Method for classifying knee cartilage damage and storage medium thereof

Abstract

The application provides a knee cartilage injury grading method and a storage medium thereof. The knee cartilage MRI image of a patient is acquired, and image region segmentation is performed on the knee cartilage MRI image to obtain a first image representing a patellofemoral joint region, a second image representing a medial femorotibial joint region, and a third image representing a lateral femorotibial joint region in the patient. Feature extraction is performed on the first image, the second image, and the third image to obtain a first feature set. The first feature set is input into a radiomics model, and the first feature set is subjected to cartilage injury grading evaluation and evaluation explainability analysis by the radiomics model to obtain a knee injury score and a knee injury grade of the patient and a visual evaluation report for explaining the generation of the knee injury score and the knee injury grade. The knee cartilage injury grading accuracy and explainability evaluation are realized, and the grading accuracy and the clinical application reliability are improved.

Description

Technical Field

[0001] This application relates to the field of medical testing, and in particular to a method for grading knee cartilage damage and its storage medium. Background Technology

[0002] Knee osteoarthritis (KOA) is one of the leading causes of pain and functional impairment worldwide, with its core pathology lying in the progressive degeneration and damage of articular cartilage. Accurate and objective assessment of the severity (grading) of cartilage damage is crucial for developing individualized treatment plans, monitoring disease progression, and evaluating treatment effectiveness. Currently, MRI-based visual semi-quantitative scoring systems (such as WORMS) are widely used in clinical practice and research. However, this method is highly dependent on the experience of radiologists, and suffers from significant intra- and inter-observer variability, is time-consuming, and struggles to achieve precise quantification.

[0003] In recent years, radiomics technology has provided a new approach for the objective assessment of cartilage damage. By extracting quantitative features from medical images in high throughput and constructing machine learning models, it holds the promise of automated and standardized grading. However, existing studies often analyze knee cartilage as a whole, failing to fully consider the complex biomechanical system composed of the patellofemoral joint (PFJ), medial femorotibiofemoral joint (MFTJ), and lateral femorotibiofemoral joint (LFTJ). These regions often exhibit different degeneration patterns in the process of knee arthritis (KOA). Ignoring this regional heterogeneity may prevent the model from capturing key local pathological information, limiting its assessment accuracy. Furthermore, most radiomics models lack effective interpretability mechanisms; their decision-making process is like a "black box," making it difficult for clinicians to understand and trust the model's predictions, becoming a significant bottleneck in the translation of technology into clinical practice. Summary of the Invention

[0004] The main objective of this application is to propose a method for grading knee cartilage damage and its storage medium, which enables accurate grading and interpretable assessment of multi-regional cartilage damage in the knee joint, thereby improving grading accuracy and reliability in clinical applications.

[0005] To achieve the above objectives, a first aspect of this application proposes a method for grading knee cartilage damage, comprising: Obtain MRI images of the patient's knee joint; Image region segmentation was performed on the knee joint MRI images to obtain a first image representing the patellofemoral joint region, a second image representing the medial femorotibiofemoral joint region, and a third image representing the lateral femorotibiofemoral joint region in the patient. Feature extraction is performed on the first image, the second image, and the third image to obtain a first feature set, which includes the feature vectors of the first image, the second image, and the third image. The first feature set is input into the radiomics model. The radiomics model is used to perform cartilage damage grading assessment and assessment interpretability analysis on the first feature set to obtain the patient's knee joint injury score and knee joint injury grade, as well as a visual assessment report used to interpret the generated knee joint injury score and knee joint injury grade. Among them, the radiomics model is a multi-class machine learning model built based on radiomics scores.

[0006] Furthermore, in some embodiments, the radiomics model includes a feature selector, a multi-class probability determiner, and a SHAP visualization interpreter. The configuration parameters of the feature selector include multiple key features determined through training and the weight coefficients corresponding to the key features. The first feature set is input into the radiomics model. The radiomics model performs interpretability analysis on the first feature set to obtain the patient's knee injury score and knee injury grade, as well as a visual assessment report used to interpret the generated knee injury score and grade, including: The feature selector selects feature vectors that match the key features from the first feature set to obtain the second feature set; By using the weight coefficients corresponding to each feature vector in the second feature set, a linear weighted sum is performed on each feature vector in the second feature set to obtain the total radiomics score of the knee joint MRI image. The total radiomics score includes the patellofemoral radiomics score of the first image, the lateral femoral-tibial radiomics score of the second image, and the lateral femoral-tibial radiomics score of the third image. The total radiomics score is calculated using a multi-class probability determiner to output a probability distribution vector belonging to each preset damage level. Calculate and determine the knee injury score and knee injury level based on the probability distribution vector; Based on the SHAP visualization interpreter, interpretability analysis of knee joint injury scores and knee joint injury grades is performed to obtain a visual assessment report.

[0007] Furthermore, in some embodiments, based on the SHAP visualization interpreter, interpretability analysis is performed on knee joint injury scores and knee joint injury grades to obtain a visual assessment report, including: The SHAP visualization interpreter calculates the SHAP value of each feature vector in the second feature set during the multi-class probability determiner operation based on the total radiomics score data. Based on the SHAP value corresponding to each feature vector in the second feature set, determine the feature contribution of each feature vector in the second feature set to the output probability distribution vector of the multi-class probability determiner; Based on the feature contribution of each feature vector in the second feature set, a global feature importance ranking map and an individual prediction interpretation map are generated, and the global feature importance ranking map and the individual prediction interpretation map are integrated into a visual evaluation report.

[0008] Furthermore, in some embodiments, the process of constructing a radiomics model includes the following steps: Acquire knee joint MRI training images from multiple patient samples; Image region segmentation was performed on multiple knee joint MRI training images to obtain multiple patellofemoral images representing the patellofemoral joint region, multiple medial femorotibia images representing the medial femorotibia joint region, and multiple lateral femorotibia images representing the lateral femorotibia joint region. Feature extraction was performed on multiple patellofemoral images, multiple medial femoral-tibia images, and multiple lateral femoral-tibia images to obtain patellofemoral feature sets, medial femoral-tibia feature sets, and lateral femoral-tibia feature sets. Standardization preprocessing was performed on the patellofemoral feature set, medial femoral-tibia feature set, and lateral femoral-tibia feature set to obtain the patellofemoral key feature set, medial femoral-tibia key feature set, and lateral femoral-tibia key feature set, respectively. Among them, the patellofemoral key feature set includes the patellofemoral key features in each patellofemoral image and the weight coefficient of each patellofemoral key feature set; the medial femoral-tibia key feature set includes the medial femoral-tibia key features in each medial femoral-tibia image and the weight coefficient of each medial femoral-tibia key feature set; and the lateral femoral-tibia key feature set includes the lateral femoral-tibia key features in each lateral femoral-tibia image and the weight coefficient of each lateral femoral-tibia key feature set. The patellofemoral key features in each patellofemoral image, the medial femoral-tibia key features in each medial femoral-tibia image, and the lateral femoral-tibia key features in each lateral femoral-tibia image were linearly weighted and summed to obtain the patellofemoral radiomics score corresponding to each patellofemoral image, the lateral femoral-tibia radiomics score corresponding to each medial femoral-tibia image, and the lateral femoral-tibia radiomics score corresponding to each lateral femoral-tibia image. A multi-class probability determiner was constructed and trained based on the patellofemoral radiomics scores of each patellofemoral image, the medial femoral-tibia radiomics scores of each medial femoral-tibia image, and the lateral femoral-tibia radiomics scores of each lateral femoral-tibia image. The patellofemoral key feature set, the medial femoral-tibia key feature set, and the lateral femoral-tibia key feature set are used as configuration parameters for the feature selector and combined with the multi-class probability determiner to form a radiomics model.

[0009] Furthermore, in some embodiments, a multi-class probability determiner is constructed and trained based on the patellofemoral radiomics score of each patellofemoral image, the medial femorotibia radiomics score of each medial femorotibia image, and the lateral femorotibia radiomics score of each lateral femorotibia image, including the following steps: The patellofemoral radiomics scores of the patellofemoral images, the medial femoral-tibial radiomics scores of each medial femoral-tibial image, and the lateral femoral-tibial radiomics scores of each lateral femoral-tibial image are merged in the same image dimension to obtain the total radiomics score of each knee joint MRI training image. Using the total radiomics score of each knee joint MRI training image as input and the knee joint injury level of the sample patients as label, multiple patellofemoral candidate classification models, multiple medial femoral-tibia candidate classification models, and multiple lateral femoral-tibia candidate classification models were constructed and trained using various statistical classification algorithms. The performance of multiple patellofemoral candidate classification models, multiple medial femoral-tibia candidate classification models, and multiple lateral femoral-tibia candidate classification models was evaluated using knee MRI images from multiple patient samples. The model with the best performance evaluation among multiple patellofemoral candidate classification models was selected as the official patellofemoral classification model. The model with the best performance evaluation among multiple medial femoral tibia candidate classification models was selected as the formal medial femoral tibia classification model. The model with the best performance among multiple candidate lateral femoral tibia classification models was selected as the official lateral femoral tibia classification model. The formal classification models of patellofemoral, medial femoral tibia, and lateral femoral tibia are combined to form a multi-class probability determinist.

[0010] Furthermore, in some embodiments, the statistical classification algorithms include logistic regression algorithms and support vector machine algorithms; Performance evaluation metrics should include at least one of the following: accuracy, precision, recall, F1 score, and integral value of the quantized ROC curve.

[0011] Furthermore, in some embodiments, the patellofemoral feature set, the medial femorotibia feature set, and the lateral femorotibia feature set are respectively subjected to standardized preprocessing to obtain the patellofemoral key feature set, the medial femorotibia key feature set, and the lateral femorotibia key feature set, including: By using the variance thresholding method, features with variances less than or equal to a predetermined threshold are removed from the intervertebral disc in the patellofemoral feature set, the medial femoral-tibia feature set, and the lateral femoral-tibia feature set, respectively, to obtain the first patellofemoral screening feature set, the first medial femoral-tibia screening feature set, and the first lateral femoral-tibia screening feature set. Using a univariate feature selection method, the first patellofemoral screening feature set, the first medial femoral-tibia screening feature set, and the first lateral femoral-tibia screening feature set were analyzed to retain features that were significantly correlated with the knee joint injury grade classification results, thus obtaining the second patellofemoral screening feature set, the second medial femoral-tibia screening feature set, and the second lateral femoral-tibia screening feature set. By using the LASSO algorithm with nested cross-validation, regression analysis was performed on the second patellofemoral screening feature set, the second medial femoral-tibia screening feature set, and the second lateral femoral-tibia screening feature set, respectively, and features with non-zero weight coefficients and their corresponding weight coefficients were retained to obtain the patellofemoral key feature set, the medial femoral-tibia key feature set, and the lateral femoral-tibia key feature set.

[0012] Furthermore, in some embodiments, the predetermined threshold is 0.75, the univariate feature selection method is the SelectKBest method, the criterion for determining that the knee joint injury level classification result is significantly correlated is that the p-value of the hypothesis test between the feature and the classification result is less than 0.05, and the LASSO algorithm uses 5-fold cross-validation.

[0013] Furthermore, in some embodiments, the type of feature vectors within the first feature set includes at least one of the following: first-order features, three-dimensional shape features, gray-level co-occurrence matrix, gray-level size region matrix, gray-level run-length matrix, neighborhood gray-level tone difference matrix, and gray-level dependency matrix.

[0014] To achieve the above objectives, a second aspect of the present application provides a storage medium, which is a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the knee cartilage injury grading method of the first aspect embodiment described above.

[0015] The embodiments of the first aspect of this application have the following beneficial effects: By dividing the cartilage into three independent functional areas—patellofemoral joint, medial and lateral femorotibiofemoral joints—based on the biomechanical characteristics of the knee joint and extracting and modeling regional features, accurate capture of the regional heterogeneity of joint degeneration is achieved. This avoids information loss caused by coarse assessment of only a single area or the whole, enabling a more comprehensive detection of the knee joint's degenerative state under different biomechanical conditions, and enhancing the accuracy and consistency of injury severity results. Furthermore, by constructing a multi-machine classification learning model based on radiomics scoring, automatic fusion and grading of multi-region image features are achieved, reducing reliance on manual image reading and improving assessment efficiency and repeatability. Moreover, an interpretability analysis mechanism is introduced to focus on and visualize the feature contribution of the model output, intuitively revealing the contribution of different regional features to specific grading decisions. This overcomes the "black box" dilemma of traditional artificial intelligence models, helping clinicians understand the relationship between different cartilage regions and key imaging features in injury grading, thus providing more reliable technical support for the staging assessment, treatment decision-making, and prognosis of knee osteoarthritis. Attached Figure Description

[0016] Figure 1 This is an optional flowchart of the knee cartilage damage grading method provided in the embodiments of this application; Figure 2 This is an optional schematic diagram of the knee cartilage injury grading method provided in the embodiments of this application; Figure 3 This is provided by the embodiments of this application. Figure 1 An optional flowchart for step S104; Figure 4 This is a schematic diagram of the pixel-level features of two patients provided in an embodiment of this application; Figure 5 This is provided by the embodiments of this application. Figure 2 An optional schematic diagram of step S205; Figure 6 This is an optional schematic diagram of the global feature importance ranking graph provided in the embodiments of this application; Figure 7 This is an optional schematic diagram of the individual prediction interpretation diagram provided in the embodiments of this application; Figure 8 This is an optional flowchart of training a radiomics model provided in the embodiments of this application; Figure 9 This is a schematic diagram of the radiomics feature set with non-zero coefficients obtained after standardized preprocessing, as provided in the embodiments of this application. Figure 10 This is provided by the embodiments of this application. Figure 8 An optional schematic diagram of step S406; Figure 11 This is provided by the embodiments of this application. Figure 8 An optional schematic diagram of step S404; Figure 12 This is an optional schematic diagram of the AUC results of the model provided in this application with and without the balancing processing strategy; Figure 13 This is an optional confusion matrix diagram of the AUC results of the model provided in this application embodiment with and without the balancing processing strategy; Figure 14 This is an optional schematic diagram showing the precision, recall, and f1-score of the model provided in this application with and without the balanced processing strategy; Figure 15 This is a schematic diagram of the hardware structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] In the description of this application, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0019] It should also be noted that in the description of this application, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. If the terms "first" and "second" are used, they are only for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0022] In recent years, radiomics technology has provided a new approach for the objective assessment of cartilage damage. By extracting quantitative features from medical images in high throughput and constructing machine learning models, it holds the promise of automated and standardized grading. However, existing studies often analyze knee cartilage as a whole, failing to fully consider the complex biomechanical system composed of the patellofemoral joint (PFJ), medial femorotibiofemoral joint (MFTJ), and lateral femorotibiofemoral joint (LFTJ). These regions often exhibit different degeneration patterns in the process of knee arthritis (KOA). Ignoring this regional heterogeneity may prevent the model from capturing key local pathological information, limiting its assessment accuracy. Furthermore, most radiomics models lack effective interpretability mechanisms; their decision-making process is like a "black box," making it difficult for clinicians to understand and trust the model's predictions, becoming a significant bottleneck in the translation of technology into clinical practice.

[0023] To address these issues, this application proposes a method for grading knee cartilage damage and its storage medium, which enables accurate grading and interpretable assessment of multi-regional cartilage damage in the knee joint, improving grading accuracy and reliability in clinical applications.

[0024] This application provides a method for grading knee cartilage damage and its storage medium, which is specifically illustrated through the following embodiments.

[0025] Firstly, referring to Figure 1 and Figure 2 As shown, Figure 1 This is an optional flowchart of the knee cartilage damage grading method provided in the embodiments of this application. Figure 2 This is an optional schematic diagram of the knee joint cartilage damage grading method provided in the embodiments of this application. The method may include, but is not limited to, steps S101 to S102.

[0026] Step S101: Obtain MRI images of the patient's knee joint.

[0027] The knee MRI images were acquired using standard clinical scanning sequences to fully present the morphological information of the knee joint cartilage and surrounding structures, providing basic imaging data for subsequent analysis. During acquisition, necessary format standardization and quality checks were performed on the images to ensure image clarity and data consistency. In one embodiment, the original knee MRI images were in DICOM format.

[0028] Step S102: Perform image region segmentation on the knee joint MRI image to obtain a first image representing the patellofemoral joint region, a second image representing the medial femorotibiofemoral joint region, and a third image representing the lateral femorotibiofemoral joint region in the patient.

[0029] Step S103: Extract features from the first image, the second image, and the third image to obtain the first feature set.

[0030] The first feature set includes feature vectors from the first image, the second image, and the third image. The feature vectors in the first feature set are of at least one of the following types: first-order features, three-dimensional shape features, gray-level co-occurrence matrix, gray-level size region matrix, gray-level run-length matrix, neighborhood gray-level tone difference matrix, and gray-level dependency matrix.

[0031] It should be noted that first-order features are based on the statistics of the image gray-level histogram, do not involve spatial relationships, and reflect the global distribution characteristics of pixel values; 3D shape features describe the three-dimensional geometric properties of the region of interest (ROI) and need to be calculated based on the segmented mask; gray-level co-occurrence matrix counts the gray-level co-occurrence probability of pixel pairs at specific directions and distances, reflecting the local texture regularity; gray-level size region matrix counts the size distribution of continuous gray-level regions, capturing the continuity of texture; gray-level run matrix counts the continuous length (run) of pixels with the same gray level, quantifying the directionality and roughness of texture; neighborhood gray-level tone difference matrix compares the difference between a pixel and the mean of its neighborhood, describing the roughness and contrast of texture; gray-level dependency matrix counts the dependency relationship between the central pixel and neighboring pixels, combining the characteristics of GLCM and GLRLM.

[0032] Step S104: Input the first feature set into the radiomics model, and use the radiomics model to perform cartilage damage grading assessment and assessment interpretability analysis on the first feature set to obtain the patient's knee joint injury score and knee joint injury grade, as well as a visual assessment report used to interpret the generated knee joint injury score and knee joint injury grade.

[0033] Among them, the radiomics model is a multi-class machine learning model built based on radiomics scores.

[0034] In one possible embodiment, the scoring criteria for the above-mentioned knee joint injury assessment are as follows: 0 points: indicating that the thickness and signal of the knee joint tissue are normal; 1 point: indicating that the thickness of the knee joint tissue remains normal, but the signal is increased on PD-FS weighted images; 2.0 points: indicating that there is a focal defect of partial thickness in the knee joint tissue, and the maximum width of the defect is less than 1 cm; 2.5 points: indicating that there is a focal defect of the entire thickness in the knee joint tissue, and the maximum width of the defect is less than 1 cm; 3 points: indicating that there are multiple partial thickness defects in the knee joint (meeting the 2.0 point standard). 4 points: The knee joint tissue shows diffuse partial thickness defects, with the defect area accounting for 75% or more of the area; 5 points: The knee joint has multiple full thickness defects (meeting the 2.5 point standard), or the width of the full thickness (2.5 point) lesion is greater than 1 cm, but less than 75% of the total area; 6 points: The knee joint tissue has diffuse full thickness defects, with the defect area accounting for 75% or more of the area.

[0035] In steps S101 to S104, the cartilage is divided into three independent functional areas—patellofemoral joint, medial and lateral femorotibiofemoral joint—based on the biomechanical characteristics of the knee joint, and regional feature extraction and modeling are performed. This achieves accurate capture of the regional heterogeneity of joint degeneration, avoiding information loss caused by coarse assessment of only a single area or the whole, and enabling a more comprehensive detection of the knee joint's degenerative state under different biomechanical conditions, thus enhancing the accuracy and consistency of injury severity results. Simultaneously, a multi-machine classification learning model based on radiomics scoring is constructed to achieve automatic fusion and grading of multi-region image features, reducing reliance on manual image reading and improving assessment efficiency and repeatability. Furthermore, an interpretability analysis mechanism is introduced to focus on and visualize the feature contribution of the model output, intuitively revealing the contribution of different regional features to specific grading decisions. This overcomes the "black box" dilemma of traditional artificial intelligence models, helping clinicians understand the relationship between different cartilage regions and key imaging features in injury grading, thereby providing more reliable technical support for the staging assessment, treatment decision-making, and prognosis of knee osteoarthritis.

[0036] Furthermore, the radiomics model includes a feature selector, a multi-class probability determiner, and a SHAP visualization interpreter. The configuration parameters of the feature selector include multiple key features determined through training and the weight coefficients corresponding to these key features. (See reference...) Figure 3 As shown, Figure 3 This is provided by the embodiments of this application. Figure 1 An optional flowchart for step S104, the method may include, but is not limited to, steps S201 to S203.

[0037] Step S201: Using a feature selector, select feature vectors that match the key features from the first feature set to obtain the second feature set.

[0038] In a specific implementation, a feature selector selects feature vectors that match key features from the first feature set to obtain the second feature set. The feature selector is a screening tool pre-built during the radiomics model training phase. It is determined using statistical learning algorithms based on training sample data and is used to identify a subset of features from high-dimensional features that have significant discriminative power for grading cartilage damage. The first feature set contains all radiomics feature vectors extracted from the first image of the patellofemoral joint region, the second image of the medial femorotibiofemoral joint region, and the third image of the lateral femorotibiofemoral joint region. The feature selector filters the first feature set according to the key feature index or rules learned during the training phase, retaining only the feature vectors corresponding to the key features. The key features are selected through a feature screening process during training, for example, by sequentially using variance thresholding, univariate selection methods, and minimum absolute contraction and selection operator regression analysis, ensuring high stability and low redundancy. The second feature set consists of feature vectors of these key features, with a significantly reduced feature dimension, thus providing concise and effective input data for subsequent radiomics scoring calculations and multi-class probability determiners, improving model computational efficiency and the reliability of grading assessment.

[0039] Step S202: Using the weight coefficients corresponding to each feature vector in the second feature set, perform a linear weighted summation on each feature vector in the second feature set to obtain the total radiomics score of the knee joint MRI image.

[0040] The total radiomics score includes the patellofemoral radiomics score of the first image, the lateral femoral-tibial radiomics score of the second image, and the lateral femoral-tibial radiomics score of the third image.

[0041] Specifically, the second feature set contains multiple key feature vectors retained after filtering by a feature selector, each of which has been standardized. Each key feature vector corresponds to a weight coefficient predetermined during model training. This weight coefficient is learned through methods such as minimum absolute contraction and selection operator regression analysis, and its magnitude and sign represent the degree and direction of contribution of the corresponding feature to the severity of cartilage damage. During calculation, the value of each standardized feature vector in the second feature set is multiplied by its corresponding weight coefficient to obtain the weighted result of that feature. Then, the weighted results of all features are added to an intercept term determined simultaneously during training; the sum is the total radiomics score of the knee MRI image. This score is a continuous quantitative value that comprehensively reflects the joint information from multiple key imaging features originating from the patellofemoral joint, medial femorotibiofemoral joint, and lateral femorotibiofemoral joint, serving as a single, highly discriminative input indicator for the subsequent multi-class probability determiner to determine the final cartilage damage level.

[0042] The formula for calculating the total radiomics score is as follows: .

[0043] Where x is the feature vector of the second feature set. These are the weight coefficients of the feature vector (obtained through model training). For the intercept term of the model, This represents the total number of features selected for the model (i.e., the total number of features in the second feature set).

[0044] Step S203: Calculate the total radiomics score using a multi-class probability determiner to output a probability distribution vector belonging to each preset damage level.

[0045] Specifically, the multi-class probability determiner is a machine learning model pre-built based on sample data during the training phase, using the total radiomics score as a single input feature. Internally, the model contains a discriminant function or decision rule learned for each preset injury level (e.g., levels 0 to 4 in the aforementioned scoring criteria). During computation, the model first converts the input total radiomics score into initial decision values ​​or logistic values ​​corresponding to each injury level based on its internal parameters. Then, it processes these initial values ​​using a normalization function (e.g., the softmax function or the sigmoid function under a one-to-many strategy), transforming them into a series of probability values ​​between 0 and 1. These probability values ​​collectively form a probability distribution vector, where each element sequentially corresponds to a preset injury level, and its value represents the posterior probability estimate of the input total radiomics score belonging to that specific level. The sum of all elements is 1, thus fully representing the probability of knee cartilage injury belonging to each level in a quantified probabilistic form.

[0046] Step S204: Calculate and determine the knee injury score and knee injury level based on the probability distribution vector.

[0047] Specifically, the knee injury score is calculated based on a linear combination of a probability distribution vector and a set of preset weight values. Each weight value in this set is pre-assigned and corresponds to a specific preset injury level, and the magnitude of the weight value is positively correlated with the clinical severity represented by that level. During calculation, each element in the probability distribution vector (i.e., the probability of belonging to the corresponding level) is multiplied by its associated preset weight value, and all products are summed to obtain a continuous knee injury score. This score is a comprehensive quantitative indicator, and its value directly reflects the overall severity of cartilage damage. Secondly, the knee injury level is determined by identifying the maximum value in the probability distribution vector. The preset injury level corresponding to this maximum value is initially determined as the knee injury level.

[0048] Step S205: Based on the SHAP visualization interpreter, perform interpretability analysis on the knee joint injury score and knee joint injury level to obtain a visualization assessment report.

[0049] Specifically, to better explain the model's prediction results, the SHAP (SHapley Additive Explanations) interpreter was used for visualization. SHAP, as a model interpretation method, is theoretically based on the Shapley value concept in game theory. It possesses powerful capabilities, accurately quantifying the contribution of each input feature to the model's prediction. Through this quantification, the embodiments of this application can delve into and understand the complex decision-making logic behind the model. In practical applications, SHAP visualization plays a crucial role in ranking feature importance. Using SHAP values, the influence of each feature on the model's prediction results can be clearly and intuitively identified, quickly pinpointing those features with the most significant impact. SHAP values ​​not only effectively explain the overall situation but also have the ability to explain individual predictions. SHAP values ​​can provide detailed analysis of the prediction results for individual data points, helping to deeply understand the reasons behind the model's specific predictions.

[0050] Furthermore, in this embodiment of the application, to enhance the interpretability of the model and support clinical decision-making, the features within the second feature set can be visualized at the pixel level. The feature values ​​of a single image are mapped to each pixel or voxel, and their local distribution is displayed as a heatmap. Two patients with different cartilage grades are randomly selected for demonstration: one patient has low-grade cartilage (grade 0-2) in all three regions, and the other patient has high-grade cartilage (grade 3-4). (Refer to...) Figure 4 As shown, Figure 4 This is a schematic diagram illustrating the visualization of pixel-level features of two patients provided in this application embodiment. A representative two-dimensional sagittal slice of the knee joint is selected for detailed pixel-level feature visualization. By detecting the degree of change in pixel-level features between high and low grades, a better understanding of the distribution changes of features across different grades can be achieved. Figure 4 As can be seen, low-grade cartilage lesions show a more uniform pixel color distribution with no obvious missing areas. High-grade cartilage lesions, on the other hand, have richer pixel colors and show some missing areas.

[0051] Specifically, in steps S201 to S205, key features are selected from the first feature set to form a second feature set. A linear weighted summation of the second feature set yields a total score including radiomics scores for the patellofemoral, medial femorotibia, and lateral femorotibia. This achieves deep fusion and precise quantification of multi-regional imaging information, significantly improving the objectivity and repeatability of the assessment. Subsequently, a multi-class probability determiner processes the total score and outputs a probability distribution vector. This not only provides discrete damage level determination but also calculates continuous knee joint injury scores based on this vector, thereby achieving a gradient and refined description of the severity of cartilage damage and enhancing the monitoring ability of subtle pathological changes. Most importantly, the SHAP visualization interpreter performs interpretability analysis on the score and level results and generates a visual assessment report, making the model's decision-making basis transparent and traceable. This effectively solves the "black box" problem of traditional models, significantly improving the credibility of the results in clinical practice and physician acceptance, providing reliable technical support for the accurate diagnosis and individualized management of knee cartilage damage.

[0052] Furthermore, refer to Figure 5 As shown, Figure 5 This is provided by the embodiments of this application. Figure 2 An optional schematic diagram of step S205 is provided, and the method may include, but is not limited to, steps S301 to S303.

[0053] Step S301: Based on the total radiomics score data, the SHAP value of each feature vector in the second feature set is calculated by the SHAP visualization interpreter during the multi-class probability determiner operation.

[0054] Specifically, the SHAP visualization interpreter is built upon the Shapley value principle of cooperative game theory, treating the predicted output of the multi-class probability determiner as the result of a collaborative "game" involving numerous features. During calculation, the interpreter first analyzes the data composition of the total radiomics score, identifying which specific feature vectors in the second feature set and their corresponding weight coefficients are generated by a linearly weighted summation. Then, using a model-independent approximation algorithm, given the multi-class probability determiner and the current input sample, the interpreter systematically simulates and evaluates the marginal impact of each feature vector in the second feature set on the probability distribution vector of the model's final output when various feature combinations are "present" or "absent." By exhaustively enumerating or approximating all possible feature combinations and fairly allocating contributions, a specific SHAP value is calculated for each feature vector in the second feature set. This value is a real number; its sign indicates the direction (positive or negative) of the feature vector's contribution to pushing the model output towards a specific damage level prediction, while its absolute value quantifies the magnitude of the feature vector's contribution.

[0055] Step S302: Determine the feature contribution of each feature vector in the second feature set to the output probability distribution vector of the multi-class probability determiner based on the SHAP value corresponding to each feature vector in the second feature set.

[0056] Specifically, this process is based on the calculated SHAP values ​​of each feature vector. For a given input sample, the SHAP value corresponding to each feature vector directly quantifies the model output offset caused by that feature in a specific prediction, relative to the average prediction of all features taking the baseline value. To determine its overall feature contribution to the output probability distribution vector, the following method is used: for the prediction of the current sample, the absolute value of the SHAP value of each feature vector is taken. The magnitude of this absolute value characterizes the strength of the feature vector's influence on the model's prediction result, regardless of its direction of influence. Subsequently, based on the set of absolute values ​​of all feature vector SHAP values, a quantified feature contribution index can be assigned to each feature vector in the second feature set by sorting or calculating relative proportions (e.g., the percentage of a single feature's SHAP absolute value to the sum of all feature SHAP absolute values). This index intuitively reveals and ranks the magnitude of the role of different image features in generating the current probability distribution vector, and thus in the final knee injury score and grade determination.

[0057] Step S303: Based on the feature contribution of each feature vector in the second feature set, generate a global feature importance ranking map and an individual prediction interpretation map, and integrate the global feature importance ranking map and the individual prediction interpretation map into a visual evaluation report.

[0058] Specifically, when generating the global feature importance ranking map, firstly, on the validation set or representative sample set, the feature contribution of each feature vector in the second feature set of each sample is calculated. Then, for each feature vector, its feature contribution across all samples is averaged to obtain a value representing its global average influence. Finally, all feature vectors are sorted from highest to lowest according to their average contribution value, and a bar chart is generated in this order. This chart visually displays the relative importance of different radiomics features in the overall model decision. When generating the individual prediction interpretation map, for the prediction results of a specific patient, the SHAP values ​​of each feature vector in the second feature set of that patient are extracted and calculated in this prediction. This data is used to generate a force map or decision map, where directional arrows visually represent the direction and magnitude of each feature vector's contribution to raising or lowering the model's base output value to the final prediction result. Finally, the global feature importance ranking chart reflecting the overall pattern of the model and the individual prediction interpretation chart for the decision-making logic of specific cases, together with the necessary patient identification, total radiomics score, final damage level and score and other key information, are compiled and output into a structured visual assessment report document for clinical review and decision-making reference.

[0059] To gain a deeper understanding of the model's decision-making mechanism and verify its feature discrimination ability, refer to Figure 6 and Figure 7 As shown, Figure 6 This is an optional schematic diagram of the global feature importance ranking graph provided in an embodiment of this application. Figure 7This is an optional schematic diagram of the individual prediction interpretation diagram provided in the embodiments of this application. The global feature importance ranking diagram shows the overall influence of each feature on the model prediction. The diagram is arranged in descending order according to the average absolute value of the SHAP value corresponding to each feature. The length of the bar directly represents the importance of the feature. The analysis found that the features that have the most significant impact on the prediction results are all first-order statistical features. For example, the feature ranked first is lbp-2D_firstorder_Mean. The lower its value, the greater the positive contribution to the prediction result of high-level damage. This suggests that the decrease in the mean of local binary pattern texture in the subchondral bone region may be related to a more severe damage level. Meanwhile, the individual prediction interpretation plot (also known as the beehive plot) reveals the details of the contribution of each feature in a single prediction. Each point in the plot represents a sample, and the horizontal position of the point indicates the SHAP value of the corresponding feature in that sample (right for positive contribution, left for negative contribution). The color indicates the level of feature value (blue for low value, red for high value), and the density distribution of the points intuitively reflects the clustering of feature SHAP values. For example, for the aforementioned lbp-2D_firstorder_Mean feature, the blue points are mainly concentrated on the right side, confirming that lower feature values ​​tend to drive the model to make predictions of higher-level lesions. This analysis significantly enhances the transparency and clinical credibility of the model's decision-making and helps to discover potential radiological biomarkers.

[0060] Secondly, to evaluate the characterization ability of the screened features, this application embodiment also employs t-SNE dimensionality reduction visualization technology to process the feature data of three locations: the patellofemoral joint, the medial femorotibiofemoral joint, and the lateral femorotibiofemoral joint. For example... Figure 7 As shown, after mapping the features of all samples belonging to different injury levels (levels 0 to 4) into a two-dimensional space, the individual prediction interpretation map shows a clear clustering trend among sample points of the same level, while there are clear separation boundaries between point sets of different levels. This result intuitively confirms that the feature set selected by the method of this application can effectively distinguish cartilage injury states of different severity, and verifies the effectiveness of the model's feature representation and classification feasibility from the perspective of data distribution.

[0061] Furthermore, refer to Figure 8 As shown, Figure 8 This is an optional flowchart of training a radiomics model provided in the embodiments of this application. The process of this method may include, but is not limited to, steps S401 to S407.

[0062] Step S401: Obtain knee joint MRI training images from multiple sample patients.

[0063] Step S402: Perform image region segmentation on multiple knee joint MRI training images to obtain multiple patellofemoral images representing the patellofemoral joint region, multiple medial femorotibia images representing the medial femorotibia joint region, and multiple lateral femorotibia images representing the lateral femorotibia joint region.

[0064] Step S403: Extract features from multiple patellofemoral images, multiple medial femoral-tibia images, and multiple lateral femoral-tibia images to obtain patellofemoral feature sets, medial femoral-tibia feature sets, and lateral femoral-tibia feature sets.

[0065] Specifically, this step is performed separately for the patellofemoral joint region image sets, medial femorotibialis joint region image sets, and lateral femorotibialis joint region image sets obtained from different patient samples through image region segmentation. Using a pre-defined radiomics feature extraction algorithm, such as the PyRadiomics library in Python or an equivalent tool, a set of quantitative image features is independently calculated and derived from each patellofemoral joint region image. The features extracted from all patellofemoral images are then aggregated to form a patellofemoral feature set. Similarly, the same feature extraction process is performed on each medial femorotibialis joint region image and each lateral femorotibialis joint region image, respectively aggregating them to form medial femorotibialis and lateral femorotibialis feature sets. Each feature set is structurally a two-dimensional data matrix, with rows corresponding to different sample images and columns corresponding to multiple radiomics features extracted from a single image. Feature types include, but are not limited to, first-order statistical features, 3D shape features, and texture features. This process provides a structured raw data foundation for subsequent feature selection and modeling analysis for each joint region.

[0066] Step S404: Perform standardization preprocessing on the patellofemoral feature set, medial femoral-tibia feature set, and lateral femoral-tibia feature set respectively to obtain the patellofemoral key feature set, medial femoral-tibia key feature set, and lateral femoral-tibia key feature set.

[0067] The patellofemoral key feature set includes the patellofemoral key features in each patellofemoral image and the weight coefficients of each patellofemoral key feature set; the medial femoral-tibia key feature set includes the medial femoral-tibia key features in each medial femoral-tibia image and the weight coefficients of each medial femoral-tibia key feature set; and the lateral femoral-tibia key feature set includes the lateral femoral-tibia key features in each lateral femoral-tibia image and the weight coefficients of each lateral femoral-tibia key feature set.

[0068] In a preferred embodiment, a total of 1409 initial radiomics features were extracted from images of the patellofemoral joint region, the medial femorotibialis joint region, and the lateral femorotibialis joint region, and respectively assigned to feature sets corresponding to the three regions (patellofemoral feature set, medial femorotibialis feature set, and lateral femorotibialis feature set). Subsequently, multi-level feature screening was performed on each feature set. In the first stage, a variance thresholding method was applied to remove low-variance features with variance below a preset threshold to exclude features with scarce information. In the second stage, the univariate feature selection method SelectKBest was used to further eliminate redundant features based on the statistical correlation between features and injury grade labels. In the third stage, a minimum absolute contraction and selection operator regression algorithm was applied for fine screening. In this stage, the optimal regularization parameter λ in the LASSO regression model was determined by 5-fold cross-validation. This λ value corresponds to the point with the minimum mean standard error of cross-validation (see attached figure). Finally, the optimal λ value was used to fit the LASSO model, retaining only features with non-zero coefficients. Following the complete process described above, 97 non-zero coefficient features from the patellofemoral feature set, 93 non-zero coefficient features from the medial femoral-tibia feature set, and 56 non-zero coefficient features from the lateral femoral-tibia feature set were selected and retained from the patellofemoral feature set, the medial femoral-tibia feature set, and the lateral femoral-tibia feature set. These features constitute the final key feature sets for each site and are used for subsequent radiomics scoring calculations. (Refer to...) Figure 9 As shown in Table 1, Figure 9 This is a schematic diagram of the radiomics feature set with non-zero coefficients obtained after standardized preprocessing according to an embodiment of this application. Table 1 is a data table of the radiomics feature set with non-zero coefficients obtained after standardized preprocessing according to an embodiment of this application. Figure 9 Each colored line represents the trajectory of the weight coefficients corresponding to each key feature. Finally, features with 97 non-zero coefficients were retained for the patellofemoral key feature set, features with 93 non-zero coefficients were retained for the lateral femoral-tibia key feature set, and features with 56 non-zero coefficients were retained for the lateral femoral-tibia key feature set. These features were used to calculate the radiomics score Rad-score and to build a radiomics model based on the radiomics score Rad-score.

[0069] Table 1 Feature selection process and results Table 1 Feature selection process and results

[0070] Step S405: Perform linear weighted summation on the patellofemoral key features in each patellofemoral image, the medial femoral-tibia key features in each medial femoral-tibia image, and the lateral femoral-tibia key features in each lateral femoral-tibia image to obtain the patellofemoral radiomics score corresponding to each patellofemoral image, the lateral femoral-tibia radiomics score corresponding to each medial femoral-tibia image, and the lateral femoral-tibia radiomics score corresponding to each lateral femoral-tibia image.

[0071] Specifically, for each patellofemoral image, each feature value in the standardized preprocessed patellofemoral key feature set is multiplied by the weight coefficient determined for that feature during model training using minimum absolute contraction and selection operator regression. All products are then summed along with the regression intercept term to obtain the patellofemoral radiomics score for that image. For each medial femorotibia and each lateral femorotibia image, the exact same calculation method is used. The medial femorotibia key feature set and its corresponding weight coefficients, and the lateral femorotibia key feature set and its corresponding weight coefficients, are used to independently calculate and generate the medial femorotibia radiomics score for each medial femorotibia image and the lateral femorotibia radiomics score for each lateral femorotibia image, respectively. The radiomics score for each location is a continuous numerical value, comprehensively representing the quantitative information of the selected key imaging features within that specific joint region, and serving as the basic input data for subsequently constructing the overall evaluation model.

[0072] Step S406: Construct and train a multi-class probability determiner based on the patellofemoral radiomics score of each patellofemoral image, the medial femoral-tibia radiomics score of each medial femoral-tibia image, and the lateral femoral-tibia radiomics score of each lateral femoral-tibia image.

[0073] Specifically, the process begins with data preparation. Patellofemoral radiomics scores, medial femorotibia radiomics scores, and lateral femorotibia radiomics scores from the same knee MRI sample are combined to form a three-dimensional feature vector, which serves as the input feature for the multi-class probability determiner. The overall knee injury level, as determined by experts, corresponding to this sample is used as the label for supervised learning. Next, a machine learning algorithm suitable for multi-class problems is selected as the base model for the probability determiner, such as logistic regression, support vector machine, or random forest. Subsequently, on a training set consisting of multiple samples, cross-validation is used to perform grid search and optimization on the key hyperparameters of the selected algorithm to minimize the preset loss function and improve the model's generalization performance. After training, the model outputs a probability distribution vector for each input three-dimensional feature vector, where each element represents the predicted probability that the input sample belongs to a preset injury level. Finally, the trained model structure, optimized parameters, and necessary probability calibration modules are saved, forming a multi-class probability determiner that can be used to predict new samples.

[0074] Step S407: Use the patellofemoral key feature set, medial femoral-tibia key feature set, and lateral femoral-tibia key feature set as configuration parameters for the feature selector, and form a radiomics model with the multi-class probability determiner.

[0075] In practice, after the model training phase is completed, the patellofemoral key feature set, medial femorotibia key feature set, and lateral femorotibia key feature set (including the specific feature names, indices, and corresponding standardized parameters of each set) representing the final screening results for each site are solidified and stored. This set of data constitutes the core configuration parameters of the feature selector. During the application phase, based on this configuration, the feature selector can automatically and accurately identify and extract a subset of key features completely consistent with those from the original features of new input samples. Subsequently, this feature selector is sequentially connected to the trained multi-class probability determiner, and together they are encapsulated into an independently runnable software module or algorithm pipeline. This integrated unit constitutes the final radiomics model, whose workflow is as follows: first, the original image features are processed through the embedded feature selector and its configuration parameters, outputting standardized key feature vectors; then, radiomics scores for each site are calculated and combined, and input into the multi-class probability determiner to finally generate the damage level and score. In this way, the complete analytical logic and key parameters of the model are solidified and integrated, ensuring the consistency, repeatability, and convenient clinical deployment capabilities of the assessment process.

[0076] In steps S401 to S407, a radiomics model constructed through a systematic training process enables more accurate, reliable, and interpretable automated assessment of knee cartilage damage. First, by segmenting the knee joint into three independent functional zones—patellofemoral, medial femorotibialis, and lateral femorotibialis—and extracting and filtering key features for each zone, the model can capture and quantify specific degenerative information from different biomechanical regions, significantly improving the anatomical specificity and accuracy of the grading. Second, the step of linearly weighting and summing the key features of each site using weight coefficients to generate a radiomics score condenses multidimensional imaging information into a single quantitative indicator with clear physical meaning. This not only simplifies model input and enhances noise resistance but also provides an intuitive numerical interpretation basis for the results. Furthermore, a multi-class probability determiner constructed and trained based on multi-site scores can comprehensively assess the overall damage state and output refined results that combine discrete grades with continuous probability distributions. Finally, the key feature set and weights obtained from training are solidified into the configuration parameters of the feature selector and integrated with the probability determiner into a complete model. This ensures that the entire process from feature processing to final decision-making is standardized, reproducible, and easy to deploy. It effectively overcomes the shortcomings of traditional methods, such as strong subjectivity, poor consistency, and the "black box" operation of existing machine learning models, and provides clinical practice with an objective, quantitative, and transparent intelligent auxiliary diagnostic tool.

[0077] further, Figure 10 This is provided by the embodiments of this application. Figure 8 An optional schematic diagram of step S406 is provided, and the method may include, but is not limited to, steps S501 to S505.

[0078] Step S501: Merge the patellofemoral radiomics scores of the patellofemoral images, the medial femorotibia radiomics scores of each medial femorotibia image, and the lateral femorotibia radiomics scores of each lateral femorotibia image into the same image dimension to obtain the total radiomics score of each knee joint MRI training image.

[0079] Step S502: Using the total radiomics score of each knee joint MRI training image as input and the knee joint injury level of the sample patients as a label, construct and train multiple patellofemoral candidate classification models, multiple medial femoral-tibia candidate classification models, and multiple lateral femoral-tibia candidate classification models through various statistical classification algorithms.

[0080] The types of statistical classification algorithms include logistic regression, support vector machine, random forest, etc., and this application does not make specific limitations.

[0081] In practice, the total radiomics score refers to a comprehensive feature vector formed by combining the radiomics scores of the patellofemoral, medial femorotibia, and lateral femorotibia joints from the same image. This feature vector is used as input features, and its corresponding overall knee joint injury level, determined by the gold standard, is used as a label for supervised learning, forming the training sample set. Subsequently, at least two statistical classification algorithms with fundamental differences, such as logistic regression and support vector machine algorithms, are selected as the basic learners. For the three evaluation sites—patellofemoral, medial femorotibia, and lateral femorotibia—each of the selected statistical classification algorithms is independently applied to train the model on the training sample set. Thus, for the patellofemoral joint, multiple patellofemoral candidate classification models based on different algorithms are generated; for the medial and lateral femorotibia joints, corresponding numbers of medial and lateral femorotibia candidate classification models are also generated, providing multiple alternative, trained machine learning models for each site.

[0082] Step S503: Using knee MRI verification images from multiple patient samples, the performance of multiple patellofemoral candidate classification models, multiple medial femoral-tibia candidate classification models, and multiple lateral femoral-tibia candidate classification models is evaluated.

[0083] Specifically, a validation image set, independent of the training set, is first prepared. This set contains knee MRI images of multiple patients, with corresponding patellofemoral, medial femorotibialis, and lateral femorotibialis regions obtained through image region segmentation. Then, using the same feature extraction process and parameters as in the training phase, radiomics features are extracted from the images of each region in the validation set. Standardized key feature vectors for each region are obtained using the standardized parameters and feature selectors saved from the training phase, and the total radiomics score for each validation sample is calculated. Next, the calculated total radiomics score is sequentially input into each candidate patellofemoral, medial femorotibialis, and lateral femorotibialis classification model to obtain the damage level prediction results for all validation samples. Then, the prediction results are compared with the known gold standard damage level labels for the validation samples. A set of preset performance metrics, including accuracy, precision, recall, F1 score, and the integral value of the receiver operating characteristic (ROC) curve, are used to independently calculate the evaluation values ​​for each candidate model. Finally, a complete set of performance evaluation parameters for each candidate model on an independent validation set is obtained, providing a quantitative basis for subsequent selection of the optimal classification model for each part.

[0084] Step S504: Select the model with the best performance evaluation from multiple patellofemoral candidate classification models as the official patellofemoral classification model; select the model with the best performance evaluation from multiple medial femoral-tibia candidate classification models as the official medial femoral-tibia classification model; and select the model with the best performance evaluation from multiple lateral femoral-tibia candidate classification models as the official lateral femoral-tibia classification model.

[0085] Step S505: Combine the formal classification models of patellofemoral, medial femoral tibia and lateral femoral tibia to form a multi-class probability determinist.

[0086] Among multiple patellofemoral candidate classification models, the model with the best overall performance index is selected based on its performance evaluation parameters obtained on the independent validation set and determined as the final official patellofemoral classification model. Using the same selection principle, one optimal model is selected from multiple medial femoral-tibia candidate classification models and one optimal model from multiple lateral femoral-tibia candidate classification models, serving as the official medial femoral-tibia classification model and the official lateral femoral-tibia classification model, respectively. Subsequently, the selected official patellofemoral, medial femoral-tibia, and lateral femoral-tibia classification models are integrated to form the final multi-class probability determiner. When this determiner operates, for a new input sample, it calls the three official classification models in parallel to perform predictions, with each model outputting a probability distribution vector. Then, through a preset integration strategy (such as weighted averaging or taking the maximum value of the corresponding probability values ​​of the three vectors), these intermediate results are fused to generate a unified, final probability distribution vector, thereby completing the comprehensive determination of the overall injury level of the knee joint.

[0087] It should be noted that in steps S501 to S505, radiomics scores from different regions of the same image are merged dimensionally to generate an overall score that integrates information from multiple sites, thereby achieving a comprehensive characterization of the overall condition of the knee joint and providing a more comprehensive and integrated data foundation for subsequent analysis. Multiple statistical classification algorithms are trained in parallel to generate multiple candidate models for each site, effectively avoiding model bias that may exist with a single algorithm. Through competition and complementarity among algorithms, multiple potential high-performance solutions are reserved for each site. Furthermore, a systematic performance evaluation of all candidate models is conducted using an independent validation image set. This objective screening mechanism ensures that the official classification model with the best discriminative ability for each anatomical site is selected based on empirical data, maximizing the predictive accuracy and robustness of each site's sub-model. Finally, the three selected official classification models are combined to construct a multi-class probability determiner. This ensemble strategy not only retains the discriminative advantages of each region-specific model but also enhances the overall grading system's comprehensive judgment ability and generalization reliability for complex multi-regional lesions through collaborative decision-making among models, thereby significantly improving the overall performance and clinical applicability of the automated grading of knee cartilage damage.

[0088] Furthermore, refer to Figure 11 As shown, Figure 11 This is provided by the embodiments of this application. Figure 8 An optional schematic diagram of step S404 is provided, and the method may include, but is not limited to, steps S601 to S603.

[0089] Step S601: Using the variance thresholding method, features with variances less than or equal to a predetermined threshold are removed from the intervertebral disc in the patellofemoral feature set, the medial femoral-tibia feature set, and the lateral femoral-tibia feature set, respectively, to obtain the first patellofemoral screening feature set, the first medial femoral-tibia screening feature set, and the first lateral femoral-tibia screening feature set.

[0090] Specifically, using the variance thresholding method, features with variances less than or equal to a predetermined threshold are removed from the patellofemoral feature set, the medial femoral-tibia feature set, and the lateral femoral-tibia feature set, respectively, to obtain the first patellofemoral screening feature set, the first medial femoral-tibia screening feature set, and the first lateral femoral-tibia screening feature set. In specific implementation, the predetermined threshold is set to 0.75. For each feature set, the variance of each feature is calculated. Then, the variance value of each feature is compared with the threshold of 0.75, and all feature entries with variance values ​​less than or equal to 0.75 are removed. The remaining features with variances greater than 0.75 constitute the first screening feature set for the corresponding feature set.

[0091] Step S602: Using a univariate feature selection method, analyze the first patellofemoral screening feature set, the first medial femoral-tibia screening feature set, and the first lateral femoral-tibia screening feature set to retain features that are significantly related to the knee joint injury level classification results, and obtain the second patellofemoral screening feature set, the second medial femoral-tibia screening feature set, and the second lateral femoral-tibia screening feature set.

[0092] In practice, the SelectKBest method was used for univariate feature selection. During analysis, for each site, the correlation between each feature in the first selected feature set and the knee injury grade label was independently calculated using ANOVA (F-test). The criterion for retaining features was that the p-value of the hypothesis test between the feature and the classification result was less than 0.05. For each feature set, only features with p-values ​​less than 0.05 were retained, and all features with p-values ​​greater than or equal to 0.05 were removed. The features retained after this step constituted the second selected feature set for the corresponding site. This stage aims to further eliminate features lacking a significant statistical association with the injury grade based on their independent discriminative ability, providing a more promising feature subset for subsequent multivariate feature selection.

[0093] Step S603: Using the LASSO algorithm with nested cross-validation, regression analysis is performed on the second patellofemoral screening feature set, the second medial femoral-tibia screening feature set, and the second lateral femoral-tibia screening feature set, respectively. Features with non-zero weight coefficients and their corresponding weight coefficients are retained to obtain the patellofemoral key feature set, the medial femoral-tibia key feature set, and the lateral femoral-tibia key feature set.

[0094] In practice, a separate LASSO regression model is constructed for each site's second selection feature set. During model training, a 5-fold cross-validation method is used to determine the optimal regularization strength parameter λ: the training data is randomly divided into five parts, and the LASSO model is fitted to four of these parts sequentially as training subsets, while the model performance is evaluated on the remaining validation subset, repeated five times. The average validation error corresponding to different λ values ​​is calculated, and the λ value that minimizes the average validation error is selected as the optimal parameter. Subsequently, the LASSO model is retrained on the entire training set using this optimal λ value. After training, only features with non-zero weight coefficients are retained, as these features are considered to make a substantial contribution to predicting knee joint injury levels. Each retained feature and its specific non-zero weight coefficient are recorded together to form the final key feature set for the corresponding site, which will be directly used to calculate the radiomics score for each site.

[0095] Furthermore, in the model construction and validation process of this application, the impact of two commonly used processing strategies on model performance was compared and analyzed to address the imbalanced distribution of cartilage gradations in the training data. One strategy is to balance the dataset before model training, and the other is to directly use the original imbalanced data for training. Referring to Table 2, which is an optional data table showing the AUC results of the model provided in the embodiments of this application with or without the balancing strategy, the comparison results show that regardless of whether support vector machine or logistic regression is used as the classifier, the impact of data balancing on the final model performance is not significant (the p-values ​​for inter-group comparisons between SVM and SVM_balanced, and between LR and LR_balanced are all greater than 0.05). At the same time, there is no statistically significant performance difference between the two classifiers, support vector machine and logistic regression (p-value greater than 0.05). Therefore, the subsequent model preferably uses the balanced logistic regression algorithm (LR_balanced) for construction.

[0096] Table 2. AUC results for models using a balancing strategy. Table 2 AUC results with or without the balanced processing strategy

[0097] Reference Figure 12 As shown, Figure 12 This is an optional schematic diagram illustrating the AUC results of the model provided in this application with and without the balancing strategy. This preferred model exhibits good classification performance on the training set, internal test set, and external test set (e.g., ...). Figure 2-8(As shown). In the internal test set, the model achieved micro / macro AUC values ​​of 0.821 / 0.785, 0.857 / 0.810, and 0.804 / 0.705 for the patellofemoral joint, medial femorotibiofemoral joint, and lateral femorotibiofemoral joint, respectively. In the external test set, the corresponding AUC values ​​were 0.806 / 0.801, 0.865 / 0.843, and 0.770 / 0.732, respectively, demonstrating good generalization ability of the model. Specifically, for each injury grade, the model showed the highest diagnostic efficacy for grade 4 cartilage lesions in the internal test set, with AUC values ​​ranging from 0.911 to 0.946 for the three sites. In the external test set, the patellofemoral joint and medial femorotibiofemoral joint had the highest AUC values ​​for grade 4 lesions (0.894 and 0.940, respectively), while the lateral femorotibiofemoral joint had the highest AUC value for grade 0 lesions (0.954).

[0098] Reference Figure 13 and Figure 14 As shown, Figure 13 This is an optional confusion matrix diagram showing the AUC results of the model with and without the balancing processing strategy provided in this application embodiment. Figure 14 This is an optional schematic diagram illustrating the precision, recall, and F1-score of the model provided in this application with and without the balanced processing strategy. Further analysis of model performance using a confusion matrix shows that in the internal test set, the recall rate for grade 0 lesions of the lateral femorotibiofemoral joint was the highest (0.807), while the recall rate for grade 2 lesions of the medial femorotibiofemoral joint was the lowest (0.444). In the external test set, the recall rate for grade 4 lesions of the medial femorotibiofemoral joint was the highest (0.808), while the recall rate for grade 2 lesions of the patellofemoral joint was the lowest (0.353). Regarding precision, the highest precision rates in the internal test set were for grade 4 lesions of the patellofemoral joint (0.778), grade 0 lesions of the medial femorotibiofemoral joint (0.896), and grade 0 lesions of the lateral femorotibiofemoral joint (0.848); the highest precision rates in the external test set were 0.800 (grade 4 patellofemoral joint), 0.840 (grade 4 medial femorotibiofemoral joint), and 0.792 (grade 0 lateral femorotibiofemoral joint), respectively. The trend of F1 scores is similar to that of accuracy (e.g.) Figure 13 (As shown). It should be noted that in the internal test set, the precision and F1 score for grade 2 lateral femorotibial joint lesions were the lowest (0.167 and 0.250, respectively); in the external test set, the corresponding values ​​for grade 3 patellofemoral joint lesions were the lowest (precision = 0.333, F1 score = 0.385). The above results fully validate the effectiveness and stability of the model constructed in this application.

[0099] The knee cartilage damage grading method provided in this application exhibits significant advantages in several aspects. First, based on the biomechanical characteristics of the knee joint, the method divides the cartilage into three independent assessment regions: the patellofemoral joint, the medial femorotibiofemoral joint, and the lateral femorotibiofemoral joint, constructing an improved whole-joint analysis system. This partitioning strategy better reflects the pathological characteristics of multiple biomechanical systems involved in knee osteoarthritis, and is conducive to capturing specific degeneration patterns in different cartilage subregions. In external validation, the model achieved micro / macro average AUC values ​​of 0.806 / 0.801, 0.865 / 0.843, and 0.770 / 0.732 for the three regions, respectively, confirming the model's good generalization ability. Furthermore, the feature contribution map generated by SHAP interpretability analysis can intuitively display the degenerative correlation between features in different regions, providing a new perspective for understanding the disease progression mechanism.

[0100] In terms of technical implementation, the method constructs an efficient feature selection process. From a total of 1409 initial radiomics features, 23 key features (LASSO coefficient greater than 0.15) were ultimately selected through screening methods such as LASSO regression, effectively controlling the risk of overfitting while improving the interpretability of the model. The t-SNE dimensionality reduction visualization shows that the selected features can clearly distinguish different damage levels. Crucially, the method systematically integrates the SHAP interpretability framework, which can quantify and reveal, for example, the dominant role of gray-level co-occurrence matrix features in high-level damage discrimination, thereby enhancing the transparency of model decision-making. Box plots of radiomics scores (Radscore) calculated for different damage levels also show significant statistical differences in scores between levels, further validating the discriminative power of the evaluation index.

[0101] In clinical applications, this method demonstrates significant value. Its diagnostic efficacy for high-grade injuries (e.g., grade 4) is outstanding (AUC between 0.90 and 0.93), providing a reliable reference for treatment decisions. The automated analysis process significantly shortens assessment time and improves diagnostic efficiency. The SHAP-driven visualization report greatly enhances the interpretability and clinical credibility of the model results. Although the current method is developed based on conventional sagittal PD-FS MRI sequences, future development by incorporating multi-parameter, multi-planar sequences and clinical data holds promise for further improving the model's robustness and decision-making value. Furthermore, combining it with emerging imaging technologies such as articular cartilage UTE can optimize the detection capability of early microstructural changes. In summary, this method provides an effective solution for the standardized, precise, and intelligent assessment of knee cartilage injuries.

[0102] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for grading knee cartilage damage. This electronic device can be any smart terminal, including mobile phones, tablets, and in-vehicle computers.

[0103] Please see Figure 15 , Figure 15 This is a schematic diagram of the hardware structure of an electronic device provided in one embodiment of this application. The electronic device includes: The processor 1501 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the knee cartilage damage grading method provided in the embodiments of this application. The memory 1502 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1502 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1502 and is called and executed by the processor 1501 using the knee cartilage injury grading method provided in the embodiments of this application. The input / output interface 1503 is used to implement information input and output; The communication interface 1504 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 1505 transmits information between various components of the device (e.g., processor 1501, memory 1502, input / output interface 1503, and communication interface 1504); The processor 1501, memory 1502, input / output interface 1503 and communication interface 1504 are connected to each other within the device via bus 1505.

[0104] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, provides a method for grading knee cartilage damage according to this application.

[0105] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0106] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0107] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0109] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0110] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0111] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0112] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0113] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0115] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-accessible storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0116] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for grading knee cartilage damage, characterized in that, include: Obtain MRI images of the patient's knee joint; The knee joint MRI image was segmented to obtain a first image representing the patellofemoral joint region, a second image representing the medial femorotibiofemoral joint region, and a third image representing the lateral femorotibiofemoral joint region in the patient. Feature extraction is performed on the first image, the second image, and the third image to obtain a first feature set, which includes the feature vectors of the first image, the second image, and the third image. The first feature set is input into the radiomics model, and the radiomics model is used to perform cartilage damage grading assessment and assessment interpretability analysis on the first feature set to obtain the patient's knee joint injury score and knee joint injury level, as well as a visual assessment report for interpreting the generated knee joint injury score and knee joint injury level. The radiomics model is a multi-class machine learning model built based on radiomics scores.

2. The method for grading knee cartilage damage according to claim 1, characterized in that, The radiomics model includes a feature selector, a multi-class probability determiner, and a SHAP visualization interpreter. The configuration parameters of the feature selector include multiple key features determined through training and the weight coefficients corresponding to the key features. The step of inputting the first feature set into the radiomics model, performing interpretability analysis on the first feature set through the radiomics model, and obtaining the patient's knee injury score and knee injury grade, as well as a visual assessment report for interpreting and generating the knee injury score and knee injury grade, includes: The feature selector selects feature vectors that match the key features from the first feature set to obtain a second feature set. By using the weight coefficients corresponding to each feature vector in the second feature set, a linear weighted sum is performed on each feature vector in the second feature set to obtain the total radiomics score of the knee joint MRI image. The total radiomics score includes the patellofemoral radiomics score of the first image, the lateral femoral-tibial radiomics score of the second image, and the lateral femoral-tibial radiomics score of the third image. The total radiomics score is calculated using the multi-class probability determiner to output a probability distribution vector belonging to each preset damage level. The knee injury score and the knee injury level are calculated and determined based on the probability distribution vector. Based on the SHAP visualization interpreter, the interpretability analysis of the knee joint injury score and the knee joint injury level is performed to obtain the visualization assessment report.

3. The method for grading knee cartilage damage according to claim 2, characterized in that, The interpretability analysis of the knee joint injury score and the knee joint injury level, based on the SHAP visualization interpreter, is performed to obtain the visualization assessment report, including: The SHAP visualization interpreter calculates the SHAP value of each feature vector in the second feature set during the multi-class probability determiner operation based on the total radiomics score data. Based on the SHAP value corresponding to each feature vector in the second feature set, determine the feature contribution of each feature vector in the second feature set to the output probability distribution vector of the multi-class probability determiner; Based on the feature contribution of each feature vector in the second feature set, a global feature importance ranking map and an individual prediction interpretation map are generated, and the global feature importance ranking map and the individual prediction interpretation map are integrated into the visualization evaluation report.

4. The method for grading knee cartilage damage according to claim 2, characterized in that, The process of constructing the radiomics model includes the following steps: Acquire knee joint MRI training images from multiple patient samples; Image region segmentation was performed on multiple knee joint MRI training images to obtain multiple patellofemoral images representing the patellofemoral joint region, multiple medial femorotibia images representing the medial femorotibia joint region, and multiple lateral femorotibia images representing the lateral femorotibia joint region. Feature extraction is performed on the multiple patellofemoral images, the multiple medial femoral-tibia images, and the multiple lateral femoral-tibia images to obtain patellofemoral feature sets, medial femoral-tibia feature sets, and lateral femoral-tibia feature sets; The patellofemoral feature set, the medial femoral-tibia feature set, and the lateral femoral-tibia feature set are respectively subjected to standardized preprocessing to obtain the patellofemoral key feature set, the medial femoral-tibia key feature set, and the lateral femoral-tibia key feature set. The patellofemoral key feature set includes patellofemoral key features in each of the patellofemoral images and the weight coefficients of each of the patellofemoral key features. The medial femoral-tibia key feature set includes medial femoral-tibia key features in each of the medial femoral-tibia images and the weight coefficients of each of the medial femoral-tibia key features. The lateral femoral-tibia key feature set includes lateral femoral-tibia key features in each of the lateral femoral-tibia images and the weight coefficients of each of the lateral femoral-tibia key features. The patellofemoral key features in each of the patellofemoral images, the medial femoral-tibia key features in each of the medial femoral-tibia images, and the lateral femoral-tibia key features in each of the lateral femoral-tibia images are linearly weighted and summed to obtain the patellofemoral radiomics score corresponding to each of the patellofemoral images, the lateral femoral-tibia radiomics score corresponding to each of the medial femoral-tibia images, and the lateral femoral-tibia radiomics score corresponding to each of the lateral femoral-tibia images. The multi-class probability determiner is constructed and trained based on the patellofemoral radiomics scores of each patellofemoral image, the medial femoral-tibia radiomics scores of each medial femoral-tibia image, and the lateral femoral-tibia radiomics scores of each lateral femoral-tibia image. The patellofemoral key feature set, the medial femoral-tibia key feature set, and the lateral femoral-tibia key feature set are used as configuration parameters of the feature selector, and together with the multi-class probability determiner, they form the radiomics model.

5. The method for grading knee cartilage damage according to claim 4, characterized in that, The process of constructing and training the multi-class probability determiner based on the patellofemoral radiomics scores of each of the patellofemoral images, the medial femoral-tibia radiomics scores of each of the medial femoral-tibia images, and the lateral femoral-tibia radiomics scores of each of the lateral femoral-tibia images includes the following steps: The patellofemoral radiomics scores of the patellofemoral images, the medial femoral-tibial radiomics scores of each of the medial femoral-tibial images, and the lateral femoral-tibial radiomics scores of each of the lateral femoral-tibial images are merged in the same image dimension to obtain the total radiomics score of each of the knee joint MRI training images. Using the total radiomics score of each knee joint MRI training image as input and the knee joint injury level of the sample patients as a label, multiple patellofemoral candidate classification models, multiple medial femoral-tibia candidate classification models, and multiple lateral femoral-tibia candidate classification models were constructed and trained using various statistical classification algorithms. The performance of the multiple patellofemoral candidate classification models, the multiple medial femoral-tibia candidate classification models, and the multiple lateral femoral-tibia candidate classification models were evaluated using knee MRI verification images from multiple patient samples. The model with the best performance evaluation among the multiple patellofemoral candidate classification models is selected as the official patellofemoral classification model. The model with the best performance evaluation among the multiple medial femoral tibia candidate classification models is selected as the formal medial femoral tibia classification model. The model with the best performance evaluation among the multiple candidate classification models for lateral femoral tibia was selected as the official classification model for lateral femoral tibia. The patellofemoral formal classification model, the medial femoral tibia formal classification model, and the lateral femoral tibia formal classification model are combined to form the multi-class probability determiner.

6. The method for grading knee cartilage damage according to claim 5, characterized in that, The statistical classification algorithms include logistic regression and support vector machine algorithms; The performance evaluation metrics include at least one of the following: accuracy, precision, recall, F1 score, and the integral value of the quantized ROC curve.

7. The method for grading knee cartilage damage according to claim 4, characterized in that, The standardization preprocessing of the patellofemoral feature set, the medial femoral-tibia feature set, and the lateral femoral-tibia feature set respectively yields the patellofemoral key feature set, the medial femoral-tibia key feature set, and the lateral femoral-tibia key feature set, including: By using the variance thresholding method, features with variances less than or equal to a predetermined threshold are removed from the intervertebral disc in the patellofemoral feature set, the medial femoral-tibia feature set, and the lateral femoral-tibia feature set, respectively, to obtain the first patellofemoral screening feature set, the first medial femoral-tibia screening feature set, and the first lateral femoral-tibia screening feature set. Using a univariate feature selection method, the first patellofemoral screening feature set, the first medial femoral-tibia screening feature set, and the first lateral femoral-tibia screening feature set were analyzed to retain features that were significantly related to the knee joint injury level classification results, thus obtaining the second patellofemoral screening feature set, the second medial femoral-tibia screening feature set, and the second lateral femoral-tibia screening feature set. The LASSO algorithm with nested cross-validation is used to perform regression analysis on the second patellofemoral screening feature set, the second medial femoral-tibia screening feature set, and the second lateral femoral-tibia screening feature set, respectively, and retain the features with non-zero weight coefficients and their corresponding weight coefficients to obtain the patellofemoral key feature set, the medial femoral-tibia key feature set, and the lateral femoral-tibia key feature set.

8. The method for grading knee cartilage damage according to claim 7, characterized in that, The predetermined threshold is 0.75, the univariate feature selection method is the SelectKBest method, the criterion for determining that the knee joint injury level classification result is significantly correlated is that the p-value of the hypothesis test between the feature and the classification result is less than 0.05, and the LASSO algorithm uses 5-fold cross-validation.

9. The method for grading knee cartilage damage according to claim 1, characterized in that, The feature vectors in the first feature set include at least one of the following types: first-order features, three-dimensional shape features, gray-level co-occurrence matrix, gray-level size region matrix, gray-level run-length matrix, neighborhood gray-level tone difference matrix, and gray-level dependency matrix.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a processor-executable program, which, when executed by a processor, is used to implement the knee cartilage injury grading method as described in any one of claims 1 to 9.

Images