Lumbar fracture bone cement distribution prediction method based on artificial intelligence

Abstract

The invention relates to the technical field of medical treatment, and discloses a lumbar fracture bone cement distribution prediction method based on artificial intelligence, and the method comprises the following steps: S1, data acquisition and screening; s2, image preprocessing and format standardization; s3, sketching a region of interest; s4, image omics feature extraction; s5, constructing and training a prediction model; s6, model verification and prediction; and S7, clinical decision assistance. According to the artificial intelligence-based lumbar fracture bone cement distribution prediction method, the high-dimensional image omics characteristics are extracted from the three-dimensional CT image, and the machine learning model is combined, so that the bone cement distribution can be accurately predicted before an operation, the empirical operation depending on fluoroscopy in the operation is reduced, and the risk of complications such as bone cement leakage is remarkably reduced; meanwhile, clinical decision making and operation scheme optimization are assisted, an objective prediction tool is provided for a doctor, the doctor is helped to evaluate the operation scheme and select the feasibility of unilateral vertebroplasty before an operation, and high-risk patients are recognized in advance, so that a treatment strategy is optimized.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to an artificial intelligence-based method for predicting the distribution of bone cement in lumbar vertebral fractures. Background Technology

[0002] Lumbar vertebral fractures are common osteoporotic fractures, especially prevalent among the elderly. Minimally invasive vertebral body reinforcement surgery (such as unilateral vertebral body reinforcement surgery) is a common method for treating lumbar vertebral fractures. This involves injecting bone cement into the vertebral body to stabilize the fracture site. However, current techniques have significant drawbacks: the distribution characteristics of bone cement within the vertebral body cannot be accurately predicted preoperatively, and intraoperative observation of bone cement diffusion relies solely on C-arm X-ray fluoroscopy. This leads to surgical plans depending on the surgeon's experience, a lack of objective predictive tools, and a high risk of uneven bone cement distribution or leakage during surgery, which may cause serious complications such as nerve compression and pulmonary embolism, affecting patient prognosis.

[0003] In recent years, artificial intelligence (AI) and radiomics technologies have been applied in the medical field, such as extracting features from CT or MRI images to assist in disease diagnosis. However, existing technologies are mostly focused on tumor or general fracture analysis and have not yet covered the prediction of bone cement distribution after lumbar fracture surgery. Traditional methods lack in-depth mining of three-dimensional CT images and cannot quantify the key features of bone cement distribution from multidimensional data. In addition, due to limited resources, primary hospitals need a simple and efficient prediction tool to guide clinical decision-making.

[0004] Therefore, there is an urgent need for an AI-based predictive method that can accurately predict bone cement distribution before surgery, optimize surgical plans, reduce intraoperative risks, and improve patient prognosis. Summary of the Invention

[0005] Given that the existing technologies mentioned above are mostly focused on tumor or general fracture analysis and have not yet addressed the prediction of bone cement distribution after lumbar fracture surgery, and that traditional methods lack in-depth mining of three-dimensional CT images and cannot quantify the key features of bone cement distribution from multidimensional data, this invention is proposed.

[0006] Therefore, the purpose of this invention is to provide an artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures. The aim is to improve prediction accuracy and surgical safety: by extracting high-dimensional radiomics features from three-dimensional CT images and combining them with machine learning models, the distribution of bone cement can be accurately predicted preoperatively, reducing the reliance on fluoroscopy-based experience during surgery and significantly reducing the risk of complications such as bone cement leakage.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence, comprising the following steps: S1: Data Acquisition and Screening: Acquire preoperative lumbar spine 3D CT images of patients with osteoporotic compression fractures of the lumbar spine, and screen the patient cohort based on preset inclusion and exclusion criteria; S2: Image preprocessing and format standardization: The three-dimensional CT images are preprocessed to unify the image resolution and scale, and the image format is converted into NIFTI format suitable for feature extraction; S3: Region of Interest Delineation: On the preprocessed 3D CT image, delineate the region of interest for the target vertebra, generate the corresponding ROI mask, and perform data augmentation to increase sample diversity; S4: Radiomics Feature Extraction: Based on the ROI mask, extract a quantitative set of radiomics features from the 3D CT image; S5: Predictive Model Construction and Training: Using machine learning algorithms, feature selection and model training are performed on the radiomics feature set to construct a bone cement distribution prediction model; S6: Model Validation and Prediction: The trained prediction model is validated in an independent validation queue, and the 3D CT data of the patient to be predicted is input into the validated model to output the predicted results of bone cement distribution. S7: Clinical Decision Support: The predicted results are used to assist doctors in formulating surgical plans for unilateral vertebral body enhancement surgery, assessing leakage risks, and predicting patient prognosis.

[0008] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar fractures according to the present invention, the preprocessing in S2 includes resampling the three-dimensional CT image using a B-spline interpolation algorithm to unify the image voxel resolution to 1mm×1mm×1mm.

[0009] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures according to the present invention, wherein: the region of interest (ROI) in step S3 is drawn manually or semi-automatically using 3D Slicer or ITK-SNAP software tools, and the ROI covers the entire cancellous bone region within the vertebral body where bone cement is expected to be injected.

[0010] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures according to the present invention, the data augmentation operation in S3 includes one or more transformations of the ROI mask and its corresponding original image, such as random rotation, flipping, translation, or adding noise, to enhance the generalization ability of the model.

[0011] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures according to the present invention, wherein: the radiomics feature set in S4 is extracted using the PyRadiomics open-source library, and the feature set includes, but is not limited to, first-order statistical features, shape features, gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level size region matrix features, and neighborhood gray-level difference matrix features.

[0012] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures according to the present invention, wherein: S5 specifically includes: Feature selection: The LASSO regression model was used to select the subset of features with non-zero coefficients that were most correlated with the distribution of bone cement from the radiomics feature set; Model training: Using the aforementioned feature subset, a regression or classification prediction model is constructed using the XGBoost ensemble learning algorithm; Parameter optimization: During model training, five-fold cross-validation is used in combination with grid search or Bayesian optimization methods to tune the model hyperparameters in order to determine the optimal model.

[0013] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar fractures according to the present invention, wherein: before S5, the sample dataset is randomly divided into a training queue and an internal validation queue according to a preset ratio, wherein the training queue is used for model building and parameter optimization, and the internal validation queue is used for preliminary evaluation of model performance.

[0014] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar fractures according to the present invention, wherein: the performance verification evaluation index in S6 includes one or more of the following: mean squared error, mean absolute error, area under receiver operating characteristic curve, or decision curve analysis.

[0015] As a preferred embodiment of the artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures according to the present invention, wherein: the prediction results in S7 are presented in the form of a visualization atlas and / or a quantitative report, the visualization atlas including the predicted bone cement diffusion morphology and distribution range simulated and displayed on a three-dimensional reconstruction model.

[0016] The method is integrated into a software system or platform, which includes an image uploading module, an automated processing module, a result display module, and a clinical report generation module.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: 1. This invention extracts high-dimensional radiomics features from three-dimensional CT images and combines them with machine learning models to accurately predict bone cement distribution preoperatively, reduce reliance on fluoroscopy during surgery, significantly reduce the risk of complications such as bone cement leakage, and assist in clinical decision-making and optimization of surgical plans. It provides doctors with objective predictive tools to help them assess surgical plans preoperatively, select the feasibility of unilateral vertebral body enhancement surgery, and identify high-risk patients in advance, thereby optimizing treatment strategies.

[0018] 2. This invention utilizes automated image processing and feature extraction technology to reduce manual intervention and shorten preoperative planning time. It is suitable for promotion in primary hospitals, improving the diagnosis and treatment level of elderly people with osteoporotic fractures. At the same time, it deeply mines the potential features in images through radiomics, quantifies the distribution pattern of bone cement that cannot be identified by the naked eye, provides data support for personalized treatment of lumbar fractures, promotes the application of artificial intelligence in the field of orthopedics, and is applicable to complex lumbar fractures. It can reduce the cost of postoperative revision and complication management, serving an aging society and has high clinical application value and market prospects. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall framework of the artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures according to the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Reference Figure 1 This paper presents an artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures. This method includes the following: S1: Data Acquisition and Screening: Acquire preoperative lumbar spine 3D CT images of patients with osteoporotic compression fractures of the lumbar spine, and screen the patient cohort based on preset inclusion and exclusion criteria; S2: Image preprocessing and format standardization: The three-dimensional CT images are preprocessed to unify the image resolution and scale, and the image format is converted into NIFTI format suitable for feature extraction; The preprocessing described in S2 includes resampling the three-dimensional CT image using the B-spline interpolation algorithm to unify the image voxel resolution to 1mm×1mm×1mm; S3: Region of Interest Delineation: On the preprocessed 3D CT image, delineate the region of interest for the target vertebra, generate the corresponding ROI mask, and perform data augmentation to increase sample diversity; The region of interest described in S3 is drawn manually or semi-automatically using 3D Slicer or ITK-SNAP software tools. The ROI covers the entire cancellous bone region within the vertebral body where bone cement is to be injected. The data augmentation operation described in S3 includes performing one or more transformations on the ROI mask and its corresponding original image, such as random rotation, flipping, translation, or adding noise, to enhance the generalization ability of the model. S4: Radiomics Feature Extraction: Based on the ROI mask, extract a quantitative set of radiomics features from the 3D CT image; The radiomics feature set mentioned in S4 is extracted using the PyRadiomics open-source library. The feature set includes, but is not limited to, first-order statistical features, shape features, gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level size region matrix features, and neighborhood gray-level difference matrix features. S5: Predictive Model Construction and Training: Utilizing machine learning algorithms to perform feature selection and model training on the aforementioned radiomics feature set, a bone cement distribution prediction model is constructed. S5 specifically includes: Feature selection: The LASSO regression model was used to select the subset of features with non-zero coefficients that were most correlated with the distribution of bone cement from the radiomics feature set; Model training: Using the aforementioned feature subset, a regression or classification prediction model is constructed using the XGBoost ensemble learning algorithm; Parameter optimization: During model training, five-fold cross-validation is used in combination with grid search or Bayesian optimization methods to tune the model hyperparameters in order to determine the optimal model; Before S5, the sample dataset is randomly divided into a training queue and an internal validation queue according to a preset ratio. The training queue is used for model building and parameter optimization, while the internal validation queue is used for preliminary evaluation of model performance. S6: Model Validation and Prediction: The trained prediction model is validated in an independent validation queue, and the 3D CT data of the patient to be predicted is input into the validated model to output the predicted results of bone cement distribution. The performance verification evaluation metrics described in S6 include one or more of the following: mean squared error, mean absolute error, area under the receiver operating characteristic curve, or decision curve analysis. S7: Clinical Decision Support: The predicted results will be used to assist doctors in developing surgical plans for unilateral vertebral body enhancement surgery, assessing leakage risks, and predicting patient prognosis. The prediction results described in S7 are presented in the form of visualization maps and / or quantitative reports, wherein the visualization maps include simulations of the predicted bone cement diffusion morphology and distribution range on a three-dimensional reconstruction model. Example

[0022] First, data from 200 patients diagnosed with osteoporotic compression fractures of the lumbar vertebrae by MRI and who underwent unilateral vertebral body augmentation were retrospectively collected from the hospital's PACS system. All patients met the following criteria: Inclusion criteria: ① Age greater than 50 years; ② Clinically diagnosed with osteoporotic fracture of the lumbar spine; ③ Possession of complete preoperative lumbar spine 3D CT thin-slice scan data (slice thickness <1mm) and 3D CT data within 24 hours postoperatively for verification.

[0023] Exclusion criteria: ① History of spinal surgery; ② Presence of bone metastases from malignant tumors; ③ Suffering from diseases such as Parkinson's disease or ankylosing spondylitis that may affect bone metabolism or spinal morphology.

[0024] The collected preoperative 3D CT data in DICOM format were imported into the computer system. Subsequently, the images were preprocessed using the SimpleITK library in Python. Resampling: The B-spline interpolation algorithm is used to resample the voxel resolution of all 3D CT images to a uniform 1mm×1mm×1mm to eliminate scale differences caused by different scanning equipment and parameters and ensure the consistency of feature extraction.

[0025] Format conversion: The preprocessed DICOM image sequence is converted into an NIFTI format file for each vertebra, which is more suitable for subsequent radiomics feature extraction and machine learning processing.

[0026] II. Region of Interest Delineation and Data Augmentation Using the open-source software 3D Slicer, two experienced radiologists independently manipulated pre-processed preoperative 3D CT images: For the target vertebral body requiring surgery, a region of interest (ROI) is manually delineated. The ROI precisely covers the entire cancellous bone region within the vertebral body where bone cement is expected to be injected, carefully avoiding cortical bone and the vertebral foramen, thereby generating a binarized ROI mask.

[0027] This process takes an average of about 15 minutes per vertebra. The ROIs drawn by the two physicians were tested for consistency using the Descein coefficient, and the ROI masks with high consistency were used for subsequent analysis.

[0028] To increase data diversity and improve model robustness, data augmentation was performed on the delineated ROI masks and their corresponding original images. Specific operations included: Random rotation (within ±10°), horizontal flipping, and small translation (within ±5 voxels) are performed in three-dimensional space.

[0029] In addition, Gaussian noise was added to some of the images.

[0030] Ultimately, the original dataset was augmented by approximately four times, effectively preventing model overfitting.

[0031] III. Radiomics Feature Extraction High-throughput feature extraction was performed using the mature open-source Python library PyRadiomics (version 3.0). NIFTI format images and corresponding ROI mask files were input into PyRadiomics, which automatically extracted a large number of quantitative features from each ROI. The feature set mainly includes: First-order statistical characteristics: describe the distribution of voxel intensity within an ROI, such as energy, entropy, skewness, kurtosis, etc.

[0032] Shape characteristics: Describe the three-dimensional geometric properties of the ROI, such as volume, surface area, sphericity, etc.

[0033] Second-order and higher-order texture features include gray-level co-occurrence matrix features (contrast, correlation), gray-level run-length matrix features (long run advantage), gray-level size region matrix features (size region variability), and neighborhood gray-level difference matrix features (busyness, roughness).

[0034] In this embodiment, an average of about 1200 radiomics features were extracted from each ROI, which constituted the basic feature set for subsequent model training.

[0035] IV. Construction and Training of Predictive Models The 200 case samples were randomly divided in a 4:1 ratio, with 160 cases serving as the training cohort and 40 cases serving as an independent validation cohort.

[0036] The model building and training process is as follows: Feature filtering: First, the approximately 1200 features extracted from the training queue are standardized (Z-score).

[0037] Then, using the actual bone cement distribution volume on postoperative CT scans as a continuous variable label, a LASSO regression model was employed for feature selection. The optimal penalty coefficient λ was determined through 10-fold cross-validation, ultimately selecting 25 features with non-zero coefficients. These features were considered to have the strongest correlation with bone cement distribution, constituting the optimal feature subset.

[0038] Model training: Using 25 selected features, a regression prediction model was constructed using the XGBoost ensemble learning algorithm to predict the distribution volume and spatial pattern of bone cement.

[0039] Parameter optimization: During model training, five-fold cross-validation was used, combined with grid search, to fine-tune the key hyperparameters of the XGBoost model (such as learning rate, maximum depth, subsampling ratio, etc.) in order to determine the model with the best performance.

[0040] V. Model Validation and Prediction The trained optimal model was then validated on a reserved queue of 40 independent validation cases.

[0041] The preoperative CT data of patients in the validation set were processed through the same S1-S4 steps as the training set and then input into the model to obtain the predicted value of bone cement distribution.

[0042] The predicted values ​​were compared with the actual bone cement distribution measured by postoperative CT scans. Evaluation metrics showed a high degree of agreement between the model's predictions and the actual values: the mean squared error was 0.15, and the mean absolute error was 0.28 mL. Furthermore, when the prediction task was transformed into a classification problem of "whether clinically significant leakage will occur," the model's area under the ROC curve reached 0.91, demonstrating extremely high discriminative power. Decision curve analysis also indicated that the model has significant clinical net benefits over a wide threshold range.

[0043] VI. Clinical Decision Support The model's predictions are presented to doctors through an integrated software platform.

[0044] The results display module generates a quantitative report of the prediction results, including the predicted bone cement volume, distribution uniformity index, and leakage risk probability.

[0045] Meanwhile, the platform uses a visualization atlas to simulate and display the predicted bone cement diffusion morphology and distribution range on the three-dimensional reconstruction model of the patient's vertebral body with different colors and transparency, intuitively showing the "cement cloud".

[0046] Clinicians can use this information before surgery: Develop a surgical plan: Assess whether unilateral puncture is sufficient or whether the puncture path needs to be adjusted.

[0047] Assess leakage risk: Prepare contingency plans in advance for high-risk patients, such as preparing for balloon dilation or adjusting cement viscosity.

[0048] Assessing patient prognosis: Based on the predicted distribution, more accurately explain the surgical outcome and risks to patients and their families.

[0049] In summary, through the above specific implementation steps, this invention successfully constructed an accurate and reliable bone cement distribution prediction model and seamlessly integrated it into the clinical workflow, providing a powerful decision support tool for the precise surgical treatment of lumbar fractures.

[0050] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence, characterized in that, Including the following: S1: Data Acquisition and Screening: Acquire preoperative lumbar spine 3D CT images of patients with osteoporotic compression fractures of the lumbar spine, and screen the patient cohort based on preset inclusion and exclusion criteria; S2: Image preprocessing and format standardization: The three-dimensional CT images are preprocessed to unify the image resolution and scale, and the image format is converted into NIFTI format suitable for feature extraction; S3: Region of Interest Delineation: On the preprocessed 3D CT image, delineate the region of interest for the target vertebra, generate the corresponding ROI mask, and perform data augmentation to increase sample diversity; S4: Radiomics Feature Extraction: Based on the ROI mask, extract a quantitative set of radiomics features from the 3D CT image; S5: Predictive Model Construction and Training: Using machine learning algorithms, feature selection and model training are performed on the radiomics feature set to construct a bone cement distribution prediction model; S6: Model Validation and Prediction: The trained prediction model is validated in an independent validation queue, and the 3D CT data of the patient to be predicted is input into the validated model to output the predicted results of bone cement distribution. S7: Clinical Decision Support: The predicted results are used to assist doctors in formulating surgical plans for unilateral vertebral body enhancement surgery, assessing leakage risks, and predicting patient prognosis.

2. The method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence according to claim 1, characterized in that: The preprocessing in S2 includes resampling the three-dimensional CT image using a B-spline interpolation algorithm to unify the image voxel resolution to 1mm×1mm×1mm.

3. The method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence according to claim 1, characterized in that: The region of interest (ROI) described in S3 is drawn manually or semi-automatically using 3D Slicer or ITK-SNAP software tools. The ROI covers the entire cancellous bone region within the vertebral body where bone cement is to be injected.

4. The method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence according to claim 3, characterized in that: The data augmentation operation in S3 includes performing one or more transformations on the ROI mask and its corresponding original image, such as random rotation, flipping, translation, or adding noise, to enhance the generalization ability of the model.

5. The method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence according to claim 1, characterized in that: The image omics feature set mentioned in S4 is extracted using the PyRadiomics open-source library. The feature set includes, but is not limited to, first-order statistical features, shape features, gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level size region matrix features, and neighborhood gray-level difference matrix features.

6. The method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence according to claim 1, characterized in that: S5 specifically includes: Feature selection: The LASSO regression model was used to select the subset of features with non-zero coefficients that were most correlated with the distribution of bone cement from the radiomics feature set; Model training: Using the aforementioned feature subset, a regression or classification prediction model is constructed using the XGBoost ensemble learning algorithm; Parameter optimization: During model training, five-fold cross-validation is used in combination with grid search or Bayesian optimization methods to tune the model hyperparameters in order to determine the optimal model.

7. The artificial intelligence-based method for predicting bone cement distribution in lumbar vertebral fractures according to claim 6, characterized in that: Before step S5, the sample dataset is randomly divided into a training queue and an internal validation queue according to a preset ratio. The training queue is used for model building and parameter optimization, while the internal validation queue is used for preliminary evaluation of model performance.

8. The method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence according to claim 7, characterized in that: The performance verification evaluation metrics in S6 include one or more of the following: mean squared error, mean absolute error, area under the receiver operating characteristic curve, or decision curve analysis.

9. The method for predicting bone cement distribution in lumbar vertebral fractures based on artificial intelligence according to claim 8, characterized in that: The prediction results in S7 are presented in the form of a visualization atlas and / or a quantitative report, wherein the visualization atlas includes a simulation display of the predicted bone cement diffusion morphology and distribution range on a three-dimensional reconstruction model.

10. The method according to claim 1, characterized in that, The method is integrated into a software system or platform, which includes an image uploading module, an automated processing module, a result display module, and a clinical report generation module.

Images