Total knee arthroplasty demand prediction method and system

Abstract

The invention discloses a total knee arthroplasty demand prediction method and system, and relates to the field of clinical auxiliary decision making. The method comprises the following steps: acquiring a knee MRI image, and automatically segmenting a joint gap key structure through a deep convolutional neural network to obtain a volume of interest (VOI); using an SERA tool to extract iconomics characteristics of the IBSI standard; and finally constructing a joint space and clinical feature radiomics model (JSC-RM) by combining a knee osteoarthritis outcome investigation (KOOS) score, so as to predict whether the TKR is accepted or not in the future 1-4 years. According to the method, LASSO feature screening and superficial layer neural network modeling are adopted, a threshold value is set in combination with a Youden index, the prediction performance is superior to that of an existing model, the consistency and accuracy of TKR opportunity judgment of resident doctors are remarkably improved, and preoperative intelligent risk intervention is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of clinical auxiliary decision-making, and in particular to a total knee replacement requirement prediction method and system. BACKGROUND

[0002] Knee osteoarthritis is one of the most common degenerative diseases in the elderly population, and it usually needs to undergo total knee replacement when it progresses to the end stage. Total knee replacement can effectively relieve pain and improve function, but it is a high-risk, high-cost and irreversible surgical procedure. Current clinical decision-making for total knee replacement is highly dependent on physician experience and subjective judgment, and lacks a unified, objective and quantifiable prediction tool.

[0003] The prior art does not integrate multi-structure MRI and KOOS score, and cannot realize quantitative early warning. Although some prior art has prediction of total knee replacement, the model is a "black box" structure, there is no structure segmentation, no symptom score input, and the clinical interpretability is poor. The prediction depth and accuracy are limited, the resolution and tissue level information of MRI are lacking, the details of cartilage and meniscus degeneration cannot be fully described, and subjective scoring or structural interpretable mechanisms are not introduced. SUMMARY

[0004] In order to solve the technical problem of lack of quantitative standard for predicting total knee replacement in the prior art, the present application provides a total knee replacement requirement prediction method and system. The present application proposes an artificial intelligence model combining MRI structure recognition, image feature analysis and clinical symptom score fusion, for realizing structure-level, individualized and quantifiable prediction of total knee replacement (TKR).

[0005] The present application is realized by the following technical solutions: A total knee replacement requirement prediction method, comprising: Step S1, image acquisition and preprocessing, acquiring a magnetic resonance image of the knee joint, automatically segmenting the structure into several parts, and obtaining the volume of interest of each structure; Step S2, image feature extraction, using SERA image feature extraction tool to extract image features from the volume of interest region; Step S3, clinical score data acquisition and integration, acquiring clinical symptom score data of the corresponding patient, including each score in the KOOS scale of the knee osteoarthritis result survey; Step S4, prediction model establishment and output, fusing image feature and KOOS score feature, inputting the prediction model to output the probability value of future total knee replacement; Step S5, prediction output and visualization, based on the set risk threshold, the output total knee replacement probability value is classified, and is visualized in the form of color coding, structure heat map or shap explanation diagram.

[0006] Further, the automatic segmentation is based on a three-dimensional U-Net structure of a convolutional neural network model, optimized using a Dice loss function, and supports joint segmentation of multiple structures.

[0007] Further, the several parts include four types of tissues, including femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus.

[0008] Further, the radiomics features include first-order statistical features, gray level co-occurrence matrix features, gray level run length matrix features and shape features.

[0009] Further, the radiomics features are screened by LASSO method, and the L1 regularization term is added to the feature coefficient, so that the coefficient of irrelevant or redundant features is automatically shrunk to 0 in the optimization process.

[0010] Further, the KOOS scale includes four indicators, including pain score, symptom score, motor function score and quality of life score.

[0011] Further, the prediction model is a shallow neural network model, which includes two hidden layers, adopts ReLU activation function, and uses Sigmoid function in the output layer to generate total knee replacement probability value.

[0012] Further, the set risk threshold is obtained by Youden index calculation, which is used to balance the sensitivity and specificity of the model.

[0013] The application also provides a total knee replacement demand prediction system based on the total knee replacement demand prediction method as described above, which comprises: An image acquisition and preprocessing module for receiving or loading knee MRI image data and performing segmentation; An image feature extraction module for extracting radiomics features based on the SERA code package in the volume of interest; A clinical information acquisition and integration module for obtaining KOOS score scale results of patients, including four dimensions of pain, symptoms, function and quality of life; A prediction modeling construction module for fusing radiomics features and clinical score features into a prediction model to output the probability value of whether to accept total knee replacement in the future; A prediction output and visualization module for classifying the prediction results by risk and outputting a visual interface, including structure heat map, SHAP structure contribution explanation diagram and total knee replacement probability value color coding prompt.

[0014] In addition, to achieve the above object, the application further provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions of a total knee arthroplasty prediction method, and the program instructions of the total knee arthroplasty prediction method can be executed by one or more processors to realize the steps of the total knee arthroplasty prediction method.

[0015] Compared with the prior art, the application has the following beneficial effects: Full-process automatic image analysis is realized: by constructing a convolutional neural network model, three-dimensional automatic segmentation of key structures (femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus) in the joint space of the knee magnetic resonance image is automatically completed, subjective bias and low efficiency problems of traditional manual delineation are avoided, and the accuracy and repeatability of structure recognition are significantly improved.

[0016] Fusion of imageomics and clinical score information improves prediction accuracy: based on the extraction of imageomics features, KOOS score data is introduced, and finally a joint space plus clinical feature imageomics model (JSC-RM) is constructed, which has higher prediction ability and stronger clinical interpretability compared with a model relying only on MOAKS structure score or a clinical scale model.

[0017] Supporting preoperative medium and long-term risk prediction and advancing intervention window: the model output is the individualized risk probability of receiving TKR in the future 1-4 years, which helps clinicians to identify potential progression patients in advance, optimize the timing of preoperative intervention, and realize active diagnosis and treatment management.

[0018] Providing a visual decision support interface to enhance the explainability experience of doctors: the model output result is intuitively presented in the form of a structure heat map, TKR probability color coding and SHAP contribution interpretation diagram, which helps resident physicians to improve the consistency and accuracy of TKR judgment, and has good generalizability and practical value.

[0019] Significantly better performance than traditional models: the AUC of the model of the application is significantly better than that of the MOAKS score model and the clinical model in the independent test set, and the sensitivity, specificity and Kappa consistency index are significantly improved in the doctor assisted decision making experiment, which shows a high clinical application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is a total knee arthroplasty prediction method flowchart according to an embodiment of the present application; Figure 2 is a knee nuclear magnetic resonance deep learning automatic segmentation map according to an embodiment of the present application; Figure 3 is a joint space feature heat map of a control group and a case group according to an embodiment of the present application; Figure 4 is a result of total knee arthroplasty predicted by a final model according to an embodiment of the present application; Figure 5 is an ability of a surgeon to predict total knee arthroplasty improved with assistance of a final model according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] The embodiments of the present application will be described in detail below with reference to the drawings.

[0022] The above examples are merely used to illustrate the present application and should not be interpreted as limiting the scope of the application. The present application can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. The claims should not be considered as limited to the embodiments described above, but should be considered as including all embodiments which would be understood by one of ordinary skill in the art to come within the scope of the claims.

[0023] It should be noted that the drawings provided in the following embodiments are only schematic and are intended to provide the basic understanding of the application. In the drawings, the shape, size, and number of components shown in the drawings are not intended to represent the actual shape, size, and number of the components in actual implementation, and the shape, size, and number of the components in actual implementation can be changed arbitrarily, and the layout of the components can be more complex.

[0024] Referring to Figure 1 A prediction method of total knee arthroplasty, comprising the following steps: S1: image acquisition and preprocessing; The knee MRI image of the KOA patient is acquired by using SAG-3D-DESS-WE sequence, and the three-dimensional input image is generated after resampling, gray scale normalization and registration processing.

[0025] The imaging parameters are as follows:

[0026] The sequence has high spatial resolution (X resolution 0.365 mm, Y resolution 0.456 mm), can clearly show the interarticular space tissue such as cartilage and meniscus, and provides a reliable data basis for subsequent three-dimensional structure segmentation and imageomics analysis.

[0027] The structure is automatically segmented, a deep learning algorithm based on three-dimensional CNN is used to automatically segment the joint space region in the MRI image, and the segmentation targets include: femoral cartilage; tibial cartilage; medial meniscus; lateral meniscus, and the volume of interest (VOI) of each structure is accurately obtained.

[0028] As Figure 2 The knee joint MRI structure automatic segmentation scheme shown in the embodiment of the application shows that on the SAG-3D-DESS-WE sequence MRI image, the deep model is used to realize high-precision automatic segmentation of key anatomical structures of the joint. The figure is the core intermediate step of the model to realize the structuring of the image data, and is the basis for subsequent imageomics feature extraction and risk modeling. The segmentation structures shown in the figure include the following four types of joint space structures: femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus.

[0029] The segmentation model used in the application 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, each channel corresponds to the above four types of anatomical structures. The model training process uses the Dice loss function to optimize, combined with the structure gold standard mask manually labeled by the OAI database, to ensure accurate identification of anatomical boundaries.

[0030] In the figure, each structure is encoded in different colors and superimposed on the MRI background image, clearly showing the automatic structure recognition ability of the model for the joint space region. Among them: the cartilage structure is completely covered on the bone end surface; the medial and lateral menisci are located between the tibial platforms, and the structure contour is clear and the signal distribution is uniform.

[0031] The average DSC of the structure automatic segmentation realized by the embodiment is more than 0.80 in the six structures, indicating that it can provide high-quality VOI for subsequent imageomics feature extraction, and improve the stability of the features and the prediction reliability of the model.

[0032] As Figure 3 The joint space feature heat map diagram generated by the embodiment of the application shows the visualization image formed after quantifying the gray scale features of the key structure region in the knee joint MRI image. The left figure is the control group, and the right figure is the case group. The figure calculates the average signal intensity of each pixel in the femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus structure, uses color heat map to represent the gray value difference of different tissue regions, and reflects the structure degeneration degree.

[0033] The implementation principle is as follows: (1) Data source: high-resolution knee MRI images collected using the SAG-3D-DESS-WE sequence, which has good contrast for cartilage and meniscal tissue.

[0034] (2) Structure segmentation: automatically segment the MRI images through a CNN model to obtain the corresponding masks for each structure, including: femoral cartilage, tibial cartilage, medial meniscus, and lateral meniscus.

[0035] (3) Feature definition: The joint space feature heat map refers to calculating the average signal intensity (i.e., the mean of pixel values) within each tissue structure region and color-coding it. This average value can represent tissue density or water content status, with strong degeneration sensitivity.

[0036] (4) Image construction: replace the pixel values of each structure region with the average value within the structure, and display them in the form of a heat map, visualizing the gray level differences between different structures, and forming a three-dimensional joint space feature heat map.

[0037] Figure 3 Different colors represent the average gray values of different structures; red / yellow areas represent higher signal intensity, which may correspond to inflammatory changes or fluid accumulation; blue / dark areas represent signal attenuation, indicating thinning of the cartilage or degeneration of the meniscus; it can be used as a structural level imageomics interpretation visualization tool to support doctors in understanding the basis of model prediction.

[0038] S2: Image feature extraction Using the imageomics analysis standard environment (SERA) code package, three-dimensional imageomics features based on the IBSI standard are extracted in the four segmented volumes of interest (VOI), including: first-order statistical features (such as mean gray level, standard deviation, skewness, etc.); second-order texture features (such as GLCM, GLRLM, etc.); three-dimensional shape features (such as volume, surface area, aspect ratio, etc.).

[0039] Features are extracted for the following structures: femoral cartilage (FC), tibial cartilage (TC), medial meniscus (MM), and lateral meniscus (LM); feature types: first-order statistics, gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and three-dimensional shape features; output format: default output is a Python dict, which can also be saved as a csv, xlsx, or json format file for subsequent LASSO modeling or neural network use.

[0040] Feature selection and model construction: The application extracts imageomics features from four structures of femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus respectively to form four sets of feature sets. In order to reduce the dimension, reduce the redundancy and improve the stability of the model, the application adopts LASSO for feature selection. LASSO regression adds an L1 regularization term to the feature coefficient, and automatically shrinks the coefficient of irrelevant or redundant features to 0 in the optimization process, so as to realize the sparse selection. Its advantage lies in retaining the features with the strongest correlation with the target variable and improving the generalization ability of the model.

[0041] The feature selection process includes: (1) input features: SERA output features of each structure, preferably about 200-400 dimensions for each VOI, a total of four groups; (2) standardization: Z-score standardization is used for all features; (3) label variable processing: binary selection is performed on the total knee arthroplasty event label, wherein 0=not receiving total knee arthroplasty, and 1=has received total knee arthroplasty; (4) LASSO regression modeling: cross-validation (10-fold CV) is used in the training set to determine the regularization parameter λ; the non-zero weight features are selected as the final input; and the non-zero weight features and the coefficients are output as the structural level explanation basis.

[0042] Modeling method: an independent shallow neural network prediction model is constructed for each structure (femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus), forming four single-structure imageomics models; each model is used to analyze the correlation between the degeneration characteristics of the corresponding structure and the occurrence of total knee arthroplasty.

[0043] After the above steps are completed, a number of most relevant features can be output for each structure, and used for subsequent single-model performance evaluation (AUC, sensitivity, specificity) and joint modeling input.

[0044] S3: Clinical score data collection and integration The KOOS score table of the corresponding patient is collected synchronously, including four dimensions of pain, symptoms, function and quality of life, which is used as an input variable reflecting the subjective functional state of the patient. The above scores are derived from the questionnaire filled out by the patient at the time of MRI examination, representing the subjective perception of the knee function state.

[0045] S4: Prediction model establishment and output The MRI imageomics features and the KOOS scores are spliced and Z-score standardized, and input as an input vector into a shallow neural network for training and prediction.

[0046] The fusion strategy is as follows: (1) Feature concatenation: the image feature vectors of each structure after LASSO screening are merged; and four KOOS score items are concatenated into a joint feature vector; all features are standardized by Z-score to ensure uniform dimension and consistent dimension.

[0047] (2) Model structure: a shallow feedforward neural network is used as a fusion modeler; the network structure includes two hidden layers, uses a ReLU activation function, and finally outputs a Sigmoid node representing the probability of total knee arthroplasty; a cross-entropy loss function is used for optimization.

[0048] (3) Output model: a joint space plus clinical feature imageomics model (JSC-RM) is constructed, which realizes the effective fusion of structural level image quantitative information and symptom level function score.

[0049] (4) Training and verification: the model is trained in the existing total knee arthroplasty osteoarthritis sub-cohort (the Pivotal Osteoarthritis initiative Magnetic resonance imaging Analyses Total Knee Replacement osteoarthritis cohort, POMA-TKR), and 10-fold cross-validation is used to evaluate stability; the prediction performance is evaluated on an independent test set, including AUC, sensitivity, specificity, and Kappa value; the performance of the JS-RM model without KOOS score and the traditional MOAKS model is compared.

[0050] Specifically, the nested cohort key MRI image in the OAI database is used to analyze the total knee arthroplasty osteoarthritis sub-cohort as the training and test data source, 10-fold cross-validation and independent test are performed, and model performance indicators including AUC, sensitivity, specificity, etc. are evaluated to ensure model stability and generalization ability.

[0051] The present application evaluates the performance of the joint model JSC-RM by constructing an independent test set, and verifies its effectiveness, stability and superiority in predicting TKR risk.

[0052] (1) Test data design: the total knee arthroplasty osteoarthritis sub-cohort in the OAI database is used, and MRI samples at baseline and 3 years, 2 years, and 1 year before follow-up are selected to form a multi-time point test set; each time point test set includes "total knee arthroplasty group" and "non-total knee arthroplasty group" cases, and is set in a 1:1 ratio; all images and KOOS score information are not used in the model building stage to ensure independence.

[0053] (2) Evaluation index and method: AUC: measure the model's ability to distinguish; Sensitivity: the proportion of true total knee arthroplasty individuals identified by the model; Specificity: the proportion of non-total knee arthroplasty individuals correctly identified by the model; Kappa value: measure the consistency of the model's prediction and the true label; 95% confidence interval and DeLong test: used to evaluate the statistically significant difference with other models.

[0054] (3) Compare JSC-RM model with existing technology:

[0055] (4) Evaluation results: In the total test set, the AUC of JSC-RM model predicting TKR is 0.85 (95% confidence interval: 0.82~0.88), which is significantly higher than MOAKS model (JS-MOM: AUC=0.67) and pure clinical model (AUC=0.62); In different time point test sets, the sensitivity of JSC-RM is between 75%~80%, and the specificity is 72%~79%; DeLong test shows that JSC-RM AUC is significantly better than JS-RM (p<0.05), and also better than JSC-MOM and Clinical model (p<0.001); Kappa value (representing prediction consistency) is 0.46~0.61, which is higher than all comparison models; The performance of JSC-RM is stable in different time point test sets, which reflects its good generalization ability.

[0056] S5: Prediction output and visualization The model supports structure level importance ranking output; visualizes the contribution of structures to prediction (SHapley Additive exPlanations, SHAP); and provides probability value of needing total knee arthroplasty and recommended threshold to assist clinical judgment.

[0057] The model output predicts the probability value of an individual receiving total knee arthroplasty in the future, and generates structure level risk contribution map or heat map, providing visual aid for doctors to make judgments.

[0058] The method can be embedded in hospital image platform, PACS system, or deployed in orthopedic outpatient, follow-up evaluation system or clinical trial platform in the form of Web Application Programming Interface (Web API) / local software, realizing interpretable and scalable TKR risk prediction support service.

[0059] Visualization of output: Receiver Operating Characteristic Curve is used to show the prediction performance of each model on different test sets; Confusion Matrix is used to show the number of true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) to assist clinical interpretation; Structural contribution map is provided by combining SHAP interpretation module to improve the transparency of prediction results.

[0060] To further verify the auxiliary decision-making ability of the final joint model (JSC-RM) constructed by the application in clinical application, the application designs a simulation resident physician prediction experiment, and uses the model result to assist the doctor in judging whether the KOA patient will receive TKR in the future.

[0061] Experimental setup: 7 orthopedic resident physicians with 1-4 years of clinical experience are recruited; the physicians independently predict whether the patients in the test set will receive TKR based on MRI images and KOOS scores; two rounds of prediction are divided: the first round: without using the JSC-RM model, only relying on their own experience; the second round: the JSC-RM model prediction output is provided, and the doctor corrects the result.

[0062] Decision threshold setting: To balance sensitivity and specificity, the application sets the prediction probability threshold in the model training set using the Youden Index to find the optimal balance point 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".

[0063] Result visualization and color coding: The TKR probability output by the model is displayed in color-coded form for each case: red indicates high risk (TKR probability ≥ 0.52); green indicates low risk (TKR probability < 0.52); structural importance scores and recommended interpretation information (structural heat map and SHAP contribution map) are also output; all information is displayed on a unified electronic interface for easy visual assessment by doctors.

[0064] Auxiliary effect evaluation: The results show that under the assistance of the JSC-RM model, the prediction ability of the resident physician is significantly improved, with sensitivity and specificity increasing by more than 30% compared to without model support, and the Kappa value is improved to the level of moderate consistency, indicating that the model of the application has significant clinical practicability in the auxiliary judgment of resident physicians.

[0065]

[0066] As Figure 4The prediction ability of the various prediction models constructed in the embodiments of the present application on TKR risk at different follow-up time points is demonstrated, with AUC as the main evaluation index. The figure contains the performance of the following five models at multiple test time points: the dark blue line represents JS-RM: only fusion of MRI multi-structural imaging features; the yellow line represents JS-MOM: structural imaging model of MOAKS score; the red line represents JSC-RM: the final fusion model proposed in the present application; the light blue line represents JSC-MOM: traditional imaging model of MOAKS score + KOOS score; and the green line represents CM: traditional clinical model containing only KOOS score.

[0067] Time point setting: Figure 4 There are five subgraphs in total, corresponding to the test queues of the five time nodes: baseline, three years from TKR, two years from TKR, one year from TKR, and the summary of all the above time points.

[0068] In each subgraph, the receiver operating characteristic curve of each model is plotted, and the AUC value and 95% confidence interval are marked in the graph for comparison of the risk prediction ability of different models.

[0069] Result summary: Among all the time points, the JSC-RM model always shows the highest AUC value (0.83~0.89), and the AUC value of the JS-RM model is also high (0.83~0.91). There is no significant difference in the AUC value between the JSC-RM model and the JS-RM model, but the AUC value of the JSC-RM model is higher than that of the JS-RM model in the total test set. Therefore, as the final model, the AUC value of the JSC-RM model is significantly better than that of the JSC-MOM (0.69~0.72), JS-MOM (0.66~0.67) and clinical model (0.62~0.67); in the time points close to TKR (1 year from TKR, 2 years from TKR), the sensitivity and specificity of JSC-RM are most balanced; this result shows that the fusion model constructed in the present application has the characteristics of strong forward-looking, high timeliness and strong individualization; the DeLong test is used for comparison of the AUC difference of the models, and the result shows that JSC-RM is significantly better than the control models JS-MOM and CM (p<0.001), and is not significantly better than the JS-RM model.

[0070] Figure 5 The improvement of the model in assisting the resident doctors to predict TKR ability is evaluated. The figure illustrates the positive effect in actual prediction performance.

[0071] The left graph is an average prediction performance comparison graph, that is, all test data are merged. The final model (JSC-RM) is compared with the average prediction performance of doctors in the test set, wherein the red line represents the prediction result of the JSC-RM model, the blue line represents the prediction result of the JSC-MOM model, the blue dot represents the average sensitivity / specificity of the doctor without model assistance, the red dot represents the result after using the JSC-RM, and the black arrow represents that the prediction ability of the resident is obviously improved with the assistance of the model, and the AUC and the diagnostic accuracy are significantly increased.

[0072] The right graph is a prediction performance change graph of seven doctors respectively. The sensitivity and specificity of each doctor under the condition of no model (hollow graph) and with the assistance of the JSC-RM (solid graph) are compared; the dotted line connection represents the prediction change track of the same doctor under the two conditions; all doctors show performance improvement, which embodies the comprehensive enhancement effect of the JSC-RM model on the judgment ability of young doctors.

[0073] The shallow neural network model (JSC-RM) constructed in the application fuses the MRI image feature and the KOOS clinical score, and realizes high-precision prediction of TKR risk in the OAI real queue in the future 4 years, and the AUC of the test set can reach 0.89 at most, which is significantly better than the traditional MOAKS model (AUC is about 0.66) and the single clinical model (AUC is about 0.62).

[0074] Unlike the black-box deep learning model, the application performs image feature analysis based on the automatically segmented cartilage and meniscus structure, can output structure contribution ranking and risk heat map, enables doctors to intuitively understand which joint structure degeneration is related to TKR risk, thereby improving the explainability and clinical adoption rate of the model, and the application firstly fuses the KOOS score and the image feature of the MRI image in a unified standardized manner, considers the objective structure degeneration and integrates the subjective symptom experience, improves the recognition ability of the model to individual differences, and enhances the adaptability of the model in a real clinical scene. Through the reading experiment verification, the model constructed in the application can significantly improve the recognition ability of the resident doctor to the TKR risk, the sensitivity is increased from 42% to 72%, and the specificity is increased from 45% to 78%, thereby providing a quantitative support tool for young orthopedic doctors, and helping to make preoperative intervention plan. The system is flexible in deployment, has wide application prospects, the application can be embedded into a hospital PACS system or deployed in the form of a Web platform, is suitable for various scenes such as outpatient management, surgical decision support, rehabilitation follow-up and clinical trial enrollment of osteoarthritis, and has good popularization and transformation value.

[0075] In the embodiment, the following are achieved: (1) Adopt SAG-3D-DESS-WE sequence MRI image to collect knee joint image, compared with two-dimensional X-ray image, it has higher soft tissue contrast and structural resolution, and can accurately reflect the microscopic morphological changes of femoral cartilage, tibial cartilage and meniscus and other key tissues, and provide high-quality basic data for TKR risk modeling.

[0076] (2) The CNN is used for automatic segmentation of the structures in the four joint spaces in the MRI image, including femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus, which solves the problems of low efficiency, large error and dependence on manual operation of the traditional segmentation method, and improves the accuracy and repeatability of structure recognition.

[0077] (3) The SERA code package is used to extract imageomics features of the VOIs of the four structures in accordance with the IBSI standard, covering first-order statistical features, second-order texture features and shape features, which can comprehensively quantify the gray distribution, texture complexity and spatial morphology of the tissues, and help to reveal the potential structure degeneration signals.

[0078] (4) The KOOS clinical score is introduced as a subjective symptom input variable, and the score data of pain, symptoms, function and quality of life are integrated, so that the model can consider the structure state and symptom perception of the patient at the same time, and improve the individualized prediction ability.

[0079] (5) A shallow neural network model is constructed to model the fused features and output the TKR occurrence probability, the model structure is simple and the calculation efficiency is high, which is convenient for deployment in a clinical environment, and has strong fitting ability and can accurately predict the probability of needing TKR surgery in the future 1-4 years.

[0080] (6) The structure level risk score and visualized results (such as heat map or probability bar chart) are output, which helps doctors to understand the model basis and the contribution degree of each structure, realizes the transformation from “prediction black box” to “explainable assistance”, and improves the trust degree of doctors on the prediction results of the model and the willingness of clinical adoption.

[0081] In summary, the joint technical path of structure segmentation + imageomics + symptom score + shallow neural network modeling + visualized output not only significantly improves the TKR prediction performance, but also enhances the structure explainability and clinical practicability of the prediction results, and provides an effective auxiliary tool for accurate intervention of KOA.

[0082] The embodiment of the application also provides a prediction system for total knee arthroplasty, based on the prediction method for total knee arthroplasty as described above, comprising: An image acquisition and preprocessing module is configured to receive or load knee MRI image data and perform segmentation; specifically, the image acquisition adopts sagittal position selective water excitation three-dimensional double echo steady-state sequence (SAG-3D-DESS-WE); a trained convolutional neural network model is called to perform three-dimensional automatic segmentation on femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus in the MRI image, and to output a mask map and a voxel region (VOI) of each structure; An image feature extraction module is configured to extract imageomic features from the VOI region based on a SERA code package, the feature types including shape features, first-order statistical features, gray level co-occurrence matrix (GLCM) features, gray level run length matrix (GLRLM) features, etc., and conforming to IBSI standards; A clinical information acquisition and integration module is configured to obtain KOOS score table results of a patient, including four dimensions of pain, symptoms, function and quality of life; A prediction modeling construction module is configured to fuse imageomic features and clinical score features into a prediction model to output a risk probability value of whether to receive TKR in the future, the model being a shallow neural network structure supporting LASSO screening and training set cross-validation; A prediction output and visualization module is configured to classify the prediction results by risk and output a visualization interface, including a structure heat map, a SHAP structure contribution explanation diagram and TKR probability color coding prompts, to assist doctors in making preoperative individualized decisions.

[0083] In addition, the embodiment of the present application also proposes a computer readable storage medium, the computer readable storage medium stores the program instructions of the total knee replacement prediction method, and the program instructions of the total knee replacement prediction method can be executed by one or more processors to realize the steps of the total knee replacement prediction method as described above.

[0084] The above-described embodiments only describe the preferred embodiments of the present application, and do not limit the scope of the present application, and various modifications and improvements to the technical solutions of the present application made by those skilled in the art without departing from the design spirit of the present application shall fall within the protection scope of the claims of the present application.

Claims

1. A method for predicting demand for total knee replacement surgery, characterized in that: include: Step S1, image acquisition and preprocessing, acquiring a MRI image of the knee joint, automatically segmenting the structure into several parts, and obtaining a volume of interest for each structure; Step S2, image feature extraction, using the SERA radiomics feature extraction tool to extract radiomics features from the volume of interest; Step S3, clinical score data collection and integration, obtaining the clinical symptom score data of the corresponding patients, including each score in the Knee Osteoarthritis Outcome Survey (KOOS) scale; Step S4, prediction model establishment and output, integrating radiomics features with KOOS score features, inputting the prediction model to output the probability value of receiving total knee replacement in the future; Step S5, prediction output and visualization, classifies the output probability value of total knee replacement based on the set risk threshold, and visualizes it in the form of color coding, structural heat map or Shapley interpretation map.

2. The method for predicting demand for total knee replacement surgery according to claim 1, wherein: The automatic segmentation is based on a convolutional neural network model with a three-dimensional U-Net structure, is optimized using a Dice loss function, and supports joint segmentation of multiple structures.

3. The method for predicting demand for total knee replacement surgery according to claim 2, wherein: The several parts include four types of tissues: femoral cartilage, tibial cartilage, medial meniscus and lateral meniscus.

4. The method for predicting demand for total knee replacement surgery according to claim 1, wherein: The imaging omics features include first-order statistical features, gray-level co-occurrence matrix features, gray-level run length matrix features and shape features.

5. The method for predicting demand for total knee replacement surgery according to claim 4, wherein: The imaging omics features are screened using the LASSO method, and by adding an L1 regularization term to the feature coefficients, the coefficients of irrelevant or redundant features are automatically shrunk to 0 during the optimization process.

6. The method for predicting demand for total knee replacement surgery according to claim 4, wherein: The KOOS scale includes four indicators: pain score, symptom score, motor function score and quality of life score.

7. The method for predicting demand for total knee replacement surgery according to claim 6, wherein: The prediction model is a shallow neural network model, which includes two hidden layers, adopts ReLU activation function, and the output layer uses Sigmoid function to generate the probability value of total knee replacement.

8. The method for predicting demand for total knee replacement surgery according to claim 6, wherein: The set risk threshold is obtained by calculating the Youden index and is used to balance the sensitivity and specificity of the model.

9. A total knee replacement demand forecasting system based on the total knee replacement demand forecasting method according to any one of claims 1 to 8, characterized in that: include: Image acquisition and preprocessing module, used to receive or load knee joint MRI image data and perform segmentation; Image feature extraction module, used to extract radiomics features from the volume of interest region based on the SERA code package; The clinical information collection and integration module is used to obtain the patient's KOOS scoring scale results, including pain, symptoms, function, and quality of life; A predictive modeling module is used to integrate radiomics features with clinical scoring features into a predictive model to output the probability of receiving total knee replacement in the future; The prediction output and visualization module is used to classify the prediction results into risk categories and output a visualization interface, including a structural heat map, a SHAP structural contribution interpretation diagram, and a color-coded prompt for the probability value of total knee replacement.

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

Images