A radiomics and deep learning nomogram developed and validated for predicting no-collapse survival in patients with osteonecrosis after multiple drilling

Abstract

Purpose

Identifying patients who may benefit from multiple drilling are crucial. Hence, the purpose of the study is to utilize radiomics and deep learning for predicting no-collapse survival in patients with femoral head osteonecrosis.

Methods

Patients who underwent multiple drilling were enrolled. Radiomics and deep learning features were extracted from pelvic radiographs and selected by LASSO-COX regression, radiomics and DL signature were then built. The clinical variables were selected through univariate and multivariate Cox regression analysis, and the clinical, radiomics, DL and DLRC model were constructed. Model performance was evaluated using the concordance index (C-index), area under the receiver operating characteristic curve (AUC), net reclassification index (NRI), integrated discrimination improvement (IDI), calibration curves, and Decision Curve Analysis (DCA).

Results

A total of 144 patients (212 hips) were included in the study. ARCO classification, bone marrow edema, and combined necrotic angle were identified as independent risk factors for collapse. The DLRC model exhibited superior discrimination ability with higher C-index of 0.78 (95%CI: 0.73–0.84) and AUC values (0.83 and 0.87) than other models. The DLRC model demonstrated superior predictive performance with a higher C-index of 0.78 (95% CI: 0.73–0.84) and area under the curve (AUC) values of 0.83 for 3-year survival and 0.87 for 5-year survival, outperforming other models. The DLRC model also exhibited favorable calibration and clinical utility, with Kaplan–Meier survival curves revealing significant differences in survival rates between high-risk and low-risk cohorts.

Conclusion

This study introduces a novel approach that integrates radiomics and deep learning techniques and demonstrates superior predictive performance for no-collapse survival after multiple drilling. It offers enhanced discrimination ability, favorable calibration, and strong clinical utility, making it a valuable tool for stratifying patients into high-risk and low-risk groups. The model has the potential to provide personalized risk assessment, guiding treatment decisions and improving outcomes in patients with osteonecrosis of the femoral head.

Background

Nontraumatic osteonecrosis of the femoral head (NONFH) arises from various factors that impact the blood supply to the femoral head, leading to the necrosis of bone cells and marrow tissue [1, 2]. In China, there are approximately 8.12 million ONFH patients, with an annual incidence of new cases ranging from 75,000 to 150,000, primarily affecting individuals aged between 30 and 50 years [3, 4]. If left untreated, approximately 80% of NONFH cases may progress to collapse, necessitating total hip arthroplasty (THA) [5]. However, outcomes following THA in younger NONTH patients are often suboptimal, frequently requiring subsequent revision surgeries. Consequently, the primary objective is to preserve the femoral head and prevent collapse [6]. Although the optimal approach to preserving the femoral head remains controversial, core decompression (CD) currently ranks among the most performed hip preservation surgeries in both the United States and China [7,8,9,10,11,12]. Studies have indicated an increased risk of further collapse and fractures following conventional CD procedures. However, the utilization of multiple drilling (MD) has been associated with a reduction in these risks, offering a simpler surgical procedure while demonstrating comparable success rates [13].

However, due to variations in research methodologies, differing outcome definitions, and variations in follow-up, the effectiveness of MD remains debatable [13,14,15]. While some studies suggest that clinical symptoms can be improved after MD [13, 14], others, such as the study conducted by Liu et al., suggest that MD may not mitigate the rate of THA, and patients with predisposing factors such as extensive lesions and bone marrow edema may face an elevated risk of collapse [15]. Hence, predicting the risk of femoral head collapse and survival time following MD, as well as identifying patients who may benefit from it, are crucial endeavors. While several studies have identified risk factors associated with postoperative collapse after MD, these factors remain unvalidated [16]. Two studies have developed scoring systems or prediction models, yet one of them lacks further validation, while the other underwent validation within the same cohort [17, 18]. Radiomics and deep learning (DL) are novel methods that are more commonly applied in the diagnosis and prognosis research within oncology, and they have shown promising results [19]. However, their application is less common in the field of orthopedics. Therefore, the primary objective of this study is to employ radiomics and deep learning to develop and validate a novel model for predicting postoperative no-collapse survival in non-traumatic femoral head osteonecrosis who underwent multiple drilling.

Methods

Study design

This observational retrospective study was approved by the Ethics Committee of the First Hospital of Jilin University, and patient informed consent was waived in accordance with the Helsinki Declaration. Patients with early stage nontraumatic osteonecrosis of the femoral head who underwent multiple drilling core decompression in our hospital from January 2015 to January 2022 and followed for at least 6 months were included. The flow chart of the study methodology is depicted in Fig. 1a.

Fig. 1

Model construction. a Flow chart of model construction of different models. b The DLRC Nomogram predicting no-collapse survival in patients after multiple drilling. DLRC: Deep Learning-Radiomics-Clinical

The inclusion criteria were as follows: (1). Diagnosis of nontraumatic osteonecrosis of the femoral head. (2). Age between 18 and 75 years. (3). Patients with ARCO 3A who refused THA. (4). No prior treatment for osteonecrosis of the femoral head. (5). A minimum follow-up period of 6 months. (6). Availability of a complete preoperative X-ray in DICOM (Digital Imaging and Communications in Medicine) format, taken one month prior to surgery. The exclusion criteria were as follows: (1). Traumatic osteonecrosis of the femoral head. (2). Classified as ARCO IIIB or above, indicating collapse of the femoral head. (3). Presence of contraindications to zoledronic acid, such as severe renal insufficiency. (4). Complications including fracture, infections, and deep vein thrombosis (DVT).

The staging adhered to the 2019 ARCO staging system based on a comprehensive radiological evaluation conducted before surgery, including an anterior–posterior pelvic radiograph, a frog-leg lateral radiograph, and magnetic resonance imaging (MRI) of the hip [20]. Lesion location was categorized into four types (Type A, B, C1, and C2) based on the midcoronal T1-weighted MRI images, following the classification described by Sugano [21]. Given that sagittal slices were not included in the MRI images of several studies, the arc of the necrosis surface on the femoral head was measured on midcoronal and mid-axis T1-weighted MRI images, documented as A and B, respectively. We then calculated the combined necrotic angle by summing the values of A and B and determined the necrosis degree index using the formula (A/180) * (B/180) * 100, as outlined by Cherian [22]. Additionally, bone marrow edema was documented, with assessment criteria based on the T2-weighted MRI images. Clinical data, comprising age, gender, body mass index (BMI), etiologies (corticosteroid-induced, alcohol-related, and idiopathic) and the duration of symptoms were extracted from the medical records.

Surgical procedures and bisphosphonate medication

The patients were placed in a supine position, with the affected hip slightly elevated and the lower extremity rotated inward by 15 degrees. Three to four Kirschner wires, each with a diameter of 3.0 mm, were carefully inserted into the femoral head and neck. The wires were advanced until they reached the necrotic area, typically located 0.5–1.0 cm below the articular surface. After complete decompression, the Kirschner wires were removed. Patients were allowed to ambulate with crutches, and weight-bearing restrictions were imposed for six months. They were advised to avoid excessive weight-bearing, jumping, and strenuous activities for one year. Additionally, some patients received a 5 mg intravenous infusion of zoledronic acid one month after surgery, as recommended by the medical team. They were also prescribed daily supplements of calcium (500–1000 mg) and vitamin D3 (400–800 units). Patients who received a weekly dose of 70 mg of alendronate sodium for a three-month duration were also included in the study, as this regimen provides an equivalent dose to the intravenous injection of 5 mg of zoledronic acid [23].

Follow-up and assessment

Patients were followed up at 6 months, 12 months, and then annually, and whenever they experienced worsening hip pain or limited activity. Diagnostic imaging, including pelvic radiographs and hip MRIs, was performed at each follow-up. The Harris Hip Score was used to assess the functional recovery of the hip joint before the surgery and at 12 and 24 months postoperatively. The main outcome of interest was no-collapse survival (femoral head depression < 2 mm), determined by pelvic radiographs or MRIs, and a collapse greater than 2 mm was considered a failure [24]. The imaging evaluation was conducted by two evaluators (Zhang and Li), with any disagreements resolved by a third senior surgeon (Gu). Secondary outcome measures included complications related to core decompression, such as intertrochanteric or femoral neck fractures and deep vein thrombosis. Additionally, adverse reactions to intravenous zoledronic acid, which could manifest as fever, flu-like symptoms, and musculoskeletal pain, were also monitored. Patients with missing data were excluded from the study analysis.

Image preprocessing and extraction of radiomics and DL features

Anterior–posterior pelvic radiographs were obtained from the Picture Archiving and Communication System (PACS) in the Digital Imaging and Communications in Medicine (DICOM) format after anonymization. Subsequently, these radiographs were converted to the Neuroimaging Informatics Technology Initiative (NIfTI) format using the dicom2nii package, a Python-based tool. Following image intensity normalization, experienced orthopedic surgeons Zhang and Li, each possessing over 10 years of expertise, manually delineated regions of interest (ROI) within the femoral head in the ITK-SNAP software (version 3.4, http://www.itksnap.org) [25]. A variety of radiomics features, including first-order statistical features, shape-based features, texture features, and high-order features, were extracted from each patient using the PyRadiomics package [26]. To incorporate deep learning features, the pretrained ResNet18 architecture was included and deep learning (DL) features were extracted from the average pooling layer.

Feature selection and radiomics and DL signature establishment

The feature selection process for radiomics and DL features involved the following steps: (1). Features with missing values or outliers were eliminated; (2). Data standardization was carried out through Z-score transformation to achieve a mean of 0 and a variance of 1; (3). Features with intraclass and interclass correlation coefficients below 0.75 were excluded; (4). A Mann–Whitney U test was conducted, and only features with a P-value < 0.05 were kept; (5). Pearson's correlation coefficients were calculated among the features, and one of the features was retained when the correlation coefficient exceeded 0.9 between any two features; (6). The Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression model was applied to using tenfold cross-validation. Features with non-zero coefficients were selected based on the minimum lambda value. These selected features, along with their respective weights, were retained to construct a radiomics and DL signature for all patients.

Model construction and evaluation

Univariate and multivariate Cox regression analyses were conducted on the clinical variables to identify risk factors. Subsequently, a clinical model was constructed using the identified factors. Furthermore, a Radiomics model was developed by integrating the radiomic signature with clinical variables. Simultaneously, a Deep Learning (DL) model was created by incorporating the deep learning signature along with clinical factors. Following these steps, a Deep Learning-Radiomics-Clinical model (DLRC) was formulated, which amalgamated the DL signature, radiomics signature, and clinical variables.

Model performance was assessed using the concordance index (C-index) and receiver operating characteristic (ROC) curve analysis to calculate the area under the ROC curve (AUC). Additionally, the net reclassification index (NRI) and integrated discrimination improvement (IDI) were computed to compare the model performance. Calibration efficiency of the model was evaluated through the generation of calibration curves, while assessment of the clinical utility of the predictive models involved Decision Curve Analysis (DCA).

Statistical analysis

Statistical analyses were performed using R software (version 4.3.1, http://www.r-project.org). The normality of continuous variables was evaluated using the Kolmogorov–Smirnov test. Continuous variables following a normal distribution are presented as mean ± standard deviation, while those not conforming to normality are described using the median and interquartile range (Q25-Q75). Group differences were assessed using Student's t-test or the Mann–Whitney U test, depending on the distribution of the data. Categorical data were reported as frequency (percentage) and analyzed using the Chi-square test or Fisher's exact test. When conducting Cox regression analysis, continuous variables were discretized into categorical variables based on their median values. Covariates with a significance level of p < 0.1 in the univariate Cox regression analysis were subsequently included in the multivariate Cox regression analysis using the backward elimination approach. Risk scores were computed for each patient according to the nomogram. Subsequently, using the median score, the samples were stratified into high-risk and low-risk groups. Kaplan–Meier survival curves were constructed and statistical differences were assessed using log-rank tests. The level of statistical significance was set at p < 0.05.

Results

Baseline characteristics

A total of 144 patients (212 hips) underwent multiple drilling procedures and were included in this study. The median follow-up period was 29.55 months (19.48—40.83). At the end of the study, 64 hips (43.24%) experienced collapse. The overall no-collapse survival rate was 39.75% (95% CI, 28.22%—55.98%). The postoperative 3-year no-collapse survival rate was 71.59% (95% CI, 64.50%—79.46%), and the 5-year no-collapse survival rate was 44.72% (95% CI, 34.73%—57.57%). Among the patients, 80.19% (170/212) were male, with a median Body Mass Index BMI of 23.71 (21.58–26.04). The median age at surgery was 42.00 years (33.00–49.00), and the median symptom duration before surgery was 2.00 months (1.00–5.00). Bisphosphonate medication was administered to 45.75% (97/212) of the hips. Simultaneous contralateral THA was performed in 10.85% (23/212) of the hips, while bilateral MD was performed in 62.26% (132/212). We identified 58 hips (27.36%) attributed to corticosteroid use, 33 hips (15.57%) associated with alcohol consumption, and 121 hips (57.08%) categorized as idiopathic. Apart from Bone Marrow Edema, Combined Necrotic Angle, and Necrosis Degree Index, no statistically significant differences were observed between the collapse and no-collapse groups in other baseline characteristics (Table 1).

Table 1 Baseline characteristics of all patients in collapse and no-collapse group

Selection of clinical factors

In the analysis, twelve clinical variables were subjected to univariate and multivariate Cox regression analyses. The results identified ARCO classification (hazard ratio [HR], 2.12; 95% confidence interval [CI], 1.25—3.58; P = 0.037), bone marrow edema (BME) (HR, 2.13; 95% CI, 1.19—3.81; P = 0.010), and combined necrotic angle (HR, 2.41; 95% CI, 1.43—4.07; P < 0.001) as independent risk factors for femoral head collapse after MD. It is noteworthy that the administration of bisphosphonate (HR, 1.14; 95% CI, 0.69—1.89; P = 0.602) was not a significant risk factor. The criteria for selecting significant variables were based on a P-value threshold of < 0.05 in the multivariate Cox regression model. These findings emphasize the importance of ARCO classification, bone marrow edema, and combined necrotic angle in predicting femoral head collapse risk. Detailed results of the univariate and multivariate Cox regression analyses are presented in Table 2.

Table 2 Cox proportional hazards regression analysis of factors predicting no-collapse survival following multiple drilling

Model construction and evaluation

For each patient, a total of 1562 radiomics features and 512 deep learning (DL) features were extracted. Feature selection was performed using LASSO regression with cross-validation to determine the optimal lambda values. A total of 14 radiomics features with non-zero coefficients were identified at an optimal lambda value of 0.054, while 21 DL features with non-zero coefficients were identified at an optimal lambda value of 0.712 (Fig. 2 a-d). The radiomics and DL signatures were computed as linear combinations of these selected features, weighted by their respective coefficients.

Fig. 2

Radiomics and DL feature selection using the LASSO-Cox regression model. a The partial likelihood deviance was plotted versus log (lambda) with radiomics features; (b) LASSO coefficient profiles of the 184 radiomics features; (c) The partial likelihood deviance was plotted versus log (lambda) with DL features; (d) LASSO coefficient profiles of the 512 DL features

The performance of the radiomics, DL, and DLRC models surpassed that of the clinical model, as demonstrated by both C-index and AUC metrics. The C-indices for the radiomics, DL, and DLRC models were 0.73 (95% CI: 0.70–0.82), 0.76 (95% CI: 0.70–0.82), and 0.78 (95% CI: 0.73–0.84), respectively, compared to the clinical model's C-index of 0.64 (95% CI: 0.56–0.71) (Table 3). For the 3-year AUC values (Table 3, Fig. 3a), the clinical model achieved 0.65, whereas the radiomics, DL, and DLRC models achieved 0.80, 0.80, and 0.83, respectively. Similarly, for the 5-year AUC values (Table 3, Fig. 3b), the clinical, radiomics, DL, and DLRC models yielded 0.77, 0.87, 0.84, and 0.87, respectively. Notably, the DLRC model demonstrated superior predictive performance, achieving the highest C-index and 3-year AUC value among all models.

Table 3 Model performance of different models

Fig. 3

Model evaluation of different models. a Receiver operating characteristic curves for predicting 3-year no-collapse survival after multiple drilling; (b) Receiver operating characteristic curves for predicting 5-year no-collapse survival after multiple drilling

According to the IDI values (Supplemental Table 1), both the radiomics, DL, and DLRC models outperformed the clinical model in predicting no-collapse survival after MD. Moreover, the DLRC model exhibited superior predictive performance for overall no-collapse survival compared to the radiomics model. Based on the NRI values (Supplemental Table 1) the radiomics, DL, and DLRC models exhibited superior predictive capacities for 3-year no-collapse survival compared to the clinical model, with no notable discrepancies among these three models. For the 5-year NRI, no significant differences were detected among the clinical, radiomics, DL, and DLRC models. The calibration curves (Fig. 4) demonstrated favorable agreement between the predicted and actual survival outcomes for the DLRC model at both 3-year and 5-year no-collapse survival after multiple drilling, highlighting the model's robustness in survival prediction.

Fig. 4

Model validation of different models. a Calibration curves for 3-year no-collapse survival after multiple drilling; (b) Calibration curves for 5-year no-collapse survival after multiple drilling

Clinical utility

The radiomics, DL, and DLRC models demonstrated superior clinical utility compared to the clinical model for 3-year no-collapse survival. Notably, the clinical model exhibited a negative net benefit at higher threshold risks (Fig. 5a). For risk thresholds below 80%, the radiomics, DL, and DLRC models achieved higher net benefits compared to the clinical model (Fig. 5b). Among these, the DLRC model showed the most consistent performance across varying thresholds and is presented as a nomogram (Fig. 1b). The Kaplan–Meier survival curves (Fig. 5c) highlight significant differences in no-collapse survival outcomes between high-risk and low-risk cohorts as predicted by the DLRC model. For the low-risk cohort, the 3-year no-collapse survival rate reached 93.96%, compared to 51.2% in the high-risk cohort. Similarly, the 5-year no-collapse survival rates were 85.7% for the low-risk group and 15.64% for the high-risk group.

Fig. 5

Clinical utility of different models. a The decision curves indicated that the radiomics, DL and DLRC model showed better clinical utility than the clinical model for 3-year no-collapse survival; (b) The decision curves indicated that the radiomics, DL and DLRC model showed better clinical utility than the clinical model for 5-year no-collapse survival when the risk threshold was less than 80%. At higher risk thresholds, the curves did not intersect with the X-axis, which may be due to factors such as the relatively small sample size, the lower 5-year survival rate, and the reduced number of high-risk individuals at 5 years.; (c) No-collapse survival comparison between patients classified by the DLRC nomogram into high-risk and low-risk cohorts

Discussion

In this retrospective study, a Deep Learning-Radiomics-Clinical (DLRC) predictive model was developed and validated using radiomics and deep learning signatures, along with three clinical characteristics that independently predict femoral head no-collapse survival after multiple drilling. The DLRC model demonstrated good discrimination ability, with an AUC of 0.83 for 3-year no-collapse survival and 0.87 for 5-year no-collapse survival. The C-index was 0.78 (95% CI: 0.73—0.84), outperforming other models. The model exhibited favorable calibration and demonstrated clinical utility. The calibration curve showed good concordance between predicted and observed outcomes across various probability ranges. Moreover, the DCA curve illustrated the model's clinical usefulness by demonstrating a positive net benefit across a range of threshold probabilities.

To the best of our knowledge, this is the first radiomics and deep learning model developed to predict no-collapse survival in patients with osteonecrosis after multiple drilling. Previous studies have mainly focused on clinical characteristics and radiographic parameters [10, 16, 27].Wei et al. developed a prognostic system for avascular necrosis of the femoral head after CD and identified seven independent risk factors [17]. Although the system achieved an AUC of 0.935, it is noteworthy that the prognostic system was constructed and evaluated within the same patient cohort and survival time was not accounted for [17]. Zhao et al. proposed a predictive model for the collapse of avascular necrosis of the femoral head after CD, but this study solely conducted internal validation and the C statistic was 0.82 [18]. We established a combined nomogram that incorporates the radiomics and DL signature and clinical factors for prognostic prediction, which is a novel approach compared to previous studies. The radiomics, DL, and DLRC models demonstrated superior performance compared to the clinical model, according to the IDI results. This finding aligns with previously reported prognostic models. Despite the absence of significant differences between DLRC and DL models, the DLRC model outperforms the radiomics model.

This study identified disease stage, bone marrow edema, and the combined necrotic angle as independent predictors of femoral head collapse, aligning with findings from previous research. Based on these factors, a clinical model was subsequently developed. Mont et al. [28] and Song et al. [13] observed a higher success rate in Ficat stage I and II compared to stage III after MD. Lieberman et al. [12] observed that in pre-collapse hips, there was a 19% conversion to total hip arthroplasty (THA) and a 31% radiographic progression, and in post-collapse hips, these rates were 30% and 49%, respectively. Within our study, the risk of collapse in ARCO stage IIIA (HR 2.12, 95% CI: 1.25—3.58) was higher compared to ARCO stage I, with no significant difference observed between ARCO stage I and ARCO stage II. However, controversy exists regarding the best classification system due to suboptimal interobserver and intraobserver agreement [20]. In the present study, the 2019 ARCO classification is adopted due to its simplicity, utilizing a 2 mm threshold to define both early and late collapse.

The extent of necrosis, size and location were identified as risk factors for collapse in previous researches. Various methods have been employed to quantify necrotic lesions, including the Steinberg classification, the Kerboul combined necrotic angle, the index of necrosis, the modified index of necrosis, and 3-D MRI measurements [22, 29]. However, consensus regarding the most reliable and valid method remains elusive [30]. Studies have suggested a higher success rate in cases with a smaller affected area (< 15%) or a necrotic angle less than 200°. In this study, we chose to use the combined necrotic angle and the index of necrosis as alternatives. The combined necrotic angle (HR 2.41, 95% CI:1.43—4.07, p < 0.05) was identified as an independent risk factor, whereas the index did not prove to be an independent risk factor in the multivariate Cox model (p > 0.05). A plausible reason for this discrepancy may be attributed to the insufficient precision of our measurement method. Research has indicated the significance of the necrotic location in relation to the weight-bearing zone, with the JIC proving to be a reliable method. Specifically, JIC C1 and JIC C2 have been associated with higher rates of collapse. However, in the current study, necrosis location did not emerge as an independent risk factor (p > 0.05). A possible explanation for this may include inadequacies in sample size.

Clinical factors such as age, BMI, etiology, and laterality were not identified as risk factors for collapse in this study. However, Wei et al. identified age, male gender, etiology, and a prolonged disease duration as prognostic factors influencing the outcome following CD [17]. Disparities in findings may stem from various factors, including differences in sample sizes used for Cox regression analysis and variations in patient characteristics such as alcohol or corticosteroid intake. Moreover, discrepancies in the accuracy of patients' recollection of symptom duration could also contribute to differing results. Some studies have suggested that bisphosphonates may play a role in preventing femoral head collapse [31, 32]. Agarwala et al. found that alendronate administration resulted in satisfactory outcomes for most patients with early-stage avascular necrosis. The results revealed that 364 hips (92.2%) achieved satisfactory clinical outcomes, obviating the need for surgical intervention [31]. Similarly, Kang et al. reported successful pain relief and delayed progression of necrosis with a combination of multiple drilling and systematic alendronate treatment [32]. However, our study did not find bisphosphonates to have a significant impact on prognosis after MD. This discrepancy may be attributed to our focus on radiological progression as an outcome measure, whereas prior research primarily relied on clinical assessments.

There are several strengths to this study. As a rapidly evolving discipline in medical imaging, radiomics and deep learning facilitates high-throughput analysis of medical image data [33]. Both radiomics and deep learning play pivotal roles in aiding diagnosis, prognosis, and prediction, with significant advancements observed in oncology studies [34]. Yet, their application in orthopedic research has been rarely reported. In the present study, various radiomics and deep learning features were extracted, and we employed several dimensionality reduction techniques to select the top 14 radiomics features and 21 DL features. Furthermore, multivariate Cox analysis demonstrated that both the radiomics and deep learning signatures were identified as independent risk factors, exhibiting a higher Hazard Ratio (Supplemental Table 2). Radiomics and deep learning methodologies may offer superior approaches for discerning the extent of necrotic regions and identifying lesion heterogeneity. DL directly generated and extracted optimal features from raw data, whereas radiomics relies on predefined features [35]. In this study, no significant differences were observed between the radiomics and DL models, which contrasts with findings from other studies. This distinction may be attributed to the specific DL architecture utilized in this investigation.

This model can stratify patients into high-risk and low-risk groups with significant statistical differences, thus enabling prediction of no-collapse survival. This study confirms the applicability of radiomics and deep learning in the field of orthopedics, thereby aiding in treatment and prognosis. Personalized assessment of the collapse risk after osteonecrosis of the femoral head (ONFH) can guide treatment decisions, enabling high-risk patients to consider adjunctive therapies such as cell therapy.

This study has several limitations. First, it is a retrospective analysis with a relatively small patient cohort from a single center, which inherently limits the diversity of the patient population. The small sample size reduces the statistical power of the analysis, increasing the risk of Type II errors and hindering the detection of subtle or less common risk factors. Additionally, the limited sample size increases the likelihood of overfitting. This overfitting may result in inflated performance during internal validation but significantly reduced predictive accuracy on independent datasets. The single-center nature of the study further restricts generalizability, as the patient cohort may reflect regional clinical practices and demographic characteristics that are not representative of broader populations. The lack of external validation exacerbates these issues, making it challenging to robustly evaluate the model's applicability across diverse patient populations. Second, the postoperative follow-up data are incomplete, lacking detailed information on parameters such as activity levels, scoring metrics, and continued exposure to preoperative etiological factors. These omissions limit the depth of the analysis and may obscure important associations. Third, there are limitations related to the imaging data. Radiographic parameters were measured using sagittal-plane MRI rather than coronal-plane MRI, which differs from approaches commonly used in prior research. Moreover, radiomics analysis was conducted using X-ray imaging because MRI images in DICOM format were unavailable for some patients from external institutions. Finally, the ResNet18 architecture employed in this study may have inherent limitations, underscoring the need for refinement in future deep learning frameworks to improve performance. Future research should address these limitations by incorporating multi-center data and external validation cohorts to enhance the model’s generalizability and robustness. In addition, it is essential to include more comprehensive follow-up data, such as postoperative activity levels, scoring metrics, and continued exposure to preoperative etiological factors, to provide deeper insights into patient outcomes. Further efforts should also focus on integrating advanced imaging modalities, such as MRI, to extract richer radiomic and deep learning features. Finally, exploring more sophisticated deep learning architectures could optimize predictive accuracy and improve the model’s overall efficiency.

Conclusions

This study introduces a novel approach that integrates radiomics and deep learning techniques to predict femoral head collapse following multiple drilling for osteonecrosis. The DLRC model, which combines radiomics and deep learning features with clinical factors, demonstrated superior predictive performance compared to traditional clinical model. These findings highlight the potential of personalized risk assessment in guiding treatment decisions and improving outcomes for patients with osteonecrosis of the femoral head.