A method and system for predicting the demand for total knee arthroplasty

Abstract

This invention discloses a method and system for predicting the demand for total knee arthroplasty (TKR), relating to the field of clinical decision support. The method includes: acquiring knee MRI images; automatically segmenting key structures of the joint space using a deep convolutional neural network to obtain the volume of interest (VOI); extracting radiomics features according to the IBSI standard using the SERA tool; and finally constructing a joint space plus clinical features radiomics model (JSC-RM) by combining the Knee Osteoarthritis Outcome Survey (KOOS) score to predict whether TKR will be performed within the next 1-4 years. The method employs LASSO feature selection and shallow neural network modeling, combined with Youden's index to set thresholds, resulting in predictive performance superior to existing models. It significantly improves the consistency and accuracy of resident physicians' judgments on the timing of TKR, achieving intelligent preoperative risk intervention.

Description

Technical Field

[0001] This invention relates to the field of clinical decision support, and in particular to a method and system for predicting the demand for total knee arthroplasty. Background Technology

[0002] Knee osteoarthritis is one of the most common degenerative diseases among middle-aged and elderly people, and in its end-stage, it usually requires total knee replacement. Although total knee replacement can effectively relieve pain and improve function, it is a high-risk, high-cost, and irreversible surgical procedure. Currently, clinical decisions regarding total knee replacement rely heavily on physician experience and subjective judgment, lacking unified, objective, and quantifiable predictive tools.

[0003] Current technologies do not integrate multi-structure MRI with KOOS scores, making quantitative early warning impossible. In addition, while some existing technologies can predict total knee replacement, their models are "black box" structures, lacking structural segmentation, symptom score input, and clinical interpretability. They also lack radiomics analysis or fusion of clinical scores, limiting prediction depth and accuracy. Furthermore, they lack MRI resolution and tissue-level information, failing to fully characterize the details of cartilage and meniscus degeneration, and do not introduce subjective scoring or structurally interpretable mechanisms. Summary of the Invention

[0004] To address the lack of quantitative standards for predicting total knee replacement (TKR) in existing technologies, this invention provides a method and system for predicting TKR demand. This invention proposes an artificial intelligence model that integrates MRI structural recognition, radiomics analysis, and clinical symptom scoring to achieve structural, individualized, and quantifiable predictions of TKR.

[0005] This invention is achieved through the following technical solution:

[0006] A method for predicting the demand for total knee arthroplasty includes:

[0007] Step S1, Image Acquisition and Preprocessing: Acquire MRI images of the knee joint, automatically segment the structure into several parts, and obtain the volume of interest for each structure;

[0008] Step S2, Image feature extraction: Use the SERA radiomics feature extraction tool to extract radiomics features from the volume of interest.

[0009] Step S3: Clinical scoring data collection and integration, obtaining the corresponding patient's clinical symptom scoring data, including various scores in the KOOS scale for knee osteoarthritis outcomes survey;

[0010] Step S4: Predictive model building and output. The fusion of radiomics features and KOOS score features is input into the predictive model to output the probability value of future total knee replacement.

[0011] Step S5, Predicted Output and Visualization: Based on the set risk threshold, the probability values ​​of total knee replacement are classified and visualized in the form of color coding, structural heatmaps, or Shapley interpretation diagrams.

[0012] Furthermore, the automatic segmentation is based on a convolutional neural network model with a 3D U-Net structure, optimized using the Dice loss function, and supports joint segmentation of multiple structures.

[0013] Furthermore, the aforementioned components include four types of tissues: femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus.

[0014] Furthermore, the radiomics features include first-order statistical features, gray-level co-occurrence matrix features, gray-level run-length matrix features, and shape features.

[0015] Furthermore, the radiomics features are screened using the LASSO method. By adding an L1 regularization term to the feature coefficients, the coefficients of irrelevant or redundant features are automatically reduced to 0 during the optimization process.

[0016] Furthermore, the KOOS scale includes four indicators: pain score, symptom score, motor function score, and quality of life score.

[0017] Furthermore, the prediction model is a shallow neural network model containing two hidden layers, employing the ReLU activation function, and the output layer uses the Sigmoid function to generate the probability value of total knee replacement.

[0018] Furthermore, the set risk threshold is obtained through the Youden index and is used to balance the sensitivity and specificity of the model.

[0019] This invention also provides a total knee arthroplasty demand prediction system, based on the total knee arthroplasty demand prediction method described above, comprising:

[0020] The image acquisition and preprocessing module is used to receive or load knee joint MRI image data and perform segmentation.

[0021] The image feature extraction module is used to extract radiomics features from regions of interest based on the SERA code package.

[0022] The clinical information collection and integration module is used to obtain the patient's KOOS score scale results, including four dimensions: pain, symptoms, function, and quality of life.

[0023] The predictive modeling module is used to fuse radiomics features and clinical scoring features into the predictive model and output the probability value of whether to undergo total knee replacement in the future.

[0024] The prediction output and visualization module is used to classify the prediction results for risk and output a visualization interface, including structural heatmaps, SHAP structural contribution interpretation diagrams, and color-coded prompts for the probability values ​​of total knee replacement.

[0025] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing program instructions for a total knee arthroplasty prediction method, the program instructions for the total knee arthroplasty prediction method being executable by one or more processors to implement the steps of the total knee arthroplasty prediction method as described above.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] It achieves fully automated image analysis: by constructing a convolutional neural network model, it automatically completes the three-dimensional automatic segmentation of key structures in the joint space (femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus) in the MRI images of the knee joint, avoiding the subjective bias and inefficiency of traditional manual sketching, and significantly improving the accuracy and repeatability of structure recognition.

[0028] Integrating radiomics and clinical scoring information to improve prediction accuracy: Based on radiomics feature extraction, this invention introduces KOOS scoring data to finally construct a joint space plus clinical feature radiomics model (JSC-RM). Compared with models that rely solely on MOAKS structural scores or clinical scales, it has higher predictive ability and stronger clinical interpretability.

[0029] Supports mid- to long-term preoperative risk prediction and advances the intervention window: The model outputs an individualized risk probability of undergoing TKR in the next 1 to 4 years, which helps clinicians identify patients with potential progression in advance, optimize the timing of preoperative intervention, and achieve proactive diagnosis and treatment management.

[0030] It provides a visual decision support interface to enhance the interpretability experience for doctors: the model output results are presented intuitively in the form of structural heatmaps, TKR probability color coding and SHAP contribution interpretation graphs, which help resident physicians improve the consistency and accuracy of TKR judgments and have good scalability and practical value.

[0031] Significantly superior performance compared to traditional models: The model of this invention has significantly better AUC than the MOAKS scoring model and clinical models on the independent test set, and significantly improved sensitivity, specificity and Kappa consistency index in physician-assisted decision-making experiments, demonstrating high clinical application prospects. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0033] Figure 1 This is a schematic flowchart of a prediction method for total knee replacement according to an embodiment of this application;

[0034] Figure 2 This is an automatic segmentation map of knee MRI using deep learning, according to an embodiment of this application.

[0035] Figure 3 These are joint space characteristic heatmaps of the control group and case group according to embodiments of this application;

[0036] Figure 4 The final model in the embodiments of this application predicts the result of total knee replacement.

[0037] Figure 5 The final model based on the embodiments of this application is used to assist surgeons in predicting the improvement of total knee replacement capability. Detailed Implementation

[0038] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0039] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0041] See Figure 1 A predictive method for total knee replacement includes the following steps:

[0042] S1: Image acquisition and preprocessing;

[0043] Knee MRI images of patients with KOA were acquired using the SAG-3D-DESS-WE sequence. After resampling, grayscale normalization, and registration, three-dimensional input images were generated.

[0044] The imaging parameters are as follows:

[0045]

[0046] This sequence has high spatial resolution (0.365 mm X resolution, 0.456 mm Y resolution), which can clearly display the tissues in the joint space such as cartilage and meniscus, providing a reliable data foundation for subsequent three-dimensional structural segmentation and radiomics analysis.

[0047] The structure is automatically segmented using a deep learning algorithm based on 3D CNN to automatically segment the joint space region in MRI images. The segmentation targets include: femoral cartilage; tibial cartilage; medial meniscus; and lateral meniscus, accurately obtaining the volume of interest (VOI) for each structure.

[0048] like Figure 2 This diagram illustrates an automatic segmentation scheme for knee joint MRI structures according to an embodiment of the present invention. It demonstrates the high-precision automatic segmentation of key anatomical structures of the joint using a depth model on SAG-3D-DESS-WE sequence MRI images. This diagram represents a core intermediate step in the model's implementation of image data structuring, and forms the basis for subsequent radiomics feature extraction and risk modeling. The segmented structures shown in the diagram include the following four types of joint space structures: femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus.

[0049] The segmentation model used in this invention is based on a three-dimensional convolutional neural network (3D U-Net). The input is MRI volume data, and the output is a multi-channel segmentation mask, with each channel corresponding to one of the four anatomical structures mentioned above. The model training process uses the Dice loss function for optimization, combined with manually annotated gold-standard structural masks from the OAI database, to ensure accurate anatomical boundary recognition.

[0050] In the image, each structure is coded with a different color and superimposed on the MRI background, clearly demonstrating the model's ability to automatically identify structures in the joint space region. Specifically, the cartilage structures are shown to completely cover the bone end face; the medial and lateral menisci are located between the tibial plateau, with clear structural outlines and uniform signal distribution.

[0051] The DSC mean of the automatic structure segmentation achieved by this implementation scheme exceeds 0.80 in all six structures, indicating that it can provide high-quality VOI for subsequent radiomics feature extraction, thereby improving feature stability and model prediction reliability.

[0052] like Figure 3 This is a schematic diagram of the joint space feature heatmap generated in an embodiment of the present invention. It shows the visualization image formed after quantifying the grayscale features of key structural regions in a knee joint MRI image. The left image is the control group, and the right image is the case group. This image calculates the average signal intensity of each pixel in structures such as the femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus, and uses a color heatmap to represent the grayscale value differences in different tissue areas, reflecting the degree of structural degeneration.

[0053] The implementation principle is as follows:

[0054] (1) Data source: High-resolution knee MRI images acquired using the SAG-3D-DESS-WE sequence, which provides good contrast for cartilage and meniscus tissue.

[0055] (2) Structural segmentation: The MRI image is automatically segmented using a CNN model to obtain the masks corresponding to each structure, including: femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus.

[0056] (3) Feature definition: The joint space feature heatmap refers to the calculation of the average signal intensity (i.e., the average pixel value) and color encoding within each tissue structure region. This average value can represent the tissue density or water content status and has strong degeneration sensitivity.

[0057] (4) Image construction: Replace the pixel values ​​of each structural region with the average value within the structure and display them in the form of a heatmap to visualize the grayscale differences of different structures, thereby forming a three-dimensional joint space feature heatmap.

[0058] Figure 3 Different colors represent the average grayscale values ​​of different structures; red / yellow areas indicate higher signal intensity, which may correspond to inflammatory changes or fluid accumulation; blue / dark areas indicate signal attenuation, suggesting cartilage thinning or meniscus degeneration; it can be used as a visualization tool for interpreting structural radiomics, supporting doctors in understanding the basis of model predictions.

[0059] S2: Image Feature Extraction

[0060] Using the Standard Environment for Radiomics Analysis (SERA) code package, three-dimensional radiomics features based on the IBSI standard are extracted from the four segments of Volume of Interest (VOI). These features include: first-order statistical features (such as average gray level, standard deviation, skewness, etc.); second-order texture features (such as GLCM, GLRLM, etc.); and three-dimensional shape features (such as volume, surface area, aspect ratio, etc.).

[0061] Features are extracted from the following structures: femoral cartilage (FC), tibial cartilage (TC), medial meniscus (MM), and lateral meniscus (LM); feature types extracted: first-order statistical features, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and three-dimensional shape features; output format: the default output is a Python dict, but it can also be saved as a csv, xlsx, or json file for subsequent LASSO modeling or neural network use.

[0062] Feature selection and model building:

[0063] This invention extracts radiomics features from four structures: femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus, forming four feature sets. To reduce dimensionality, redundancy, and improve model stability, this invention employs LASSO for feature selection. LASSO regression automatically shrinks the coefficients of irrelevant or redundant features to 0 during optimization by adding an L1 regularization term to the feature coefficients, thereby achieving sparsity selection. Its advantage lies in retaining features with the strongest correlation to the target variable, improving the model's generalization ability.

[0064] The feature selection process includes:

[0065] (1) Input features: SERA output features of each structure, preferably, each VOI has about 200 to 400 dimensions, and there are four groups in total;

[0066] (2) Standardization: All features are standardized using Z-score;

[0067] (3) Label variable processing: The event label of total knee arthroplasty is selected by binary classification, where 0 = not having undergone total knee arthroplasty, and 1 = having undergone total knee arthroplasty;

[0068] (4) LASSO regression modeling: use cross-validation (10-fold CV) in the training set to determine the regularization parameter λ; select non-zero weight features as the final input; output non-zero weight features and coefficients as the basis for structural level interpretation.

[0069] Modeling approach: Independent shallow neural network prediction models were constructed for each structure (femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus), forming four single-structure radiomics models; each model was used to analyze the correlation between the degenerative characteristics of the corresponding structure and the occurrence of total knee arthroplasty.

[0070] After the above steps are completed, each structure can output several most relevant features, which can be used for subsequent single-model performance evaluation (AUC, sensitivity, specificity) and joint modeling input.

[0071] S3: Clinical scoring data collection and integration

[0072] The KOOS score scale for corresponding patients was collected simultaneously, including four dimensions: pain, symptoms, function, and quality of life. These scores were used as input variables to reflect the patients' subjective functional status. All scores were obtained from the questionnaires completed by the patients at the time of the MRI examination and represent their subjective perception of knee joint function.

[0073] S4: Predictive Model Building and Output

[0074] The MRI radiomics features and KOOS scores are concatenated and standardized using Z-score, and then used as input vectors to train and predict a shallow neural network.

[0075] The fusion strategy is as follows:

[0076] (1) Feature stitching: The image omics feature vectors of each structure after LASSO screening are merged; and stitched with the four KOOS scoring items into a joint feature vector; all features are Z-score standardized to ensure uniformity of dimensions and consistent units.

[0077] (2) Model structure: A shallow feedforward neural network is used as the fusion modeler; the network structure contains two hidden layers, uses the ReLU activation function, and finally outputs a 1 Sigmoid node, which represents the probability of needing total knee replacement; the cross-entropy loss function is used for optimization.

[0078] (3) Output model: The joint space plus clinical feature radiomics model (JSC-RM) was constructed. This model effectively integrates the quantitative information of images at the structural level with the functional scores at the symptom level.

[0079] (4) Training and validation: The model was trained in the existing total knee replacement osteoarthritis sub-cohort (the Pivotal Osteoarthritis initiative Magnetic resonance imaging Analyses Total Knee Replacement osteoarthritis cohort, POMA-TKR), and stability was evaluated using 10-fold cross-validation. Predictive performance, including AUC, sensitivity, specificity, and Kappa value, was evaluated on an independent test set. The performance of the model without KOOS score (i.e., the JS-RM model) and the traditional MOAKS model were compared.

[0080] Specifically, the nested cohort key MRI image analysis sub-cohort for total knee arthroplasty osteoarthritis from the OAI database was used as the data source for training and testing. 10-fold cross-validation and independent testing were conducted to evaluate model performance metrics, including AUC, sensitivity, and specificity, to ensure model stability and generalization ability.

[0081] This invention evaluates the performance of the joint model JSC-RM by constructing an independent test set, verifying its effectiveness, stability, and superiority in predicting TKR risk.

[0082] (1) Test data design: The OAI database was used to select the baseline and MRI samples from 3 years, 2 years and 1 year of follow-up to form a multi-time point test set. Each time point test set included cases of "total knee arthroplasty group" and "non-total knee arthroplasty group" in a 1:1 ratio. All images and KOOS score information were not used in the model building stage to ensure the independence of the validation.

[0083] (2) Evaluation indicators and methods: AUC: measures the model’s ability to distinguish; Sensitivity: the proportion of individuals who were actually undergoing total knee arthroplasty identified by the model; Specificity: the proportion of individuals who were correctly identified by the model as not undergoing total knee arthroplasty; Kappa value: measures the consistency between the model’s prediction and the actual label; 95% confidence interval and DeLong test: used to evaluate statistical significance differences with other models.

[0084] (3) Compare the JSC-RM model with existing technologies:

[0085]

[0086] (4) Evaluation results: In the overall test set, the AUC of the JSC-RM model for predicting TKR was 0.85 (95% confidence interval: 0.82~0.88), which was significantly higher than that of the MOAKS model (JS-MOM: AUC=0.67) and the pure clinical model (AUC=0.62). In the test sets at different time points, the sensitivity of JSC-RM was between 75% and 80%, and the specificity was between 72% and 79%. The DeLong test showed that the AUC of JSC-RM was significantly better than that of JS-RM (p<0.05), and also better than that of JSC-MOM and the Clinical model (p<0.001). The Kappa value (indicating predictive consistency) was 0.46~0.61, which was higher than that of all the comparison models. The model performance was stable in the test sets at different time points, reflecting its good generalization ability.

[0087] S5: Predicted Output and Visualization

[0088] The model supports outputting a ranking of structural importance; it visualizes the contribution of structures to predictions (SHapley Additive exPlanations, SHAP); and it provides a probability value and recommended threshold for requiring total knee arthroplasty to assist in clinical judgment.

[0089] The model outputs a predicted probability value for an individual to undergo total knee replacement in the future, and generates a structural risk contribution map or heat map to provide doctors with a visual aid for judgment.

[0090] The method can be embedded in hospital imaging platforms, PACS systems, or deployed in orthopedic outpatient clinics, follow-up assessment systems, or clinical trial platforms as a Web Application Programming Interface (Web API) / local software to provide interpretable and scalable TKR risk prediction support services.

[0091] Visualization output: Receiver operating characteristic curves are used to show the predictive performance of each model on different test sets; a confusion matrix is ​​used to show the number of true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) to assist in clinical interpretation; and a structural contribution map is provided in conjunction with the SHAP interpretation module to improve the transparency of prediction results.

[0092] To further verify the decision-making support capability of the final joint model (JSC-RM) constructed in this invention in clinical applications, this invention designs a simulated resident physician prediction experiment, using the model results to assist physicians in determining whether KOA patients will receive TKR in the future.

[0093] Experimental setup: Seven orthopedic residents with 1-4 years of clinical experience were recruited. Based on MRI images and KOOS scores, physicians independently predicted whether patients in the test set would undergo TKR. The prediction was divided into two rounds: Round 1: The JSC-RM model was not used, and the physicians relied solely on their own experience. Round 2: The JSC-RM model prediction output was provided, and the physicians made corrections based on the results.

[0094] Decision threshold setting: To balance sensitivity and specificity, this invention uses the Youden index in the model training set to set the prediction probability threshold, which is used to find the optimal balance between sensitivity and specificity from the TKR probability output by the model; the formula is: Youden Index = Sensitivity + Specificity − 1; in the JSC-RM model, the optimal TKR prediction probability threshold is set to 0.52; if the prediction probability ≥ 0.52, it indicates "high TKR risk"; if < 0.52, it indicates "low risk".

[0095] Results visualization and color coding: The TKR probability output by the model is displayed in color coding for each case: red indicates high risk (TKR probability ≥ 0.52); green indicates low risk (TKR probability < 0.52); structural importance score and recommendation explanation information (structural heatmap and SHAP contribution map) are also output; all information is displayed on a unified electronic interface for easy and intuitive evaluation by doctors.

[0096] Evaluation of auxiliary effect: The results showed that with the assistance of the JSC-RM model, the predictive ability of resident physicians was significantly improved. The sensitivity and specificity were both increased by more than 30% compared with the absence of model support, and the Kappa value was improved to a moderate consistency level, indicating that the model of the present invention has significant clinical applicability in assisting resident physicians in judgment.

[0097]

[0098] like Figure 4 This paper demonstrates the predictive capabilities of various prediction models constructed in the embodiments of the present invention for TKR risk at different follow-up time points, with AUC as the primary evaluation indicator. The figure includes the performance of the following five models at multiple test time points: Dark blue line represents JS-RM: only fusion of MRI multi-structural radiomics features; Yellow line represents JS-MOM: structural imaging model with MOAKS score; Red line represents JSC-RM: the final fusion model proposed in this invention; Light blue line represents JSC-MOM: traditional imaging model with MOAKS score + KOOS score; Green line represents CM: traditional clinical model with only KOOS score.

[0099] Time point settings: Figure 4It includes five subgraphs, corresponding to the test queues at five time points: baseline, three years from TKR, two years from TKR, one year from TKR, and summarizing all the above time points.

[0100] In each subplot, the receiver operating characteristic curve for each model is plotted, and the AUC value and 95% confidence interval are labeled in the plot to compare the risk prediction capabilities of different models.

[0101] Results Summary: At all time points, the JSC-RM model consistently exhibited the highest AUC value (0.83–0.89), while the JS-RM model also had a high AUC value (0.83–0.91). There was no significant difference in AUC values ​​between the JSC-RM and JS-RM models. However, the JSC-RM model had a higher AUC value on the overall test set than the JS-RM model. Therefore, the JSC-RM model was chosen as the final model, and its AUC value was significantly superior to that of JSC-MOM (0.69–0.72) and JS-RM. OM (0.66~0.67) and clinical model (0.62~0.67); at time points close to TKR (1 year from TKR, 2 years from TKR), JSC-RM showed the most balanced sensitivity and specificity; this result indicates that the fusion model constructed in this invention has the characteristics of strong prospectivity, high timeliness, and strong individualization ability; using the DeLong test to compare the differences in model AUC, the results showed that JSC-RM was significantly better than the control models such as JS-MOM and CM (p<0.001), but not significantly better than the JS-RM model.

[0102] Figure 5 The improvement of model assistance in resident physicians' ability to predict TKR was evaluated. The figure illustrates the positive effect on actual predictive performance.

[0103] The left figure is a comparison of average predictive performance, i.e., all test data combined. It shows the comparison between the final model (JSC-RM) and the average predictive performance of physicians in the test set. The red line represents the prediction results of the JSC-RM model, and the blue line represents the prediction results of the JSC-MOM model. The blue dots represent the average sensitivity / specificity of physicians without model assistance, and the red dots represent the results after using JSC-RM. The black arrows indicate that the predictive ability of resident physicians is significantly improved with model assistance, and the AUC and diagnostic accuracy are significantly increased.

[0104] The right figure shows the changes in prediction performance for seven doctors. It compares the sensitivity and specificity of each doctor under no model (hollow graph) and with JSC-RM assistance (solid graph); the dashed lines connect to represent the prediction trajectory of the same doctor under the two conditions; all doctors show performance improvement, demonstrating the comprehensive enhancement of the judgment ability of young doctors by the JSC-RM model.

[0105] The shallow neural network model (JSC-RM) constructed in this invention integrates MRI radiomics features and KOOS clinical scores to achieve high-precision prediction of TKR risk within the next 4 years in the OAI real cohort. The highest AUC on the test set can reach 0.89, which is significantly better than the traditional MOAKS model (AUC of about 0.66) and the single clinical model (AUC of about 0.62).

[0106] Unlike black-box deep learning models, this invention performs radiomics analysis based on automatically segmented cartilage and meniscus structures. It outputs structural contribution rankings and risk heatmaps, enabling physicians to intuitively understand which joint structures' degeneration is associated with TKR risk, thereby improving model interpretability and clinical adoption. Furthermore, this invention is the first to unify and standardize the KOOS score and radiomics features of MRI images into a unified model, considering both objective structural degeneration and subjective symptom experience, improving the model's ability to identify individual differences and enhancing its adaptability in real clinical scenarios. Through image interpretation experiments, the model constructed in this invention significantly improves the ability of hospitalized orthopedic surgeons to identify TKR risk, increasing sensitivity from 42% to 72% and specificity from 45% to 78%. It provides young orthopedic surgeons with a quantitative support tool, aiding in the development of preoperative intervention plans. The system is flexibly deployable and has broad application prospects. This invention can be embedded in hospital PACS systems or deployed as a web platform, suitable for various scenarios such as outpatient management of osteoarthritis, surgical decision support, rehabilitation follow-up, and clinical trial enrollment, possessing good scalability and translational value.

[0107] In this implementation, the following is achieved:

[0108] (1) Knee joint images were acquired using SAG-3D-DESS-WE sequence MRI images. Compared with two-dimensional X-ray images, it has higher soft tissue contrast and structural resolution, and can accurately reflect the microscopic morphological changes of key tissues such as femoral cartilage, tibial cartilage and meniscus, providing high-quality basic data for TKR risk modeling.

[0109] (2) Automatic segmentation of the four joint spaces in MRI images is achieved by using CNN, including femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus. This solves the problems of low efficiency, large error and reliance on manual labor in traditional segmentation methods, and improves the accuracy and repeatability of structure recognition.

[0110] (3) Using the SERA code package, image omics features conforming to the IBSI standard are extracted from the VOIs of the above four structures, covering first-order statistical features, second-order texture features and shape features. This can comprehensively quantify the gray-scale distribution, texture complexity and spatial morphology of the tissue, which helps to reveal potential structural degradation signals.

[0111] (4) Introducing the KOOS clinical score as a subjective symptom input variable, by integrating the score data of pain, symptoms, function and quality of life, the model can simultaneously consider the patient's structural state and symptom perception, thereby improving the individualized prediction ability.

[0112] (5) Construct a shallow neural network model to model the fused features and output the probability of TKR occurrence. The model has a simple structure, high computational efficiency, and is easy to deploy in the clinical environment. It also has a strong fitting ability and can accurately predict the probability of needing TKR surgery in the next 1 to 4 years.

[0113] (6) Output structural risk scores and visualization results (such as heatmaps or probability bar charts) to help doctors understand the basis of the model and the contribution of each structure, realize the transformation from "prediction black box" to "explainable assistance", and improve doctors' trust in the model prediction results and their willingness to adopt it in clinical practice.

[0114] In summary, this invention, through a combined technical approach of structural segmentation, radiomics, symptom scoring, shallow neural network modeling, and visualization output, not only significantly improves the predictive performance of TKR but also enhances the structural interpretability and clinical applicability of the prediction results, providing an effective auxiliary tool for the precise intervention of KOA.

[0115] This invention also proposes a prediction system for total knee arthroplasty, based on the prediction method for total knee arthroplasty described above, including:

[0116] The image acquisition and preprocessing module is used to receive or load knee joint MRI image data and perform segmentation. Specifically, image acquisition uses a sagittal selective water-excited three-dimensional dual-echo steady-state sequence (SAG-3D-DESS-WE). A trained convolutional neural network model is called to perform three-dimensional automatic segmentation of the femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus in the MRI image, and outputs the mask map and voxel region (VOI) of each structure.

[0117] The image feature extraction module is used to extract radiomics features from the above-mentioned VOI region based on the SERA code package. The feature types include shape features, first-order statistical features, gray-level co-occurrence matrix (GLCM) features, gray-level run-length matrix (GLRLM) features, etc., and conform to the IBSI standard.

[0118] The clinical information collection and integration module is used to obtain the patient's KOOS score scale results, including four dimensions: pain, symptoms, function, and quality of life.

[0119] The predictive modeling module is used to fuse radiomics features and clinical scoring features into the predictive model and output the risk probability value of whether to receive TKR in the future. The model is a shallow neural network structure that supports LASSO screening and cross-validation of the training set.

[0120] The prediction output and visualization module is used to classify the prediction results for risk and output a visualization interface, including structural heatmaps, SHAP structural contribution interpretation diagrams, and TKR probability color coding prompts, to assist doctors in making individualized preoperative decisions.

[0121] Furthermore, this invention also proposes a computer-readable storage medium storing program instructions for a total knee arthroplasty prediction method. These program instructions can be executed by one or more processors to implement the steps of the total knee arthroplasty prediction method as described above.

[0122] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the demand for total knee arthroplasty, characterized in that, include: Step S1, Image Acquisition and Preprocessing: Acquire MRI images of the knee joint, automatically segment the structure into several parts, and obtain the volume of interest for each structure; Step S2, Image feature extraction: Use the SERA radiomics feature extraction tool to extract radiomics features from the volume of interest. Step S3: Clinical scoring data collection and integration, obtaining the corresponding patient's clinical symptom scoring data, including various scores in the KOOS scale for knee osteoarthritis outcomes survey; Step S4: Predictive model building and output. The fusion of radiomics features and KOOS score features is input into the predictive model to output the probability value of future total knee replacement. Step S5, Predicted Output and Visualization: Based on the set risk threshold, the probability values ​​of total knee replacement are classified and visualized in the form of color coding, structural heatmap or Shapley interpretation diagram. The automatic segmentation is based on a convolutional neural network model with a 3D U-Net structure, optimized using the Dice loss function, and supports joint segmentation of multiple structures; The image omics features include first-order statistical features, gray-level co-occurrence matrix features, gray-level run length matrix features, and shape features.

2. The method for predicting the demand for total knee arthroplasty according to claim 1, characterized in that, The aforementioned components include four types of tissues: femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus.

3. The method for predicting the demand for total knee arthroplasty according to claim 2, characterized in that, The radiomics features were screened using the LASSO method. By adding an L1 regularization term to the feature coefficients, the coefficients of irrelevant or redundant features were automatically reduced to 0 during the optimization process.

4. The method for predicting the demand for total knee arthroplasty according to claim 2, characterized in that, The KOOS scale includes four indicators: pain score, symptom score, motor function score, and quality of life score.

5. The method for predicting the demand for total knee arthroplasty according to claim 4, characterized in that, The prediction model is a shallow neural network model containing two hidden layers, employing the ReLU activation function, and the output layer uses the Sigmoid function to generate the probability value of total knee replacement.

6. The method for predicting the demand for total knee arthroplasty according to claim 4, characterized in that, The set risk threshold is obtained by calculating the Youden index and is used to balance the sensitivity and specificity of the model.

7. A total knee arthroplasty demand prediction system, based on the total knee arthroplasty demand prediction method as described in any one of claims 1 to 6, characterized in that, include: The image acquisition and preprocessing module is used to receive or load knee joint MRI image data and perform segmentation. The image feature extraction module is used to extract radiomics features from regions of interest based on the SERA code package. The clinical information collection and integration module is used to obtain the patient's KOOS score scale results, including four dimensions: pain, symptoms, function, and quality of life. The predictive modeling module is used to fuse radiomics features and clinical scoring features into the predictive model and output the probability value of whether to undergo total knee replacement in the future. The prediction output and visualization module is used to classify the prediction results for risk and output a visualization interface, including structural heatmaps, SHAP structural contribution interpretation diagrams, and color-coded prompts for the probability values ​​of total knee replacement.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions for a total knee arthroplasty demand prediction method, which can be executed by one or more processors to implement the steps of the total knee arthroplasty demand prediction method as described in any one of claims 1 to 6.

Images