Patellofemoral joint osteoarthritis prediction method based on knee joint lateral position x-ray imaging omics

Abstract

This application discloses a method for predicting patellofemoral arthritis based on lateral X-ray radiomics of the knee joint. First, lateral X-ray images of the knee joint and clinical information of the patient to be tested are acquired. Then, the X-ray images and clinical information are input into a clinical radiomics prediction model. The model identifies regions of interest (ROIs) in the X-ray images and extracts features from these ROIs to obtain candidate radiomics features. These candidate features are then filtered according to dimensionality reduction rules to obtain target radiomics features. The types of radiomics features include at least grayscale features, shape features, texture features, and wavelet transform features. This application, by using a prediction model to extract and filter features from ROIs in X-ray images, accurately aligns with the diagnostic criteria for patellofemoral arthritis defined by MRI, improving the reliability of the prediction. Simultaneously, by using low-cost X-ray examinations to replace high-cost MRI, patients can detect the condition earlier without compromising diagnostic effectiveness, thus reducing the burden of examinations.

Description

Technical Field

[0001] The embodiments of this application relate to, but are not limited to, the field of smart medical technology, and in particular to a method for predicting patellofemoral arthritis based on lateral X-ray radiomics of the knee joint. Background Technology

[0002] Patellofemoral osteoarthritis (PFOA), a common knee joint disease, relies heavily on early diagnosis for effective intervention and treatment. Currently, early PFOA diagnosis in clinical practice primarily relies on magnetic resonance imaging (MRI), a technique that can clearly detect subtle lesions such as early cartilage damage and bone changes, making it a crucial tool for assessing early PFOA. However, MRI examinations have significant limitations: firstly, the cost of equipment and examinations is high, making it difficult for primary healthcare institutions or patients with limited financial resources to use as a routine screening method; secondly, the availability of MRI equipment is low, especially in primary healthcare facilities where resources are relatively scarce, resulting in many potential early-stage patients not receiving timely examination opportunities. These factors make MRI-based early PFOA diagnosis difficult to implement in primary healthcare settings, limiting the efficiency of early disease identification and intervention. Summary of the Invention

[0003] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0004] This application provides a method for predicting patellofemoral osteoarthritis based on lateral knee X-ray radiomics, which can accurately predict early patellofemoral osteoarthritis as defined by MRI by using low-cost lateral knee X-rays combined with clinical information.

[0005] This application provides a method for predicting patellofemoral arthritis based on lateral knee X-ray radiomics. The method includes: acquiring lateral knee X-ray images and clinical information of a patient to be tested; inputting the X-ray images and clinical information into a pre-trained clinical radiomics prediction model; determining regions of interest (ROIs) in the X-ray images using the model, and extracting features from the ROIs to obtain candidate radiomics features; filtering the candidate radiomics features according to a preset dimensionality reduction rule using the model to obtain target radiomics features, wherein the types of target radiomics features include at least grayscale features, shape features, texture features, and wavelet transform features; the grayscale features are used to describe the distribution and statistical characteristics of voxel intensity within the ROIs; the shape features are used to describe the geometric characteristics of the ROIs; the texture features are used to describe the spatial distribution information of voxels within the ROIs; the wavelet transform features are used to obtain multi-resolution image description information by wavelet transforming the image of the ROIs; and the model makes predictions based on the types of features in the target radiomics features and the clinical information to obtain prediction results.

[0006] In one embodiment of this application, determining the region of interest in the X-ray image includes: determining the superior and posterior margins of the patellofemoral joint in the femoral region of the X-ray image; determining the lower margins of the medial and lateral femoral condyles in the X-ray image; determining the boundary of the patellar region in the X-ray image; and obtaining the region of interest in the X-ray image based on the superior margin of the patellofemoral joint, the posterior margin of the patellofemoral joint, the lower margin of the medial femoral condyle, the lower margin of the lateral femoral condyle, and the boundary of the patellar region.

[0007] In one embodiment of this application, determining the superior and posterior margins of the patellofemoral joint in the femoral region of the X-ray image includes: determining a first intersection point in the X-ray image between the midline of the femoral shaft and the line connecting the anterior and posterior margins of the distal femoral joint; determining a second intersection point in the X-ray image between the anterior and superior margins of the femoral joint; connecting the first and second intersection points to obtain the superior margin of the patellofemoral joint; determining a third intersection point in the X-ray image between the trochlear base line and the projection line at the bottom of the intercondylar fossa; and connecting the first and third intersection points to obtain the posterior margin of the patellofemoral joint.

[0008] In one embodiment of this application, determining the lower edge of the medial femoral condyle articular surface and the lower edge of the lateral femoral condyle articular surface in the X-ray image includes: determining the reverse extension line of the intercondylar projection line in the X-ray image; taking the first intersection point of the reverse extension line and the medial femoral condyle articular surface as the lower edge of the medial femoral condyle articular surface; and taking the second intersection point of the reverse extension line and the lateral femoral condyle articular surface as the lower edge of the lateral femoral condyle articular surface.

[0009] In one embodiment of this application, the step of extracting features from the region of interest to obtain candidate image omics features includes: extracting corresponding candidate grayscale features, candidate shape features, candidate texture features, and candidate wavelet transform features from the region of interest; and obtaining candidate image omics features based on the candidate grayscale features, the candidate shape features, the candidate texture features, and the candidate wavelet transform features.

[0010] In one embodiment of this application, the step of filtering the candidate radiomics features according to a preset dimensionality reduction rule to obtain target radiomics features includes: using a variance thresholding method to initially filter the candidate radiomics features to obtain variance-effective features; using a univariate selection method to filter the variance-effective features to obtain association features related to patellofemoral arthritis; and using the LASSO algorithm to filter the association features to obtain target radiomics features.

[0011] In one embodiment of this application, the training steps of the pre-trained clinical radiomics prediction model include: acquiring lateral X-ray images of the knee joint and clinical information of patients in the training set; delineating regions of interest in the X-ray images of the training set and extracting features from the delineated regions of interest to obtain training candidate radiomics features; performing dimensionality reduction and screening on the training candidate radiomics features to obtain training target radiomics features; fusing the training target radiomics features with the corresponding clinical information to construct training samples; training the initial prediction model using the training samples based on the logistic regression algorithm, and adjusting the parameters of the initial prediction model through iterative optimization to obtain the pre-trained clinical radiomics prediction model.

[0012] In one embodiment of this application, after obtaining the prediction result, the method further includes: visualizing the prediction result based on a nomogram using the model, and marking the contribution of target radiomics features and clinical information to the prediction result in the nomogram.

[0013] In one embodiment of this application, after adjusting the initial prediction model parameters through iterative optimization, the method further includes: acquiring lateral X-ray images of the test knee joint and clinical information of the test set patients; testing the iteratively optimized model using the lateral X-ray images of the test knee joint and the clinical information to obtain test prediction results; and adjusting the model performance based on the test prediction results and the actual disease status of the test set patients.

[0014] In one embodiment of this application, the clinical information includes the age and gender of the patient to be tested.

[0015] This application provides a method for predicting patellofemoral arthritis based on lateral X-ray radiomics of the knee joint. First, lateral X-ray images of the knee joint and clinical information of the patient to be tested are acquired. Then, the X-ray images and clinical information are input into a pre-trained clinical radiomics prediction model. The model identifies regions of interest (ROIs) in the X-ray images and extracts features from these ROIs to obtain candidate radiomics features. These candidate features are then filtered according to a preset dimensionality reduction rule to obtain target radiomics features. The target radiomics features include at least grayscale features, shape features, texture features, and wavelet transform features. Grayscale features describe the distribution and statistical characteristics of voxel intensity within the ROI; shape features describe the geometric characteristics of the ROI; texture features describe the spatial distribution of voxels within the ROI; and wavelet transform features obtain multi-resolution image description information by performing wavelet transform on the ROI image. Subsequently, the model makes predictions based on the various types of features in the target radiomics features and the clinical information to obtain the prediction results. This application embodiment uses a clinical radiomics prediction model to extract and screen features of regions of interest in X-ray images, which can accurately match the diagnostic criteria for patellofemoral arthritis defined by MRI, improving the reliability of the prediction. At the same time, by using low-cost X-ray examination to replace high-cost MRI, patients can detect their condition earlier without reducing the diagnostic effect, thus reducing the burden of examination. Attached Figure Description

[0016] Figure 1 This is a flowchart of the patellofemoral arthritis prediction method provided in the embodiments of this application;

[0017] Figure 2 This is a flowchart of the training process for the clinical radiomics prediction model provided in the embodiments of this application; Figure 3 This is a schematic diagram of the region of interest segmentation on a lateral X-ray of the right knee joint provided in an embodiment of this application; Figure 4 This is provided by the embodiments of this application. Figure 1 Flowchart for step 130; Figure 5 This is provided by the embodiments of this application. Figure 1 Flowchart for step 140; Figure 6(a) is a partial schematic diagram of the feature selection process provided in an embodiment of this application; Figure 6(b) is a schematic diagram of another part of the feature screening process provided in the embodiments of this application; Figure 6(c) is another schematic diagram of the feature selection process provided in the embodiments of this application; Figure 7(a) shows the ROC curves of the clinical factor model provided in the embodiments of this application on different datasets; Figure 7(b) shows the ROC curves of the radiomics model provided in the embodiments of this application on different datasets; Figure 7(c) shows the ROC curves of the comprehensive model provided in the embodiments of this application on different datasets; Figure 8 This is a Nomogram of PFOA prediction based on MRI definition provided in one embodiment of this application; Figure 9(a) is a schematic diagram of the calibration curve of the Nomogram in the training set provided in the embodiments of this application; Figure 9(b) is a schematic diagram of the calibration curve of the Nomogram in the internal test set provided in the embodiments of this application; Figure 9(c) is a schematic diagram of the calibration curve of the Nomogram in the external test set provided in the embodiments of this application; Figure 10(a) is a schematic diagram of DCA curves for different models in the training set provided in the embodiments of this application; Figure 10(b) is a schematic diagram of DCA curves for different models in the internal test set provided in the embodiments of this application; Figure 10(c) is a schematic diagram of DCA curves for different models in the external test set provided in the embodiments of this application. Detailed Implementation

[0018] 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.

[0019] It should be noted that although the flowchart shows a logical order, in some cases, the steps shown or described may be performed in a different order than that shown in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the structures, proportions, sizes, etc., depicted in the drawings are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and purposes achieved by this application, should still fall within the scope of the technical content disclosed in this application. Similarly, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are only for clarity of description and are not used to limit the scope of implementation of this application. Changes or adjustments in their relative relationships, without substantially altering the technical content, should also be considered within the scope of implementation of this application.

[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] Patellofemoral osteoarthritis (PFOA) is a common knee joint disease, and its early diagnosis is crucial for the effectiveness of subsequent intervention and treatment. Currently, in clinical practice, the diagnosis of early PFOA mainly relies on magnetic resonance imaging (MRI) technology. This technology can clearly detect subtle lesions such as early cartilage damage and bone changes in the joint, and is an important means of assessing early PFOA.

[0022] However, MRI examinations have significant limitations: on the one hand, the cost of equipment and examinations is high, making it difficult for primary healthcare institutions or patients with limited financial resources to use as a routine screening method; on the other hand, the availability of MRI equipment is low, especially in primary healthcare institutions where resources are relatively scarce, resulting in many potential early-stage patients not being able to access examinations in a timely manner. These factors make it difficult to widely implement MRI-based early PFOA diagnosis in primary healthcare settings, limiting the efficiency of early disease identification and intervention.

[0023] In view of this, this application provides a method for predicting patellofemoral arthritis based on lateral X-ray radiomics of the knee joint. First, lateral X-ray images of the knee joint and clinical information of the patient to be tested are acquired. Then, the X-ray images and clinical information are input into a pre-trained clinical radiomics prediction model. The model determines the region of interest (ROI) in the X-ray image and extracts features from the ROI to obtain candidate radiomics features. The candidate radiomics features are then filtered according to a preset dimensionality reduction rule to obtain target radiomics features. The types of target radiomics features include at least grayscale features, shape features, texture features, and wavelet transform features. Grayscale features describe the distribution and statistical characteristics of voxel intensity within the ROI; shape features describe the geometric characteristics of the ROI; texture features describe the spatial distribution information of voxels within the ROI; and wavelet transform features obtain multi-resolution image description information by performing wavelet transform on the ROI image. Subsequently, the model makes predictions based on the various types of features in the target radiomics features and the clinical information to obtain the prediction results. This application embodiment uses a clinical radiomics prediction model to extract and screen features of regions of interest in X-ray images, which can accurately match the diagnostic criteria for patellofemoral arthritis defined by MRI, improving the reliability of the prediction. At the same time, by using low-cost X-ray examination to replace high-cost MRI, patients can detect their condition earlier without reducing the diagnostic effect, thus reducing the burden of examination.

[0024] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0025] Reference Figure 1 , Figure 1 This is a flowchart of a method for predicting patellofemoral arthritis based on lateral knee X-ray radiomics, provided in an embodiment of this application. The process may specifically include, but is not limited to, steps 110 to 150.

[0026] Step 110: Obtain lateral X-ray images and clinical information of the patient's knee joint; Step 120: Input X-ray images and clinical information into the pre-trained clinical radiomics prediction model; Step 130: Determine the region of interest in the X-ray image using the model, and extract features from the region of interest to obtain candidate radiomics features; Step 140: The candidate image omics features are filtered according to the preset dimensionality reduction rules by the model to obtain the target image omics features. The types of target image omics features include at least grayscale features, shape features, texture features, and wavelet transform features. Grayscale features are used to describe the distribution and statistical characteristics of voxel intensity within the region of interest. Shape features are used to describe the geometric characteristics of the region of interest. Texture features are used to describe the spatial distribution information of voxels within the region of interest. Wavelet transform features are used to obtain multi-resolution image description information by performing wavelet transform on the image of the region of interest. Step 150: The model makes predictions based on various types of features and clinical information in the target radiomics features, and obtains the prediction results.

[0027] In a feasible embodiment, when acquiring lateral X-ray images of the knee joint of the patient to be tested (i.e., DR lateral examination images), it can be first confirmed whether the patient meets the basic conditions for image acquisition, namely, no acute phase of knee fracture, no metal internal fixation shadow obscuring the patellofemoral joint area, and no severe joint deformity affecting the position, to ensure that the acquired images can clearly show the patellofemoral joint structure. Then, the patient is instructed to lie on their side on the radiography table, with the knee joint of the examined side close to the bed surface, and the contralateral lower limb naturally extended forward and upward to avoid obstruction; the affected knee joint is kept at a 20°-35° flexion angle, with the lateral side of the knee joint close to the detector, and the lower edge of the patella is precisely aligned with the center of the projection field; the center line is adjusted to be perpendicular to the upper end of the tibia, and during the imaging, it is ensured that the medial and lateral femoral condyles coincide in the image to avoid unclear display of joint structures due to positional deviation.

[0028] In one feasible embodiment, after image acquisition is completed, the quality of the X-ray image can be preliminarily verified to check whether there are obvious artifacts, blurring or occlusion in the patellofemoral joint area, and to confirm that the edge contours of key structures such as the distal femur, patella, and trochlea are clearly distinguishable and meet the requirements for subsequent region of interest delineation and feature extraction. If the image quality is not up to standard (such as the presence of structural overlap), the above process can be repeated until a qualified lateral X-ray image of the knee joint is obtained.

[0029] In one feasible embodiment, the clinical information of the patient to be tested refers to basic clinical data related to the onset and disease assessment of patellofemoral arthritis (PFOA), specifically including two core types of information: the patient's age and gender.

[0030] In one feasible embodiment, the pre-trained clinical radiomics prediction model can be directly used for PFOA prediction analysis of the patients to be tested. In step 130, the model can first locate the region of interest (i.e., the joint region that needs to be analyzed) in the image related to patellofemoral osteoarthritis (PFOA), and then extract a large number of initial features from the region. These unfiltered features are candidate radiomics features.

[0031] In a feasible embodiment, given the large number of candidate features and the potential for redundancy, in step 140, the model can perform dimensionality reduction and screening of the candidate radiomics features according to preset rules, ultimately retaining the target radiomics features with the highest correlation to PFOA. These target features include at least four categories: grayscale features describing voxel intensity distribution, shape features describing regional geometry, texture features describing voxel spatial arrangement, and wavelet transform features that obtain multi-resolution information through wavelet transform.

[0032] In one feasible embodiment, the model can combine the screened core imaging features with clinical information, and output a prediction result after comprehensive calculation to determine whether the patient has PFOA or related risks.

[0033] The entire process from steps 110 to 150 achieves automated analysis from low-cost X-ray images to PFOA prediction through a pre-trained model. This avoids the problems of high cost and difficulty in popularizing MRI examinations at the grassroots level, and can accurately extract key features associated with the disease. Ultimately, it can assist in the early identification of PFOA in an economical and convenient way, and is especially suitable for promotion and application in primary healthcare institutions.

[0034] In one feasible embodiment, such as Figure 2 As shown, the training process of the pre-trained clinical radiomics prediction model may include at least steps 210 to 250.

[0035] Step 210: Obtain lateral X-ray images of the knee joint and clinical information of the patients in the training set; Step 220: Delineate regions of interest in the X-ray images of the training set, and extract features from the delineated regions of interest to obtain training candidate image omics features. Step 230: Perform dimensionality reduction and screening on the training candidate image omics features to obtain the target image omics features of the training set; Step 240: Fuse the target radiomics features of the training set with the corresponding clinical information to construct training samples; Step 250: Based on the logistic regression algorithm, train the initial prediction model using training samples, and adjust the parameters of the initial prediction model through iterative optimization to obtain a pre-trained clinical radiomics prediction model.

[0036] In a feasible embodiment, in step 210, to ensure the effectiveness and relevance of the training data, the patients in the training set must meet clear inclusion and exclusion criteria: Inclusion criteria: (1) Chinese males or females aged 35 years or older; (2) no bone tumors or inflammatory osteoarthritis (such as rheumatoid arthritis, gouty arthritis, tuberculous arthritis, etc.); (3) no history of knee fracture or internal fixation with metal fixation; (4) no congenital malformation of the knee joint; (5) have undergone at least one knee joint magnetic resonance imaging (MRI) examination and DR lateral view examination at the same time, with the time interval between the two examinations not exceeding 6 months, and have not undergone knee joint surgery during the period.

[0037] Exclusion criteria: (1) Incomplete MRI image sequence or poor image quality (such as severe artifacts or occlusions) that cannot meet the diagnostic requirements; (2) Poor quality of lateral X-ray images of the knee joint (such as insufficient clarity or positional deviation) that affect subsequent feature extraction.

[0038] In a feasible embodiment, the equipment, coils, and related parameters used in MRI scanning, which serve as the gold standard for determining the patient's actual condition during model training and testing, are shown in Table 1. Here, TR is the repetition time, TE is the echo time, IR is the reversal time, ETL is the echo train length, and FOV is the field of view.

[0039] Table 1 Knee joint imaging scan parameters

[0040] The specific procedure for the lateral X-ray examination of the knee joint used in this embodiment is as follows: The patient lies on his / her side on the radiography table, with the lower limb of the examined side close to the bed surface and the lower limb of the opposite side extended forward and upward; the knee joint of the affected side is kept at a flexion angle of 20°-35°, with the outer side of the knee joint close to the detector; the lower edge of the patella is placed in the center of the projection field, the center line is aligned with the upper end of the tibia and is injected vertically, ensuring that the medial and lateral femoral condyles coincide in the image, so as to ensure the standardization and usability of the captured image.

[0041] In a feasible embodiment, in step 220, during the region of interest delineation stage of the model training phase, to ensure the accuracy of the region extraction and data consistency of the pre-trained clinical radiomics prediction model, standardized patellofemoral joint (PFJ) region of interest (ROI) delineation data can be prepared for the model first. The specific process is as follows: First, determine the delineation range of the region of interest: such as... Figure 3 As shown, the upper edge of the PFJ in the femoral region is the line connecting "the intersection point a of the line connecting the midline of the femoral shaft and the upper edge of the anterior and posterior margin of the distal femoral articular surface" and "the intersection point b of the anterior and superior margin of the femoral articular surface"; the posterior edge of the PFJ is the line connecting intersection point a and "the intersection point c of the projection line of the trochlear base line and the bottom of the intercondylar fossa"; the lower edge of the medial and lateral femoral condylar articular surfaces is the intersection point of the reverse extension line of the intercondylar fossa projection line and the articular surface; the patellar region is delineated along the natural edge of the patella, and delineation deviation is avoided through a unified range standard. Based on this, all knee lateral X-ray images from the training set were imported into ITKSNAP (3D and 4D biomedical image segmentation software). This software assisted physicians in delineating and labeling the PFJ region according to the above definition. Specifically, the delineation operation was performed by a radiologist with 3 years of experience in musculoskeletal imaging diagnosis under the guidance of a radiologist with 6 years of experience in musculoskeletal imaging diagnosis. Finally, the delineation results were reviewed and confirmed by a senior physician with over 20 years of experience in musculoskeletal imaging diagnosis to ensure reliability. After delineation, the image with the labeled region of interest was input into the model for subsequent extraction of training candidate radiomics features.

[0042] In a feasible embodiment, considering that the X-ray images in the training set may come from different models, to eliminate the impact of equipment differences on subsequent feature extraction, all lateral X-ray images of the knee joint can be preprocessed first. Specifically, this includes: standardizing the radiomics features extracted from the images to a normal distribution of z-scores, and truncating the feature values ​​within the range of 1%-99%, thereby ensuring the comparability of feature data from images taken by different devices. After completing the above preprocessing, feature extraction is then performed on the patellofemoral joint (PFJ) region of interest that has been delineated and verified using ITKSNAP software. The resulting initial features, without dimensionality reduction screening, are the training candidate radiomics features.

[0043] In a feasible embodiment, in step 230, in order to avoid model overfitting and multicollinearity, this step uses a third-order dimensionality reduction method to screen the training candidate radiomics features: (1) Variance threshold screening: set the variance threshold to 0.8, screen out low-information features with variance below the threshold, and retain features with high variance and strong discriminative power; (2) Univariate selection screening: screen out features that are not significantly associated with patellofemoral arthritis (PFOA) (p>0.05) through statistical tests, and retain features with strong association; (3) Minimum absolute contraction and selection operator (LASSO) algorithm screening: for the features screened in the first two steps, LASSO model fitting is performed based on the lesion type, and 10-fold cross-validation is used to avoid overfitting. Finally, the features most related to cartilage damage are screened out, which are the target radiomics features of the training set.

[0044] In a feasible embodiment, the target image omics features obtained from the screening training set are roughly divided into four categories: (1) First-order features (i.e., gray-scale features): features that describe the distribution of voxel intensity, reflecting the density information of lesions, the degree of regional disorder, etc.; (2) Shape features: features that describe the geometric characteristics of lesions, reflecting the size, shape, surface roughness, etc. of lesions; (3) Texture features: features that describe the spatial distribution information of lesion voxels, reflecting the spatial heterogeneity features of image gray-scale changes, granularity, roughness, etc.; (4) Wavelet transform-based features (i.e., wavelet transform features): multi-resolution image description information obtained by wavelet transforming the original image.

[0045] In a feasible embodiment, in step 240, the target radiomics features of the training set obtained in step 230 are fused with the clinical information of the corresponding patients collected in step 210 to form a combined data of target radiomics features and clinical information. The combined data of each patient is a training sample, and all training samples together constitute the dataset for model training.

[0046] In a feasible embodiment, step 250 may employ a logistic regression (LR) algorithm for model training, including: training the initial prediction model using the training samples constructed in step 240 as input; during training, continuously adjusting model parameters through iterative optimization, and evaluating model performance using various metrics, such as evaluating model predictive efficacy through the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, and accuracy; evaluating the consistency between the model's prediction results and actual conditions through a calibration curve; and evaluating the model's clinical applicability through a clinical decision curve. When the model performance metrics reach a preset standard (e.g., AUC ≥ 0.85), iterative optimization is stopped, ultimately yielding a pre-trained clinical radiomics prediction model.

[0047] In a feasible embodiment, after iteratively optimizing and adjusting the initial prediction model parameters, the training method further includes a process for performance testing and verification of the iteratively optimized model. Specific steps include: acquiring lateral X-ray images of the test knee joint and clinical information of the test set patients; testing the iteratively optimized model using the lateral X-ray images of the test knee joint and clinical information to obtain test prediction results; and adjusting the model's performance based on the test prediction results and the actual disease status of the test set patients. Specifically, firstly, lateral X-ray images of the test knee joint and clinical information of the test set patients are acquired. The inclusion and exclusion criteria for the test set patients are consistent with those for the training set patients in step 210; the acquisition of the lateral X-ray images of the test knee joint follows the aforementioned examination criteria; the test clinical information also includes the patient's age and gender to ensure consistency and comparability between the test data and the training data, avoiding the impact of data differences on the objectivity of the test results. Next, following the same processing procedure as the training set data, the acquired lateral X-ray images of the test knee joint and the test clinical information were processed: First, all lateral X-ray images of the test knee joint were imported into ITKSNAP software, and the patellofemoral joint (PFJ) region of interest was delineated according to the aforementioned definition. Then, the delineation was reviewed by a radiologist of the same qualifications. Subsequently, features were extracted from the delineated region of interest, and the extracted test candidate radiomics features were preprocessed (standardized to a normal distribution of z-scores, truncated within the range of 1%-99%). Then, dimensionality reduction and screening were performed sequentially using the variance thresholding method, univariate selection method, and LASSO algorithm to obtain the target radiomics features of the test set. Finally, the target radiomics features of the test set were fused with the corresponding test clinical information to form the test sample. Then, the constructed test samples are input one by one into the iteratively optimized model. The model performs comprehensive calculations based on the target radiomics features and clinical information of the test set in the test samples, and outputs the test prediction results for each patient in the test set. The test prediction results include the judgment of whether the patient has patellofemoral arthritis (PFOA) and the corresponding probability value and risk level (e.g., low risk, intermediate risk, high risk). Subsequently, the actual disease status of all patients in the test set is collected. The gold standard can be used to confirm whether the patient actually has PFOA and the degree of lesion, using the knee MRI examination results combined with the clinical diagnosis conclusion. The test prediction results output by the model are compared and analyzed one by one with the actual disease status. In this process, the same performance evaluation indicators as in the model training phase can be used, namely, the area under the receiver operating characteristic curve (ROC curve) (AUC), prediction accuracy, sensitivity, and specificity are used to evaluate the predictive discrimination ability and reliability of the model, the calibration curve is used to evaluate the consistency between the test prediction results and the actual disease status, and the clinical decision curve is used to evaluate the application value of the model in actual clinical scenarios.If the evaluation results show that the model's performance metrics (such as AUC, prediction accuracy, etc.) do not meet the preset qualification standards (e.g., AUC < 0.85), then based on the comparison results of the test set, the relevant parameters of the model are fine-tuned. If all performance metrics of the model meet the preset qualification standards and stably meet the requirements for clinical application, then the model performance is confirmed to be up to standard, and the entire model training and testing verification process is completed. Through the above testing steps, the generalization ability and clinical applicability of the iteratively optimized model can be comprehensively verified, ensuring that the model can maintain stable and accurate prediction results when faced with test data from different sources, providing a reliable guarantee for subsequent practical application in PFOA prediction of patients.

[0048] In one feasible embodiment, such as Figure 4 As shown, the execution flow of determining the region of interest in the X-ray image by the model in step 130 may include, but is not limited to, steps 410 to 440.

[0049] Step 410: Determine the superior and posterior margins of the patellofemoral joint in the femoral region on the X-ray image; Step 420: Determine the lower margins of the medial and lateral femoral condyles in the X-ray image; Step 430: Determine the boundaries of the patellar region in the X-ray image; Step 440: Based on the upper edge of the patellofemoral joint, the posterior edge of the patellofemoral joint, the lower edge of the medial femoral condyle articular surface, the lower edge of the lateral femoral condyle articular surface, and the boundary of the patellar region, obtain the region of interest in the X-ray image.

[0050] In a feasible embodiment, the execution process of step 410 includes: determining the first intersection point of the line connecting the midline of the femoral shaft and the upper edge of the anterior and posterior margin of the distal femoral joint in the X-ray image; determining the second intersection point of the anterior and upper margin of the femoral joint in the X-ray image; connecting the first and second intersection points to obtain the upper margin of the patellofemoral joint; determining the third intersection point of the trochlear base line and the projection line of the bottom of the intercondylar fossa in the X-ray image; and connecting the first and third intersection points to obtain the posterior margin of the patellofemoral joint.

[0051] In a feasible embodiment, the execution of step 420 includes: determining the reverse extension line of the intercondylar projection line in the X-ray image; taking the first intersection point of the reverse extension line with the medial femoral condyle articular surface as the lower edge of the medial femoral condyle articular surface; and taking the second intersection point of the reverse extension line with the lateral femoral condyle articular surface as the lower edge of the lateral femoral condyle articular surface.

[0052] It should be noted that the specific implementation process of determining the region of interest in the X-ray image in steps 410 to 440 can be performed with reference to the region of interest delineation specifications described above, including the definition of the range of the patellofemoral joint region of interest (see...). Figure 3The auxiliary usage methods of ITKSNAP, as well as the outline and review process, will not be elaborated here.

[0053] In a feasible embodiment, the process of extracting features from the region of interest to obtain candidate radiomics features specifically includes: for the delineated and confirmed patellofemoral joint region of interest, extracting four types of initial features from the image data of the region, namely, candidate grayscale features reflecting the voxel intensity distribution pattern, candidate shape features describing the geometric shape of the region, candidate texture features reflecting the spatial arrangement characteristics of voxels, and candidate wavelet transform features corresponding to the multi-resolution information obtained by wavelet transform; integrating and summarizing these four types of extracted initial features to form a feature set that has not undergone subsequent dimensionality reduction screening, which is the candidate radiomics feature.

[0054] In one feasible embodiment, such as Figure 5 As shown, the execution flow of step 140, which involves dimensionality reduction and screening of candidate image omics features according to preset rules, to obtain target image omics features may include, but is not limited to, steps 510 to 530.

[0055] Step 510: Use the variance thresholding method to perform preliminary screening of candidate image omics features to obtain variance-effective features; Step 520: Use univariate selection to screen variance-efficient features to obtain association features related to patellofemoral arthritis; Step 530: Use the LASSO algorithm to filter the associated features to obtain the target image omics features.

[0056] In one feasible embodiment, the candidate radiomics features contain a large number of initially extracted features, some of which have small numerical fluctuations (i.e., low variance). These features contribute very little to distinguishing patellofemoral arthritis, or are even invalid information. Step 510 can filter out low-information features with variances below a fixed variance threshold (such as 0.8 as mentioned above) by setting a fixed variance threshold, and retain only features with high variances that can reflect data differences. These retained features are the variance-effective features, thus achieving preliminary simplification of the candidate features.

[0057] In a feasible embodiment, some of the initially screened variance-efficient features may still have a weak association with PFOA. Step 520 analyzes the correlation between each variance-efficient feature and the PFOA disease status using univariate statistical tests (such as t-tests), filters out statistically insignificant features (such as p>0.05), and retains only features with a strong correlation with the onset and severity of PFOA, i.e., features associated with patellofemoral arthritis.

[0058] In a feasible embodiment, the associated features may have problems such as mutual redundancy and multicollinearity, which can easily lead to model overfitting. Step 530 uses the LASSO algorithm to perform regularization screening of the associated features. By introducing a penalty term to compress the weights of redundant features (or even setting some feature weights to zero to remove them), and combining it with 10x cross-validation to avoid the risk of overfitting, the core features that are most valuable and non-redundant for PFOA prediction are finally selected, which are the target image omics features.

[0059] In a feasible embodiment, after obtaining the prediction results, the model can visualize the results using a nomogram, and annotate the contribution of target radiomics features and clinical information to the prediction results. Specifically, the prediction results can be presented intuitively through a visualized nomogram, which transforms the prediction weights corresponding to the target radiomics features and clinical information into clear visual scoring scales. In clinical applications, doctors only need to find the scores corresponding to each feature and clinical information in the nomogram based on the patient's specific data and sum the total scores to quickly retrieve the corresponding prediction results. The prediction results include at least the following: a judgment on whether the patient has patellofemoral arthritis (PFOA) as defined by MRI; the corresponding probability value, which directly reflects the likelihood of the patient having the disease; the risk level classification result, which divides the risk into low, intermediate, and high risk levels based on a preset probability threshold, providing clear guidance for clinical decision-making; and a list of the 2-3 target radiomics features (such as grayscale features and texture features) that contribute the most to the prediction results, helping doctors understand the prediction logic.

[0060] The advantages of this clinical radiomics prediction model are further illustrated below with specific experimental data and results. The specific implementation process and results are as follows: I. Preparation of Experimental Data This embodiment incorporates two datasets for model validation. The first dataset contains 208 patients (corresponding to 249 knee joints), including 66 males (aged 36-83 years) and 142 females (aged 37-87 years). Using MRI results as the gold standard, 182 knee joints (73.09%) were diagnosed with PFOA based on MRI definition, while 69 knee joints (26.91%) were diagnosed without PFOA. This dataset was randomly divided into a training set (198 knee joints, including 145 PFOA cases) and an internal test set (51 knee joints, including 37 PFOA cases) at an 8:2 ratio. The second dataset selected 56 patients as an independent external test set, including 10 males (aged 41-74 years) and 46 females (aged 36-76 years), from whom 60 knee joint imaging data were obtained (42 of which were PFOA cases).

[0061] The results of statistical analysis of the demographic characteristics of the above datasets are shown in Table 2. There was a statistically significant difference in age between the training set and the internal test set (P < 0.05), but no statistically significant difference in gender (P > 0.05). Furthermore, there were no statistically significant differences in age and gender among the three datasets (training set, internal test set, and external test set) (age P = 0.917, gender P = 0.088, both P > 0.05), ensuring the comparability between the datasets and providing a guarantee for the objectivity of model validation.

[0062] Table 2 Information Statistical Results

[0063] II. Radiomics Feature Extraction and Screening For 249 lateral X-ray images of the knee joint in the training set, after completing the region of interest delineation and feature extraction as described above, a total of 1688 initial radiomics features were obtained. Using the three-order feature selection method described above, 15 core radiomics features with the strongest correlation to PFOA were finally selected (see Figures 6(a) to 6(c) and Table 3 for the relevant selection results). Among them, Figure 6(a) uses 10-fold cross-validation to determine the optimal tuning parameter -log(α)=1.8; Figure 6(b) selects 15 radiomics features with non-zero coefficients based on this optimal parameter λ (each colored line represents the changing trend of the corresponding feature coefficient); Figure 6(c) shows the coefficient weight distribution of the 15 core features. Table 3 clarifies the types of the 15 features, including gray-level run-length matrix features (glrlm), gray-level co-occurrence matrix features (glcm), first-order statistical features (firstorder), gray-level dependency matrix features (gldm), and shape features.

[0064] Table 3. 15 radiomics features selected by the LASSO algorithm radiomics features Coefficients values lbp-2D_glrlm_ShortRunEmphasis 0.079462464 lbp-2D_glrlm_ShortRunHighGrayLevelEmphasis 7.04E-16 lbp-2D_glrlm_ShortRunLowGrayLevelEmphasis 1.79E-17 wavelet-LLL_glcm_ClusterProminence -0.031327492 original_firstorder_Uniformity 0.008365147 wavelet-HHL_gldm_DependenceNonUniformity 0.035596597 square_glcm_ClusterShade -0.033399702 wavelet-HLH_glszm_LargeAreaLowGrayLevelEmphasis 0.025044065 wavelet-LLL_glcm_Imc2 -0.01587826 original_shape_MinorAxisLength 0.058120961 square_gldm_GrayLevelNonUniformity 0.01027014 wavelet-LLL_firstorder_Kurtosis 0.070759746 wavelet-HHH_glcm_SumSquares 0.02107181 wavelet-LLH_glcm_Imc2 0.032830663 wavelet-HHL_firstorder_Median 0.004529677 III. Model Construction and Performance Evaluation Based on 15 selected core radiomics features, combined with clinical information such as patient age and gender, three types of models were constructed using the logistic regression (LR) algorithm: a clinical factor-only model, a radiomics-only model, and a comprehensive model combining radiomics features with clinical information (i.e., the clinical radiomics prediction model described in this protocol). The predictive performance of the three types of models on the training set, internal test set, and external test set was comprehensively evaluated using the following metrics: 1. The evaluation indicators for predicting radiological PFOA based on MRI definition are detailed in Table 4. Among them, Model 1 is the clinical factor model; Model 2 is the radiomics model; and Model 3 is the comprehensive model. ROC curves are shown in Figures 7(a) to 7(c).

[0065] Table 4. Predictive performance of the three models on the training set, internal test set, and external test set.

[0066] The results show: In the internal test set, the AUC value of the integrated model was 0.865, which was higher than that of the clinical factor-only model (AUC=0.752) and the radiomics-only model (AUC=0.764). The Delong test confirmed that the predictive power of the integrated model was statistically different from that of the latter two. In the external test set, the AUC value of the integrated model was 0.741, which was higher than that of the clinical factor-only model (AUC=0.689) and similar to that of the radiomics-only model (AUC=0.742). The Delong test showed that there was no statistically significant difference in predictive power between the two models (P=0.676).

[0067] 2. To evaluate the predictive consistency of the integrated model, the LR integrated model was visualized by plotting a Nomogram (see...). Figure 8 The relationship between predicted probabilities and actual observed probabilities was analyzed using the Hosmer-Lemeshow test and calibration curves (see Figures 9(a) to 9(c)). The results showed that the Hosmer-Lemeshow test p-values ​​for the internal test set and the external test set were 0.173 and 0.826, respectively (both p > 0.05), indicating that the model predictions were in good agreement with the actual disease status of patients.

[0068] 3. Referring to Figures 10(a) to 10(c), the DCA curves (i.e., clinical decision curves) of the three datasets show that the Nomogram based on the integrated model has a significantly better clinical net benefit than the individual clinical factor model and radiomics model, indicating that the model has higher application value in actual clinical scenarios.

[0069] In summary, through multi-dataset validation and multi-dimensional performance evaluation, it can be seen that compared with individual clinical factor models and individual radiomics models, the clinical radiomics prediction model constructed based on the LR algorithm in this scheme has better predictive efficacy and clinical net benefit in predicting PFOA based on MRI definition, and has good stability and consistency, making it more suitable for clinical practice in early screening and risk assessment of PFOA.

[0070] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting patellofemoral arthritis based on lateral knee radiomics, characterized in that, The method includes: Obtain lateral X-ray images and clinical information of the knee joint of the patient to be tested; The X-ray images and clinical information are input into a pre-trained clinical radiomics prediction model; The model is used to determine the region of interest in the X-ray image, and features are extracted from the region of interest to obtain candidate radiomics features. The model filters the candidate imagemic features according to a preset dimensionality reduction rule to obtain target imagemic features. The types of target imagemic features include at least grayscale features, shape features, texture features, and wavelet transform features. The grayscale features are used to describe the distribution pattern and statistical characteristics of voxel intensity within the region of interest. The shape features are used to describe the geometric characteristics of the region of interest. The texture features are used to describe the spatial distribution information of voxels within the region of interest. The wavelet transform features are used to obtain multi-resolution image description information by performing wavelet transform on the image of the region of interest. The model makes predictions based on the various types of features in the target radiomics and the clinical information, and the prediction results are obtained.

2. The method for predicting patellofemoral arthritis according to claim 1, characterized in that, Determining the region of interest in the X-ray image includes: Determine the superior and posterior margins of the patellofemoral joint in the femoral region of the X-ray image; Determine the lower margins of the medial and lateral femoral condyles in the X-ray image; Determine the boundaries of the patellar region in the X-ray image; The region of interest in the X-ray image is obtained based on the upper edge of the patellofemoral joint, the posterior edge of the patellofemoral joint, the lower edge of the medial femoral condyle articular surface, the lower edge of the lateral femoral condyle articular surface, and the boundary of the patellar region.

3. The method for predicting patellofemoral arthritis according to claim 2, characterized in that, Determining the superior and posterior margins of the patellofemoral joint in the femoral region of the X-ray image includes: In the X-ray image, determine the first intersection point of the line connecting the midline of the femoral shaft and the superior position of the anterior and posterior margins of the distal femoral articular surface. Determine the second intersection point of the anterior and superior margins of the femoral joint in the X-ray image; Connect the first intersection point and the second intersection point to obtain the upper edge of the patellofemoral joint; Determine the third intersection point of the trochlear baseline and the projection line at the bottom of the intercondylar fossa in the X-ray image; Connect the first intersection point and the third intersection point to obtain the posterior edge of the patellofemoral joint.

4. The method for predicting patellofemoral arthritis according to claim 2, characterized in that, Determining the lower edge of the medial femoral condyle articular surface and the lower edge of the lateral femoral condyle articular surface in the X-ray image includes: Determine the reverse extension of the intercondylar concave projection line in the X-ray image; The first intersection point of the reverse extension line and the medial femoral condyle articular surface is taken as the lower edge of the medial femoral condyle articular surface; The second intersection point of the reverse extension line and the lateral femoral condyle articular surface is taken as the lower edge of the lateral femoral condyle articular surface.

5. The method for predicting patellofemoral arthritis according to claim 1, characterized in that, The step of extracting features from the region of interest to obtain candidate image omics features includes: Extract the corresponding candidate grayscale features, candidate shape features, candidate texture features, and candidate wavelet transform features from the region of interest; Candidate image omics features are obtained based on the candidate grayscale features, the candidate shape features, the candidate texture features, and the candidate wavelet transform features.

6. The method for predicting patellofemoral arthritis according to claim 1, characterized in that, The step of filtering the candidate radiomics features according to a preset dimensionality reduction rule to obtain the target radiomics features includes: The variance thresholding method was used to initially screen candidate radiomics features to obtain variance-effective features; The variance-efficient features were screened using a univariate selection method to obtain the association features related to patellofemoral arthritis; The LASSO algorithm was used to filter the associated features to obtain the target image omics features.

7. The method for predicting patellofemoral arthritis according to claim 1, characterized in that, The training steps for the pre-trained clinical radiomics prediction model include: Acquire lateral X-ray images of the knee joints and clinical information of patients in the training set; Regions of interest are delineated from the X-ray images in the training set, and features are extracted from the delineated regions of interest to obtain the training candidate radiomics features. The training candidate image omics features are subjected to dimensionality reduction and screening to obtain the training set target image omics features; The target radiomics features of the training set are fused with corresponding clinical information to construct training samples; Based on the logistic regression algorithm, the initial prediction model is trained using the training samples, and the parameters of the initial prediction model are adjusted through iterative optimization to obtain a pre-trained clinical radiomics prediction model.

8. The method for predicting patellofemoral arthritis according to claim 6, characterized in that, After obtaining the prediction result, the method further includes: The model visualizes the prediction results using a nomogram, and the contribution of target radiomics features and clinical information to the prediction results is marked in the nomogram.

9. The method for predicting patellofemoral arthritis according to claim 7, characterized in that, After adjusting the initial prediction model parameters through iterative optimization, the method further includes: Acquire lateral X-ray images of the test knee joint and clinical information of the test patients; The model, which has been iteratively optimized, is tested using the lateral X-ray image of the tested knee joint and the clinical information of the test to obtain the test prediction results; The model's performance is adjusted based on the test prediction results and the actual disease status of patients in the test set.

10. The method for predicting patellofemoral arthritis according to claim 1 or 7, characterized in that, The clinical information includes the age and gender of the patient to be tested.

Images