According to literature reports, hip preservation failure has been defined as hip replacement within 3 years after HPS, or Harris score < 90 with progressive collapse of the femoral head on imaging. From January 2009 to December 2019, 137 patients (168 hips) with pathologically confirmed ONFH who underwent HPS&FA at our institution were enrolled. The criteria for inclusion and exclusion were as follows: Inclusion criteria (1) All patients got the same surgical intervention (performed by same surgeon); (2) Preoperative CT image data were available; (3) Patients had no history of hip surgery prior to HPS&FA; (4) The duration of follow-up was greater than three years. Exclusion criteria: (1) Insufficient CT picture quality for radiomics analysis; (2) Postoperative cancer, hip tumor, bone tuberculosis, and other malignant disorders. This study used the hip as a unit and comprised 138hips from 112 patients. Then, the entire dataset was randomly divided into a training cohort (n = 96) and a validation cohort (n = 42) with a ratio of 7:3 using computer-generated random numbers [15, 16]. The Institutional Review Board and Human Ethics Committee approved this retrospective study and waived the requirement to obtain written informed consent. The case selection process is shown in Fig. 1a, and the flowchart of study is shown in Fig. 1b.

a Flowchart of study enrollment, b Flowchart of the study

General data (age, gender, affected side, disease duration, body mass index (BMI), pathogenic factors), preoperative examination indexes (D-dimer, alkaline phosphatase (ALP), white blood cell count (WBC), neutrophil percentage(N), α-L-fucosidase (AFU)), ARCO stage, JIC classification, and preoperative Harris score were gathered and recorded. Following HPS&FA, patients were followed up to see if they continued use of glucocorticoids or alcohol, non-weightbearing time, and whether they underwent hip replacement. Additionally, postoperative Harris score and X-ray were assessed every 3 months for the first year and then, every 6 months after that. Ultimately, on June 30, 2022, every piece of data was reviewed.

The Picture Archiving and Communication (PCAS) system was utilized to acquire images. In addition, all patients underwent CT examinations of the hip utilizing the Philips 128-row Brilliance CT and the GE 64-row LightSpeed VCT. Scan parameters include tube voltage 110–150 kV, tube current 220–680 mA, exposure time 240–800 ms, slice thickness 1.0–3.0 mm, slice spacing 1.0–3.0 mm, and reconstruction matrix 512 × 512.

We then segmented and extracted features from CT images. Bilinear interpolation was used for resampling, with layer thickness and spacing of 1 mm, imported into 3D Slicer (https://www.slicer.org, V5.0.2) as NII files. This study focused on femoral head necrosis, defined as fractures of the trabecular bone, texture disorders, sclerosis zones surrounding low-density areas, and cystic degeneration on CT images (sagittal, coronal and horizontal). Reader 1 (Xin Liu) and Reader 2 (Bin Du) outlined this ROI (Fig. 2) in the bone window of CT images. (Hounsfiled Unit (HU)) value was set at 1500HU, window level at 400U. Pyradiomics plugin automatically extracted 851 imaging features. Radiomics features extracted by two readers were evaluated using ICC. Consistency is deemed satisfactory when ICC > 0.75. Reader 1 segmented 30 CT pictures twice within one month to compute intra-observer ICC. Reader 2 segmented selected images separately to calculate inter-observation ICC. We calculated intra-observer and inter-observer ICC. There was no statistically significant difference between Reader 1 and Reader 2. The intra- and inter-observer ICC exceeded 0.75. Thus, both intra- and inter-observer feature extraction exhibited high repeatability. Eventually, all results were based on measurements made by Reader 1. Then, Z-score normalization was applied to guarantee repeatability.

Representative CT images for failed a HPS&FA and successful b HPS&FA. c Comparison of basic information and risk factors. Red arrow: necrosis volume was large and involved the lateral column in (a). Green arrow: necrosis volume was relatively small and the lateral column was not accumulated in (b)

Both clinical and CT radiomics data were screened. Using univariate and multivariate analysis, clinical predictors were quickly extracted from 16 clinical data. In addition, the LASSO was utilized to determine the optimal radiomics features among CT radiomics features. The optimal radiomics features with nonzero coefficients were ultimately linearly combined to yield a Rad-score for classification analysis.

Clinical data and Rad-score were utilized for modeling by using R statistical software, and then, two prediction models were established, namely CPM and RPM. Subsequently, in order to improve model prediction performance whenever possible, a comprehensive model was developed by integrating clinical predictors with Rad-score as the predictive model for HPS&FA, namely CRPM. At last, using DeLong test, the significance of differences in AUC between models was determined.

ROC and AUC were plotted to analyze the diagnostic efficacy of model. Then, to visualize the relationship between the variables in the prediction model, a nomogram based on CRPM for individualized efficacy prediction was constructed. Furthermore, a calibration curve was developed to evaluate the calibration utility of nomogram. To evaluate the medical benefit of nomogram under different risk thresholds, DCA was employed. The model was finally validated using the validation cohort.

For statistical analysis and model development, SPSS 26.0 and R statistical software (version 1.2.5042) were employed. Using SPSS, both univariate and multivariate analyses were conducted. The R software "glmnet" package was used to perform LASSO Using the "survminer" package for proportional hazards model (COX) survival analysis in order to visualize the relationship between variables and determine the cut-off value. ROC and AUC were then plotted using the "pROC" package, and a nomogram was constructed using the "rms" package. We drew calibration and decision curves for the accuracy and clinical utility of prediction models, respectively, using the "rmda" package.