A clinical–radiomics model based on multiparametric MRI for discriminating solitary primary spinal tumors from solitary spinal metastases

Abstract

Objective

To develop and validate a clinical-radiomics model based on multiparametric MRI for differentiating solitary primary spinal tumors from solitary spinal metastases.

Methods

This dual-center retrospective study included 510 patients with pathologically confirmed spinal tumors, randomly split into training (n = 328), internal validation (n = 82), and external test (n = 100) cohorts. Radiomics and deep learning features were extracted from T1-weighted, T2-weighted, T2-weighted fat-suppressed, and T1-weighted contrast-enhanced sequences. Three models were constructed and compared: a radiomics model (Rad-M), a deep learning-radiomics fusion model (DRad-M), and a clinical-radiomics model (CRad-M) that integrated radiomics features with patient age. The performance of seven machine learning classifiers was evaluated for each model.

Results

The CRad-M, utilizing a logistic regression (LR) classifier, demonstrated superior performance, achieving areas under the curve (AUCs) of 0.909 and 0.824 on the internal and external test sets, respectively. It significantly outperformed both the Rad-M and DRad-M models (all p < 0.05). The incorporation of deep learning features did not yield a significant improvement over the radiomics-only model. Calibration and decision curve analyses confirmed the robust clinical utility of the CRad-M.

Conclusion

The proposed LR-based CRad-M is an effective non-invasive tool for the preoperative differentiation of solitary primary spinal tumors and solitary spinal metastases, with its performance enhanced by the integration of clinical data (age) alongside radiomic features.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02448-3.

Introduction

Bone tumors are a rare and diverse group of neoplasms [1]. Based on their differentiation characteristics and biological properties, bone tumors can be classified as benign, intermediate, or malignant, each requiring different therapeutic strategies [2]. Early diagnosis of bone tumors is crucial, as it directly influences patient prognosis and cure rates [3].

Metastasis is the most common type of malignant tumor involving the skeletal system, with the bones being the third most frequent site of metastatic cancer, following the lungs and liver. Virtually all malignant tumors can metastasize to the bones [4]. The spine is the most common site for bone metastasis, accounting for approximately 68% of cases, and many patients experience pathological fractures leading to neuronal injury and spinal cord compression [5]. While most spinal metastases are multifocal, solitary bone metastases do occur and pose a considerable diagnostic dilemma, as they can mimic the imaging appearance of solitary primary spinal tumors [6]. Although histopathological confirmation via bone biopsy is a critical component of the diagnostic strategy [4], there is a clear need for reliable non-invasive methods to differentiate these entities.

Magnetic resonance imaging (MRI) plays a critical role in the initial diagnosis of bone tumors [2]. MRI aids in assessing intramedullary and intraosseous extensions, matrix and soft tissue involvement, and vascular infiltration, playing a significant role in detecting skip lesions and tumor staging [7]. Specifically, multiparametric MRI protocols, including T1-weighted (T1), T2-weighted (T2), T2-weighted fat-suppressed (T2FS), and T1-weighted contrast-enhanced (T1CE) sequences, provide complementary information for characterizing tumor tissue and its microenvironment. Nevertheless, the overlap in morphological and imaging characteristics between solitary metastases and solitary primary spinal tumors often impedes reliable visual differentiation [8].

The rise of artificial intelligence (AI) in medical imaging offers a promising avenue to address this challenge. Recent studies have classified the benign and malignant nature, histological subtype, or tissue origin of bone tumors based on images such as radiographs [9–14], CT [15, 16], or MRI [11, 17]. Previous research has explored radiomics based on T1 and T2FS sequences for distinguishing solitary primary spinal tumors from solitary metastases [18]; however, the study is constrained by small sample sizes and a lack of validation in independent external cohorts, limiting the generalizability of their findings. Furthermore, the potential of more comprehensive multiparametric MRI protocols remains underexplored. In addition, it employed only a radiomics model and did not investigate hybrid or clinical-radiomics approaches. To address these gaps, this study aimed to develop and compare three distinct models for differentiating solitary spinal metastases from solitary primary spinal tumors: a radiomics model, a hybrid model combining radiomic and deep-learning (DL) features, and a clinical–radiomics model, all based on multiparametric spinal MRI.

Materials and methods

Patients

The study was approved by the Medical Science Research Ethics Committee of our institution (registration number M2024226), which waived the need for written informed consent due to its retrospective nature.

This dual-center retrospective study included consecutive patients from Institution A between January 2017 and December 2024 as the internal dataset, and consecutive patients from Institution B between January 2017 and March 2025 as the independent external dataset (Fig. 1).

Fig. 1.

Flowchart showing the patient selection process. All included cases, including solitary spinal metastases, were confirmed by histopathological examination of spinal lesion specimens obtained through surgery or image-guided biopsy

The inclusion criteria were: (1) patients with solitary primary spinal tumors or solitary spinal metastases confirmed by histopathological examination of spinal lesion specimens obtained through surgery or image-guided biopsy; (2) Absence of additional skeletal lesions on whole-body bone scintigraphy, also referred to as emission computed tomography (ECT), ensuring the spinal lesion was the only detectable bone lesion; and (3) contrast-enhanced MRI performed within 15 days before the surgery or biopsy collection.

The exclusion criteria were: (1) poor image quality, missing MRI sequences (sagittal T1-weighted, T2-weighted, T2FS, and T1CE images), or lack of sagittal sequences; (2) patients who had undergone surgery or treatment before spinal contrast-enhanced MRI; and (3) failed preprocessing.

The dataset from Institution A (n = 410 patients) was randomly divided into training (n = 328) and internal validation (n = 82) sets at an 8:2 ratio. The 100 patients from Institution B served as an independent external test cohort.

Image acquisition and tumor segmentation

Figure 2 depicts the study design. The MRI protocol is detailed in Supplementary Appendix S1. Manual segmentation of spinal tumor lesions for all patients was performed using the open-source software 3D Slicer (https://www.slicer.org/; version 4.8.0, MA, USA). The initial delineation was conducted by a junior musculoskeletal radiologist (Radiologist1, with three years of experience) on the T1CE sequence, with reference to other contemporaneous MRI sequences to accurately determine lesion boundaries. Four weeks later, 150 patients were randomly selected from the entire cohort and independently segmented by radiologist 1 and radiologist 2 (with 6 years of experience) to assess intra- and inter-observer reproducibility.

Fig. 2.

Study workflow. The pipeline consisted of tumor segmentation on multiparametric MRI, extraction and selection of radiomic and deep-learning features, and construction of distinct predictive radiomics, deep learning–radiomics, and clinical–radiomics models

The regions of interest (ROIs) obtained from the T1CE sequence were propagated to the T1, T2, and T2FS sequences via rigid registration using Mattes Mutual Information as the metric and gradient descent optimization. Following registration, all transformed ROIs underwent visual inspection to verify their alignment with anatomical boundaries across sequences. Tumor ROIs for all sequences (ROI_T1, ROI_T2, ROI_T2FS, and ROI_T1CE) were saved in NIfTI format for subsequent analysis.

Radiomic feature extraction

Medical imaging analysis followed a standardized radiomics workflow: (1) isotropic voxel resampling (1.0³ mm³) using B-spline interpolation; (2) feature extraction via the Python Pyradiomics module (version 3.0.1) from original images and 17 filtered variants; (3) acquisition of 1,197 radiomic features across eight categories covering intensity statistics, shape-based, filter and wavelet, and texture features. Feature extraction was performed for all patients using an identical preprocessing pipeline. All data-driven preprocessing and feature-selection steps were fitted exclusively in the training cohort, as described in the Model Construction section. Detailed information is provided in Supplementary Appendix S2.

Deep-learning feature extraction

For deep learning-based feature extraction, a ResNet50 architecture initialized with ImageNet-pretrained weights was used. For each MRI sequence, two-dimensional tumor-centered images were generated from the slice showing the largest tumor area. The tumor region was cropped according to the ROI bounding box with an additional 10% margin, clipped at the image boundary, resized to 224 × 224 pixels, and normalized using z-score normalization. Because the pretrained ResNet50 expects three-channel input, the single-channel MR image was duplicated into three channels. The final fully connected layer was replaced with a one-neuron fully connected classification head, and the network was fine-tuned using only the training cohort. The model was trained using binary cross-entropy loss and the Adam optimizer, with an initial learning rate of 1 × 10^-4, a batch size of 16, and a maximum of 50 epochs. Early stopping was applied based on the validation loss from an internal split of the training cohort. The internal validation cohort and external test cohort were not used for network training, hyperparameter tuning, or early stopping. After fine-tuning, the classification head was removed, and the 2,048-dimensional feature vector from the global average pooling layer was extracted for each MRI sequence. The four sequence-specific vectors from T1, T2, T2FS, and T1CE images were concatenated to generate an 8,192-dimensional deep-learning feature vector for each patient. Further details regarding the deep-learning framework and feature extraction process are provided in Supplementary Appendix S3.

Feature selection

To avoid data leakage, all data-driven preprocessing, feature-selection, and model-training procedures were fitted exclusively within the training cohort and then applied unchanged to the internal validation and external test cohorts. Radiomic and deep-learning features were extracted for all patients using identical feature-extraction pipelines. ComBat harmonization was performed using scanner model as the batch variable to mitigate scanner-related variability, and Z-score normalization was performed using the mean and standard deviation estimated from the training cohort. The training-derived ComBat and normalization parameters were then applied to the internal validation and external test cohorts. The distribution of scanner models across institutions is provided in Supplementary Table S1.

Feature selection was performed only within the training cohort. Radiomic features were first filtered by intra- and inter-observer intraclass correlation coefficients (ICCs). In the reproducibility analysis, the intra-observer ICCs of radiomic features were 0.88 ± 0.06 (range:0.70–0.99), and the inter-observer ICCs were 0.85 ± 0.07 (range:0.68–0.98). Only radiomic features with both intra- and inter-observer ICCs > 0.80 were retained for subsequent feature selection. ICC filtering was not applied to deep-learning features because they were extracted from ResNet50-derived latent representations rather than handcrafted ROI-based measurements. Both radiomic and deep-learning features were then subjected to variance filtering, Spearman correlation-based redundancy reduction, and Least Absolute Shrinkage and Selection Operator (LASSO) selection. Features with variance < 0.01 were removed, and one feature from each pair with Spearman correlation coefficient > 0.90 was excluded. LASSO was performed with ten-fold cross-validation within the training cohort to determine the optimal regularization parameter (λ), and features with non-zero coefficients were retained. The selected feature set, training-derived preprocessing parameters, and trained classifiers were fixed before evaluation in the internal validation and external test cohorts. The number of initial features and the number of features retained after feature-selection steps are summarized in Supplementary Table S2.

Model construction

We trained and evaluated seven machine learning classifiers on the training set: Logistic Regression (LR), Naive Bayes, Support Vector Machine (SVM), Random Forest, eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Gradient Boosting. Using the selected radiomic features, a radiomics model (Rad-M) was constructed. The fused deep learning–radiomics model (DRad-M) was built by combining the optimal deep-learning features with the radiomic features. Additionally, a clinical–radiomics model (CRad-M) was developed using clinical and radiomic features. As a supplementary analysis, an expanded clinical model was constructed using logistic regression with age, sex, lesion location, and maximum lesion diameter as predictors to evaluate whether these variables provided additional predictive value beyond age. As supplementary analyses, subgroup analyses were performed in the external test cohort according to age (≤ 50years vs.>50years) and lesion location (cervical, thoracic, and lumbar/sacrum-coccyx). The final trained models were applied without refitting.

Model performance was evaluated based on accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, F1 score, and area under the curve (AUC). Calibration curves and decision curve analysis (DCA) were used to assess clinical utility. The calibration curve evaluated the goodness-of-fit of the model with the best discriminative performance.

Statistical analysis

Statistical analysis was performed using Python (version 3.11). The Mann–Whitney U test was employed to assess age differences between metastatic and primary tumor groups within the training and internal validation sets. For the external test set, where normality assumptions were met, the independent-samples t-test was used. Differences in age across the three datasets (training, internal validation, and external test sets) were evaluated using the Kruskal–Wallis H test. Associations for categorical variables (sex and lesion location) were analyzed using Pearson’s chi-squared test, while Fisher’s exact test was applied where expected cell counts were less than 5. The 95% confidence intervals (CIs) for AUC values were calculated using DeLong’s non-parametric method. DeLong’s test was also used to evaluate diagnostic performance among the different models. We set the threshold for statistical significance at P < 0.05.

Results

Clinical characteristics of the patients

Table 1 summarizes the baseline characteristics of the 510 patients included in this study. The mean age was 46.6 ± 17.6 years, and 56.5% of patients were male. Within both the metastatic and primary tumor subgroups, age, sex, lesion location, and maximum lesion diameter were comparable across the training, internal validation, and external test cohorts. In each cohort, metastatic tumor patients were significantly older than primary tumor patients, whereas sex, lesion location, and maximum lesion diameter did not differ significantly between groups. Detailed pathological distributions of primary spinal tumors and primary sites of metastatic tumors are shown in Supplementary Table S3. Institution-level comparisons of age, sex, lesion location, diagnosis group, maximum lesion diameter, and the benign/intermediate/malignant distribution of primary spinal tumors are provided in Supplementary Table S4-S5.

Table 1.

Clinical characteristics of patients in the training, internal validation, and external test cohorts

	Training cohort (n = 328)			Internal validation cohort (n = 82)				External test cohort (n = 100)			Training vs. internal; validation vs. external test; P-value
	Metastasis	Primary	P-value	Metastasis	Primary	P-value		Metastasis	Primary	P-value	Metastasis	Primary
	109	219	-	27	55	-		31	69	-
Age	58.35 ± 12.31	40.64 ± 16.97	< 0.001	56.44 ± 13.06	40.84 ± 19.77	< 0.001		54.45 ± 13.79	44.14 ± 15.89	0.002	0.35	0.35
Sex			0.64			0.64				0.71	0.61	0.74
Male	60 (55.0%)	128 (58.4%)		12 (44.4%)	29 (52.7%)			16 (51.6%)	40 (58.0%)
Female	49 (45.0%)	91 (41.6%)		15 (55.6%)	26 (47.3%)			15 (48.4%)	29 (42.0%)
Location			0.14			0.68				0.82	0.36	0.15
Cervical	45 (41.3%)	109 (49.8%)		9 (33.3%)	26 (47.3%)			10 (32.3%)	20 (29.0%)
Thoracic	26 (23.8%)	46 (21.0%)		7 (25.9%)	12 (21.8%)			10 (32.3%)	20 (29.0%)
Lumbar	34 (31.2%)	48 (21.9%)		8 (29.6%)	12 (21.8%)			7 (22.6%)	22 (31.9%)
Sacrum-coccyx	4 (3.7%)	16 (7.3%)		3 (11.1%)	5 (9.1%)			4 (12.9%)	7 (10.1%)
Lesion Size	46.12 ± 19.23	47.49 ± 19.19	0.53	43.44 ± 14.74	45.71 ± 17.08	0.61		50.08 ± 21.36	45.81 ± 16.69	0.48	0.56	0.86

The Mann-Whitney U test was employed to assess age differences between the metastatic and primary tumor groups within the training and internal validation sets. For the external test set, the independent-samples t-test was used, as the assumptions of normality were met. Differences in age across datasets (training, internal validation, and external test sets) were evaluated using the Kruskal–Wallis H test. Group associations with categorical variables (sex and lesion location) were analyzed using the Pearson’s chi-squared test, while Fisher’s exact test was applied where the expected cell count was less than 5

Feature selection and construction of Rad-M, DRad-M, and CRad-M

A total of 1,197 radiomic features and 2,048 deep-learning features were extracted from each MRI sequence. Following feature selection, 15 features were retained to construct the Rad-M—three from T1, five from T2, two from T2FS, and five from T1CE. The DRad-M incorporated 17 features—two from DL, one from T1, six from T2, five from T2FS, and three from T1CE. The CRad-M integrated 18 features, including age and 17 radiomic features—three from T1, four from T2, four from T2FS, and six from T1CE. Notably, the T2 and T1CE sequences demonstrated relatively high predictive weighting across all models (Fig. 3c, and Supplementary Figures S1c and S2c).

Fig. 3.

Feature Selection and Model Performance for the Clinical–Radiomics Model (CRad-M). (a) Tuning parameter (λ) selection in the Lasso logistic model was chosen using tenfold cross-validation based on the minimum criteria. The optimal values of the LASSO tuning parameter (λ) are indicated by the vertical dashed lines. (b) LASSO coefficient profiles of the features. A coefficient profile plot was generated versus the selected log (λ) value using tenfold cross-validation. (c) Composition and coefficients of the final feature set, comprising patient age and 17 radiomic features, including three from T1, four from T2, four from T2FS, and six from T1CE sequences. (d) Comparative performance of the seven machine learning classifiers implemented with the CRad-M features, summarized by an accuracy distribution histogram

The CRad-M, which combined clinical information (age) with radiomic features, exhibited strong performance across all seven classifiers (Table 2 and Fig. 3d). Among these, the LR classifier achieved AUCs of 0.884 (95% CI: 0.848–0.921), 0.909 (95% CI: 0.848–0.971), and 0.824 (95% CI: 0.727–0.920) in the training, internal validation, and external test sets, respectively, with corresponding accuracy values of 0.802 (95% CI: 0.755–0.841), 0.866 (95% CI: 0.776–0.923), and 0.830 (95% CI: 0.745–0.891). The detailed feature selection and performance for Rad-M, DRad-M, and CRad-M are illustrated in Fig. 3 and Supplementary Figures S1 and S2, while the mathematical formulations of the models are provided in Table S6. The expanded clinical model did not outperform the age-only model, suggesting limited incremental value from the additional available variables (Supplementary Table S7). In supplementary subgroup analyses, the CRad-M maintained comparable discriminative performance across age strata and lesion-location subgroups (Supplementary Table S8).

Table 2.

Performance of various machine learning classifiers based on clinical-radiomics models

Classifier	Training cohort		Internal validation cohort		External test cohort
	ACC (95% CI)	AUC (95% CI)	ACC (95% CI)	AUC (95% CI)	ACC (95% CI)	AUC (95% CI)
Logistic Regression	0.802 (0.755–0.841)	0.884 (0.848–0.921)	0.866 (0.776–0.923)	0.909 (0.848–0.971)	0.830 (0.745–0.891)	0.824 (0.727–0.920)
NaiveBayes	0.729 (0.678–0.774)	0.847 (0.804–0.890)	0.695 (0.589–0.784)	0.802 (0.705–0.899)	0.690 (0.594–0.772)	0.714 (0.610–0.818)
SVM	0.686 (0.634–0.734)	0.812 (0.766–0.857)	0.817 (0.720–0.886)	0.887 (0.814–0.960)	0.640 (0.542–0.727)	0.738 (0.636–0.839)
Random Forest	0.988 (0.969–0.995)	0.999 (0.998-1.000)	0.756 (0.653–0.836)	0.822 (0.731–0.912)	0.700 (0.604–0.781)	0.703 (0.593–0.813)
XGBoost	1.000 (0.988-1.000)	1.000 (1.000–1.000)	0.793 (0.693–0.866)	0.859 (0.779–0.939)	0.780 (0.689–0.850)	0.768 (0.660–0.876)
LightGBM	0.866 (0.825–0.899)	0.959 (0.941–0.977)	0.841 (0.747–0.905)	0.879 (0.793–0.965)	0.700 (0.604–0.781)	0.703 (0.593–0.813)
Gradient Boost	0.887 (0.848–0.917)	0.945 (0.921–0.968)	0.854 (0761-0.914)	0.862 (0.774–0.950)	0.740 (0.646–0.816)	0.780 (0.686–0.874)

ACC, accuracy; AUC, area under the receiver operating characteristic curve; CI, confidence interval; SVM, support vector machine; XGBoost, extreme gradient boosting; LightGBM, light gradient boosting machine. The 95%CIs for AUCs were calculated using DeLong’s non-parametric method, and the 95%CIs for accuracy were calculated using the Wilson score method

Performance comparison among Rad-M, DRad-M, and CRad-M

Table 3; Fig. 4 present the performance of Rad-M, DRad-M, and CRad-M in the internal validation and external test sets.

Table 3.

Performance comparison of logistic regression-based Rad-M, DRad-M, and CRad-M with pairwise DeLong test results

	Training cohort		Internal test cohort		External test cohort
	ACC (95% CI)	AUC (95% CI)	ACC (95% CI)	AUC (95% CI)	ACC (95% CI)	AUC (95% CI)
Rad-M	0.729(0.678–0.774)	0.802(0.754–0.849)	0.732(0.627–0.816)	0.870(0.794–0.946)	0.780(0.689–0.850)	0.737(0.626–0.848)
DRad-M	0.732(0.681–0.777)	0.805(0.758–0.852)	0.756(0.653–0.836)	0.873(0.797–0.948)	0.720(0.625–0.799)	0.716(0.607–0.824)
CRad-M	0.802(0.755–0.841)	0.884(0.848–0.921)	0.866(0.776–0.923)	0.909(0.848–0.971)	0.830(0.745–0.891)	0.824(0.727–0.920)
Delong test	Standard error	P-value	Standard error	P-value	Standard error	P-value
Rad-M vs. DRad-M	0.009(-0.021-0.014)	0.70	0.015(-0.033-0.027)	0.86	0.017(-0.013-0.055)	0.23
CRad-M vs. Rad-M	0.040(0.030–0.185)	0.007	0.050(-0.059-0.137)	0.44	0.034(0.021–0.153)	0.01
CRad-M vs. DRad-M	0.030(0.020–0.139)	0.009	0.050(-0.061-0.134)	0.47	0.037(0.036–0.180)	0.003

The AUCs among models were compared using the DeLong test. The 95%CIs for AUCs were calculated using DeLong’s non-parametric method, and the 95%CIs for accuracy were calculated using the Wilson score method. ACC, accuracy; AUC, area under the receiver operating characteristic curve; CI, confidence interval; Rad-M, radiomics model; DRad-M, the fused deep learning-radiomics model; CRad-M, the fused clinical-radiomics model

Fig. 4.

Diagnostic Performance of the Rad-M, DRad-M, and CRad-M. (a, b) Receiver operating characteristic (ROC) curves of the three models using the Logistic Regression classifier, evaluated on the internal validation (a) and external test (b) sets. Calibration curves (c, d) and decision curve analysis (e, f) for the corresponding models and datasets

The LR-based Rad-M achieved AUCs of 0.870 (95% CI: 0.794–0.946) and 0.737 (95% CI: 0.626–0.848) in the internal validation and external test sets, respectively (Table S9). The incorporation of deep-learning features to construct DRad-M did not improve the performance significantly, with AUCs of 0.873 (95% CI: 0.797–0.948) and 0.716 (95% CI: 0.607–0.824) in the internal validation and external test sets, respectively (Table S10).

Given that the inclusion of deep-learning features did not enhance model performance, we further integrated patient age with radiomic features to develop the CRad-M. This model, using the LR classifier, demonstrated the best predictive performance, with AUCs of 0.909 (95% CI: 0.848–0.971) and 0.824 (95% CI: 0.727–0.920) in the internal validation and external test sets, and corresponding accuracy values of 0.866 (95% CI: 0.776–0.923) and 0.830 (95% CI: 0.745–0.891) (Fig. 4a and b). DeLong’s test results indicated that CRad-M significantly outperformed both Rad-M and DRad-M in the training and external test sets, whereas Rad-M and DRad-M were statistically comparable (all P > 0.05; Table 3).

Calibration curve analysis showed reasonably good agreement between predicted probabilities and observed outcomes for CRad-M (Brier scores = 0.120 and 0.147 for the internal validation and external test sets, respectively), with the curve closely approximating the ideal diagonal (Fig. 4c and d). Decision curve analysis demonstrated a consistent net clinical benefit across all threshold probabilities when using the CRad-M (Fig. 4e and f), supporting its potential for real-world clinical implementation. Comprehensive evaluation metrics for the three models across all seven machine learning classifiers are detailed in Supplementary Tables S11–S13.

Performance comparison between clinical model and CRad-M

We observed that age alone was a significant distinguishing factor between the two groups, with a notable difference in mean age between the metastasis group and the primary tumor group (Table 1). A univariate logistic regression model based exclusively on age yielded AUCs of 0.739 (95% CI: 0.629–0.847) and 0.688 (95% CI: 0.577–0.800) on the internal and external test sets, respectively (Table 4). To determine whether radiomics features offered additional diagnostic value beyond age, we compared the performance of the CRad-model against the age-only baseline using the DeLong test. The CRad-model demonstrated significantly superior discriminative ability, achieving AUCs of 0.909 (95% CI: 0.848–0.971) and 0.824 (95% CI: 0.727–0.920) on the internal and external test sets (Table 3). The DeLong test produced a p-value < 0.001, indicating that incorporating radiomics features significantly enhanced predictive performance relative to using age alone. These findings confirm that the radiomics component provides meaningful incremental value beyond simple patient demographics.

Table 4.

Performance comparison of logistic regression-based clinical model and CRad-M with pairwise DeLong test results

Metrics	Clinical model
Training cohort
ACC (95% CI)	0.695 (0.643–0.742)
AUC (95% CI)	0.769 (0.718–0.820)
P-value	< 0.001
Internal test cohort
ACC (95% CI)	0.659 (0.551–0.752)
AUC (95% CI)	0.739 (0.629–0.847)
P-value	< 0.001
External test cohort
ACC (95% CI)	0.710 (0.615–0.790)
AUC (95% CI)	0.688 (0.577-0.800)
P-value	< 0.001

The AUCs among models were compared using the DeLong test. The 95%CIs for AUCs were calculated using DeLong’s non-parametric method, and the 95%CIs for accuracy were calculated using the Wilson score method. ACC, accuracy; AUC, area under the receiver operating characteristic curve; CI, confidence interval

Discussion

Discriminating solitary primary spinal tumors from metastases based on preoperative imaging remains a significant clinical challenge, yet it is critical for guiding treatment decisions. Radiomics offers a promising solution by extracting quantitative features from medical images to aid in diagnosis and classification, as demonstrated in various studies [18–22]. Deep learning, particularly convolutional neural networks (CNNs), has gained prominence for its ability to autonomously learn discriminative features from large datasets [10, 23, 24].

This study developed and externally validated a clinical–radiomics model for differentiating solitary primary spinal tumors from solitary spinal metastases. Integrating deep-learning features did not yield a significant performance improvement over the radiomics model alone, whereas incorporating age formed the CRad-M and achieved the best overall performance, with AUCs of 0.884, 0.909, and 0.824 in the training, internal validation, and external test cohorts, respectively. This performance was superior to that of the radiomics-only model (AUCs: 0.802, 0.870, and 0.737). The contribution of age is consistent with clinical epidemiology: primary benign spinal tumors predominantly occur in patients aged 31–50years (59.90%), primary malignant tumors peak in the 51–60-year group (27.34%), whereas 94.53% of metastatic spinal tumors occur in patients older than 50 years, with 76.95% occurring in those older than 60 years [25].

However, age should be interpreted as a stable and readily available clinical variable rather than a tumor-specific imaging biomarker. The lower AUC in the external test cohort may be related to inter-institutional heterogeneity, including differences in scanner settings, imaging protocols, patient composition, pathological subtype distribution, metastatic primary sites, and lesion characteristics. Because radiomic and deep-learning features are image-derived, they may be more sensitive to these variations, which may explain the larger external performance decrease of Rad-M and DRad-M. In contrast, age may have contributed to the better external stability of CRad-M. Importantly, CRad-M remained superior to the age-only model in both validation cohorts, indicating that radiomic features provided additional diagnostic information beyond demographic differences. Although age contributed substantially to model performance, the model also maintained acceptable discrimination across age-stratified and lesion-location subgroups, suggesting that its performance was not solely driven by age. These findings support the use of age as the only clinical variable in the final CRad-M to maintain model parsimony, although more comprehensive clinical information may further improve performance in future studies.

Beyond age, the selected radiomic features may provide interpretable information related to tumor heterogeneity. The most influential features were mainly derived from T2, T2FS, and T1CE sequences, which are clinically important for evaluating marrow replacement, edema, necrosis, and contrast-enhancement patterns in spinal tumors. These findings suggest that the radiomics component of CRad-M captures quantitative imaging information beyond demographic variables, although the biological interpretation of individual radiomic features should remain cautious and requires further radiologic–pathologic validation (Table S14).

The integration of deep-learning features did not demonstrate a significant advantage over radiomic features alone. First, although 8192 ResNet50-derived features were initially extracted for each patient, the training sample size was relatively small compared with the feature dimensionality, and only two deep-learning features were ultimately retained after feature selection. This suggests limited independent discriminative value of most deep-learning features in this cohort. Deep-learning models are highly dependent on large-scale, well-annotated datasets [26–28], and the limited size of our training cohort may have restricted the generalizability of the fine-tuned ResNet50 features [29–31]. Second, ResNet50 was used as a two-dimensional slice-based feature extractor, whereas spinal tumors are three-dimensional lesions with complex craniocaudal extension and heterogeneous internal architecture. Therefore, two-dimensional deep-learning features may not fully capture volumetric tumor heterogeneity. Future studies with larger multicenter datasets may benefit from three-dimensional deep-learning architectures and more specialized medical imaging pretraining strategies.

Among the evaluated classifiers, logistic regression achieved the best overall performance. In contrast, several complex tree-based ensemble classifiers, such as Random Forest, XGBoost, LightGBM, and Gradient Boosting, showed near-perfect performance in the training cohort but declined substantially in the validation cohorts, suggesting potential overfitting. The simpler linear decision boundary and inherent regularization of LR may provide a better bias–variance balance for datasets with limited sample size and high-dimensional features [32, 33].

Our work demonstrates clear advancements over previous studies. We established a robust clinical cohort of 510 patients with independent external validation, far exceeding the scale of earlier works (n = 135) [18]. Technically, our model leverages a richer multiparametric MRI dataset than previous study [18], incorporating not only T1 and T2FS but also T2 and T1CE sequences. Methodologically, we conducted a systematic exploration that included conventional radiomic and deep-learning features, thereby providing a more exhaustive analysis for differentiating spinal tumors from metastases.

This study has several limitations. First, despite the implementation of strict inclusion and exclusion criteria, the retrospective design inherently carries a risk of selection bias that cannot be fully eliminated. Second, the manual delineation of ROIs remains a subjective and time-consuming process. Although intra- and inter-observer reproducibility analyses were performed to retain stable radiomic features, manual segmentation may still introduce variability in practical clinical settings, particularly across radiologists with different levels of experience or across institutions with heterogeneous imaging protocols. This may affect the reproducibility and efficiency of model deployment. Therefore, future development of robust automatic or semi-automatic segmentation algorithms would help improve the reproducibility, efficiency, and objectivity of the proposed workflow. Third, although our model is based on information-rich, multi-sequence contrast-enhanced MRI, computed tomography (CT) offers unique advantages in evaluating bone matrix mineralization and periosteal reaction, particularly advantageous for lesions located in complex anatomical structures [7, 34]. CT was not incorporated due to considerations regarding radiation exposure and the practical challenges of acquiring and co-registering multimodal preoperative images; this represents a potential direction for further enhancing model performance. Fourth, this study did not include a reader study comparing CRad-M with radiologists’ visual assessment or evaluating whether model assistance could improve radiologists’ diagnostic performance. Therefore, although CRad-M showed promising discrimination in the external test cohort, its incremental value in real-world radiology workflow remains to be validated. Previous studies have suggested that artificial intelligence models may support radiologists in musculoskeletal tumor assessment [10, 35]. Future prospective reader studies are warranted to determine whether CRad-M can improve diagnostic accuracy, confidence, or efficiency in clinical practice. Finally, the study is limited by the inherent rarity of both solitary primary spinal bone tumors and solitary spinal metastases, resulting in a relatively small sample size, particularly for the metastasis cohort. This may affect the generalizability of the model, necessitating future validation through multicenter collaborations to expand the cohort size.

Conclusion

In conclusion, we developed and validated a clinical–radiomics model that demonstrated robust performance in differentiating solitary primary spinal tumors from solitary spinal metastases. With high accuracy confirmed in an independent external test set, this model may serve as a valuable non-invasive tool to support preoperative decision-making and facilitate personalized treatment planning for patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

Rad-M

Radiomics Model

DRad-M

Deep learning-Radiomics Model

CRad-M

Clinical-Radiomics Model

LASSO

Least Absolute Shrinkage and Selection Operator

LR

Logistic Regression

SVM

Support Vector Machine

LightGBM

Light Gradient Boosting Machine

XGBoost

eXtreme Gradient Boosting

CNN

Convolutional Neural Network

ROI

Region of Interest

T1

T1-weighted

T1CE

T1-weighted Contrast-Enhanced

T2

T2-weighted

T2FS

T2-weighted Fat-Suppressed

AUC

Area Under the Curve

DCA

Decision Curve Analysis

NPV

Negative Predictive Value

PPV

Positive Predictive Value

Author contributions

CW and JX contributed equally to this work. CW and JX conceived the study, curated the data, and performed the primary analyses. CW and MZ contributed to data collection and verification. CW and DH assisted with methodology and statistical analysis. CW and DH contributed to interpretation of results and manuscript revision. NL supervised the study, provided critical revisions, and is the corresponding author. All authors read and approved the final manuscript.

Funding

This study has received funding by the National Natural Science Foundation of China (grant number 82371921).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Ethics Committee of Peking University Third Hospital, and the requirement for written informed consent was waived. All procedures performed in this study were conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.