe short head).

[0053] Table 2 Short Head / Scaphoid Head Assessment Scale

[0054]

[0055] Example 1: This example uses a 2-year-old male child's head CT image as an example to automatically evaluate the data using the method of this invention. The specific process is as follows: First, the original head CT image of the child is acquired; using a pre-trained deep learning key point detection model, the voxel coordinates of the nasal root point, left tragus point, and right tragus point are automatically identified; a standard anatomical coordinate system is constructed with the midpoint of the left and right tragus points as the origin, and the original image is corrected through affine transformation to eliminate positional deviation; using the three points as seed points, an adaptive threshold range is set to extract the three-dimensional connected components to obtain the main area of ​​the child's head; then, two-dimensional morphological closure is performed layer by layer to obtain a binary mask of the complete skull; then, the Level 0 plane is determined by the three points; then, a complementary strategy of roundness detection and volume centroid method is used to identify the top direction of the head to determine the Level 10 plane; the Level 3 plane is calculated through linear interpolation; the midpoint M of the two tragus points is used to determine the Level 10 plane; and the Level 20 plane is determined by the midpoint M of the two tragus points. Centered on the projection of the 3-plane, rays are emitted in multiple directions to search for collision points with the binary mask of the skull. All collision points are collected to form a collision point set. The pixel distances of the collision point set are converted into physical lengths, and the radial lengths in preset directions are extracted. The skull asymmetry index and skull proportions are calculated. The calculation results are compared with a predefined clinical grading scale to generate a final automated assessment report. Figure 5 As shown.

[0056] It can be seen that the anteroposterior diameter (90° direction) of the skull of the 2-year-old male patient is 129.88 mm, the lateral diameter (0° / 180° direction) is 109.57 mm, and the long diagonal (60° / 240° direction) is... =126.37mm, short diagonal (120° / 300° direction) =134.57mm, calculated cranial asymmetry index CVAI=6.1%, cranial proportion CR=0.84, compared with Table 1, CR compared with Table 2, the assessment result is that there is mild cranial asymmetry.

[0057] Furthermore, in terms of detection efficiency, traditional manual observation of a single imaging examination takes 15-30 minutes, physical measurement (head shape ruler / soft ruler) takes 10-20 minutes, and laser measurement (equipment debugging and pressure processing) takes 5-10 minutes. This invention, however, requires only one click to upload, automatically generating a report within 10 seconds, significantly shortening the detection time. Secondly, regarding measurement accuracy, using manual observation by experienced radiologists as the gold standard, the mean square error (MSE) of traditional physical measurements for CVAI and CR is 0.1-0.2, while the MSE of the output results of this invention is only 0.03-0.06, demonstrating a significant improvement in accuracy.

[0058] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. An automatic assessment method for cranial morphology based on pediatric CT images, characterized in that, Includes the following steps: S1: Acquire raw pediatric head CT images ; S2: Using a pre-trained deep learning keypoint detection model, identify raw pediatric head CT images. Voxel coordinates of the middle nasal root point, left tragus point, and right tragus point; S3: Based on the voxel coordinates of the nasal root point, left tragus point, and right tragus point, a standard anatomical coordinate system is constructed, and the original pediatric head CT images are transformed using affine transformation. Redirecting to this coordinate system yields the corrected head image; S4: Perform three-dimensional connected component extraction and morphological closure operations on the corrected skull image to obtain a binary mask of the complete skull. S5: Based on a binary mask of the complete skull, perform planar localization and skull contour extraction for pediatric skull measurement to obtain a collision point set; S6: Convert the pixel distance of each collision point in the collision point set into physical length, calculate the cranial asymmetry index and cranial proportion, and then compare it with a predefined clinical grading scale to generate the assessment results.

2. The method for automatic assessment of cranial morphology based on pediatric CT images according to claim 1, characterized in that, The specific steps for S3 are as follows: S3.1: Calculate the midpoint M of the two tragus points based on the coordinates of the left and right tragus points, and use it as the origin of the standard anatomical coordinate system; S3.2: Construct a standard anatomical coordinate system based on three anatomical landmarks, including the Y-axis vector. Z-axis vector X-axis vector ; S3.3: For the Y-axis vector Z-axis vector X-axis vector Normalize and construct the rotation matrix. ; S3.4: Using a rotation matrix Construct an affine transformation matrix, as shown in Equation (1), to transform the original pediatric head CT image. Reprojecting the image onto the standard anatomical coordinate system yields the uprighted skull image. ; （1） in: This is a head image after the procedure was completed. Original pediatric head CT images, Let the coordinates be the center of rotation. It is a rotation matrix.

3. The method for automatic assessment of cranial morphology based on pediatric CT images according to claim 1, characterized in that, The specific steps for S4 are as follows: S4.1: Using the nasal root point, left tragus point, and right tragus point detected by S2 as seed points, and setting an adaptive threshold range. ,in, The average gray level of the image. The standard deviation of the noise level is used to extract the connected components in three-dimensional space to obtain the main region of the child's skull; S4.2: Perform two-dimensional morphological closure operations layer by layer on the extracted main area of ​​the child's skull to obtain a binary mask of the complete skull.

4. The method for automatic assessment of cranial morphology based on pediatric CT images according to claim 1, characterized in that, The specific steps for S5 are as follows: S5.1: Determine the Level 0 plane based on the nasal root point, left tragus point, and right tragus point; S5.2: Using a complementary strategy of roundness detection and volume centroid method, the top direction of the child's head is identified, the physical position of the top of the head is determined along the determined top direction, a plane parallel to the Level 0 plane is drawn, and this plane is determined as the Level 10 top plane. S5.3: Divide the vertical distance between the Level 0 plane and the Level 10 overhead plane into 10 equal parts using linear interpolation, and calculate the z-coordinate of the Level 3 plane; S5.4: Using the midpoint of both tragus Centered on the projection on the Level 3 plane, rays are emitted in multiple directions to detect collision points with the binary mask of the skull and extract the collision point set.

5. The method for automatic assessment of cranial morphology based on pediatric CT images according to claim 4, characterized in that, The specific steps for identifying the top direction of a child's head using a complementary strategy of roundness detection and volume centroid method, as described in S5.2, are as follows: S5.2.1: Extracting raw pediatric head CT images In the non-background region of the whole slice sequence, determine the Z-axis span of the brain tissue; S5.2.2: Extract slice images at 1 / 4 and 3 / 4 of the span respectively, perform edge detection and cranial ring repair on the extracted images to obtain the outermost closed contour of the skull; S5.2.3: Calculate the roundness value of the two slice images using a roundness detection algorithm; S5.2.4: Compare the roundness values ​​of the two slice images. If the slice image at 1 / 4 is close to a regular circle or ellipse, while the slice image at 3 / 4 is irregular, then the lower index direction is determined to be the top of the head; otherwise, the higher index direction is determined to be the top of the head. S5.2.5: If the judgment in S5.2.4 fails, calculate the geometric centroid of the z-axis for all slices containing brain tissue. S5.2.6: Compare the preset z-coordinate of Level 0 with the geometric centroid of the z-axis. If the z-coordinate of Level 0 is greater than the geometric centroid of the z-axis, then Level 10 is located at the low index end of the slice sequence; otherwise, Level 10 is located at the high index end of the slice sequence.

6. The method for automatic assessment of cranial morphology based on pediatric CT images according to claim 4, characterized in that, The specific steps in S5.4 are as follows: S5.4.1: Set the sampling angle range θ∈[0,2π), and emit rays from the center point along multiple preset angles; S5.4.2: Perform a collision point search for each ray; S5.4.3: Collect all collision points of the rays to form a collision point set.

7. The method for automatic assessment of cranial morphology based on pediatric CT images according to claim 6, characterized in that, The specific steps in S5.4.2 are as follows: S5.4.2.1: For each angle... The ray, setting the initial distance scalar ; S5.4.2.2: Calculate the unit direction vector As shown in equation (2), the coordinates of the current sampling point are updated. As shown in equation (3): （2） in: It is a unit direction vector. For ray index sequence number; （3） in: This represents the x-coordinate of the current sampling point in the real-time voxel of the Level 3 axis plane. This represents the ordinate of the current sampling point in the real-time voxel on the Level 3 axis plane. Midpoint of both tragus The x-coordinate of the orthographic projection onto the Level 3 plane. Midpoint of both tragus The ordinate of the orthographic projection onto the Level 3 plane. To start from the origin The cumulative distance scalar along the ray direction from the starting point; S5.4.2.3: Query the value of the binary mask of the complete skull at the current sampling point (x,y) Mask(x,y). If Mask(x,y)=1, record this point as the first collision point and terminate the extension of the current ray; if Mask(x,y)=0, continue stepping and repeat S5.4.2.2-S5.4.2.3 until each ray finds a collision point.

8. The method for automatic assessment of cranial morphology based on pediatric CT images according to claim 1, characterized in that, The specific steps for S6 are as follows: S6.1: Set the collision points The pixel distance is converted into physical length to obtain the preset skull feature value; The preset skull feature values ​​are: anteroposterior diameter, lateral diameter, long diagonal, and short diagonal; S6.2: Based on the preset skull feature values ​​obtained in S6.1, calculate the skull asymmetry index and skull proportion, and then compare them with the predefined clinical grading scale to generate the assessment results.

Images