CN-122023324-A - Vertebral bone quality detection and evaluation method for osteoporosis fracture

CN122023324ACN 122023324 ACN122023324 ACN 122023324ACN-122023324-A

Abstract

The invention provides a vertebral bone quality detection and evaluation method for osteoporotic fracture, which comprises the steps of automatically extracting a vertebral bone region through an advanced depth segmentation network, analyzing a bone trabecular main direction based on a Hessian matrix, constructing an anisotropic guide graph, forming a physical constraint convolution kernel to simulate biomechanical response, carrying out low-dimensional coding on a transformation energy density field by utilizing a transformation self-encoder, combining image features extracted by the convolution network, realizing dynamic fusion of images and biomechanical features through a double-channel gating fusion network, and finally predicting equivalent bone density by adopting linear regression and optimizing and improving performance from end to end.

Inventors

Tian Houze
HUO JIAQI
LU XINGBAO
WU SHENHAO

Assignees

克拉玛依市中心医院

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. The vertebral bone quality detection and evaluation method for the osteoporosis fracture is characterized by comprising the following steps of: s1, acquiring three-dimensional medical image data of a centrum bone, executing segmentation processing, generating a three-dimensional segmentation mask of the centrum bone based on the medical image data, and outputting a three-dimensional voxel matrix of a region of interest of the centrum bone; s2, based on local structure tensor analysis of a Hessian matrix, calculating a main direction field of a bone trabecula by using the three-dimensional voxel matrix, and generating an anisotropic guide map; S3, constructing a physical constraint convolution kernel, simulating anisotropic response characteristics of the force lines conducted along the bone trabeculae according to the local bone trabeculae orientation dynamic rotation convolution weight, and inputting the anisotropic guide map into the physical constraint convolution kernel to generate a biomechanical response map; s4, estimating strain energy density distribution under a virtual load based on the biomechanical response map, and training a variation self-encoder in an unsupervised mode to perform low-dimensional embedding on an energy field and output a biomechanical latent variable set; S5, extracting traditional image features including HU mean values, gray level co-occurrence matrix texture entropy and cortical thickness features through a standard convolution network to generate image feature vectors; S6, constructing a double-channel gating fusion unit, respectively inputting the image feature vector and the biomechanical latent variable into parallel LSTM branches, calculating cross-modal time sequence alignment weight through an attention mechanism, dynamically adjusting the contribution ratio of the two types of features, and outputting a combined feature vector; and S7, performing linear regression based on the joint feature vector, predicting the equivalent bone density value of the vertebral body, and performing end-to-end back propagation optimization by adopting a mean square error loss function.
2. The method for detecting and assessing vertebral bone quality for osteoporotic fracture according to claim 1, wherein the step S1 specifically includes: Acquiring CT or MRI three-dimensional tomographic image data of a spine region of a subject based on a DICOM standard to obtain an original medical image voxel matrix; Performing N4 offset field correction and isotropic resampling preprocessing operation on the original medical image voxel matrix to generate a standardized three-dimensional medical image voxel matrix; constructing a vertebral bone segmentation network model based on a U-Net architecture, wherein the segmentation network model completes pre-training on a public medical image data set containing a labeling vertebral region in a migration learning mode; Inputting the standardized three-dimensional medical image voxel matrix into a pre-trained centrum bone segmentation network model, outputting probability distribution of each voxel belonging to a centrum bone region based on a Softmax classification function, and generating a preliminary centrum bone segmentation probability map; performing connected domain analysis and morphological closing operation on the vertebral bone segmentation probability map, removing isolated noise areas and smoothing segmentation boundaries to generate a binary vertebral bone three-dimensional segmentation mask; and performing voxel-by-voxel logical AND operation on the three-dimensional segmentation mask of the centrum bone and the voxel matrix of the original medical image, and extracting a three-dimensional voxel matrix corresponding to the region of interest of the centrum bone.
3. The vertebral bone quality detection and evaluation method for osteoporotic fracture according to claim 2, wherein the N4 bias field correction is performed by using parameters of 50 iterations, 0.001 convergence threshold and 50mm B-spline control point spacing, and each step of optimization improves voxel gray level uniformity based on a log domain minimization energy function optimization mode.
4. The method for detecting and assessing vertebral bone quality for osteoporotic fracture according to claim 1, wherein the step S2 specifically includes: performing local structure tensor estimation on the three-dimensional voxel matrix, and constructing a second-order symmetrical tensor field based on gray gradient information of the vertebral bone region of interest; Performing eigenvalue decomposition on the local structure tensor, and calculating three orthogonal eigenvectors and corresponding eigenvalues of the local structure tensor, wherein the corresponding direction of the maximum eigenvalue is identified as the main direction of the bone trabecula; Constructing a main direction field of the bone trabecula based on the orthogonal feature vectors, mapping the main direction vector of each voxel point into a polar angle and an azimuth angle under a spherical coordinate system, and generating a local orientation angle parameter matrix; Smoothing the local orientation angle parameter matrix, and carrying out weighted average on the main directions of adjacent voxel points by adopting a Gaussian check; An anisotropic guide map is generated based on the smoothed principal direction field, the anisotropic guide map storing local orientation information of each voxel point in a three-dimensional matrix form.
5. The method for detecting and assessing vertebral bone quality for osteoporotic fracture according to claim 1, wherein the step S3 specifically includes: constructing a rotatable physical constraint convolution kernel template based on local bone trabecular orientation angle parameters in the anisotropic guide diagram so as to simulate anisotropic response characteristics of the force lines conducted along the bone trabecular direction and output a rotated convolution kernel weight matrix; Performing local coordinate system alignment transformation on a vertebral bone region voxel matrix of interest divided in a three-dimensional medical image in combination with a bone trabecula main direction field provided by an anisotropic guide image, rotating each local region image block to a reference coordinate system with the bone trabecula main direction being consistent, and outputting an aligned local image block set; Performing point-by-point convolution operation on the aligned local image block set based on the rotated physical constraint convolution kernel weight matrix, and performing weighted summation on each image block to obtain a biomechanical response intensity value of a local area and output a preliminary biomechanical response map of the biomechanical response map; Performing nonlinear activation processing on the preliminary biomechanical response map, mapping response intensity values, and outputting an enhanced biomechanical response map; and carrying out normalization processing on the enhanced biomechanical response map, calculating a normalization coefficient based on the statistical distribution of the local response values, mapping the response values of all voxels in the response map to a uniform scale range, and outputting the normalized biomechanical response map.
6. The method for detecting and evaluating vertebral bone quality for osteoporotic fracture according to claim 5, wherein the step S3 further comprises the steps of processing the preliminary biomechanical response map by a ReLU function, setting the activation threshold to zero, setting the range of the sparse coefficient to 1-3, setting the size of the filter kernel to 3 x 3, and uniformly mapping the response value to the 0-1 interval by adopting a local statistical normalization mode.
7. The method for detecting and assessing vertebral bone quality for osteoporotic fracture according to claim 1, wherein the step S4 specifically includes: Virtual load loading of finite element simulation is executed based on biomechanical response maps, linear elasticity hypothesis modeling is carried out on the internal stress-strain relation of the vertebral bone, and a strain energy density field under the action of anisotropic load is generated; Performing normalization processing on the strain energy density field, and performing linear scaling on the basis of the local maximum value and the local minimum value to obtain a standardized energy distribution map; Constructing a three-dimensional convolution variation self-encoder architecture, wherein an encoder part of the three-dimensional convolution self-encoder architecture comprises three stacked three-dimensional convolution layers and a batch normalization layer, and a subsequent parameterization layer is used for carrying out feature encoding on the standardized energy distribution map; Based on KL divergence constraint and reconstruction loss function joint optimization variation self-encoder, performing parameter estimation on potential distribution output by the encoder by adopting an unsupervised learning mode to obtain prior distribution of low-dimensional biomechanical hidden variables; Forward reasoning is performed on the trained variable self-encoder portion, and a normalized strain energy density field is input into the encoder to obtain a set of differentiable biomechanical latent variables.
8. The method for detecting and assessing the bone quality of a vertebral body for an osteoporotic fracture according to claim 7, wherein in the step S4, the encoder section of the variant-self encoder comprises three sets of 3 x 3 convolutional layers and a batch normalization layer, the optimization loss consists of a weighted sum of KL divergence and mean square reconstruction loss, and the weight parameter is set to 0.001.
9. The method for detecting and assessing vertebral bone quality for osteoporotic fracture according to claim 1, wherein the step S5 specifically includes: acquiring a three-dimensional voxel matrix of a vertebral bone region of interest, performing HU value normalization processing based on the three-dimensional voxel matrix, and outputting a normalized three-dimensional HU matrix; Performing sliding window gray level statistical analysis on the normalized three-dimensional HU matrix, calculating a HU mean value feature map based on local neighborhood gray level distribution, and outputting a HU mean value feature vector; constructing a gray level co-occurrence matrix based on the standardized three-dimensional HU matrix, constructing a multi-angle co-occurrence matrix by adopting four directions and a plurality of gray level intervals, and outputting a texture entropy feature map; Performing weighted average processing on the texture entropy feature map, and outputting texture entropy feature vectors based on weight distribution of co-occurrence matrixes in different directions; Performing cortical bone edge detection on the standardized three-dimensional HU matrix, extracting cortical bone boundaries by adopting Canny operators and morphological closing operation, calculating average cortical thickness parameters based on boundary thickness distribution, and outputting cortical thickness feature vectors; splicing the HU mean feature vector, the texture entropy feature vector and the cortex thickness feature vector to form an initial image feature vector; Performing residual convolution operation on the initial image feature vector based on ResNet modules, extracting deep image features layer by adopting batch normalization and ReLU activation functions, and outputting a deep feature map; performing global average pooling operation on the deep feature map, calculating a feature response mean value based on channel dimensions, and outputting a global feature vector; And carrying out L2 normalization processing on the global feature vector, and outputting a final image feature vector.
10. The method for detecting and evaluating vertebral bone quality for osteoporotic fracture according to claim 9, wherein the texture entropy features are formed by splicing after average values are calculated by gray level co-occurrence matrices with four directions (0 °, 45 °,90 °, 135 °) and gray level intervals of 1-3.

Description

Vertebral bone quality detection and evaluation method for osteoporosis fracture Technical Field The invention relates to the technical field of medical image analysis and bone biomechanics modeling, in particular to a vertebral body bone quality detection and evaluation method for osteoporosis fracture. Background The current evaluation field of osteoporosis fracture risk and centrum bone quality mainly depends on quantitative analysis technology based on CT or MRI three-dimensional medical images. The mainstream scheme generally adopts Quantitative CT (QCT), MRI or dual-energy X-ray (DXA) imaging means, and bone density is estimated through indexes such as gray level intensity, voxel density and the like, so that quantitative judgment on the quality of centrum bone is realized. The technology is widely applied in clinic, and is popularized in aspects of early osteoporosis screening, fracture risk prediction, drug efficacy monitoring and the like; Along with the combination of medical image analysis and artificial intelligence, some automatic bone density estimation and fracture risk modeling algorithms based on deep learning are also developed in recent years, and the methods automatically extract image features by using deep structures such as convolutional neural networks and the like, so that the evaluation efficiency and accuracy are further improved. In addition, to further enhance disease phenotype recognition capability, some studies have attempted to fuse multimodal data (e.g., QCT, MRI different sequences, clinical indices) or introduce Finite Element Analysis (FEA) based biomechanical simulations of bone tissue, combining image quantification with physical modeling to aid in bone quality assessment; However, the prior art still has the following limitations and unresolved problems: (1) The traditional scheme takes the gray level or voxel mean value of an image as a core feature, so that the real bearing and response mechanism of a bone structure to a complex physiological load is difficult to reflect, and the correlation between a fracture risk assessment result and an actual biomechanical state is limited; (2) For fusion of multi-modal data, most of the current main current methods are feature stitching or static weighting based on decision level, and the deep semantic relation among different modalities and the collaborative expression among physiological variables are ignored; (3) The finite element aided analysis can obtain a certain mechanical index, but has the defects that firstly, biomechanical modeling is usually used as a post-processing step, cannot be integrated with the front end of a bone mineral density prediction task and is difficult to optimize end to end; (4) Most of the prior art lacks dynamic and differentiable physical modeling capability, cannot introduce high-order biomechanical priors such as bone trabecula main direction, anisotropic conduction and the like in real time in a feature extraction stage, so that image features and biomechanical latent variables are fractured, a fusion mechanism is stiff, and comprehensive evaluation capability is limited. Disclosure of Invention The invention aims to solve the technical problems and provides a vertebral bone quality detection and evaluation method for osteoporosis fracture. The technical scheme of the invention is realized in that the vertebral body bone quality detection and evaluation method for the osteoporosis fracture comprises the following steps: s1, acquiring three-dimensional medical image data of a centrum bone, executing segmentation processing, generating a three-dimensional segmentation mask of the centrum bone based on the medical image data, and outputting a three-dimensional voxel matrix of a region of interest of the centrum bone; s2, based on the analysis of the local structure tensor of the Hessian matrix, calculating a main direction field of the bone trabecula by using the three-dimensional voxel matrix, and generating an anisotropic guide map, wherein the main direction of the bone trabecula is obtained by calculating the characteristic vector of the Hessian matrix, and the anisotropic guide map comprises local orientation angle parameters; S3, constructing a physical constraint convolution kernel, simulating anisotropic response characteristics of the force lines conducted along the bone trabeculae according to the dynamic rotation convolution weight of the local bone trabeculae, and inputting the anisotropic guide map into the physical constraint convolution kernel to generate a biomechanical response map, wherein the rotation angle of the convolution kernel is determined by local orientation angle parameters; S4, estimating strain energy density distribution under a virtual load based on a biomechanical response map, and training a variation self-encoder in an unsupervised mode to perform low-dimensional embedding on an energy field to output a biomechanical latent variable set, wherein an encoder p