Radiographic imaging is frequently employed in the diagnosis of knee OA. Evaluating the severity of knee OA is a challenging and subjective process and often involves a qualitative analysis of plain radiographs. The K/L grading system, the most commonly utilized scale for classifying knee OA, is limited by its subjective nature and the notable variability in agreement between different observers. Therefore, an objective and consistent approach to grading the severity of knee OA is needed. Radiomics, the process of extracting quantitative features from medical images, holds promise for enhancing the precision and reliability of grading knee OA [17]. In our study, we devised a radiomic model utilizing the LR classifier to automate the grading of knee OA severity with plain radiographs. We selected both AP&LAT radiographs of the knee joint for the extraction of radiomic features. The mean ICC for the ROI delineation conducted by two radiologists was 0.91, signifying high reproducibility. The diagnostic efficiency results of all three radiomic models showed good diagnostic performance, and the AP&LAT model performed best, with the highest overall accuracy and the highest AUC value (macro/micro) of 0.864/0.879 in the test cohort, which was statistically significant compared to the other two models. The five-class ROC curves showed that the combined AP&LAT model achieved the best grading performance with AUCs of 0.990, 0.781, 0.728, 0.913, and 0.885 from class 0 to 4. These results suggest that combining information from both AP&LAT images significantly improved the performance of our model, and the combined AP&LAT radiomics model holds promise as a valuable tool for early and accurate diagnosis of knee OA.

Numerous studies have explored the application of radiomics in the classification and grading of knee OA. For instance, the research conducted by Abdelbasset Brahim et al [18] introduced a comprehensive computer-aided diagnosis (CAD) system designed for the early detection of knee OA utilizing knee X-ray images and machine learning algorithms. The findings revealed that the system offered promising predictive capabilities in OA detection, with an accuracy of 82.98%, a sensitivity of 87.15%, and a specificity reaching 80.65%. Likewise, the study by Mahrukh Saleem et al [19] showcased a computer-vision system aimed at aiding radiologists by assessing radiological indicators in X-rays for knee OA. The outcome demonstrated that this approach could effectively identify OA, achieving an impressive detection accuracy rate of over 97%. The above studies achieved good results in the identification of knee OA, but they all used the AP X-ray images and did not add the LAT radiography to the study. Additionally, a study by Luca Minciullo et al [20] introduced a fully automated technique utilizing a Random Forest Regression Voting Constrained Local Model (RFCLM) for differentiating between radiographs of individuals with knee OA and those without. The study highlighted that the automated analysis of the LAT view yielded classification results that were on par with or superior to those obtained by applying similar methods to the frontal view. This study showed that LAT images also have information for knee OA classification, although their study did not compare the classification effect of the AP, LAT, and combined images. In our prior research, we undertook a binary classification investigation of knee OA by developing a radiomic model. Our findings indicated that among the four groups of models tested, the LR model outperformed the others, achieving an AUC value of 0.843. This result demonstrates that the radiomics model possesses a strong capacity for the accurate diagnosis of knee OA [16]. Therefore, in this study, we continued to use the LR model to automatically grade knee OA and tested whether LAT radiographs could play an important role in model establishment. The results showed that the radiomics model could indeed accurately classify knee OA in radiographs, and the LAT images provided characteristic information that was different from the AP views.

The integration of deep-learning approaches, such as convolutional neural networks (CNNs), has enhanced the efficacy of radiomic models in the grading and classification of knee OA [21,22,23]. For instance, Berk Norman et al [24] introduced a fully automated algorithm called DenseNets, designed for knee OA detection employing K/L grading scales. The reported sensitivity rates for detecting no OA, mild, moderate, and severe OA were 83.7%, 70.2%, 68.9%, and 86.0%, respectively. Aleksei Tiulpin et al [25] conducted a study that yielded an automatic technique for predicting K/L and Osteoarthritis Research Society International (OARSI) grades from knee radiographs using deep learning. This method attained an impressive AUC of 0.98 and an average precision score of 0.98 in detecting the presence of radiographic OA. Kevin A. Thomas et al [26] created an automated model to assess the severity of knee OA using radiographic images. They compared the performance of their model with that of musculoskeletal radiologists. The model demonstrated an average F1 score of 0.70 and an accuracy of 0.71 across the entire test set. One of our previous studies also showed that deep-learning techniques can accurately grade knee OA in X-ray images and we also found that multiview X-ray images and prior knowledge improved classification efficacy. The overall accuracy of the DL model with multiview images and prior knowledge was 0.96 compared to 0.86 for an experienced radiologist [27]. Despite deep-learning techniques being regarded as cutting-edge technology for image classification, their highly complex internal structures often render the models’ decision-making processes opaque to human understanding, leading to a lack of explicability [28, 29]. However, the features extracted through radiomics are more interpretable, offering a clear understanding of their outcomes [10]. In our study, the feature extraction process captures a range of traceable image information that may elude radiologists’ observation but proves to be crucial for the diagnosis of knee OA.

Currently, the focus of international research on the musculoskeletal system is predominantly on osteoporosis, bone mineral density, fractures, bone tumors, and the like. The majority of investigations into degenerative osteoarthropathy are underpinned by CT/MRI imaging [25, 30]. Plain radiographs are seldom used in isolation due to their limited informational yield. But for knee OA, X-ray is a faster and more convenient non-invasive examination method. Faster and more accurate knee OA grading will help both radiologists and clinicians in their work. Abdelbasset Brahim et al confirmed that the radiomics model can accurately distinguish the K/L classification of the knee joint in X-ray images, with an accuracy of 82.98% [18]. This is consistent with our results, indicating that radiomics model is accurate for the K/L classification of the knee joint. In this study, we conducted a comprehensive analysis of the knee joint X-ray, capturing detailed information such as the femur, medial and lateral tibial condyles, patella, and corresponding joint space width. Through the selection of radiomics features, we identified the nine most relevant features, which may not be discernible to the naked eye but play a crucial role in measuring important parameters of knee OA.

It is noteworthy that our study not only utilized AP radiographs of the knee joint but also incorporated LAT radiographs. This approach sets our study apart from many other related studies, allowing for a more comprehensive analysis of the knee joint from different perspectives [18,19,20, 31]. LAT radiographs offer enhanced density and shape information, enabling a more comprehensive assessment of knee lesions. However, it is noteworthy that most knee OA studies have primarily utilized AP radiographs. This preference can be attributed to the fact that the reference standard K/L grading for knee OA is based on evaluations conducted in the AP position [31,32,33]. Despite this, recognizing the value of LAT radiographs, it is essential to explore their potential benefits and incorporate them into future research to further enhance the evaluation of knee OA. In order to enhance the extraction of radiomic features and capture more knee image information, we made an innovative addition of LAT radiographs to our study. Interestingly, we discovered that the utilization of the LAT view model alone yielded greater overall accuracy and a higher AUC value compared to the AP view model. This finding highlights the potential superiority of LAT radiographs in improving the diagnostic performance of knee imaging analysis. We suspect there are several potential reasons for the observed differences in the effectiveness of LAT and AP radiographs. Firstly, the patella, which is one of the bones most affected by knee OA, is more prominently visible in the LAT view compared to the AP view, where it can be obscured by the femur. This improved visibility in LAT radiographs allows for a better assessment of patellar involvement in the radiomics model. Secondly, LAT radiographs provide a better display of joint space and characteristic information related to osteophytes, which may differ from the information obtained from AP images. This additional information, unique to LAT radiographs, can contribute to more accurate grading judgments by the radiomics model. These factors suggest that incorporating LAT radiographs can offer valuable insights and complementary information to enhance the performance of radiomics models in assessing knee OA. Further research and validation may help elucidate the full potential and benefits of utilizing LAT radiographs in this context. Compared to the conventional practice of delineating the ROI using rectangular shapes, our approach involves segmentation along the entire knee edge. This innovative technique enables more precise extraction of radiomic features, resulting in improved filtration of irrelevant image information. Through the utilization of this segmentation technique, our aim is to enhance the discriminative power of the radiomics model, providing more reliable and meaningful insights into the assessment of knee OA. Nonetheless, further validation and comparative studies are warranted to comprehensively evaluate the benefits and potential advantages of this segmentation methodology over existing approaches.

Finally, we have demonstrated that the radiomic model developed in this study outperformed radiologists with 4 years and 2 years of musculoskeletal diagnostic experience. This finding indicates that our AP&LAT model exhibits higher accuracy compared to junior radiologists and can conveniently offer clinicians with diagnosis and treatment guidance.

Our study is subject to several limitations that should be acknowledged. Firstly, it is important to note that this study is based on a retrospective analysis, with all radiographic data obtained solely from a single hospital. This lack of diverse external data verification may introduce potential selective bias into our findings, thereby limiting the generalizability of our results. Secondly, the radiomics analyses in our study were conducted exclusively using radiographic images. Future studies would benefit from incorporating joint analyses of multimodal datasets and incorporating additional clinical parameters. This approach would provide a more comprehensive and holistic understanding of the disease and potentially improve the accuracy and robustness of the radiomic model. Thirdly, manual segmentation of the ROI for each image was performed, which can be time-consuming and may introduce inter-observer variability. Exploring the feasibility of automatic segmentation techniques in future research could significantly enhance efficiency and reduce potential errors in ROI delineation.