Construction of a radiomics-based diagnostic nomogram for patellofemoral osteoarthritis—using lateral knee X-ray images from a South China population

Introduction

Patellofemoral osteoarthritis (PFOA) is an important but often underrecognized subtype of knee osteoarthritis (1-3). It is closely associated with anterior knee pain, difficulty with stair climbing and squatting, and impaired daily function. PFOA differs from tibiofemoral-predominant disease in its biomechanical profile and may progress under the combined influence of patellar maltracking, altered trochlear geometry, extensor mechanism imbalance, obesity, and age-related degeneration (1,2,4). However, in routine radiographic interpretation, attention is often focused on the tibiofemoral compartment, whereas patellofemoral structural changes may be less systematically assessed.

Plain radiography remains the first-line imaging modality for knee osteoarthritis because it is inexpensive, widely available, and routinely used in clinical practice. Lateral knee radiographs can demonstrate patellofemoral joint-space narrowing, osteophytes, sclerosis, and cystic changes (5-9). Nevertheless, visual assessment of PFOA on radiographs may be affected by projectional overlap, patient positioning, image quality, and reader experience. Therefore, quantitative tools that can extract additional information from standard radiographs may help improve the consistency and diagnostic yield of PFOA assessment without increasing imaging burden.

Radiomics provides a practical approach for converting medical images into quantitative features that describe intensity distribution, texture heterogeneity, and spatial patterns. In osteoarthritis research, radiomics and machine-learning methods have been applied to radiographs and magnetic resonance imaging for disease classification, progression prediction, and risk stratification (10-14). These studies suggest that routine imaging may contain clinically relevant information beyond conventional visual scoring (10,11). Compared with deep learning, radiomics-based models are relatively interpretable and can be readily combined with clinical variables in a nomogram format (15-17).

Although several studies have explored radiomics or texture analysis in knee osteoarthritis, relatively few have focused specifically on PFOA using lateral knee radiographs (18). In addition, many existing models were developed using public datasets or single-source cohorts, and their performance in Chinese clinical populations remains uncertain (19-24). Multicenter evaluation is important because radiographic protocols, patient characteristics, image quality, and disease spectrum may vary across institutions and influence model generalizability (8,16,17,22).

In this multicenter retrospective study, we developed a radiomics-based diagnostic model for PFOA using lateral knee radiographs from a primary cohort and further evaluated its performance in an independent external test set (4,19,20,25,26). We compared logistic regression (LR), k-nearest neighbors (KNN), and random forest (RF) models, selected the best-performing radiomics classifier, and constructed a radiomics-clinical nomogram incorporating age and sex (12,13,15-17,23,27-30). This study aimed to develop an objective and clinically practical tool to assist radiographic identification of PFOA in routine practice (13,23,25). We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0849/rc).

Methods

Study population

This retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Ethics Committee of The Fifth Affiliated Hospital of Sun Yat-sen University (No. ZDWY[2022] Lunzi No. K163-1). The requirement for informed consent was waived because of the retrospective design.

The primary cohort was collected from The Fifth Affiliated Hospital of Sun Yat-sen University, including 1,742 patients with 2,197 knees who underwent knee radiography between July 2017 and July 2020. After excluding 13 patients with 15 knees because of poor positioning or unqualified image quality, 1,729 patients with 2,182 knees were included for model development and internal testing. The primary cohort was randomly divided into a training set and an internal test set at a ratio of 7:3. An independent external test set was retrospectively collected from another participating center—The Fourth People’s Hospital of Guiyang—which used the same eligibility criteria and outcome definition to evaluate model generalizability.

The inclusion criteria were as follows: (I) adults aged 18 years or older with closed epiphyses; (II) no definite history of inflammatory arthritis, such as rheumatoid arthritis, gout, or tuberculosis; (III) no bone tumors; (IV) no history of knee fracture or metallic internal fixation; and (V) no congenital deformities.

The exclusion criteria were poor positioning and unqualified image quality. The participant screening process, dataset allocation, and external testing workflow are illustrated in Figure 1.

Figure 1 Study flowchart and workflow of radiomics-based model development, internal testing, and external testing for the diagnosis of PFOA using lateral knee radiographs. AUC, area under the curve; DCA, decision curve analysis; KNN, k-nearest neighbors; LR, logistic regression; PFOA, patellofemoral osteoarthritis; RF, random forest; ROC, receiver operating characteristic.

PFOA was defined on lateral knee radiographs according to the Framingham Osteoarthritis Study criteria as osteophytes of grade 2 or higher, or joint-space narrowing of grade 2 or higher accompanied by any osteophyte, sclerosis, or cyst in the same region (5-7). PFOA status was independently determined by two musculoskeletal radiologists with 3 and 6 years of experience, respectively. Disagreements were adjudicated by a senior chief radiologist with more than 20 years of musculoskeletal imaging experience. The readers who assessed outcome status were blinded to the radiomic features and model outputs.

Radiographic examination

The patient is placed in a lateral decubitus position for a lateral X-ray of the knee joint, with the side being examined closer to the detector. The contralateral lower limb was extended anteriorly and superiorly. The examined knee was flexed to approximately 20°–35°, and the lateral aspect of the knee was placed close to the detector. The inferior pole of the patella was centered within the irradiation field, and the central beam was directed vertically toward the proximal tibia. Superimposition of the femoral condyles was used to ensure an acceptable lateral projection. This standardized acquisition protocol was intended to reduce positional variability in subsequent region of interest (ROI) analysis. Because the study was retrospective and images were collected from routine clinical practice, radiographs were acquired using different digital radiography systems. To minimize acquisition-related variability, all examinations followed the standardized protocols across participating centers and patient-positioning procedure.

ROI delineation and image preprocessing

All anonymized imaging data were imported into the Radcloud platform (Huiying Medical Technology Co., Ltd., Beijing, China). ROIs were manually delineated by a radiologist with 3 years of musculoskeletal imaging experience under the supervision of another radiologist with 6 years of experience, and all delineations were reviewed by a chief radiologist with more than 20 years of experience. The ROI was defined as a rectangular box spanning the patellofemoral and adjacent periarticular structures visible on the lateral radiograph. Specifically, the superior boundary was the highest point of the distal femoral articular surface, the inferior boundary was the midpoint of the tibial tuberosity, the anterior boundary corresponded to the anterior margin of the quadriceps tendon insertion, and the posterior boundary was defined by the fibular neck. Figure 2 shows a schematic illustration of the ROI delineation.

Figure 2 Schematic illustration of region of interest delineation on lateral knee radiographs.

The use of a relatively broad rectangular region rather than a narrowly confined patellar subregion was based on the hypothesis that PFOA is not limited to isolated cartilage loss or osteophyte formation but may also be associated with image-texture changes in adjacent bone and periarticular structures. Because the original Digital Imaging and Communications in Medicine (DICOM) images were acquired on different imaging devices, preprocessing standardization was performed before model development to reduce potential scanner-related variability. Specifically, all radiomic features were standardized to z scores and truncated to the 1st–99th percentile range to reduce the influence of inter-scanner heterogeneity, feature-scale differences, and extreme values, thereby improving feature comparability across datasets.

Radiomic feature extraction and selection

The 2,182 knees were randomly divided into a training set and an internal test set at a ratio of 7:3 before feature selection and model development. A total of 1,409 radiomic features were extracted from the delineated ROIs using the Radcloud platform. These features included first-order statistics and multiple classes of texture descriptors after different image transformations, thereby capturing intensity distribution, spatial heterogeneity, and higher-order filtered patterns.

Feature reduction was carried out in a multistep manner to reduce redundancy and minimize overfitting. First, features with a variance lower than 0.8 were removed using the variance threshold method. Second, univariate feature selection was performed to exclude features that were not significantly associated with the outcome (P>0.05). Finally, least absolute shrinkage and selection operator (LASSO) regression was applied to all retained features. Ten-fold cross-validation was used to select the tuning parameter and identify the subset of features most strongly associated with PFOA. This procedure resulted in 25 radiomic features for subsequent model construction. The radiomics score (Radscore) was calculated as a weighted linear combination of the radiomics features selected by the LASSO regression model. The retained radiomics features and their corresponding coefficients are presented in Table 1. Because the model intercept was set to zero during fitting, the Radscore was calculated according to the following equation:

RadScore=∑i=1nβiXi[1]

where (n) denotes the total number of radiomics features retained by the LASSO model, (Xi) represents the value of the (i)-th radiomics feature, and (βi) represents the corresponding LASSO-derived coefficient. Positive coefficients contribute positively to the Radscore, whereas negative coefficients contribute inversely to the final score.

Model construction and evaluation

Three machine-learning algorithms were used to construct radiomics models based on the selected features: LR, KNN, and RF. These algorithms were chosen because they represent complementary modeling paradigms: a conventional interpretable linear classifier, a distance-based nonparametric method, and an ensemble tree-based learner. This design allowed us to compare whether the relationship between the selected radiomic features and PFOA was best captured by a relatively simple linear model or by more flexible nonlinear approaches.

Model discrimination was assessed by receiver operating characteristic (ROC) analysis, and the area under the curve (AUC), accuracy, sensitivity, and specificity were calculated. Pairwise DeLong tests were performed to compare the discriminative performance of the three candidate models and to identify the optimal radiomics classifier.

Clinical factors were then combined with the best-performing radiomics model to develop an integrated nomogram. Univariate and multivariate LR analyses were used to identify clinically relevant predictors. The final nomogram incorporated the radiomics model together with age and sex. Calibration curves, the Hosmer-Lemeshow test, and decision curve analysis (DCA) were used to evaluate model calibration and potential clinical utility. The overall workflow of radiomics-based model development and validation is shown in Figure 3.

Figure 3 Workflow of the radiomics-based diagnostic model for patellofemoral osteoarthritis using lateral knee radiographs, including study population selection, ROI delineation, feature extraction and selection, model development, and nomogram construction with validation. AUC, area under the curve; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; RF, random forest; ROC, receiver operating characteristic; ROI, region of interest.

External testing

An independent external test set was retrospectively collected from another participating center using the same inclusion and exclusion criteria as the primary cohort to further evaluate model generalizability. Radiomic feature extraction, preprocessing, and model evaluation were performed using the same workflow as that used in the training and internal test sets. The diagnostic performances of the LR, KNN, and RF models were further evaluated in the external test set using ROC curve analysis.

Statistical analysis

Statistical analyses were performed using SPSS (version 26.0; IBM) and R software (version 4.1.2; R Foundation for Statistical Computing, Vienna, Austria). Categorical variables were compared using the chi-square test or Fisher’s exact test. Continuous variables were analyzed using the Student’s t-test or the Mann-Whitney U test, as appropriate. The Kolmogorov-Smirnov test was used to assess normality. Normally distributed continuous variables were expressed as mean ± standard deviation and analyzed using analysis of variance (ANOVA), whereas non-normally distributed variables were expressed as median with interquartile range (IQR) and analyzed using the Mann-Whitney U test. Categorical variables were presented as frequencies. A two-sided P<0.05 was considered statistically significant. No missing data were observed for the variables used in model development and evaluation. The DeLong test was used to compare AUCs between ROC curves. Patients and the public were not involved in the design, conduct, reporting, or dissemination of this research.

Results

Patient characteristics

A total of 1,742 patients with 2,197 knees were initially assessed. After excluding 13 patients with 15 knees because of poor positioning or unqualified image quality, 1,729 patients with 2,182 knees were included in the primary cohort. Among them, 1,133 knees were classified as PFOA and 1,049 as non-PFOA. The primary cohort was randomly divided into a training set and an internal test set at a ratio of 7:3. The training set included 1,527 knees, and the internal test set included 655 knees. An independent external test set was additionally collected, including 367 patients with 472 knees. Of these, 269 knees were classified as PFOA and 203 as non-PFOA. The baseline characteristics of the training, internal test, and external test sets are shown in Table 2. In all three sets, the PFOA group was significantly older than the non-PFOA group, and sex distribution differed significantly between groups. In the external test set, the median age was 65.00 years in the PFOA group and 55.00 years in the non-PFOA group (P<0.001). The proportion of female knees was also higher in the PFOA group than in the non-PFOA group [217/269 (80.7%) vs. 136/203 (67.0%), P=0.001]. No significant differences in age or sex distribution were observed between the training and internal test sets.

Radiomic feature selection

A total of 1,409 radiomic features were extracted from the training set using the Radcloud platform. Following dimensionality reduction, 25 optimal radiomic features most relevant to PFOA were selected (Figure 4, Table 1).

Figure 4 Radiomic feature selection using the LASSO regression model. (A) Ten-fold cross-validation was used to determine the optimal tuning parameter in the LASSO model. (B) Twenty-five radiomic features with nonzero coefficients were selected at the optimal value of λ; each colored line represents the coefficient trajectory of an individual feature. (C) Coefficient weights of the 25 selected radiomic features. CV, cross-validation; LASSO, least absolute shrinkage and selection operator.

Performance of radiomics models

LR, KNN, and RF radiomics models were constructed using the 25 selected radiomic features derived from lateral knee radiographs. The diagnostic performance of these three models is summarized in Table 3, and the ROC curves are presented in Figure 5. Pairwise DeLong comparisons showed that, in the internal test set, the LR model achieved the best performance, with an AUC of 0.773, significantly outperforming both the KNN and RF models (Table 4).

Figure 5 Receiver operating characteristic curves of different radiomics models in the training set (A) and internal test set (B). AUC, area under the curve.

Nomogram construction and validation

The LR algorithm was selected for subsequent nomogram development because it showed the best overall performance. Univariate and multivariate LR analyses demonstrated that age was an independent risk factor for PFOA (Table 5). Although sex was not an independent risk factor in multivariate analysis, it has previously been reported as a risk factor for PFOA and was therefore retained in the final model together with age for clinical relevance. A visualized nomogram was subsequently constructed (Figure 6).

Figure 6 Nomogram for the diagnosis of PFOA. PFOA, patellofemoral osteoarthritis.

The calibration performance of the nomogram was assessed using the Hosmer-Lemeshow test and calibration curves (Figure 7). The Hosmer-Lemeshow test yielded P=0.827 in the training set and P=0.115 in the internal test set, indicating good model fit. DCA showed that when the threshold probability ranged from 0.1 to 1.0, the LR-based nomogram provided a favorable net benefit for the diagnosis of PFOA (Figure 8). Moreover, the nomogram demonstrated better clinical applicability than the radiomics-only model. Overall, the nomogram showed good calibration, discrimination, and potential clinical utility for identifying PFOA. In the training and internal test sets, the AUCs of the nomogram were 0.876 and 0.842, respectively (Figure 7, Table 6).

Figure 7 Calibration curves of the nomogram in the training set (A) and internal test set (B).

Figure 8 Decision curve analysis of the logistic regression radiomics model and the nomogram in the training set (A) and internal test set (B).

External testing

An independent external test set was used to further evaluate the robustness and generalizability of the radiomics models. As shown in Table 3 and Figure 9, the LR model achieved the highest AUC in the external test set among the three candidate models, with an AUC of 0.751. The KNN and RF models achieved AUCs of 0.685 and 0.696, respectively. Although the diagnostic performance was slightly lower than that observed in the internal test set, the LR model still demonstrated acceptable discrimination in the independent external test set.

Figure 9 Receiver operating characteristic curves of the logistic regression, k-nearest neighbors, and random forest radiomics models in the external test set. AUC, area under the curve.

Discussion

In this study, we developed and internally evaluated a radiomics-based diagnostic nomogram for PFOA using lateral knee radiographs from a South China population. Among the three candidate radiomics models, LR showed the best performance in the independent internal test set, and the addition of age and sex further improved discrimination, yielding a test-set AUC of 0.842. The calibration and decision curve analyses suggested that the integrated nomogram was not only statistically robust within the current dataset but also potentially useful for clinical decision support. These findings support the broader view that routine radiographs contain quantitative information beyond what is captured by visual inspection alone and that this information can be leveraged to improve recognition of underdiagnosed compartment-specific osteoarthritis phenotypes.

Our findings fit within a growing body of work showing that artificial intelligence and quantitative imaging can enhance osteoarthritis assessment across multiple modalities and clinical tasks. Deep-learning models based on radiographs have been used to automatically grade knee osteoarthritis severity, identify future total knee replacement, and predict symptom trajectories (10-13,28,29). Magnetic resonance imaging (MRI)-based studies have extended this concept to the prediction of structural progression and the quantification of tissue-specific abnormalities in cartilage, subchondral bone, and intra-articular fat pads (15,20-24). At a broader level, recent reviews indicate that osteoarthritis imaging research continues to move toward integrative pipelines in which radiographs, MRI, handcrafted radiomic features, and learned deep features are increasingly combined with clinical variables to improve diagnostic and prognostic performance (16,17,27,30). Although deep-learning approaches can achieve excellent predictive performance through automated feature learning, they generally require substantially larger datasets and are often limited by reduced interpretability. In contrast, radiomics-based approaches provide quantitatively defined imaging features that can be directly linked to underlying structural alterations and clinical characteristics. Against this background, our study contributes a focused model for PFOA diagnosis using routinely acquired lateral knee radiographs. Compared with many previous studies that primarily focused on tibiofemoral osteoarthritis or MRI-based assessment, our framework specifically targets PFOA and integrates radiomics features with clinical variables. In addition, unlike many deep-learning approaches that function as relatively opaque prediction systems, the radiomics-based nomogram provides a more interpretable framework that may facilitate clinical implementation and decision support.

One of the most relevant comparators for the present work is the study by Bayramoglu et al., who reported that patellar texture features extracted from lateral radiographs could identify PFOA and that models combining texture and clinical features performed better than those based on clinical information alone (25). Our study similarly found that the radiomics-clinical nomogram outperformed the radiomics-only model, reinforcing the idea that image-derived structural signatures and demographic risk factors carry complementary information. At the same time, our diagnostic performance was somewhat lower than that reported in some previous studies. Several explanations are plausible.

This study has several limitations. First, although an independent external test set from another center was included, the retrospective design may still introduce selection bias, and prospective validation in broader clinical settings is needed (18,24). Second, only age and sex were incorporated as clinical predictors because other potentially relevant variables, such as body mass index, pain scores, alignment parameters, and semiquantitative radiographic grades, were unavailable or incomplete (2,4,21,25,31). Third, the ROI was manually delineated as a broad rectangular region, which may have introduced redundant information and may affect feature reproducibility; future studies should evaluate automated segmentation and formal interobserver and intraobserver reproducibility (13,15,23,29). Fourth, bilateral knees from the same patient were analyzed as separate samples, and potential within-subject correlation could not be fully excluded. Finally, external testing was performed for the radiomics models, whereas further external assessment of the final nomogram, including calibration and clinical utility, would strengthen the evidence for clinical application (13,27).

Despite these limitations, the study has several strengths. First, it addresses a clinically relevant but underexplored phenotype. Much of the osteoarthritis imaging literature focuses on whole-knee Kellgren-Lawrence grading, total knee replacement, or composite progression endpoints, whereas relatively fewer studies specifically target PFOA on plain radiographs. Because patellofemoral disease may contribute disproportionately to anterior knee pain and functional limitation, a dedicated PFOA-oriented model is clinically justified. Second, the model is built on lateral radiographs, which are inexpensive and routinely available in real-world practice. This increases the likelihood that a future validated version could be translated into workflow. Third, the final model is visualized as a nomogram, which improves interpretability compared with a purely black-box system and may facilitate clinical adoption in settings where transparent decision support is preferred.

The current findings also have implications for the design of future musculoskeletal imaging models. Recent studies suggest that radiography-based machine-learning models can remain highly competitive even in an era dominated by MRI and deep learning, particularly when the target use case is screening or routine outpatient assessment (11,13,27,29,30). At the same time, MRI-based radiomics and deep-learning studies consistently show that tissue-specific imaging biomarkers may detect early or pre-radiographic disease-related changes (20-22,24). Rather than viewing radiography-based and MRI-based models as competing approaches, they may be more usefully conceptualized as complementary tools serving different points in the care pathway. In this context, a radiomics-based nomogram for lateral radiographs could be used as an accessible first-line tool to flag knees with probable PFOA, whereas MRI-based quantitative models might be reserved for more detailed phenotyping or progression-risk stratification.

Several further steps are needed before the present model can be considered ready for clinical deployment. Although an independent external test set was performed in the present study, prospective multicenter studies in independent Chinese and non-Chinese populations remain essential to further evaluate model robustness and clinical generalizability. In addition, future investigations should assess the integration of the nomogram into routine radiology workflows and determine whether its use improves diagnostic consistency, clinical decision-making, and workflow efficiency in real-world practice. Future studies should also test whether automated or semiautomated ROI delineation can reduce workload and improve reproducibility. In addition, broader predictor sets should be explored, including body mass index, symptom scores, conventional radiographic grades, alignment metrics, and possibly biochemical or longitudinal imaging markers. It will also be important to assess whether knee-level analysis should be supplemented by modeling strategies that account for within-patient correlation when bilateral knees are available. Finally, reporting the full model equation or radiomics score in supplementary material would enhance reproducibility and enable third-party evaluation, an important consideration for prediction-model transparency.

In summary, our results indicate that radiomic features extracted from lateral knee radiographs can support the diagnosis of PFOA and that combining these features with simple clinical factors improves predictive performance. While the nomogram does not replace expert radiological assessment, it may serve as a useful adjunct in settings where PFOA is easily overlooked on routine reading. More broadly, the study supports continued development of quantitative imaging tools tailored not only to whole-knee osteoarthritis but also to clinically meaningful compartment-specific phenotypes.

Conclusions

In this multicenter retrospective study, we developed and evaluated a PFOA radiomics diagnostic map based on lateral knee X-rays using clinical and imaging data from a population in South China. The model demonstrated good discrimination, calibration, and potential clinical utility, with improved performance after integration of radiomic features with age and sex. These findings suggest that radiomics may provide objective imaging biomarkers for PFOA on routine radiographs. Further multicenter external test set and methodological refinement are warranted before clinical implementation.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0849/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0849/dss

Funding: This study was supported by the National Natural Science Foundation of China (No. 82502304) and the Guangdong Medical Science and Technology Research Fund Project (No. B2025313).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0849/coif). All authors report that this study was supported by the Natural Science Foundation of China (No. 82502304) and the Guangdong Medical Science and Technology Research Fund Project (No. B2025313). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Ethics Committee of The Fifth Affiliated Hospital of Sun Yat-sen University (No. ZDWY[2022] Lunzi No. K163-1). The requirement for informed consent was waived because of the retrospective design.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Hinman RS, Lentzos J, Vicenzino B, Crossley KM. Is Patellofemoral Osteoarthritis Common in Middle‐Aged People With Chronic Patellofemoral Pain? Arthritis Care Res 2014;66:1252-7.
2. van Middelkoop M, Bennell KL, Callaghan MJ, Collins NJ, Conaghan PG, Crossley KM, Eijkenboom JJFA, van der Heijden RA, Hinman RS, Hunter DJ, Meuffels DE, Mills K, Oei EHG, Runhaar J, Schiphof D, Stefanik JJ, Bierma-Zeinstra SMA. International patellofemoral osteoarthritis consortium: Consensus statement on the diagnosis, burden, outcome measures, prognosis, risk factors and treatment. Semin Arthritis Rheum 2018;47:666-75. [Crossref] [PubMed]
3. Rathleff MS, Rathleff CR, Olesen JL, Rasmussen S, Roos EM. Is Knee Pain During Adolescence a Self-limiting Condition? Prognosis of Patellofemoral Pain and Other Types of Knee Pain. Am J Sports Med 2016;44:1165-71.
4. Lankhorst NE, Damen J, Oei EH, Verhaar JAN, Kloppenburg M, Bierma-Zeinstra SMA, van Middelkoop M. Incidence, prevalence, natural course and prognosis of patellofemoral osteoarthritis: the Cohort Hip and Cohort Knee study. Osteoarthritis Cartilage 2017;25:647-53. [Crossref] [PubMed]
5. Duncan RC, Hay EM, Saklatvala J, Croft PR. Prevalence of radiographic osteoarthritis--it all depends on your point of view. Rheumatology (Oxford) 2006;45:757-60. [Crossref] [PubMed]
6. Jones AC, Ledingham J, McAlindon T, Regan M, Hart D, MacMillan PJ, Doherty M. Radiographic assessment of patellofemoral osteoarthritis. Ann Rheum Dis 1993;52:655-8. [Crossref] [PubMed]
7. Felson DT, McAlindon TE, Anderson JJ, Naimark A, Weissman BW, Aliabadi P, Evans S, Levy D, LaValley MP. Defining radiographic osteoarthritis for the whole knee. Osteoarthritis Cartilage 1997;5:241-50. [Crossref] [PubMed]
8. Lyu L, Ren J, Lu W, Zhong J, Song Y, Li Y, Yao W. A machine learning-based radiomics approach for differentiating patellofemoral osteoarthritis from non-patellofemoral osteoarthritis using Q-Dixon MRI. Front Sports Act Living 2025;7:1535519. [Crossref] [PubMed]
9. Kornaat PR, Bloem JL, Ceulemans RY, Riyazi N, Rosendaal FR, Nelissen RG, Carter WO, Hellio Le Graverand MP, Kloppenburg M. Osteoarthritis of the knee: association between clinical features and MR imaging findings. Radiology 2006;239:811-7. [Crossref] [PubMed]
10. Li W, Xiao Z, Liu J, Feng J, Zhu D, Liao J, Yu W, Qian B, Chen X, Fang Y, Li S. Deep learning-assisted knee osteoarthritis automatic grading on plain radiographs: the value of multiview X-ray images and prior knowledge. Quant Imaging Med Surg 2023;13:3587-601. [Crossref] [PubMed]
11. Leung K, Zhang B, Tan J, Shen Y, Geras KJ, Babb JS, Cho K, Chang G, Deniz CM. Prediction of Total Knee Replacement and Diagnosis of Osteoarthritis by Using Deep Learning on Knee Radiographs: Data from the Osteoarthritis Initiative. Radiology 2020;296:584-93. [Crossref] [PubMed]
12. Tiulpin A, Saarakkala S. Automatic Grading of Individual Knee Osteoarthritis Features in Plain Radiographs Using Deep Convolutional Neural Networks. Diagnostics (Basel) 2020;10:932. [Crossref] [PubMed]
13. Schiratti JB, Dubois R, Herent P, Cahané D, Dachary J, Clozel T, Wainrib G, Keime-Guibert F, Lalande A, Pueyo M, Guillier R, Gabarroca C, Moingeon P. A deep learning method for predicting knee osteoarthritis radiographic progression from MRI. Arthritis Res Ther 2021;23:262. [Crossref] [PubMed]
14. Tan JM, Menz HB, Munteanu SE, Collins NJ, Hart HF, Donnar JW, Cleary G, O'Sullivan IC, Maclachlan LR, Derham CL, Crossley KM. Can radiographic patellofemoral osteoarthritis be diagnosed using clinical assessments? Musculoskeletal Care 2020;18:467-76. [Crossref] [PubMed]
15. Guo J, Yan P, Qin Y, Liu M, Ma Y, Li J, Wang R, Luo H, Lv S. Automated measurement and grading of knee cartilage thickness: a deep learning-based approach. Front Med (Lausanne) 2024;11:1337993. [Crossref] [PubMed]
16. Guo J, Yan P, Luo H, Ma Y, Jiang Y, Ju C, Chen W, Liu M, Lv S, Qin Y. Predicting joint space changes in knee osteoarthritis over 6 years: a combined model of TransUNet and XGBoost. Quant Imaging Med Surg 2025;15:1396-410. [Crossref] [PubMed]
17. Zhao H, Ou L, Zhang Z, Zhang L, Liu K, Kuang J. The value of deep learning-based X-ray techniques in detecting and classifying K-L grades of knee osteoarthritis: a systematic review and meta-analysis. Eur Radiol 2025;35:327-40. [Crossref] [PubMed]
18. Mayerhoefer ME, Materka A, Langs G, Häggström I, Szczypiński P, Gibbs P, Cook G. Introduction to Radiomics. J Nucl Med 2020;61:488-95. [Crossref] [PubMed]
19. Zhang J, Jiang T, Chan LC, Lau SH, Wang W, Teng X, Chan PK, Cai J, Wen C. Radiomics analysis of patellofemoral joint improves knee replacement risk prediction: Data from the Multicenter Osteoarthritis Study (MOST). Osteoarthr Cartil Open 2024;6:100448. [Crossref] [PubMed]
20. Jiang H, Peng Y, Qin SY, Chen C, Pu Y, Liang R, Chen Y, Zhang XM, Sun YB, Zuo HD. MRI-Based Radiomics and Delta-Radiomics Models of the Patella Predict the Radiographic Progression of Osteoarthritis: Data From the FNIH OA Biomarkers Consortium. Acad Radiol 2024;31:1508-17. [Crossref] [PubMed]
21. Liu Q, Chu H, LaValley MP, Hunter DJ, Zhang H, Tao L, Zhan S, Lin J, Zhang Y. Prediction models for the risk of total knee replacement: development and validation using data from multicentre cohort studies. Lancet Rheumatol 2022;4:e125-34. [Crossref] [PubMed]
22. Yu K, Ying J, Zhao T, Lei L, Zhong L, Hu J, Zhou JW, Huang C, Zhang X. Prediction model for knee osteoarthritis using magnetic resonance-based radiomic features from the infrapatellar fat pad: data from the osteoarthritis initiative. Quant Imaging Med Surg 2023;13:352-69. [Crossref] [PubMed]
23. Hu J, Zheng C, Yu Q, Zhong L, Yu K, Chen Y, Wang Z, Zhang B, Dou Q, Zhang X. DeepKOA: a deep-learning model for predicting progression in knee osteoarthritis using multimodal magnetic resonance images from the osteoarthritis initiative. Quant Imaging Med Surg 2023;13:4852-66. [Crossref] [PubMed]
24. Xue Z, Wang L, Sun Q, Xu J, Liu Y, Ai S, Zhang L, Liu C. Radiomics analysis using MR imaging of subchondral bone for identification of knee osteoarthritis. J Orthop Surg Res 2022;17:414. [Crossref] [PubMed]
25. Bayramoglu N, Nieminen MT, Saarakkala S. Machine learning based texture analysis of patella from X-rays for detecting patellofemoral osteoarthritis. Int J Med Inform 2022;157:104627. [Crossref] [PubMed]
26. Li D, Li S, Chen Q, Xie X. The Prevalence of Symptomatic Knee Osteoarthritis in Relation to Age, Sex, Area, Region, and Body Mass Index in China: A Systematic Review and Meta-Analysis. Front Med (Lausanne) 2020;7:304. [Crossref] [PubMed]
27. Tong B, Chen H, Wang C, Zeng W, Li D, Liu P, Liu M, Jin X, Shang S. Clinical prediction models for knee pain in patients with knee osteoarthritis: a systematic review. Skeletal Radiol 2024;53:1045-59. [Crossref] [PubMed]
28. Guan B, Liu F, Mizaian AH, Demehri S, Samsonov A, Guermazi A, Kijowski R. Deep learning approach to predict pain progression in knee osteoarthritis. Skeletal Radiol 2022;51:363-73. [Crossref] [PubMed]
29. Mahmoud K, Alagha MA, Nowinka Z, Jones G. Predicting total knee replacement at 2 and 5 years in osteoarthritis patients using machine learning. BMJ Surg Interv Health Technol 2023;5:e000141. [Crossref] [PubMed]
30. Hayashi D, Roemer FW, Guermazi A. Osteoarthritis year in review 2024: Imaging. Osteoarthritis Cartilage 2025;33:88-93. [Crossref] [PubMed]
31. Stefanik JJ, Duncan R, Felson DT, Peat G. Diagnostic performance of clinical examination measures and pain presentation to identify patellofemoral joint osteoarthritis. Arthritis Care Res 2018;70:157-61.