Osteoporosis prediction method and system based on centrum CT image

Abstract

The invention provides an osteoporosis prediction method and system based on a centrum CT image, and the method comprises the steps: obtaining a patient centrum CT image, and carrying out the preprocessing of the image, and obtaining standardized three-dimensional voxel data; performing spatial resampling on the original data to a uniform resolution; constructing a hierarchical feature extraction network for centrum bone structure perception, and designing a non-uniform sampling mechanism for centrum density distribution; constructing an intervertebral biomechanical conduction diagram network; designing an osteoporosis specific loss function; and outputting a grading prediction result containing confidence, generating probability distribution of each grade through a softmax function, and generating a visual thermodynamic diagram of the lesion area. Through time sequence consistency constraint, the system can analyze image changes of the same patient at different time points and evaluate the treatment effect and the disease progress. The dynamic monitoring ability provides a powerful tool for long-term management of chronic osteoporosis, and is helpful for timely adjustment of treatment schemes and improvement of prognosis of patients.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for predicting osteoporosis based on vertebral CT images. Background Art

[0002] Osteoporosis is a systemic skeletal disease characterized by decreased bone mass and degeneration of bone microarchitecture, primarily manifested by decreased bone density and increased bone brittleness, which can easily lead to brittle fractures. Approximately 200 million people worldwide suffer from osteoporosis, with postmenopausal women and the elderly being the most susceptible groups. Vertebral compression fractures are one of the most common complications of osteoporosis, which not only seriously affects the patient's quality of life but also leads to spinal deformity, chronic pain, and decreased mobility, placing a heavy economic burden on families and society. Therefore, early diagnosis and risk assessment of osteoporosis are of great clinical significance for preventing fractures and improving prognosis.

[0003] The current gold standard for clinical diagnosis of osteoporosis is dual-energy X-ray absorptiometry (DXA), which calculates areal bone density by measuring the difference in bone absorption of X-rays of different energies, and diagnoses and grades according to T-scores. However, DXA technology has many limitations. First, DXA can only provide areal bone density in two-dimensional projections and cannot reflect the three-dimensional structural information of the bone, especially the trabecular microstructure that is crucial for bone strength. Second, DXA measurement results are easily interfered with by factors such as spinal degeneration, vascular calcification, and osteophyte formation, and false-negative results are common in elderly patients. In addition, DXA is not sensitive enough for detecting early osteoporosis and mild bone loss, and abnormalities are often not detected until bone loss reaches 30% or more, missing the optimal time for intervention.

[0004] In recent years, although some studies have attempted to apply deep learning technology to the diagnosis of osteoporosis in vertebral CT images, the existing methods generally have the following technical difficulties. First, most methods use general convolutional neural networks or VisionTransformer architectures, and are not optimized for the special structure of bone tissue. Trabecular bone has obvious directional and anisotropic characteristics, and standard isotropic convolution kernels are difficult to effectively capture these structural information, resulting in insufficient feature extraction. Second, existing methods usually use a uniform sampling strategy to process the entire vertebral image, without taking into account the distribution characteristics of osteoporotic lesions in different density ranges. Studies have shown that the cancellous bone area with HU values ​​in the range of 100-200 is the most sensitive area for early changes in osteoporosis, but the uniform sampling strategy has insufficient sampling density in these key areas and is prone to missing early lesion signals.

[0005] From a clinical perspective, osteoporosis diagnosis should go beyond assessing current status and include the prediction of future fracture risk and the development of personalized treatment plans. Existing methods primarily focus on bone density measurement and grading, but lack the ability to quantitatively assess fracture risk. Integrating imaging features with clinical factors to construct more accurate fracture risk prediction models is a pressing clinical challenge. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this paper proposes a method and system for predicting osteoporosis based on vertebral CT images. This method is capable of processing CT images from different devices and with varying scanning parameters, and exhibits excellent generalization capabilities. The system seamlessly integrates with the hospital's existing PACS system, eliminating the need to alter existing clinical workflows and simplifying deployment. The automated processing reduces manual workload, eliminates subjective factors, and improves diagnostic consistency and repeatability.

[0007] To achieve the above objectives, the present invention proposes a method for predicting osteoporosis based on vertebral CT images, comprising the following steps:

[0008] S1: Acquire vertebral CT images of patients and perform preprocessing to obtain standardized 3D voxel data; spatially resample the original DICOM data to a uniform resolution, adjust the window width and window position, and perform intensity normalization processing;

[0009] S2: Constructing a hierarchical feature extraction network for vertebral bone structure perception,

[0010] S3: Design a non-uniform sampling mechanism for the cone density distribution;

[0011] S4: Constructing an intervertebral biomechanical conduction map network;

[0012] S5: Design osteoporosis-specific loss function;

[0013] S6: Output the graded prediction results including confidence, generate the probability distribution of each grade through the softmax function, and generate a visual heat map of the lesion area.

[0014] Furthermore, step S2 specifically includes:

[0015] S2.1: Divide the input voxels into overlapping 3D patches, each of size 16 × 16 × 16 voxels, with an overlap ratio of 50%;

[0016] S2.2: Calculate the three-dimensional gradient for each patch, construct a direction tensor matrix through the gradient outer product, and perform eigenvalue decomposition on the tensor matrix to obtain the main directions of the trabecular bone;

[0017] S2.3: Designing adaptive convolution kernel based on trabecular orientation:

[0018] K(θ,φ,ψ)=R(θ,φ,ψ)·K0

[0019] in:

[0020] K(θ,φ,ψ): adaptive convolution kernel;

[0021] R(θ,φ,ψ): 3D rotation matrix based on the main directions of trabecular bone;

[0022] θ, φ, ψ: the rotation angles around the x, y, and z axes, respectively, determined by the eigenvectors of the trabecular orientation tensor;

[0023] K0: standard 3D convolution kernel basis;

[0024] S2.4: Extracting direction-sensitive features through adaptive convolution operation and performing weighted summation on the convolution results in multiple directions; the trabecular direction calculation in S2.2 specifically includes:

[0025] The Hessian matrix is ​​used to analyze the local second-order derivative structure, and a 3×3 symmetric matrix is ​​constructed by calculating the second-order partial derivatives of the image intensity;

[0026] Perform eigenvalue decomposition on the Hessian matrix to obtain three eigenvalues ​​and corresponding eigenvectors;

[0027] The structural type is determined based on the relationship between the eigenvalues: when the absolute value of the smallest eigenvalue is much smaller than the other two, it is determined to be plate-like trabecular bone; when the two smaller eigenvalues ​​are close and much smaller than the largest eigenvalue, it is determined to be rod-like trabecular bone;

[0028] The main direction of the trabecular bone is determined according to the eigenvector corresponding to the maximum eigenvalue.

[0029] Furthermore, step S3 is specifically as follows:

[0030] S3.1: Calculate the HU value distribution histogram inside the vertebral body and identify the cortical bone area, cancellous bone area, and low-density area through threshold segmentation;

[0031] S3.2: Design sampling function based on density distribution:

[0032] S(x,y,z)=α·exp(-β·|HU(x,y,z)-HU mean |)

[0033] in:

[0034] S(x,y,z): sampling probability density at the spatial position (x,y,z);

[0035] α: normalization coefficient, ensuring ∫S(x,y,z)dxdydz=1;

[0036] β: density sensitivity parameter, which can be learned and the initial value is set to 0.01;

[0037] HU(x,y,z): Hounsfield unit value at position (x,y,z);

[0038] HU mean : Average HU value of the vertebral cancellous bone region, typically ranging from 150 to 200;

[0039] S3.3: Increase the sampling density by 2 times in low-density areas to generate density-adaptive feature maps;

[0040] Introduce enhanced sampling weights in low-density sensitive areas:

[0041]

[0042] in:

[0043] w(HU): sampling weight corresponding to HU value;

[0044] HU: Hounsfield unit value;

[0045] 150: center value of early changes in osteoporosis;

[0046] σ: standard deviation, set to 25, covering the 100-200 HU sensitivity range;

[0047] The final sampling probability is generated by the importance sampling algorithm, and the sampling probability is proportional to the product of the density sampling function and the weight function:

[0048] P(sample|x,y,z)∝S(x,y,z)·w(HU(x,y,z)).

[0049] Furthermore, step S4 is specifically as follows:

[0050] S4.1: Each vertebra is treated as a graph node, and the node features are constructed by connecting the direction-sensitive features, density-adaptive features and the centroid coordinates of the vertebra;

[0051] S4.2: Define edge weights based on biomechanical conduction pathways:

[0052]

[0053] in:

[0054] w ij : The edge weight from vertebra i to vertebra j;

[0055] τij =cos 2 (θ ij ): stress transmission coefficient, θ ij is the vertebral main axis angle;

[0056] intervertebral disc degeneration regulator;

[0057] SI ij : T2 signal intensity of the intervertebral disc between vertebrae i and j;

[0058] SI normal : Reference signal intensity of normal intervertebral disc (usually 250-300);

[0059] λ: distance attenuation coefficient, ranging from 0.5 to 1.0;

[0060] d ij : Euclidean distance between vertebral centroids;

[0061] h ij : disc height;

[0062] S4.3: Update node features through the message passing mechanism of the graph neural network, using the sigmoid activation function to process the linear combination of its own features and the weighted features of neighboring nodes;

[0063] The message delivery update in S4.3 is specifically as follows:

[0064] Collect neighbor node features and perform weighted aggregation based on edge weights;

[0065] Perform linear transformation on its own features and aggregated neighbor features through learnable weight matrices respectively;

[0066] The two transformation results are added together and then the updated node features are obtained through the sigmoid activation function;

[0067] Repeat the messaging process 3-5 times to fully spread the information.

[0068] Furthermore, step S5 is specifically as follows:

[0069] S5.1: Define fracture risk-sensitive losses:

[0070]

[0071] The weight function is:

[0072] ω i =1+γ·P(fracture|x i )

[0073] ω i: The loss weight of the i-th sample;

[0074] γ: risk sensitivity coefficient, ranging from [0.5 to 2.0];

[0075] y i : True bone density T value;

[0076] Predict bone density T value;

[0077] P(fracture|x i ): 10-year fracture probability based on clinical indicators, calculated by logistic regression model;

[0078] S5.2: Introduce bone density gradient consistency constraints:

[0079]

[0080] in:

[0081] Predicting spatial gradients in bone density;

[0082] spatial gradients of true bone density;

[0083] ||·||2: L2 norm;

[0084] S5.3: Classification loss uses standard cross entropy loss to calculate the osteoporosis grade classification error;

[0085] S5.4: Total loss function:

[0086] L total =L cls +μL frac +ρL grad

[0087] Where μ = 0.3, ρ = 0.2 are the balance coefficients;

[0088] Definition of the vertebral structural index to quantify the degree of osteoporosis:

[0089]

[0090] in:

[0091] SMI: structural model index, ranging from [0,3], where 0 represents plate-like and 3 represents rod-like;

[0092] BV / TV: bone volume fraction, normal value 0.15-0.25;

[0093] Tb.Sp: trabecular spacing, in mm, normal value 0.5-1.0 mm;

[0094] Tb.Th: trabecular bone thickness, unit: mm, normal value: 0.1-0.2mm.

[0095] Furthermore, the fracture probability calculation adopts an improved FRAX scoring system, which integrates multiple risk factors such as age, bone density, previous fracture history and structural index through a logistic regression model, and the output value is mapped to the range of 0-1 by a sigmoid function as the fracture probability.

[0096] Furthermore, we introduce temporal consistency constraints:

[0097] L temp =||f(I t+1 )-(f(I t )+Δf)||2

[0098] in:

[0099] f(I t ): image features at time point t;

[0100] Δf: expected change based on clinical evolution law;

[0101] Δf=g(Δt,Age,BMD t ), g is a learnable time series prediction function implemented by the gated recurrent unit GRU;

[0102] Δt: time interval (month).

[0103] Furthermore, a deep integration strategy is used to improve prediction stability:

[0104] Train 5 sub-models with different random initializations;

[0105] During prediction, the output of each sub-model is weighted averaged, and the weights are dynamically adjusted based on the performance of the validation set;

[0106] Calculate the forecast variance as a measure of uncertainty and define the trigger mechanism:

[0107]

[0108] in:

[0109] Prediction variance

[0110] θ uncertainty : Uncertainty threshold, set according to clinical needs, typical value is 0.3;

[0111] When Trigger=1, the system marks the case as requiring manual review.

[0112] A system for predicting osteoporosis based on vertebral CT images, which implements the method for predicting osteoporosis based on vertebral CT images, comprises:

[0113] Image acquisition module: receives DICOM format CT data, supports 0.625-2.5mm slice thickness, and automatically identifies lumbar spine sequences;

[0114] Bone structure perception module: performs adaptive convolution and non-uniform sampling to extract direction-sensitive features;

[0115] Biomechanical modeling module: constructs intervertebral conduction map network and models pathological propagation patterns;

[0116] Risk assessment module: calculates the fracture risk sensitivity loss and outputs the T-value grading results;

[0117] Quality control module: evaluates prediction uncertainty and marks cases for review when the variance exceeds a threshold;

[0118] Visualization module: Generates 3D heat maps to highlight low-density areas and potential lesions.

[0119] Compared with the prior art, the present invention has the following beneficial effects:

[0120] 1. The present invention provides an osteoporosis prediction method and system based on vertebral CT images. By introducing an adaptive mechanism for the direction of trabeculae, feature extraction can be performed along the natural structure of bone tissue, significantly improving the ability to identify changes in trabecular microstructure. Compared with the traditional isotropic convolution method, adaptive convolution can more accurately capture the morphological characteristics of plate-shaped and rod-shaped trabeculae, greatly improving the detection rate of early osteoporosis. The density adaptive sampling strategy ensures sensitivity to subtle lesions by enhancing the sampling density in the pathologically sensitive range, especially in low-density areas where early changes in osteoporosis are most obvious. The system can detect lesion signals that are easily missed by traditional methods.

[0121] 2. This invention provides an osteoporosis prediction method and system based on vertebral CT images. This system incorporates biomechanical modeling into osteoporosis diagnosis and, by constructing a graph network structure between vertebrae, enables a comprehensive assessment of the overall pathological state of the spine. The system not only independently analyzes the bone status of each vertebra but also quantifies the inter-vertebral influence and identifies patterns of pathological transmission.

[0122] 3. This invention provides a method and system for predicting osteoporosis based on vertebral CT images. This system provides multi-dimensional diagnostic output, including osteoporosis grading, continuous T-score prediction, fracture risk assessment, and lesion area visualization. This comprehensive diagnostic information provides physicians with a comprehensive basis for decision-making, enabling them to assess current bone status and predict future disease trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0123] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0124] Figure 1 This is a schematic flow chart of the method of the present invention DETAILED DESCRIPTION

[0125] The technical solutions of the present invention will be more clearly and completely explained below through description of preferred embodiments of the present invention in conjunction with the accompanying drawings.

[0126] like Figure 1 As shown, the osteoporosis prediction system based on vertebral CT images provided by the present invention adopts an end-to-end intelligent analysis process, from the original CT image input to the final multi-dimensional prediction output, to achieve fully automated osteoporosis risk assessment.

[0127] The system receives vertebral CT images transmitted from the hospital's PACS system. These images are stored in DICOM format and include a complete vertebral sequence from L1 to L4, with a scan layer thickness of 1 mm. After the raw image data enters the image preprocessing module, the system performs three key processing steps: spatially resampling images acquired by different devices to a uniform voxel resolution of 1 cubic millimeter; standardizing CT values, converting them to uniform Hounsfield units and applying bone window settings; and finally, extracting the region of interest for each vertebra to prepare standardized input for subsequent analysis. The preprocessed data then enters the vertebral automatic segmentation module, which uses a 3D U-Net deep network trained on large-scale data to accurately identify and locate each vertebra with a positioning accuracy of 98.5%. A 128×128×64 voxel three-dimensional image block is extracted for each vertebra to ensure that complete vertebral structural information is included.

[0128] After completing preprocessing and segmentation, the system enters a multi-level feature extraction and fusion module, which includes three feature extraction branches running in parallel. The first branch is the trabecular direction adaptation module, which analyzes the three-dimensional structural characteristics of the trabeculae by calculating the Hessian matrix of each local area. The system performs eigenvalue decomposition on the Hessian matrix, determines whether the area is a plate-like trabecular or rod-like trabecular structure based on the relative size of the eigenvalues, and extracts the main direction information of the trabeculae. Based on the detected direction information, the system dynamically generates an adaptive convolution kernel, allowing feature extraction to proceed along the natural direction of the trabeculae, significantly improving the sensitivity to changes in the microstructure of bone tissue, and ultimately outputting direction-sensitive features.

[0129] The second branch is the density-adaptive sampling module, which first analyzes the HU value distribution within the vertebral body to identify the density range most sensitive to osteoporotic changes. The system pays special attention to the sensitive range of 100 to 200 HU, which is the range where early osteoporosis changes are most obvious. In low-density areas, the system automatically increases the sampling density by 2 times and optimizes the distribution of sampling points through an importance sampling algorithm, while maintaining diagnostic accuracy and improving computational efficiency by 80%. This non-uniform sampling strategy enables the system to focus computing resources on areas where pathological changes are most likely to occur, avoiding redundant calculations in areas of normal bone tissue.

[0130] The third branch is the biomechanical graph network module, which expands the traditional independent vertebral analysis into a correlation analysis that considers the mutual influence between vertebrae. The system models each vertebra as a node in the graph structure, and the node features include the imaging features and spatial position information of the vertebra. Based on the biomechanical principle of Wolff's law, the system calculates the stress conduction coefficient between vertebrae and quantifies the mechanical influence between adjacent vertebrae. At the same time, the degree of intervertebral disc degeneration is evaluated by analyzing T2-weighted images, and the state of the intervertebral disc is included as a regulating factor in the edge weight calculation. Information is propagated through a multi-layer graph attention mechanism. The feature representation of each vertebra not only contains its own information, but also integrates the pathological state of the adjacent vertebrae, realizing the extraction of vertebral correlation features.

[0131] The features extracted by the three branches are integrated in the fusion layer to form a comprehensive feature representation of each vertebra. These features are then input into the multi-task prediction output module, and the system performs four prediction tasks in parallel. The osteoporosis grading task divides the vertebrae into three grades: normal, osteopenia, and osteoporosis, with a classification accuracy of 94.2%. The T-value prediction task outputs a continuous bone density T-value, with a mean absolute error of only 0.18 compared to the DXA gold standard. The fracture risk assessment task calculates the patient's probability of fracture in the next 10 years based on clinical risk factors and imaging features, providing an important reference for clinical decision-making. The heat map visualization task generates a three-dimensional density distribution map, annotating low-density areas and potential lesions with different colors to help doctors intuitively understand the distribution of lesions.

[0132] The entire system is trained using a specially designed loss function that combines classification and regression losses with gradient consistency constraints to ensure balanced optimization across different tasks. The system also incorporates an uncertainty quantification mechanism, calculating a confidence level for each prediction result. When the confidence level falls below a threshold, it automatically flags the prediction as requiring manual review, enhancing the safety of clinical applications. From inputting the raw CT image to outputting a complete diagnostic report, the entire process takes only 3.2 seconds, fully meeting the requirements of real-time clinical applications.

[0133] As a specific example, this system was implemented in an osteoporosis screening system at a tertiary hospital. This system uses a Siemens SOMATOM Force dual-source CT scanner, with scanning parameters set to 120kV tube voltage, 250mAs tube current, 1mm slice thickness, and a 512x512 reconstruction matrix. The system is deployed on an NVIDIA DGX A100 server, comprising eight A100 GPUs, each with 80GB of video memory, capable of processing large-scale 3D medical images in real time.

[0134] The dataset includes lumbar spine CT data from 3,500 patients between January 2020 and June 2024, including 1,200 patients in the normal group (T-score greater than -1.0), 1,500 patients in the osteopenia group (T-score between -2.5 and -1.0), and 800 patients in the osteoporosis group (T-score ≤ -2.5). All patients also had concurrent dual-energy X-ray absorptiometry (DXA) results as the gold standard. The data were divided into a training set of 2,450 patients, a validation set of 700 patients, and a test set of 350 patients in a 7:2:1 ratio.

[0135] The system first reads the original CT series in DICOM format and sorts them according to the image position information to ensure spatial continuity. The pixel spacing of the original data is usually not uniform, so three-dimensional resampling is required to standardize all data to a voxel resolution of 1 cubic millimeter. This step is crucial for subsequent quantitative analysis because the uniform spatial resolution ensures comparability between data from different patients. After resampling, the CT values ​​are converted to standard Hounsfield units (HU), and bone window settings are applied, with the window center set to 400HU and the window width set to 2000HU. The image intensity is cropped to the range of -600 to 1400HU and finally normalized to the range of 0 to 1.

[0136] Vertebral localization is automated using a pre-trained 3D U-Net network, trained on 1,000 manually annotated datasets, achieving 98.5% accuracy. For each identified vertebra, the system extracts a 128×128×64 voxel region of interest (ROI), ensuring that the entire vertebral structure and surrounding tissue within 1 cm are included, providing ample context for subsequent analysis.

[0137] The system divides each vertebral ROI into multiple overlapping blocks of 16×16×16 voxels, with a 50% overlap ratio to ensure continuity in feature extraction. A three-dimensional gradient field is calculated for each block, and a 3×3 direction tensor matrix is ​​constructed through the outer product of the gradients. Eigenvalue decomposition is performed on this tensor matrix, yielding three eigenvalues ​​and their corresponding eigenvectors. The relative magnitudes of the eigenvalues ​​are used to determine the type of local trabecular bone structure: when the absolute value of the smallest eigenvalue is much smaller than the other two, the structure is determined to be plate-like; when the two smaller eigenvalues ​​are close and both much smaller than the largest eigenvalue, the structure is determined to be rod-like. The eigenvector corresponding to the largest eigenvalue represents the principal direction of the trabeculae in that region.

[0138] Based on the detected main direction of the trabeculae, the system calculates the corresponding rotation angle θ, ψ, construct a three-dimensional rotation matrix The standard 3×3×3 convolution kernel K0 is transformed by the rotation matrix to obtain an adaptive convolution kernel that adapts to the local trabecular direction. This adaptive mechanism enables convolution operations to extract features along the natural orientation of trabeculae, significantly improving sensitivity to changes in bone microstructure. Experiments have shown that, on 200 test vertebrae, adaptive convolution improves the accuracy of trabecular structure recognition from 82% to 94% compared to standard convolution.

[0139] The system counts the HU value distribution inside the vertebral body and calculates the average value of the cancellous bone area (HU value between 100 and 300), which is measured to be approximately 175HU. The sampling probability density function S(x, y, z) is calculated based on the difference between the HU value of each position and the average value. The density sensitivity parameter β is set to 0.01, which is the best value obtained through a large number of experimental optimizations. In the sensitive range of early changes in osteoporosis (100 to 200HU), the system additionally introduces a Gaussian weight function w(HU), with the center value set to 150HU and the standard deviation σ of 25, covering the entire sensitive range. The final sampling probability is determined by the product of the density sampling function S and the weight function w, and is normalized.

[0140] Using the importance sampling algorithm, 10,000 sampling points are generated for each vertebra, improving computational efficiency by 80% compared to the 50,000 points required by traditional uniform sampling to achieve the same lesion detection rate. In a validation study of 100 osteoporosis patients, non-uniform sampling achieved a 92% lesion detection rate with just 10,000 sampling points, while uniform sampling achieved only an 85% detection rate with 50,000 points.

[0141] The graph neural network employs a three-layer graph attention mechanism for message passing, with ReLU activation functions and a dropout rate of 0.2 between each layer to prevent overfitting. The first layer maps 512-dimensional node features to a 256-dimensional hidden representation. The second layer refines the features in the latent space, and the third layer outputs a final 128-dimensional node representation. Through this message passing mechanism, the features of each vertebra not only contain its own information but also incorporate the pathological conditions of adjacent vertebrae, enabling modeling of intervertebral interactions. Experimental results show that the introduction of the biomechanical graph network improves the accuracy of predicting the association of adjacent vertebral lesions from 72% to 87%.

[0142] The fracture risk-sensitive loss function was designed based on a clinical fracture risk assessment system. The system uses a modified FRAX scoring model to calculate the 10-year fracture probability for each sample, integrating multiple risk factors such as age, bone density, previous fracture history, and structural index. The structural index is calculated by combining parameters such as the structural model index (SMI), bone volume fraction (BV / TV), trabecular spacing (Tb.Sp), and trabecular thickness (Tb.Th). Regression analysis of 5,000 clinical data points yielded weighted coefficients for each factor: 0.06 for age, -1.2 for BMD, 1.8 for previous fractures, and 2.1 for the structural index.

[0143] The loss function consists of three components. The classification loss uses the standard cross entropy to calculate the prediction error of the three osteoporosis categories (normal, osteopenia, and osteoporosis). The fracture risk sensitive loss is calculated by the sample weight ω. iThe regression loss is adjusted, with a weight consisting of a base value of 1 plus the product of the risk sensitivity coefficient γ and the fracture probability, with γ set to 1.0. The gradient consistency loss calculates the spatial gradient difference between the predicted and true bone density maps to ensure the spatial continuity and rationality of the prediction results. The three losses are weighted together, with the classification loss weighted to 1, the fracture sensitivity loss weighted to α to 0.3, and the gradient loss weighted to β to 0.2.

[0144] The model was trained using the AdamW optimizer with a learning rate of 1e-4, a weight decay of 1e-5, and a batch size of 8. A cosine annealing learning rate scheduler was used, gradually reducing the learning rate over 100 training epochs. During training, model performance steadily improved, achieving 81.2% classification accuracy and a T-mean absolute error of 0.42 at epoch 10. This reached 91.3% accuracy with a T-mean absolute error of 0.25 at epoch 50, ultimately reaching 94.2% accuracy with a T-mean absolute error of 0.18 at epoch 100.

[0145] A comprehensive evaluation on a test set of 350 cases showed that the system achieved 94.2% accuracy in the osteoporosis three-classification task, with a sensitivity of 96.3% and a specificity of 92.8% for detecting osteoporosis. The mean absolute error of T-score prediction was 0.18, and the correlation coefficient with the DXA gold standard reached 0.95. A complete analysis of a single patient took only 3.2 seconds, fully meeting the requirements of real-time clinical applications.

[0146] Using a deep ensemble strategy, five sub-models with different initializations were trained, and the output variance was calculated during prediction as a measure of uncertainty. When the prediction variance exceeded a threshold of 0.3, the system automatically flagged the case for manual review. Of the 350 cases tested, 312 were labeled as high-confidence predictions, achieving an accuracy rate of 97.8%. Of the 38 cases marked as requiring review, 35 were confirmed to be borderline cases or exhibited image quality issues, resulting in an accuracy rate of 92.1% for uncertainty labeling.

[0147] The system has been in trial operation in the hospital's orthopedics and imaging departments for six months, successfully assisting in the diagnosis of more than 2,000 patients. Through seamless integration with the hospital's PACS system, it has achieved full automation of the entire process, including automatic monitoring of newly added lumbar CT examinations, automatic analysis and processing, and automatic return of results. The structured reports generated by the system use the DICOM SR format and include detailed vertebral grading diagnosis, T-value prediction, fracture risk assessment, and visual images, providing comprehensive decision support for clinicians. Through technical means such as mixed precision training, model quantization, and batch optimization, the system's training speed has increased by 60%, the inference speed has increased by 3 times, and the GPU utilization rate has reached 95%, fully utilizing the hardware performance and ensuring the efficiency and practicality of clinical applications.

[0148] The above-described specific embodiments merely describe preferred embodiments of the present invention and do not limit the scope of protection of the present invention. Any modifications, substitutions, and improvements made to the technical solution of the present invention by a person skilled in the art based on the textual description and drawings provided herein, without departing from the design concept and spirit of the present invention, shall 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 method for predicting osteoporosis based on vertebral CT images, characterized in that: The following steps are involved: S1: Acquire vertebral CT images of patients and perform preprocessing to obtain standardized 3D voxel data; spatially resample the original DICOM data to a uniform resolution, adjust the window width and window position, and perform intensity normalization processing; S2: Constructing a hierarchical feature extraction network for vertebral bone structure perception, S3: Design a non-uniform sampling mechanism for the cone density distribution; S4: Constructing an intervertebral biomechanical conduction map network; S5: Design osteoporosis-specific loss function; S6: Output the graded prediction results including confidence, generate the probability distribution of each grade through the softmax function, and generate a visual heat map of the lesion area.

2. The osteoporosis prediction method based on vertebral CT images according to claim 1, characterized in that: Step S2 specifically includes: S2.1: Divide the input voxels into overlapping 3D patches, each of size 16 × 16 × 16 voxels, with an overlap ratio of 50%; S2.2: Calculate the three-dimensional gradient for each patch, construct the direction tensor matrix through the gradient outer product, and perform eigenvalue decomposition on the direction tensor matrix to obtain the main direction of the trabecular bone; S2.3: Designing adaptive convolution kernel based on trabecular orientation: K(θ,φ,ψ)=R(θ,φ,ψ)·K0 in: K(θ,φ,ψ): adaptive convolution kernel; R(θ,φ,ψ): 3D rotation matrix based on the main directions of trabecular bone; θ, φ, ψ: the rotation angles around the x, y, and z axes, respectively, determined by the eigenvectors of the trabecular orientation tensor; K0: standard 3D convolution kernel basis; S2.4: Extracting direction-sensitive features through adaptive convolution operation and performing weighted summation on the convolution results in multiple directions; the trabecular direction calculation in S2.2 specifically includes: The Hessian matrix is ​​used to analyze the local second-order derivative structure, and a 3×3 symmetric matrix is ​​constructed by calculating the second-order partial derivatives of the image intensity; Perform eigenvalue decomposition on the Hessian matrix to obtain three eigenvalues ​​and corresponding eigenvectors; The structural type is determined based on the relationship between the eigenvalues: when the absolute value of the smallest eigenvalue is much smaller than the other two, it is determined to be plate-like trabecular bone; when the two smaller eigenvalues ​​are close and much smaller than the largest eigenvalue, it is determined to be rod-like trabecular bone; The main direction of the trabecular bone is determined according to the eigenvector corresponding to the maximum eigenvalue.

3. The osteoporosis prediction method based on vertebral CT images according to claim 1, characterized in that: Step S3 is as follows: S3.1: Calculate the HU value distribution histogram inside the vertebral body and identify the cortical bone area, cancellous bone area, and low-density area through threshold segmentation; S3.2: Design sampling function based on density distribution: S(x,y,z)=α·exp(-β·|HU(x,y,z)-HU mean |) in: S(x, y, z): sampling probability density at the spatial position (x, y, z); α: normalization coefficient, ensuring that ∫S(x, y, z)dxdydz=1; β: density sensitivity parameter, which can be learned and the initial value is set to 0.01; HU(x,y,z): Hounsfield unit value at position (x,y,z); HU mean : Average HU value of the vertebral cancellous bone region, typically ranging from 150 to 200; S3.3: Increase the sampling density by 2 times in low-density areas to generate density-adaptive feature maps; Introduce enhanced sampling weights in low-density sensitive areas: in: w(HU): sampling weight corresponding to HU value; HU: Hounsfield unit value; 150: center value of early changes in osteoporosis; σ: standard deviation, set to 25, covering the 100-200 HU sensitivity range; The final sampling probability is generated by the importance sampling algorithm, and the sampling probability is proportional to the product of the density sampling function and the weight function: P(sample|x,y,z)∝S(x,y,z)·w(HU(x,y,z)).

4. The osteoporosis prediction method based on vertebral CT images according to claim 1, characterized in that: Step S4 is specifically as follows: S4.1: Each vertebra is treated as a graph node, and the node features are constructed by connecting the direction-sensitive features, density-adaptive features and the centroid coordinates of the vertebra; S4.2: Define edge weights based on biomechanical conduction pathways: in: w ij : The edge weight from vertebra i to vertebra j; τ ij =cos 2 (θ ij ): stress transmission coefficient, θ ij is the vertebral main axis angle; intervertebral disc degeneration regulator; SI ij : T2 signal intensity of the intervertebral disc between vertebrae i and j; SI normal : Reference signal intensity of normal intervertebral disc (usually 250-300); λ: distance attenuation coefficient, ranging from 0.5 to 1.0; d ii : Euclidean distance between vertebral centroids; h ij : disc height; S4.3: Update node features through the message passing mechanism of the graph neural network, using the sigmoid activation function to process the linear combination of its own features and the weighted features of neighboring nodes; The message delivery update in S4.3 is specifically as follows: Collect neighbor node features and perform weighted aggregation based on edge weights; Perform linear transformation on its own features and aggregated neighbor features through learnable weight matrices respectively; The two transformation results are added together and then the updated node features are obtained through the sigmoid activation function; Repeat the messaging process 3-5 times to fully spread the information.

5. The osteoporosis prediction method based on vertebral CT images according to claim 1, characterized in that: Step S5 is specifically as follows: S5.1: Define fracture risk-sensitive losses: The weight function is: oh i =1+γ·P(fracture|x i ) ω i : The loss weight of the i-th sample; γ: risk sensitivity coefficient, ranging from [0.5 to 2.0]; y i : True bone density T value; Predict bone density T value; P(fracture|x i ): 10-year fracture probability based on clinical indicators, calculated by logistic regression model; S5.2: Introduce bone density gradient consistency constraints: in: Predicting spatial gradients in bone density; spatial gradients of true bone density; ||·||2: L2 norm; S5.3: Classification loss uses standard cross entropy loss to calculate the osteoporosis grade classification error; S5.4: Total loss function: L total =L cls +μL frac +ρL grad Where μ = 0.3, ρ = 0.2 are the balance coefficients; Definition of the vertebral structural index to quantify the degree of osteoporosis: in: SMI: structural model index, ranging from [0 to 3], where 0 represents plate-like and 3 represents rod-like; BV / TV: bone volume fraction, normal value 0.15-0.25; Tb.Sp: trabecular spacing, in mm, normal value 0.5-1.0 mm; Tb.Th: trabecular bone thickness, unit: mm, normal value: 0.1-0.2mm.

6. The osteoporosis prediction method based on vertebral CT images according to claim 1, characterized in that: The fracture probability calculation adopts an improved FRAX scoring system, which integrates multiple risk factors such as age, bone density, previous fracture history and structural index through a logistic regression model. The output value is mapped to the range of 0-1 by a sigmoid function as the fracture probability.

7. The osteoporosis prediction method based on vertebral CT images according to claim 1, characterized in that: Introducing timing consistency constraints: L temp =||f(I t+1 )-(f(I t )+Δf)||2 in: f(I t ): image features at time point t; Δf: expected change based on clinical evolution law; Δf=g(Δt,Age,BMD t ), g is a learnable time series prediction function implemented by the gated recurrent unit GRU; Δt: time interval (month).

8. The osteoporosis prediction method based on vertebral CT images according to claim 1, characterized in that: Use deep integration strategies to improve prediction stability: Train 5 sub-models with different random initializations; During prediction, the output of each sub-model is weighted averaged, and the weights are dynamically adjusted based on the performance of the validation set; Calculate the forecast variance as a measure of uncertainty and define the trigger mechanism: in: Prediction variance θ uncerrtainty : Uncertainty threshold, set according to clinical needs, typical value is 0.3; When Trigger=1, the system marks the case as requiring manual review.

9. An osteoporosis prediction system based on vertebral CT images, characterized in that: Implementing the osteoporosis prediction method based on vertebral CT images as described in any one of claims 1 to 8, comprising: Image acquisition module: receives DICOM format CT data, supports 0.625-2.5mm slice thickness, and automatically identifies lumbar spine sequences; Bone structure perception module: performs adaptive convolution and non-uniform sampling to extract direction-sensitive features; Biomechanical modeling module: constructs intervertebral conduction map network and models pathological propagation patterns; Risk assessment module: calculates the fracture risk sensitivity loss and outputs the T-value grading results; Quality control module: evaluates prediction uncertainty and marks cases for review when the variance exceeds a threshold; Visualization module: Generates 3D heat maps to highlight low-density areas and potential lesions.

Images