A Deep Learning–Based Multimodal Fusion Model for Predicting Bone Cement Leakage in Percutaneous Kyphoplasty: Development and Validation

ABSTRACT

Background

There is a lack of intelligent methodologies that effectively integrate multimodal information to predict bone cement leakage (BCL) during percutaneous kyphoplasty (PKP). This study aimed to develop and validate a deep learning (DL)–based multimodal fusion model that incorporates preoperative CT, MRI, and clinical variables to predict BCL subtypes during PKP.

Methods

This study included a retrospective internal dataset for model training and validation, a prospective internal dataset, and an external dataset for independent testing. The fusion model incorporated preoperative spinal CT, MRI, and structured clinical baseline data within a two‐stage framework. The first stage consisted of target vertebra localization based on vertebral segmentation. The second stage comprised a classification module implemented using a multibranch 3D ResNet‐50 network. Performance was compared with image‐only models, single‐modality models, and spine surgeons using the area under the receiver operating characteristic curve (AUC), the area under the precision–recall curve, and other metrics.

Results

The multimodal DL model achieved AUC values ranging from 0.795 to 0.861 in the internal test set and from 0.767 to 0.848 in the external test set for predicting BCL subtypes. Type III leakage demonstrated the highest predictive performance (internal AUC, 0.861; external AUC, 0.848). Overall, the fusion model achieved the highest AUC values and showed superior accuracy and agreement compared with spine surgeons, particularly for Type I (κ: internal, 0.439 vs. 0.030–0.075; external, 0.298 vs. 0.019–0.066) and Type IV leakage (κ: internal, 0.459 vs. 0.133–0.186; external, 0.545 vs. 0.002–0.095).

Conclusion

A two‐stage multimodal fusion DL framework enables accurate, reliable, and promisingly generalizable prediction of BCL subtypes in PKP, outperforming spine surgeons and supporting individualized preoperative decision‐making.

1. Introduction

Osteoporotic vertebral compression fracture (OVCF) is a common spinal disorder among older adults, and its incidence has increased markedly with population aging [1]. OVCFs frequently cause severe pain, spinal deformity, and functional impairment, thereby substantially reducing patients' quality of life. Percutaneous kyphoplasty (PKP), a minimally invasive surgical procedure, has become one of the standard treatments for OVCFs because of its ability to rapidly relieve pain, restore vertebral height, and enhance spinal stability [2]. However, bone cement leakage (BCL) remains one of the most common and clinically significant complications associated with PKP [3, 4]. BCL is typically classified into subtypes according to the anatomical site of leakage, each of which is associated with distinct adverse outcomes [5]. Therefore, early identification of patients at high risk for BCL and the implementation of preventive strategies are of critical clinical importance.

Recently, deep learning (DL) models have emerged as powerful tools for classification tasks in healthcare. Their end‐to‐end learning architecture and superior capacity for automatic extraction of high‐level features from imaging data provide effective support for disease screening and prediction of adverse events [6, 7, 8]. Nevertheless, current prediction of BCL during PKP predominantly relies on traditional machine learning (TML) approaches [9, 10, 11, 12], which are limited by potential information loss, workflow inefficiency, and bias introduced by manual feature extraction from medical images. Our team previously developed a DL model based exclusively on preoperative computed tomography (CT) images [13]. However, this model demonstrated suboptimal performance in certain BCL subtypes, such as segmental venous leakage, likely because CT imaging provides limited visualization of soft tissues, including paravertebral veins. Therefore, incorporating more comprehensive data sources into model training may represent a promising strategy for improving predictive performance.

Multimodal fusion models integrate information from heterogeneous data sources to extract more comprehensive and robust disease‐related features [14, 15, 16, 17]. For OVCFs, CT imaging excels in visualizing bony structures and fracture morphology, whereas magnetic resonance imaging (MRI) provides superior assessment of soft tissues, including intervertebral discs and vascular structures. In addition, structured clinical variables—such as patient age, body mass index (BMI), and previous spinal fracture—offer important biological and demographic context [11]. These data sources encompass nearly all known preoperative risk factors for BCL [18, 19, 20]. Integrating these heterogeneous modalities into a unified DL framework may leverage the complementary strengths of each modality, thereby enhancing the accuracy, robustness, and interpretability of predictive models.

This study aimed to develop a DL‐based multimodal fusion model that incorporates preoperative CT, MRI, and structured clinical baseline data to predict BCL subtypes during PKP.

2. Materials and Methods

2.1. Study Design and Data Sources

A retrospective dataset was used for model development and internal validation, followed by evaluation using a prospective internal dataset and an independent external dataset. Model development and internal validation were conducted using data from a tertiary academic teaching hospital. Consecutive inpatients who underwent single‐level PKP between January 2016 and July 2023 were retrospectively screened for eligibility. The internal prospective test dataset was derived from a prospective cohort established between August 2023 and March 2024 to evaluate adjacent segment complications following PKP for the treatment of OVCFs. This cohort was registered with the Chinese Clinical Trial Registry (registration number: ChiCTR2300073507). External validation was performed using an independent retrospective dataset comprising consecutive inpatients treated at another participating center between September 2023 and February 2024. Ethical approval was obtained from the institutional review boards of both participating centers. Because the original data collection was retrospective, the informed consent requirement was waived. All study procedures were conducted in accordance with the principles of the Declaration of Helsinki. The study followed the TRIPOD+AI statement for reporting clinical prediction models (File S1). The sample size for the prospective internal validation cohort was calculated in advance using PASS 2021. The study targeted a minimum AUC of 0.80 for each BCL subtype, with a two‐sided α of 0.05 and an allowable error corresponding to a confidence interval (CI) half‐width of 0.20. Assuming an incidence of 20%–30% for each BCL subtype, the estimated sample size required for the validation cohort was 110–155 patients.

2.2. Participants

For the retrospective training and validation cohort, the inclusion criteria were as follows: (1) age ≥ 55 years; (2) T11–L5 OVCF diagnosed based on clinical manifestations and radiographic evidence; (3) undergoing single‐level PKP; (4) availability of complete imaging data within 2 weeks preoperatively, including spinal X‐ray, CT, and MRI. The exclusion criteria were as follows: (1) previous surgery at the index vertebra; (2) coexisting spinal infection, tumor, or major spinal deformity; (3) loss of key clinical data or both postoperative lumbar X‐ray and CT performed within 14 days after surgery. Among patients with multi‐level OVCFs, only those who underwent PKP at a single index vertebra were included. As postoperative CT is considered the reference standard for BCL diagnosis, its absence within 14 days was applied as an additional exclusion criterion for the internal and external test sets to ensure reliable ground‐truth labeling. All included patients had complete preoperative imaging obtained within 2 weeks before surgery, with CT and MRI scans acquired during the same preoperative period and in the supine position to ensure consistent pathological status and minimize temporal and positional discrepancies between imaging modalities. Beyond that, the test cohorts followed the same eligibility criteria as the training and validation cohort. A random 8:2 split was conducted to divide training and validation sets (Figure 1). Detailed surgical procedures are provided in File S2 and were previously described in our published study [17].

FIGURE 1.

Flowchart of patient selection and cohort grouping.

2.3. Outcome Definition and Predictors Collection

BCL during PKP was defined as cement extending beyond the boundaries of the vertebral body on postoperative CT or X‐ray. BCL was further categorized into four anatomical subtypes: Type I, leakage through the segmental vein; Type II, leakage via a cortical defect into perivertebral soft tissue; Type III, leakage into the intervertebral disc; and Type IV, leakage into the spinal canal via basivertebral veins or a cortical defect [20, 21]. Two independent spine surgeons (Y.X. and R.C.), blinded to patients' clinical information and preoperative imaging, classified the BCL subtype based on postoperative CT or X‐ray. Discrepancies were resolved by a third senior spine expert (L.Z.), and the final classification was considered the reference standard. For the retrospective datasets, including the training and validation sets, BCL status and subtype were determined using available postoperative imaging, prioritizing postoperative CT scans whenever available and radiographs otherwise. For the internal prospective and external test cohorts, only patients with postoperative CT obtained within 14 days were included, thereby maintaining a stricter reference standard for independent assessment.

Preoperative CT images and MRI were collected in the Digital Imaging and Communications in Medicine (DICOM) format. Imaging acquisition parameters for each center are detailed in File S3. Two spine surgeons (B.W. and T.W.) independently performed vertebral contour segmentation on preoperative CT and MRI using 3D Slicer (version 4.11) to generate binary masks. These manually annotated masks served as the ground truth for segmentation model training, validation, and testing. Additionally, potentially impactive demographic and baseline variables, including age, sex, BMI, level of fracture, presence of multi‐level OVCFs, history of previous OVCFs, and time from injury to surgery [18, 20, 21, 22], were extracted from the electronic medical records by two additional spine surgeons (M.L. and T.X.). To align with the study's objective of supporting preoperative surgical planning, intraoperative variables under surgeon control, such as cement volume, were excluded. All data were de‐identified before analysis to ensure patient privacy and data security.

2.4. Spine Surgeons' Predictions

Two spine surgeons (Evaluator 1, PD, and Evaluator 2, NF), each with more than 10 years of PKP experience and who had not participated in the treatment or data collection of any patients in the test cohorts, were recruited as the clinical evaluator group. The evaluators independently predicted the BCL subtype for each patient in the internal test set using preoperative bone‐window axial CT scans, T2‐weighted sagittal MRI images, and seven structured clinical variables, consistent with the inputs provided to the DL models.

2.5. Development of the DL Model

This study developed an end‐to‐end DL prediction system comprising two primary modules: a vertebral segmentation module and a leakage risk prediction module (Figure 2). The workflow begins with CT and MRI images. First, a 3D U‐Net‐based instance segmentation network delineates the thoracolumbar and lumbar vertebrae (T11–L5), identifying the three‐dimensional spatial location of each vertebra. This module can simultaneously process CT and MRI data using an encoder–decoder architecture with skip connections to capture detailed vertebral edge features, thereby providing high‐quality anatomical references for subsequent analysis.

FIGURE 2.

Workflow of the two‐stage multimodal deep learning prediction system. FC: fully connected layers; ReLU: rectified linear unit activation function. 3D U‐Net: Three‐Dimensional U‐shaped Network. ResNet‐50: Residual Network with 50 layers.

Following segmentation, the system automatically extracts a 64 × 64 × 64 voxel cube encompassing the target vertebra and inputs it into the risk prediction module (Figure S1). This module employs a dual‐branch 3D ResNet‐50 architecture to process CT and MRI data separately, extracting deep imaging features via residual learning. Each branch produces a 2048‐dimensional feature vector encoding vertebral morphology, density distribution, and surrounding soft‐tissue information. Next, features extracted from CT and MRI modalities are fused with clinical baseline data through feature concatenation to generate a comprehensive high‐dimensional representation (Figure S2). The fused feature vector is then passed through a classifier with three fully connected layers to produce prediction probabilities for the four BCL subtypes. Softmax normalization provides quantitative risk scores for each leakage subtype, thereby supporting clinicians in preoperative planning and risk management. File S4 provides detailed specifications of the DL architecture. Additionally, we evaluated four suboptimal predictive strategies: MRI + CT (Model 2), CT alone (Model 3), MRI alone (Model 4), and clinical variables alone (Model 5), which served as control models.

2.6. Statistical Analysis and Model Evaluation

All statistical analyses were conducted using Python (version 3.11) and R software (version 4.4.3). Segmentation performance of the 3D U‐Net model was evaluated using the Dice coefficient, and vertebral mismatch rates were calculated. Predictive performance of the models and surgeons was primarily assessed using the area under the receiver operating characteristic curve (AUC) and the area under the precision‐recall curve (AUPRC). Differences between AUCs were compared using the DeLong test, and post hoc analyses were adjusted using the Holm–Bonferroni method. Calibration curves were generated using nonparametric LOESS smoothing with 1000 bootstrap resamples to estimate mean observed probabilities and corresponding 95% CIs. Clinical utility was evaluated using decision curve analysis (DCA). Binary classification performance metrics, including sensitivity, specificity, precision, F1‐score, balanced accuracy (defined as the average of sensitivity and specificity), and the Matthews correlation coefficient (MCC), were calculated using the threshold corresponding to the maximum Youden index. Reliability of BCL subtype diagnosis for both evaluators and models was assessed using the simple kappa coefficient (κ), interpreted as follows: > 0.7, good agreement; 0.4–0.7, moderate agreement; and < 0.4, poor agreement. The 95% CIs for all evaluation metrics were estimated using bootstrap resampling with 1000 iterations. A two‐sided p value < 0.05 was considered statistically significant.

3. Results

3.1. Patient Characteristics and Structured Data

A total of 874 retrospective patients were included for model development, and 126 prospective patients were used for model testing. Additionally, 52 patients from an external center were included in the external test set. The demographic and clinical baseline characteristics of the study cohorts are summarized in Table 1.

TABLE 1.

The demographic and clinical baseline characteristics of the datasets.

Variables	Training set	Validation set	Internal test set	External test set
Patients, n	701	173	126	52
Age, years	71.5 ± 9.7	71.1 ± 10.6	72.6 ± 9.0	75.0 ± 10.3
Gender, n (%)
Female	525 (74.9)	127 (73.4)	90 (71.4)	30 (57.7)
Male	176 (25.1)	46 (26.6)	36 (28.6)	22 (42.3)
BMI, kg/m2	24.7 ± 13.2	24.4 ± 3.9	24.75 ± 8.5	23.0 ± 4.2
Level of fracture, n (%)
T11	65 (9.3)	16 (9.2)	22 (17.5)	4 (7.7)
T12	195 (27.8)	51 (29.5)	30 (23.8)	13 (25.0)
L1	175 (25.0)	46 (26.6)	33 (26.2)	18 (34.6)
L2	114 (16.3)	28 (16.2)	20 (15.9)	8 (15.4)
L3–5	150 (21.4)	32 (18.5)	21 (16.7)	9 (17.3)
Multi‐level OVCFs, n (%)	117 (16.7)	24 (13.9)	12 (9.5)	8 (15.4)
Previous OVCFs, n (%)	180 (25.7)	48 (27.7)	25 (19.8)	15 (28.8)
Time from injury to surgery, days a	7.0 (2.0–19.0)	4.0 (1.0–14.0)	9.0 (3.0–15.0)	5.0 (1.8–14.3)
BCL, n (%) b
Type I	154 (22.0)	36 (20.8)	39 (31.0)	8 (15.4)
Type II	143 (20.4)	30 (17.3)	32 (25.4)	14 (26.9)
Type III	165 (23.5)	39 (22.5)	52 (41.3)	9 (17.3)
Type IV	160 (22.8)	39 (22.5)	29 (23.0)	9 (17.3)

Abbreviations: BCL: bone cement leakage; BMI: body mass index; OVCFs: osteoporotic vertebral compression fractures.

^a Results are given as the median (interquartile range).

^b Sum of percent is not equal to 100% because some samples experienced more than one subtype of BCL.

3.2. Vertebra Segmentation and Localization

The segmentation model demonstrated consistently high accuracy across imaging modalities and vertebral levels. In the internal test set, the mean Dice coefficient was 0.951 ± 0.016 for CT and 0.917 ± 0.021 for MRI. In the external test set, the mean Dice coefficient was 0.936 ± 0.026 for CT and 0.900 ± 0.035 for MRI (Table S1). For vertebral localization, mismatches occurred in CT and MRI. In the internal test set, 2.8% (3/126) of vertebrae were mismatched, including 1 mismatch on CT and 2 on MRI; these cases were attributed to lumbar sacralization and severe vertebral collapse. In the external test set, 3.9% (2/52) of vertebrae were mismatched, including 1 mismatch on CT and 1 on MRI; one case involved sacral lumbarization and the other had multi‐level previous OVCFs. Cases with segmentation mismatches were excluded from subsequent BCL risk prediction analyses to minimize the impact of segmentation errors on downstream performance.

3.3. Prediction Performance for BCL Subtypes

The multi‐input DL model (Model 1) achieved AUCs ranging from 0.795 to 0.861 in the internal test set and from 0.767 to 0.848 in the external test set for predicting BCL subtypes (Figure 3). Type III leakage was the most accurately predicted subtype, with AUCs of 0.861 (95% CI: 0.791–0.921, internal test set) and 0.848 (95% CI: 0.690–0.964, external test set). Among all models (Models 1–5), the final fused model demonstrated the highest AUCs (Figure 3; Table S2). It was significantly superior to single‐modality input models (Models 3–5) for Type I and IV prediction and superior to Models 4–5 for Type II and III prediction in the internal test set (Holm–Bonferroni p < 0.05; Table S3). In the external test set, this superiority was only significant when compared with the clinical variables–only model (Model 5). To address class imbalances, the AUPRC was also evaluated (Figure S3). The model achieved the highest AUPRC for Type III in the internal test set (0.823; 95% CI: 0.720–0.903) and for Type IV in the external test set (0.700; 95% CI: 0.336–0.931). Calibration curves indicated slight overestimation in the internal test set but generally aligned with the ideal diagonal line. Overestimations were more pronounced in the external test set, except for Type IV leakage (Figure S4). Calibration performance was further quantified using the Brier score for Model 1. The Brier scores ranged from 0.131 to 0.196 in the internal test set and from 0.099 to 0.212 in the external test set, reflecting acceptable calibration across all BCL subtypes (Table S4). DCA revealed that the multimodal DL model yielded threshold‐dependent clinical net benefit that varied according to BCL subtype and validation cohort. In the internal test set, Type III and Type IV leakage showed net benefit over both treat‐all and treat‐none strategies across a broad range of threshold probabilities, whereas Type I and Type II leakage were beneficial mainly at low‐to‐moderate thresholds. In the external test set, the net benefit was generally more limited, particularly for Type I, II, and III leakage, although Type IV maintained benefit at certain thresholds (Figure S5). For one‐time binary classification, Type III was the best‐predicted subtype in the internal test set (F1‐score: 0.762 [95% CI: 0.667–0.845]; MCC: 0.585 [95% CI: 0.437–0.714]), whereas Type IV remained the best‐predicted subtype in the external test set (F1‐score: 0.632 [95% CI: 0.286–0.833]; MCC: 0.547 [95% CI: 0.146–0.806]). Confusion matrices for BCL subtype prediction are presented in Figure S6, and classification metrics with 95% CIs, including sensitivity, specificity, precision, accuracy, balanced accuracy, F1‐score, and MCC, are summarized in Figure 4 and Table S2.

FIGURE 3.

Receiver Operating Characteristic (ROC) curves of the DL models with different inputs: MRI + CT + clinical variables (Model 1), MRI + CT (Model 2), CT alone (Model 3), MRI alone (Model 4), and clinical variables alone (Model 5) for BCL subtype prediction. (A–D) Type I–IV BCL prediction in the internal test set; (E–H) Type I–IV BCL prediction in the external test set. AUC: area under the receiver operating characteristic curve.

FIGURE 4.

Radar charts displaying seven primary performance metrics for the multimodal fusion model in both internal and external test sets. (A–D) Type I–IV BCL prediction. MCC: Matthews Correlation Coefficient. Balanced accuracy: the average of sensitivity and specificity.

3.4. Comparative Performances Between DL and Surgeons

The predictive performance of the DL model was compared with that of experienced spine surgeons using the same preoperative information. Across all BCL subtypes, the DL model demonstrated higher agreement with the reference standard than human evaluators. This difference was particularly pronounced for Type I leakage (κ, internal/external test sets: 0.439 vs. 0.030–0.075/0.298 vs. 0.019–0.066) and Type IV leakage (κ, internal/external test sets: 0.459 vs. 0.133–0.186/0.545 vs. 0.002–0.095). Confusion matrices and additional performance metrics for the evaluators' predictions are provided in Figures S7 and S8; Table S2.

4. Discussion

This study developed and validated a novel multimodal fusion DL pipeline for predicting BCL subtypes during single‐level PKP using combined CT, MRI, and structured clinical data. The DL framework comprised a 3D U‐Net module for simultaneous vertebral segmentation and localization on CT and MRI, followed by a downstream multi‐branch 3D ResNet‐50 module for BCL subtype prediction. The system demonstrated a low vertebral mismatch rate and achieved high AUCs of 0.795–0.861 and 0.767–0.848 in the internal and external test sets, respectively. Moreover, the model outperformed spine surgeons in reliability, suggesting its potential as a powerful tool for preoperative risk assessment and its promising integration into the clinical workflow for OVCF treatment.

Current mainstream approaches for predicting BCL largely rely on TML models driven by identified imaging‐ or surgery‐related predictors including identified cortical disruption, intravertebral clefts, the basivertebral foramen sign, and cement volume, and so forth [19, 23]. Although these models achieve acceptable performance, they are limited by suboptimal interobserver agreement, workflow inefficiency, and the potential loss of subtle but impactful imaging details resulting from manual feature extraction. Previously, our team developed an automated DL system based solely on preoperative CT images to predict BCL subtypes, achieving internal and external AUCs of 0.734–0.831 and 0.709–0.818, respectively [20]. This two‐stage pipeline included a segmentation and localization module followed by a multi‐output classification module. However, the model underperformed in predicting segmental venous leakage (Type I BCL), highlighting the need to incorporate multimodal data to compensate for the limited visualization of soft tissue in CT.

Since both spine CT and MRI are routinely obtained before PKP, integrating features from these two imaging modalities together with essential clinical data is feasible and capable of providing complementary information for BCL prediction. The multimodal DL system proposed in this study retained the two‐stage architecture while incorporating MRI and structured clinical data as compensatory inputs. DL‐based multimodal fusion strategies are generally categorized into early, late, and intermediate fusion [24]. Intermediate fusion combines high‐level representations from different modalities within a shared fusion layer, offering greater flexibility and better suitability for heterogeneous data. Accordingly, our DL model employed an intermediate fusion approach by concatenating 2048‐dimensional image feature vectors with clinical feature vectors, enabling flexible processing of heterogeneous imaging and clinical data.

Validated using a prospective internal dataset and an external dataset, the combined system achieved the highest AUCs of 0.795–0.861 and 0.767–0.848 among all models. The system also demonstrated higher reliability than experienced spine surgeons. Although predictive performance for Type I leakage was significantly improved compared with the CT‐only model in the internal set, it remained lower than that for other subtypes. This finding suggests that additional features, such as imaging details of discs and nerves, may provide valuable information for predicting all BCL subtypes. Moreover, predicting Type I leakage remains inherently challenging, as human evaluators also demonstrate relatively low reliability for this subtype. Although the fused model achieved the highest overall AUC, its superiority was not statistically significant compared with the CT + MRI, CT‐only, or MRI‐only models in the external test set. This may be due to the limited sample size, and future external validation studies are warranted to assess the model's generalizability more robustly.

Another key factor that may have contributed to performance decline in the external test set may be domain shift caused by differences in imaging acquisition parameters between the two centers. These differences directly impact spatial resolution, image noise, and image contrast. Although intensity normalization was used during preprocessing, domain shift may still alter the subtle texture and edge features on which the 3D ResNet‐50 backbone relies for feature extraction. Such issues are more likely to occur in models developed using single‐center data [25]. Incorporating multi‐center datasets with heterogeneous imaging protocols during model training and exploring domain adaptation techniques capable of learning scanner‐invariant feature representations may help mitigate this issue [26]. Therefore, the current model should be used cautiously when imaging parameters differ substantially from the training data. In addition, because the model was trained on a relatively inter‐type balanced cohort, its performance may vary when applied to external datasets with different class distributions. Therefore, model performance should be interpreted with caution when evaluating datasets that exhibit marked class imbalance.

The fusion model does not provide perfect predictions because important intraoperative factors that influence BCL occurrence, such as cement volume, cement viscosity, injection speed, puncture trajectory, balloon expansion, and surgeon experience, were intentionally excluded from model development. Positioned as a preoperative risk assessment tool, the DL pipeline was specifically designed to support individualized surgical planning using only information available before surgery. Despite this limitation, the DL pipeline provides discriminative preoperative risk estimates for different BCL subtypes, which may help optimize surgical strategies before intervention and potentially reduce the incidence of BCL. Furthermore, we limited the clinical variables to those routinely available, improving the model's applicability in resource‐limited settings and reducing potential input errors. Future studies could enhance predictive performance by developing a dynamic prediction framework that combines preoperative risk assessment with updating intraoperative risk. Such a framework could incorporate standardized intraoperative data to improve the accuracy of BCL risk forecasts. Integration of large language models for automated extraction of electronic medical record data may further reduce manual overload, minimize input errors, and further enhance model generalizability.

This work's novelty lies in the combination of incremental technical refinement and its primary clinical significance. From a technical perspective, employing 3D U‐Net and 3D ResNet‐50 architectures is a practical choice rather than a radical algorithmic innovation. Our main technical contribution is the development of a specialized multibranch fusion framework that effectively maps and concatenates targeted vertebral data from various modalities without requiring explicit coregistration. From a clinical perspective, the key innovation lies in the study's translation into practice and its improved predictive performance. Unlike traditional or single‐modality AI models, which often lose information or perform poorly in subtype‐specific prediction, our multimodal pipeline achieved acceptable generalizability across prospective and external validation cohorts. Moreover, the model outperforms experienced spine surgeons with > 10 years of practice, particularly in predicting Type I and Type IV leakage, for which visual assessment usually fails. Thus, it demonstrates potential to support preoperative risk stratification in cases where human assessment is inherently limited, addressing a long‐standing challenge in PKP‐related complication prediction.

This study has several limitations. First, the limited sample size constrained the exploration of more data‐intensive architectures, such as Vision Transformers. Future multi‐center studies with larger cohorts are needed to evaluate the potential of these advanced models. Second, although all patients in the test sets had postoperative CT—the gold standard for BCL assessment—patients in the internal retrospective dataset may not have had postoperative CT, which could have introduced label bias. Third, to fully validate clinical utility, the model should be developed into a software tool and prospectively tested in real‐world surgical practice to compare BCL occurrence under DL‐assisted versus standard decision‐making. Fourth, because of the small test cohort, the precision and robustness of the external validation results may be limited. Finally, sources of bias should also be acknowledged. The retrospective collection may have introduced selection bias because only patients with complete clinical and imaging data who underwent single‐segment surgery were included. Although an external test cohort was included, it originated from the same city as the primary center and included similar patient profiles and imaging protocols, which may limit institutional heterogeneity and generalizability. In addition, differences in surgical workflow, surgeon experience, and intraoperative factors difficult to standardize or accurately record, such as cement viscosity, injection speed, puncture trajectory, and intraoperative fluoroscopy, may have introduced residual confounding. Therefore, larger multicenter studies across regions with standardized imaging and intraoperative data collection are warranted.

5. Conclusion

This study developed a two‐stage multimodal fusion DL pipeline that integrates preoperative CT, MRI, and clinical baseline data to predict BCL subtypes during single‐level PKP. The model demonstrated high accuracy, reliability, and promising generalizability while outperforming experienced spine surgeons. These results suggest that the system has the potential to support intelligent, personalized risk assessment of OVCFs and to facilitate individualized preoperative decision‐making.

Author Contributions

Tianyi Wang: data curation, formal analysis, writing – original draft. Yu Xi: data curation, formal analysis, writing – original draft. Ruiyuan Chen: formal analysis, methodology, writing – original draft. Xingyu Liu: writing – review and editing, formal analysis. Dong Liu: methodology. Minghui Liang: methodology. Tianlang Xie: methodology, writing – review and editing. Baodong Wang: methodology. Aobo Wang: methodology, writing – review and editing. Ning Fan: visualization, writing – review and editing. Peng Du: visualization, writing – review and editing. Shuncheng Jiao: visualization, writing – review and editing. Yiling Zhang: conceptualization, project administration, writing – review and editing. Lei Zang: conceptualization, writing – review and editing, project administration.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Proposed 3D U‐Net algorithm framework for CT and MRI vertebra segmentation. 3D U‐Net: Three‐Dimensional U‐shaped Network. Conv: convolution. ReLU: rectified linear unit. Max pooling: maximum pooling. Transposed Conv: transposed convolution. D: depth. H: height. W: width.

Figure S2: Network architectures of the multimodal 3D ResNet‐50 CNN classifier. 3D ResNet‐50: three‐dimensional 50‐layer residual network. CNN: convolutional neural network. OVCFs: Osteoporotic Vertebral Compression Fractures. Conv: convolution. Max pooling: maximum pooling. Avg pooling: average pooling. FC: fully connected layer. ReLU: rectified linear unit.

Figure S3: Precision‐recall curves (PRC) of the deep learning (DL) models with different inputs: MRI + CT + clinical variables (Model 1), MRI + CT (Model 2), CT alone (Model 3), MRI alone (Model 4), and clinical variables alone (Model 5) for BCL subtype prediction. (A–D) Type I–IV BCL prediction in the internal test set; (E–H) Type I–IV BCL prediction in the external test set. AUPRC: area under the precision‐recall curve.

Figure S4: Calibration curves of multimodal fusion model. (A) internal test set, (B) external test set.

Figure S5: Decision curves of multimodal fusion model. (A–D) Type I–IV BCL prediction.

Figure S6: Confusion matrixes of the DL models with different inputs: MRI + CT + clinical variables (Model 1), MRI + CT (Model 2), CT alone (Model 3), MRI alone (Model 4), and clinical variables alone (Model 5) for BCL subtype prediction. (A–D) Type I–IV BCL prediction in the internal test set; (E–H) Type I–IV BCL prediction in the external test set.

Figure S7: The radar charts of the binary classification prediction metrics of the spine surgeons. (A–D) Surgeon 1's prediction for Type I–IV BCL; (E–H) Surgeon 2's prediction for Type I–IV BCL. MCC: Matthews Correlation Coefficient. Balanced accuracy: the average of sensitivity and specificity.

Figure S8: Confusion matrixes of the spine surgeons for BCL subtype prediction. (A–D) Type I–IV BCL prediction in the internal test set; (E–H) Type I–IV BCL prediction in the external test set.

Table S1: Dice coefficients of vertebral segmentation in internal and external test sets.

Table S2: Prediction performance with 95% CI of all models and evaluators.

Table S3: DeLong test of AUCs between Model 1 and other models.

Table S4: Brier scores with 95% CI of Model 1 in internal and external test sets.

File S1: TRIPOD+AI statement.

File S2: Details of surgical procedures and perioperative management.

File S3: Details of imaging scanning parameters.

File S4: Details of deep learning models architecture.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.