From February 2016 to April 2024, a senior musculoskeletal radiologist at Xi’an Jiaotong University Second Hospital reviewed radiographs taken by 140 patients; 70 of these patients had osteoarthritis (OA), and 70 had osteoarthritis (ONFH). Thus, the ground truth for our investigation and the development of various machine learning models was based on these. Clinical information and potential risk factors such as a history of steroid usage or alcohol consumption, the presence of a double line sign, and MRI markers of sclerotic and necrotic bone alterations were considered when evaluating MRI results for ONFH patients. The evaluation for OA patients centered on recent indications of inflammation, MRI-confirmed synovitis, elevated subchondral bone signal, bone marrow edema, and joint effusion. Patients were not included in the study if they had any of the following conditions: a history of hip-related problems, collapsing femoral heads, bone cancers, or low-quality MRI pictures. Our hospital’s electronic health record system was the source of the data. See Figure 1 for a visual representation of the research design pipeline. Patients were not required to provide informed consent for this retrospective analysis, which was approved by our hospital’s ethics review board.

The 1.5T tesla scanner (Avanto, Siemens Healthineers; Erlangen, Germany) was used to capture the MRI photographs, and the parameters used were: In the coronal plane, in the (headfirst-spine) posture; sequence type turbo spin echo (TSE), T2-weighted with fat suppressed (FS), slice thickness 4.5mm, field of view 640*640mm, acquisition matrix 0\320\240\, echo duration 67ms, repetition time 3000ms.

The ROI on each MRI scan was manually segmented by an orthopedic surgeon with five years of expertise using ITK-SNAP version 4.2.0 (https://www.itksnap.org/). For individuals with osteoarthritis, the region of interest (ROI) included the femoral head and neck. Nevertheless, the ROI was restricted to the necrotic region for patients with ONFH.

A 7:3 ratio was used for the random case split between the training and testing groups. In order to train the machine learning models, we used the whole training dataset. To evaluate the models’ performance internally, we used instances from the testing dataset. Unaware of the ground truth diagnosis, three users with varying levels of experience were shown the same set of randomly shuffled radiographs. The first user was an orthopedic surgeon working in our orthopedic department, focusing on sports medicine. A general radiologist who works in the radiology department of the same institution was the second user. The third user, was another orthopedic surgeon from an external hospital. He is focused on OA, ONFH, and arthroplasty operations. We handled any differences in voxel spacing in this experiment using the fixed-resolution resampling approach. By resampling each image to a 1*1*1 mm size, the voxel spacing was uniform across all images. The data was finally normalized using z-scores, another name for zero-mean normalization.

This study’s feature extraction used traditional, hand-crafted radiomic features—geometry, intensity, and texture—derived from the initial radiographs. In order to extract radiomic characteristics, PyRadiomics was employed. The geometric, intensity and textural types of manually generated radiomic features are the three main groups. The geometric features relate to the three-dimensional shape of the bone cells. An analysis of the statistical distribution of voxel intensities within the femoral head is performed by the intensity features using first-order statistics. Features that characterize patterns or spatial distributions of intensities beyond the first order are indicated by the texture features. Texture features are retrieved using a variety of methodologies, such as the gray-level run length matrix (GLRLM), neighborhood gray-tone difference matrix (NGTDM), gray-level size zone matrix (GLSZM), and the gray-level co-occurrence matrix (GLCM). The nonlinear intensity of picture voxels is transformed using a number of transformations—including Square, Square Root, Logarithm, Gradient, LBP3D, and Exponential—in order to attain high-throughput features. One, two, and three are the sigma values used by the high Laplace filter. In addition, eight wavelet transform algorithms—HLL, HLH, HHL, LHH, LLL, LLH, and LHL—were used in the process of obtaining first-order statistics and texture features. The web resource at https://pyradiomics.readthedocs.io/en/latest/features.html provides a thorough explanation of all radiography features.

The features were first standardized using the z-score standardization method approach before going through three screening stages before final selection. First, all features were subjected to the Mann-Whitney U test; features with a P-value below 0.05 were kept. Then, to identify highly correlated features, the Pearson test was used. In order to be considered potentially predictive, features needed a P-value lower than 0.05. Finally, the least absolute shrinkage and selection operators (LASSO) were used to evaluate the key features in the end.

After LASSO was used to identify the key features, we passed them into various machine learning classifiers such as XGBoost, NaiveBayes (NB), Random Forests (RF), Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and others. We selected the best performer after comparing all of them to construct the final model. We employed 5-fold cross-validation in this particular instance.

We used the Python Statsmodels package (0.13.2 version) to evaluate the data, and we deemed a p-value below 0.05 statistically significant. Using ROC curves and the associated diagnostic accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), we assessed the clinical significance of the models in distinguishing between ONFH and OA. Decision curve analysis (DCA) and calibration curves were also used to evaluate the discriminative capacity of the model. To further evaluate the model’s robustness, we employed the Hosmer-Lemeshow test.

The study comprised a total of 140 patients’ MRI scans, consisting of 70 OA individuals and 70 ONFH patients. The patients were categorized into a training group of 98 individuals and a testing group of 42 individuals. Table 1 demonstrates a summary of the patient’s primary attributes.

A grand total of 1,119 Radiomics features were extracted using a tool that is specifically designed for feature analysis and is part of Pyradiomics (http://pyradiomics.readthedocs.io). There was a total of 234 First Order features, 182 GLRM features, 208 GLSLSM features, 65 NGTDM features, 286 GLCM features, and 14 shape features.

For each of the selected features, we ran a feature screening and a Mann-Whitney U test. The final count was 1006 features, with features being maintained only if their P-value was less than 0.05.

The second step was to use the Pearson correlation coefficient, a measure of feature correlation, to assess features with good repeatability. When two features had a correlation coefficient greater than 0.9, just single one of them was retained. In the end, 215 features were retained.

Step three involved using the logistic regression model (LASSO) to narrow down the feature set for the model’s construction and decrease the number of features overall. All regression coefficients are shrunk by LASSO until they approach zero, and the coefficients of insignificant features are adjusted to zero according to the regulation weight Lambda (λ). The optimal value of λ was found by performing a 10-fold cross-validation with a minimum criteria method. The final value of λ was chosen based on its ability to produce the lowest cross-validation error. After integrating the features with non-zero coefficients, a Radiomics model was built using the features used to construct a regression model. Following this, we determined a patient’s radiomics score by adding all the retained features multiplied by their corresponding model coefficients.

Twelve radiomics-related features were identified by LASSO regression modeling, which was executed using the scikit-learn package in Python. Below, you can see a plot of the LASSO models’ coefficient profiles and the mean square errors (MSE) that were derived from 10-fold validation. Every independent predictor’s changing trajectory is shown by each curve in the plot. Figure 2 (left) Describes how LASSO logistic regression is used to choose features in the Radiomics model. Figure 2 (right) Displays the mean squared error (MSE) values derived from a 10-fold cross-validation of the Radiomics model. For the purpose of interpreting and visualizing the radiomic features that were utilized in the radiomics models, the SHAP (16) approach was utilized in Figure 3 . This technique draws attention to the significance of individual features within the context of a complicated machine learning framework. It provides insights into how each feature contributes to the probability of a particular outcome. Moreover, the SHAP force plot in Figure 4 summarizes how each selected radiomic feature shifts the model’s prediction from the baseline (average) output toward the final predicted value. This means it displays the most influential features in classifying the given MRI case (Case 1 in our dataset) as either OA or ONFH. The length and direction of each red bar indicate how much that particular feature “pushes” the model’s decision. Positive SHAP values suggest that the feature’s measured value for this particular case pushes the model’s prediction toward the chosen positive class (e.g., ONFH), Whereas negative SHAP values indicate that the feature’s value nudges the model’s prediction away from the positive class, potentially supporting the alternative classification (e.g., OA).

XGBoost (17), LR (18), MLP (19), sigmoid_SVM (20), and NB (21) based models were constructed and compared; their AUCs in the testing cohort were (0.950, 0.964, 0.962, 0.962, and 0.971) respectively as shown in Figure 5 . As NB performed the best, it was the classifier of choice for in constructing the final Radiomics-model. On the other hand, the performance of User 1(AUC=0.606) and (AUC=0.632); User 2 (AUC=0.819) and (AUC=0.565); User 3 (AUC=0.886) and (AUC=0.880) in the training and testing cohorts respectively.

The significant features selected for the Radiomics-model are presented in Table 2 . The diagnostic AUC, 95%CI, accuracy, sensitivity, specificity, PPV, and NPV of all three users and the Radiomics-model are demonstrated in Table 3 . The confusion matrix for the Radiomics model is presented in Figure 6 , comprising True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN). In addition, the calibration curves showed good agreement between the Radiomics based model and the perfectly calibrated line, as shown in Figure 7 . The P-values of the Hosmer-LemeShow test in Table 4 were 0.574, and 0.329 for the Radiomics model in the training and testing cohort respectively. This indicates a good-fitting model, as all of the values were greater than 0.05. Both the CLEAR (22) and METRICS (23) checklists of this study were presented in Supplementary Data sheet 1 and Supplementary Data sheet 2 . Furthermore, the radiomic signatures extracted from each tested patient have been provided in Supplementary Data sheet 3 . Besides, the net benefit was plotted against threshold probability in Figure 8 , which displays the Decision curve analysis (DCA); it designates that the Radiomics-model has the highest net benefit in differentiating between OA and ONFH. Supplementary Figures 10 and 11 present heatmaps that display the radiomics features selected in the study.