Vertebral CT image standardization preprocessing method and system fusing multi-parameter features

Abstract

The invention provides a multi-parameter feature fused vertebral CT image standardization preprocessing method and system. The method comprises the following steps: carrying out initial segmentation on an input original vertebral CT image; carrying out enhancement processing on the image by adopting an anisotropic diffusion filtering algorithm of multi-scale structure perception; establishing a tissue specificity density mapping model of multi-parameter fusion; adopting a self-adaptive segmentation histogram mapping method to map gray distribution of an original image to a standard template; and constructing a rigid-elastic mixed registration model. According to the invention, a complete quality evaluation system is established, the processing quality can be automatically detected and reported, and the reliability of an output result is ensured. By introducing a patient individualized correction mechanism, individual difference information is fully reserved while the standardization effect is ensured, and diagnosis information loss caused by excessive standardization is avoided. The multi-parameter fusion strategy ensures complete reservation of bone tissue density, texture and structure information.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a standardized preprocessing method and system for vertebral CT images that integrates multi-parameter features. Background Technology

[0002] With the accelerating aging of the population and the continuous rise in the incidence of orthopedic diseases, vertebral CT imaging plays an increasingly important role in clinical diagnosis. CT imaging technology, with its high spatial resolution, three-dimensional reconstruction capabilities, and excellent visualization of bone tissue, has become the preferred imaging method for vertebral fracture diagnosis, spinal deformity assessment, osteoporosis screening, and surgical planning. However, in actual clinical applications, different medical institutions are equipped with various CT equipment models, including products from multiple manufacturers such as GE, Siemens, and Philips. Significant differences exist in the age of the equipment, scanning parameter settings, and reconstruction algorithm selection, resulting in obvious heterogeneity in the image quality and presentation of the acquired vertebral CT images.

[0003] This image heterogeneity manifests itself in several ways. First, there is inconsistency in grayscale distribution; the same bone tissue may exhibit grayscale values ​​differing by hundreds of HU on different devices, severely impacting the quantitative assessment of bone mineral density. Second, there are differences in contrast; images acquired by some devices lack sufficient contrast between bone and soft tissue, resulting in blurred vertebral body boundaries and unclear trabecular bone structures. Third, there are differences in spatial resolution; different devices and scanning protocols result in variations in slice thickness and pixel spacing, leading to significant differences in image detail. Furthermore, differences in image noise levels directly affect image quality; high noise not only reduces visual appeal but also interferes with subsequent automated analysis algorithms.

[0004] In recent years, artificial intelligence (AI) technology has made groundbreaking progress in the field of medical image analysis, with algorithms based on deep learning for automatic vertebral segmentation, fracture detection, and osteoporosis assessment emerging one after another. However, these AI models generally face a serious challenge: models trained on a single data source often experience a significant performance drop when applied to image data from other hospitals or equipment. Research indicates that the root cause of this performance degradation lies in the distributional differences between training and application data, primarily stemming from the aforementioned image heterogeneity problem. To enable AI models to operate stably in multi-center environments, effective standardization preprocessing of the input CT images is urgently needed.

[0005] Therefore, developing a standardized preprocessing technique for vertebral CT images that can comprehensively consider multidimensional image features, take into account the characteristics of different tissue types, and incorporate individualized patient information has become a key technological requirement. This technique must not only address the consistency issue at the image level but also ensure the complete preservation of clinically relevant information, providing high-quality standardized input for subsequent intelligent analysis. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes a standardized preprocessing method and system for vertebral CT images that integrates multi-parameter features. The entire standardization process is highly automated, with the complete processing time for a single vertebra taking only a few seconds, and it supports batch processing. Compared to traditional manual adjustment methods, the processing efficiency is improved by more than an order of magnitude. The system supports GPU parallel acceleration, enabling real-time processing on workstations equipped with high-performance graphics cards, meeting the timeliness requirements of practical clinical applications. To achieve the above objectives, this invention proposes a standardized preprocessing method for vertebral CT images that integrates multi-parameter features, comprising the following steps:

[0007] S1: Initial segmentation is performed on the input raw vertebral CT image. Coarse segmentation based on region growth and anatomical prior is used, selecting voxels with gray values ​​higher than +200 HU in the central region of the vertebral body as initial seed points. The gray value difference range of the growth threshold is limited to ±150 HU. A graph-based prior modeling strategy is introduced, constructing key anatomical points of the vertebral body as nodes in a graph neural network. Accurate segmentation is achieved by minimizing the energy function of the vertebral body topological constraints.

[0008]

[0009] in:

[0010] E topo Total energy value of cone topological constraints

[0011] V: A set of vertebral body anatomical landmarks, including key anatomical points such as the vertebral body center, endplate angle, and spinous process tip. v: A single anatomical landmark within set V.

[0012] φ(v): Single-point potential function of node v, measuring the degree of matching between the local features of this point and the vertebral body model. E: Set of connecting edges constructed based on anatomical relevance, representing the spatial relationship between anatomical points. (u,v): Edge connecting nodes u and v. ψ(u,v): Paired potential function of edge (u,v), measuring the strength of the spatial constraint between two anatomical points. d anatomy (u,v): Anatomical distance metric function, combining Euclidean distance and geodesic distance, used to quantify the anatomical correlation between two anatomical points.

[0013] S2: An anisotropic diffusion filtering algorithm with multi-scale structure awareness is used to enhance the image. By calculating the structure tensor at different scales, the local structure type is analyzed, and the diffusion direction and amplitude are dynamically adjusted. The number of iterations is set to 10-20, and the stopping criterion is that the mean square error of the image between two iterations is less than 10^-4.

[0014] S3: Establish a multi-parameter fusion tissue-specific density mapping model, and achieve accurate density normalization through the following formula:

[0015]

[0016] in:

[0017] HU norm (x,y,z): Normalized CT value at spatial location (x,y,z);

[0018] i: Tissue type index, i=1 represents cortical bone, i=2 represents cancellous bone, i=3 represents trabecular bone;

[0019] w i The weight coefficients of the i-th type of organization are obtained based on local structure tensor analysis and satisfy the following conditions: HU raw Raw CT values, in HU.

[0020] f i : The specific nonlinear mapping function ρ of the i-th type of tissue i : Density mapping intensity coefficient of the i-th type of tissue, controlling the magnitude T of the density mapping. i : Texture preservation parameter for the i-th type of organization, controlling the degree of preservation of texture details G patient Patient-specific correction function

[0021] Age: Patient's age, in years

[0022] Sex: Patient gender

[0023] Patient-specific correction function G patient Combining the Gaussian decay of age with the sinusoidal interaction effect of gender-age:

[0024]

[0025] in:

[0026] α: Age-related correction magnitude coefficient, obtained through regression analysis of large-scale clinical data, with a typical range of 0.1-0.3. τ: Standard deviation parameter of the age effect, controlling the decay rate of the age effect, in years, with a typical value of 15-20 years. β: Magnitude coefficient of the sex-age interaction effect, with a typical range of 0.05-0.15.

[0027] Sex code : Gender code, 1 for male and 0 for female

[0028] Age - 50: Age deviation value centered around 50 years old

[0029] Periodic function, reflecting the age - related physiological cycle changes. S4: Using an adaptive piece - wise histogram mapping method, map the original image gray - level distribution to a standard template. The mapping function uses sigmoid for smooth transition at the piece - wise boundaries to ensure the continuity of gray - level transformation:

[0030] When g0 ≤ g < g2:

[0031] T(g) = T1(g)·(1 - σ(g - g1))+T2(g)·σ(g - g1)

[0032] When g1 ≤ g < g3:

[0033] T(g) = T2(g)·(1 - σ(g - g2))+T3(g)·σ(g - g2)

[0034] When g2 ≤ g ≤ g max :

[0035] T(g) = T3(g)

[0036] Where:

[0037] T(g): The output value after mapping of the gray - level value g. g: Input gray - level value

[0038] T1(g), T2(g), T3(g): Local mapping functions for three gray - level intervals. g0, g1, g2, g3, g max : Piece - wise points determined adaptively, dividing the gray - level range into three intervals:

[0039] [g0, g1]: The main active interval of the first mapping function T1

[0040] [g1, g2]: The transition and mixing interval of T1 and T2

[0041] [g2, g3]: The transition and mixing interval of T2 and T3 <000​​​​​​​​

[0044] S5: Construct a rigid-elastic hybrid registration model. First, perform three-dimensional affine transformation by extracting anatomical anchor points such as the vertebral body center, posterior edge of the vertebral body, and spinous process to achieve global alignment. Then, use B-spline deformation field to describe local elastic deformation. Achieve accurate registration by minimizing the energy function containing data terms and regularization terms. In this process, bone tissue preservation constraints are introduced to ensure the fidelity of bone density information during registration.

[0045] Furthermore, in step S1, the segmented binary mask undergoes 2-3 morphological closing operations through a spherical structural element with a radius of 5mm to fill internal holes and smooth boundaries; the graph neural network adopts a multi-layer graph convolution structure and is trained on more than 500 manually labeled data.

[0046] Furthermore, in step S3, the organization type weight w i The KL divergence between the local grayscale distribution and the reference distribution of each tissue is calculated and then normalized by softmax. The cortical bone is defined as having a CT value greater than 700 HU, the cancellous bone is defined as having a CT value between 150 and 700 HU, and the trabecular bone is defined as having a CT value between 150 and 400 HU.

[0047] Furthermore, the tissue-specific mapping function f in step S3 i Defined as:

[0048]

[0049] in:

[0050] tanh: Hyperbolic tangent function, used to achieve nonlinear density compression μ. i : The average gray value of the i-th type of tissue, in HU

[0051] σ i : Standard deviation of gray level for the i-th type of organization, in HUλ i : Adjustment parameters for logarithmic transformation, controlling the degree of texture enhancement.

[0052] Furthermore, the parameters are set as follows for different tissue types:

[0053] Bone cortex (i=1): μ1=900HU, σ1=150HU, ρ1=0.8, T1=0.2, λ1=500

[0054] Cancellous bone (i=2): μ2=400HU, σ2=100HU, ρ2=0.6, T2=0.4, λ2=300

[0055] Trabecular bone (i=3): μ3=200HU, σ3=50HU, ρ3=0.4, T3=0.6, λ3=150.

[0056] Furthermore, in step S4, the standard reference template is constructed through statistical analysis of high-quality typical cases; a smoothing regularization mechanism is introduced during the mapping process to prevent overfitting of the mapping function; and the gray-level distribution of the original image is gradually made to approximate the standard template through iterative optimization.

[0057] Furthermore, in step S5, the anatomical anchor points are automatically extracted using the 3D-SIFT feature detection algorithm; the flexible registration adopts a multi-resolution strategy, gradually optimizing from coarse to fine; and the registration accuracy is evaluated using normalized mutual information as a similarity metric.

[0058] Furthermore, it also includes an image quality assessment step, which detects the signal-to-noise ratio, resolution, and integrity of the input image before processing to ensure that it meets the preset requirements; and calculates the signal-to-noise ratio of bone tissue, inter-tissue contrast, and structural similarity index after processing to evaluate the standardization effect.

[0059] Furthermore, the method supports batch processing mode, establishing spatiotemporal consistency constraints for multi-sequence, multi-temporal vertebral CT images to ensure standardized consistency of images from different periods of the same patient; it includes an adaptive parameter learning mechanism that automatically adjusts the parameter settings of each processing module according to the device type, scanning protocol, and image features of the input image. A vertebral CT image standardization preprocessing system integrating multi-parameter features, applicable to the above method, includes: an initial image segmentation module for implementing the function of step S1 in claim 1, comprising a seed point selection unit, a region growing unit, a graph neural network unit, and an energy optimization unit;

[0060] The image enhancement module, used to implement the function of step S2 in claim 1, includes a multi-scale analysis unit, a structure tensor calculation unit, and an anisotropic filtering unit.

[0061] The density modeling and normalization module is used to implement the function of step S3 in claim 1, and includes a tissue classification unit, a multi-parameter fusion mapping unit and an individualized correction unit.

[0062] The grayscale distribution mapping module is used to implement the function of step S4 in claim 1, and includes an adaptive segmentation unit, a smooth transition unit, and an iterative optimization unit.

[0063] The spatial registration module is used to implement the function of step S5 in claim 1, and includes an anchor point extraction unit, a rigid transformation unit and an elastic deformation unit.

[0064] The quality control module is used to assess the quality of input and output images and adaptively adjust parameters. The modules are interconnected through standardized data interfaces, supporting pipelined processing and parallel computing acceleration.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] 1. This invention provides a standardized preprocessing method and system for vertebral CT images that integrates multi-parameter features, significantly improving the consistency of multi-center, multi-device vertebral CT images. The variability in grayscale distribution, standard deviation, and texture features of similar vertebral images from different devices is greatly reduced. The difference in bone mineral density measurements at the same anatomical location is reduced from ±45 HU to ±12 HU, greatly improving the accuracy and repeatability of quantitative analysis. The image signal-to-noise ratio is significantly improved, and the visualization of fine structures such as trabeculae is significantly enhanced, providing clearer imaging evidence for clinical diagnosis.

[0067] 2. This invention provides a standardized preprocessing method and system for vertebral CT images that integrates multi-parameter features. Standardization provides high-quality, consistent training data for deep learning models. In automatic vertebral segmentation tasks, the model's segmentation accuracy is significantly improved. More importantly, the model's generalization ability on cross-device test sets is greatly enhanced, effectively solving the problem of performance degradation when AI models are deployed in different hospitals. In tasks such as osteoporosis screening and fracture detection, both classification accuracy and detection sensitivity are substantially improved.

[0068] 3. This invention provides a standardized preprocessing method and system for vertebral CT images that integrates multi-parameter features. This technology effectively solves the data heterogeneity problem in multi-center collaborative studies, enabling direct comparison and joint analysis of image data from different hospitals. Radiologists no longer need to frequently adjust window width and level when interpreting images, significantly improving interpretation efficiency. Improved cross-device consistency in bone mineral density measurement makes patient examination results comparable across different hospitals, which is beneficial for long-term disease follow-up and efficacy evaluation. Attached Figure Description

[0069] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0070] Figure 1 This is a flowchart of the invention. Detailed Implementation

[0071] The technical solution of the present invention will be more clearly and completely explained below with reference to the accompanying drawings and through the description of preferred embodiments of the present invention.

[0072] like Figure 1As shown, the standardized preprocessing workflow for vertebral CT images of this invention adopts a five-step cascaded processing architecture to achieve a systematic transformation from raw images to standardized images. First, the raw vertebral CT image enters the initial image segmentation module (S1), where the initial contour of the vertebral body is determined using a region growing algorithm, and the anatomical topology of the vertebral body is modeled using a graph neural network. Accurate segmentation is achieved through energy function optimization. The segmented image then enters the image enhancement module (S2), which employs multi-scale structure-aware anisotropic diffusion filtering technology to effectively suppress noise while preserving bone tissue edges and texture details.

[0073] The enhanced image then enters the density modeling and normalization module (S3), which is the core of the entire process. Through a multi-parameter fusion mapping strategy, specific mapping functions are established for the three tissues: cortical bone, cancellous bone, and trabecular bone. Individualized correction factors such as patient age and gender are introduced to ensure the accuracy of density normalization. Next, the grayscale distribution mapping module (S4) uses adaptive segmented histogram technology to adjust the image's grayscale distribution to a standard template. Segment boundaries are smoothly transitioned using the sigmoid function to avoid abrupt grayscale changes.

[0074] Finally, the spatial registration module (S5) employs a rigid-elastic hybrid registration strategy. It first performs a global affine transformation to achieve overall alignment, then finely adjusts the alignment through local elastic deformation. Simultaneously, it introduces bone tissue preservation constraints to ensure that bone density information is not distorted during registration. The entire process is equipped with a quality assessment mechanism to monitor the processing results in real time and dynamically adjust the parameters of each module through feedback channels. This ensures that the output standardized CT images maintain the integrity of clinical diagnostic information while achieving high consistency of multi-center, multi-device image data, providing reliable standardized input for subsequent artificial intelligence analysis.

[0075] In one specific embodiment of the present invention, a workstation equipped with an NVIDIA RTX 3090 graphics card was used as the processing platform. Based on the Python 3.8 environment, the entire standardized preprocessing workflow was implemented using the PyTorch deep learning framework and the SimpleITK medical image processing library. The experimental data came from 1200 lumbar spine CT images acquired by the radiology department of a tertiary hospital between 2022 and 2024, covering three different models of equipment: GE Revolution CT, Siemens SOMATOM Force, and Philips BrillianceiCT. Scanning parameters included tube voltage of 120-140kV, slice thickness of 0.625-1.25mm, and reconstruction algorithms covering both standard and bone algorithms.

[0076] In the initial image segmentation stage, the raw CT images in DICOM format are first read and converted into three-dimensional volumetric data. By analyzing the anatomical features of the vertebral bodies in the sagittal plane, voxels with a gray value greater than 200 HU are searched near the geometric center of each vertebral body as seed points. Specifically, a 3×3×3 sliding window is used to traverse the central region of the vertebral body, and the position with the highest average gray value within the window is selected as the initial seed. During region growth, a growth threshold is set to ±150 HU of the average gray value of the currently grown region to avoid overgrowth into adjacent soft tissues. After growth is completed, three morphological closure operations are performed using spherical structuring elements with a radius of 5 mm to effectively fill low-density areas inside the vertebral body, such as the medullary cavity.

[0077] Twelve key anatomical points of the vertebral body (including the four corner points of the superior and inferior endplates, the midpoints of the anterior and posterior edges of the vertebral body, the roots of the left and right transverse processes, and the tips of the spinous processes) were constructed as graph nodes. Node features included position coordinates, local gray-level statistics, and gradient orientation histograms. The graph edges were constructed based on anatomical connectivity, such as the four corner points of the superior endplate forming rectangular constraints. A three-layer graph convolutional network was trained on a labeled dataset to learn the geometric constraint patterns of the vertebral body. In practical processing, the topological constraint energy function E was minimized. topo The segmentation boundary is optimized, where the single-point potential energy φ(v) is designed as a weighted sum of the local gray-scale consistency term and the Gaussian attenuation term of the distance to the center of the vertebral body, and the paired potential energy ψ(u,v) is predefined according to the type of anatomical point pair, such as the potential energy between the corner points of the endplate is set to 0.8, and the potential energy of the anterior and posterior edge points of the vertebral body is set to 0.6.

[0078] Image enhancement processing employs multi-scale structure-aware anisotropic diffusion filtering with three scale levels and standard deviations of 1.0, 2.0, and 4.0 voxels. At each scale, a 3×3×3 neighborhood structure tensor is calculated, and the principal orientation of the local structure is obtained through eigenvalue decomposition. The diffusion process iterates 15 times, dynamically adjusting the diffusion intensity based on the eigenvalue ratio of the structure tensor in each iteration. When the eigenvalue ratio (λ1-λ2) / (λ1+λ2) is greater than 0.7, the structure is identified as an edge structure, and the diffusion coefficient is reduced to 0.1 to protect the edge. When the ratio is less than 0.3, the region is identified as uniform, and the diffusion coefficient is increased to 0.8 to enhance the denoising effect. Through this adaptive processing, fine structures such as trabeculae are effectively preserved, while background noise is significantly reduced.

[0079] Density normalization first involves fitting the gray-level distribution of the local region using a Gaussian mixture model to automatically identify three tissue types: cortical bone, cancellous bone, and trabecular bone. For each voxel, its gray-level histogram within its 26-neighborhood is calculated, and the KL divergence is calculated against the pre-established reference distributions for the three tissue types. The tissue weight w is then obtained through softmax normalization. iIn a typical case, the tissue weights in the central region of the L3 vertebral body were: cortical bone 0.15, cancellous bone 0.65, and trabecular bone 0.20. When applying the multi-parameter fusion mapping formula, for the cancellous bone region with an original CT value of 350 HU, the weights were 0.42 after processing with the tissue-specific function f2. Considering the patient was a 65-year-old male, the individualized correction function G... patient The calculated value was 1.08, and the final normalized result was 385 HU. By comparing bone mineral density measurements before and after processing, the measurement difference between different devices at the same anatomical location decreased from ±45 HU to ±12 HU, significantly improving the consistency of quantitative analysis.

[0080] The grayscale distribution mapping employs an adaptive segmentation strategy, dividing the original histogram into three main regions—background, soft tissue, and bone tissue—using K-means clustering. In this embodiment, the automatically determined segmentation points are: g1 = 50 HU (boundary between background and soft tissue), g2 = 150 HU (boundary between soft tissue and bone tissue), and g3 = 400 HU (boundary between low-density bone and high-density bone). The mapping function uses a sigmoid function with a smoothness parameter γ = 0.3 at the segment boundaries to ensure the continuity of grayscale transformation. Through iterative optimization, the JS divergence between the grayscale distribution of the processed image and the standard template is less than 0.05.

[0081] The spatial registration process begins with the extraction of stable anatomical feature points using the 3D-SIFT algorithm. Typically, 30-50 feature points can be extracted for each vertebra. The RANSAC algorithm is then used to select the 15 most reliable corresponding point pairs. Based on these point pairs, an initial affine transformation matrix is ​​calculated to achieve overall vertebral alignment. Subsequently, elastic registration is performed using a level 3 B-spline deformation field, with the control point spacing gradually refined from 16 mm to 4 mm. Bone tissue preservation constraints are introduced during registration to ensure that the Jacobian determinant of high-density bone regions remains within the range of 0.9-1.1, preventing density distortion caused by excessive deformation.

[0082] To verify the effectiveness of this invention, 100 multicenter vertebral CT images were selected for a comparative experiment. Before processing, images of the same type of vertebra acquired by different devices showed significant differences in mean grayscale, standard deviation, and texture features, with coefficients of variation of 18.5%, 22.3%, and 31.2%, respectively. After processing using the method of this invention, the coefficients of variation of the above indicators decreased to 4.2%, 5.8%, and 8.7%, respectively. In the subsequent deep learning vertebral segmentation task, the U-Net model trained using the standardized data showed an improvement in its Dice coefficient from 0.912 to 0.946, and the performance degradation on the cross-device test set decreased from 15.3% to 3.8%. In terms of processing speed, the complete processing time for a single vertebra was approximately 8.5 seconds, including 2.1 seconds for segmentation, 1.8 seconds for enhancement, 2.3 seconds for normalization, and 2.3 seconds for registration, meeting the timeliness requirements for clinical applications.

[0083] The above-described specific embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, and improvements made by those skilled in the art to the technical solutions of the present invention based on the provided textual description and drawings, without departing from the design concept and spirit of the present invention, should all fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. A standardized preprocessing method for vertebral CT images integrating multi-parameter features, characterized in that, It includes the following steps: S1: Perform initial segmentation on the input original vertebral CT image. Through rough segmentation processing based on region growing and anatomical prior, select the voxels with gray values higher than +200HU within the central region of the vertebra as the initial seed points, and limit the gray value difference range of the growth threshold to ±150HU; introduce a prior modeling strategy based on graph structure, construct the key anatomical points of the vertebra as the nodes of the graph neural network, and achieve accurate segmentation by minimizing the vertebral topology constraint energy function: Where: E topo : Total energy value of the cone topological constraint; V: Set of vertebral anatomical feature points; v: A single anatomical feature point in set V; φ(v): Single-point potential energy function of node v, measuring the matching degree between the local features of this point and the vertebral model; E: Set of connecting edges constructed based on anatomical relevance, representing the spatial relationship between anatomical points; (u, v): Edge connecting node u and node v; ψ(u, v): Pairing potential energy function of edge (u, v), measuring the spatial constraint strength between two anatomical points; d anatomy (u,v): Anatomical distance metric function, which combines Euclidean distance and geodesic distance to quantify the anatomical correlation between two anatomical points; S2: Use a multi-scale structure-aware anisotropic diffusion filtering algorithm to enhance the image. By calculating the structure tensors at different scales, analyze the local structure types, and dynamically adjust the diffusion direction and amplitude; set the number of iterations to 10 - 20 times, and the stopping criterion is that the mean square error of the image between two iterations is less than 10^-4; S3: Establish a tissue-specific density mapping model with multi-parameter fusion, and achieve accurate density normalization through the following formula: Where: HU norm (x,y,z): Normalized CT value at spatial location (x,y,z); i: Index of tissue type, i = 1 represents cortical bone, i = 2 represents cancellous bone, i = 3 represents trabecular bone; w i The weight coefficients of the i-th type of organization are obtained based on local structure tensor analysis and satisfy the following conditions: HU raw Raw CT values, in units of HU; f i : The specific nonlinear mapping function for the i-th type of tissue; ρ i : Density mapping intensity coefficient of the i-th type of tissue, which controls the magnitude of density mapping; T i : Texture preservation parameter for the i-th type of organization, which controls the degree to which texture details are preserved; G patient Patient-specific correction function; Age: Patient age, in years; Sex: Patient gender; Patient-specific correction function G patient Combining the Gaussian decay of age with the sinusoidal interaction effect of gender-age: Where: α: Age-related correction amplitude coefficient, obtained through regression of large-scale clinical data, and the typical value range is 0.1 - 0.3; τ: Standard deviation parameter affected by age, controlling the attenuation rate of the age effect, in years, and the typical value is 15 - 20 years; β: Amplitude coefficient of gender-age interaction effect, and the typical value range is 0.05 - 0.15; Sex code Gender coding: 1 for male, 0 for female; Age - 50: Age deviation value centered on 50 years old; A periodic function that reflects age-related physiological cycle changes; S4: Use an adaptive piecewise histogram mapping method to map the gray distribution of the original image to a standard template. The mapping function uses sigmoid to smoothly transition at the piecewise boundaries to ensure the continuity of gray transformation: When g0 ≤ g < g2: T(g) = T1(g)·(1 - σ(g - g1)) + T2(g)·σ(g - g1) When g1 ≤ g < g3: T(g) = T2(g)·(1 - σ(g - g2)) + T3(g)·σ(g - g2) When g2≤g≤g max hour: T(g) = T3(g) Where: T(g): Output value after mapping of gray value g; g: Input gray value; T1(g), T2(g), T3(g): Local mapping functions in three gray intervals; g0,g1,g2,g3,g max The adaptively determined segmentation points divide the grayscale range into three intervals: [g0, g1]: Main action interval of the first mapping function T1; [g1, g2]: Transition and mixing interval of T1 and T2; [g2, g3]: Transition and mixing interval of T2 and T3; [g3,g max ]: The principal region of action of the third mapping function T3; sigmoid smooth transition function; γ: Transition smoothness parameter, controlling the transition steepness at the piecewise boundaries, and the typical value is 0.1 - 0.5; S5: Construct a rigid-elastic hybrid registration model. First, global alignment is achieved by extracting anatomical anchor points and performing a three-dimensional affine transformation. Then, B-spline deformation field is used to describe local elastic deformation. Accurate registration is achieved by minimizing the energy function containing data terms and regularization terms. Bone tissue preservation constraints are introduced to ensure the fidelity of bone density information during the registration process.

2. The method for standardized preprocessing of vertebral CT images fused with multi-parameter features according to claim 1, characterized in that, In step S1, the segmented binary mask undergoes 2-3 morphological closing operations through a spherical structural element with a radius of 5mm to fill internal holes and smooth boundaries; the graph neural network adopts a multi-layer graph convolution structure and is trained on more than 500 manually labeled data.

3. The method for standardized preprocessing of vertebral CT images fused with multi-parameter features according to claim 1, characterized in that, In step S3, the organization type weight w i The KL divergence between the local grayscale distribution and the reference distribution of each tissue is calculated and then normalized using softmax. Cortical bone is defined as a CT value greater than 700 HU, cancellous bone is defined as a CT value between 150 and 700 HU, and trabecular bone is defined as a CT value between 150 and 400 HU.

4. The method for standardized preprocessing of vertebral CT images fused with multi-parameter features according to claim 1, characterized in that, The tissue-specific mapping function f in step S3 i Defined as: in: tanh: hyperbolic tangent function, used to achieve nonlinear density compression; μ i : The average gray value of the i-th type of organization, in HU; σ i : Standard deviation of gray level for the i-th type of organization, in HU; λ i : Adjustment parameters for logarithmic transformation, controlling the degree of texture enhancement.

5. The method for standardized preprocessing of vertebral CT images with fused multi-parameter features according to claim 1 or 4, characterized in that, The parameters are set for different tissue types as follows: Bone cortex (i=1): μ1=900HU, σ1=150HU, ρ1=0.8, T1=0.2, λ1=500; Cancellous bone (i=2): μ2=400HU, σ2=100HU, ρ2=0.6, T2=0.4, λ2=300; Trabecular bone (i=3): μ3=200HU, σ3=50HU, ρ3=0.4, T3=0.6, λ3=150.

6. The method for standardized preprocessing of vertebral CT images fused with multi-parameter features according to claim 1, characterized in that, In step S4, the standard reference template is constructed through statistical analysis of high-quality typical cases; a smoothing regularization mechanism is introduced during the mapping process to prevent overfitting of the mapping function; and the gray-level distribution of the original image is gradually made to approximate the standard template through iterative optimization.

7. The method for standardized preprocessing of vertebral CT images fused with multi-parameter features according to claim 1, characterized in that, In step S5, the anatomical anchor points are automatically extracted using the 3D-SIFT feature detection algorithm; Flexible registration employs a multi-resolution strategy, optimizing progressively from coarse to fine. Registration accuracy is assessed by using normalized mutual information as a similarity metric.

8. The method for standardized preprocessing of vertebral CT images fused with multi-parameter features according to claim 1, characterized in that, It also includes an image quality assessment step, which detects the signal-to-noise ratio, resolution, and integrity of the input image before processing to ensure that it meets the preset requirements; and calculates the signal-to-noise ratio of bone tissue, inter-tissue contrast, and structural similarity index after processing to evaluate the standardization effect.

9. The method for standardized preprocessing of vertebral CT images fused with multi-parameter features according to claim 1, characterized in that, The method supports batch processing mode and establishes spatiotemporal consistency constraints for multi-sequence, multi-phase vertebral CT images to ensure standardized consistency of images from different periods of the same patient. It includes an adaptive parameter learning mechanism that automatically adjusts the parameter settings of each processing module according to the device type, scanning protocol, and image features of the input images.

10. A standardized preprocessing system for vertebral CT images incorporating multi-parameter features, applicable to the standardized preprocessing method for vertebral CT images incorporating multi-parameter features as described in any one of claims 1-9, characterized in that, include: The image initial segmentation module is used to implement the function of step S1 in claim 1, and includes a seed point selection unit, a region growing unit, a graph neural network unit, and an energy optimization unit. The image enhancement module, used to implement the function of step S2 in claim 1, includes a multi-scale analysis unit, a structure tensor calculation unit, and an anisotropic filtering unit. The density modeling and normalization module is used to implement the function of step S3 in claim 1, and includes a tissue classification unit, a multi-parameter fusion mapping unit and an individualized correction unit. The grayscale distribution mapping module is used to implement the function of step S4 in claim 1, and includes an adaptive segmentation unit, a smooth transition unit, and an iterative optimization unit. The spatial registration module is used to implement the function of step S5 in claim 1, and includes an anchor point extraction unit, a rigid transformation unit and an elastic deformation unit. The quality control module is used to assess the quality of input and output images and adaptively adjust parameters. The modules are interconnected through standardized data interfaces, supporting pipelined processing and parallel computing acceleration.

Images