Development and multi-institutional validation of a deep learning algorithm for predicting cervical cord compression using dynamic cervical lateral radiographs

Abstract

Although magnetic resonance imaging (MRI) is the gold standard for diagnosing degenerative cervical myelopathy (DCM), its cost and limited availability can delay diagnosis. Deep learning (DL) models with convolutional neural networks (CNNs) may offer a screening alternative with plain radiographs. We aimed to develop a CNN-based DL algorithm to predict spinal cord compression (SCC) using dynamic cervical lateral radiographs (flexion, neutral, extension), perform multi-institutional validation, and identify potential causal features using gradient-weighted class activation mapping (Grad-CAM) analysis. 7878 patients who underwent both cervical radiographs and MRI at a tertiary center were labeled as SCC or control based on T2-weighted sagittal MR images and assigned to training (80.0%), validation (10.0%), and test (10.0%) sets. Ten ImageNet-pretrained architectures were trained on single-position, combined, and combined-plus-demographics models. External validation included 575 patients from an independent community hospital. The combined model with VGG-16 achieved the highest area under the receiver operating characteristic curve of 0.888 internally and 0.820 externally. Grad-CAM highlighted regions that may correspond to disc herniations, osteophytes, ossification of the posterior longitudinal ligament, and segmental instability. These results suggest that our algorithm may serve as a cost-effective screening tool with the potential to enhance diagnostic efficiency and clinical outcomes in DCM.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-56657-5.

Introduction

Degenerative cervical myelopathy (DCM) is a spinal cord dysfunction caused by spinal cord compression (SCC)¹. It is characterized by various neurological deteriorations, including motor weakness, neuropathic pain, functional impairment, and long-term disability, leading to reduced quality of life^2,3. Early detection and appropriate management of DCM before progression to irreversible cord dysfunction are crucial for achieving favorable outcomes⁴. However, in the early stages of DCM, patient symptoms may be subtle, leading to delays in diagnosis and worsening of symptoms before visiting a care provider^5,6.

Although magnetic resonance imaging (MRI) is the gold standard diagnostic modality for DCM^7,8, it has significant shortcomings such as high cost and limited immediate accessibility⁹. It has been estimated that approximately 66% of the global population lacks access to MRI, and MRI density varies substantially across income levels¹⁰. Furthermore, variation in MRI accessibility has been reported across healthcare settings, even within the same healthcare system¹¹. On the other hand, radiographs are a valuable screening tool for DCM because of their low cost, easy availability, and rapid operability¹². Triage for the possibility of DCM can be conducted by a physician’s interpretation of cervical radiographs, along with patients’ history and physical examination in a real clinical setting. However, while several radiological findings and parameters such as overall degenerative changes, regional or global alignment, anterior-posterior canal diameter, reduced disc height, and ossified lesions may allow physicians to suspect DCM¹, they may become reluctant to perform MRI when patients’ symptoms are subtle⁹. Therefore, patients requiring MRI for the diagnosis of DCM may not have been screened.

A convolutional neural network (CNN) is a representative deep learning (DL) model characterized by its ability to classify images and detect objects^13,14. Recently, AI-based approaches in spine imaging have expanded beyond plain radiographs to include MRI-based models for detecting canal stenosis, assessing spinal cord compression, and supporting clinical decision-making^7,15–17. These MRI-based approaches have shown promising results in evaluating spinal cord pathology. However, radiograph-based models remain clinically important as a first-line screening tool due to the accessibility and cost-effectiveness of radiographs. Several previous investigations have been conducted to predict cervical SCC or DCM from radiographs using DL algorithms^9,18. Nevertheless, these radiograph-based studies generally included relatively small sample sizes from a single institution. Moreover, these studies utilized only neutral lateral radiographs although assessing positional changes is important for diagnosing DCM as dynamic factors such as translational or angular instability can be significant contributors to SCC¹.

Therefore, the purpose of this study was to develop a DL algorithm using a CNN to identify imaging-based SCC in the cervical spine using dynamic cervical lateral radiographs and to evaluate its performance through multi-institutional validation. We also aimed to identify inspection points indicating potential causal lesions using gradient-weighted class activation mapping (Grad-CAM) analysis.

Methods

Patient cohort

The current study was approved by the Institutional Review Board of Seoul National University Hospital, and a waiver of informed consent was obtained. This study adhered to relevant guidelines and regulations, including the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) guidelines. Data were collected from all consecutive patients who underwent plain radiographs and MRI of the cervical spine between January 2007 and December 2022 at a single tertiary care referral center, with maximum 180 days between two imaging studies. A maximum interval of 180 days was allowed to capture the same clinical condition while preserving real-world applicability and adequate sample size. Cervical spine MRI was typically indicated for patients with symptoms suggestive of cervical radiculopathy or myelopathy for differential diagnosis. Patients with deformities, tumors, infections, trauma, diseases affecting bone quality or morphology, or a history of cervical spine surgery were excluded. Additionally, 202 patients with SCC at the atlantoaxial segment were excluded to focus the analysis on subaxial cervical levels. The detailed exclusion criteria are listed in Table 1.

Table 1.

Exclusion criteria.

Deformity

Infection

Trauma

Tumorous lesion

Bone metastasis

Primary bone tumor

Intramedullary tumor

Intradural extramedullary tumor

Hematologic disorders or malignancy

Diseases affecting bone quality or morphology

Spondyloarthropathy

Paget’s disease

Achondroplasia

Klippel-Feil syndrome

Chiari malformation

Hirayama disease

History of previous cervical spine surgery

Spinal cord compression at the atlantoaxial segment

Absence of at least one cervical spine lateral radiograph in flexion, neutral, or extension position

Among the 9079 identified patients, 1201 patients were excluded. The remaining 7878 patients were split into training (6302, 80%), validation (787, 10%), and test (789, 10%) sets (Fig. 1). Only the training and validation sets were used for training and searching for the most optimal set of hyperparameters, while the test set was used exclusively for final internal validation of the trained model. One-way analysis of variance assessed the difference in age distribution, and the χ2 test evaluated the heterogeneity of sex and label distributions across sets, with p-values < 0.05 indicating statistical significance.

Fig. 1.

Flow diagram illustrating the process of data collection and assignment to training, validation, and test sets.

Data preparation: acquisition of ground truth labels

Ground-truth labels were obtained from T2-weighted sagittal MR images. To ensure consistency, cervical canal stenosis was evaluated from C2-3 to C6-7 by a single spine specialist with more than 5 years of experience by evaluating MR images and referring to official radiology reports. The degree of cervical canal stenosis was classified into four grades according to the Kang system¹⁹: (1) grade 0, absence of cervical canal stenosis; (2) grade 1, cervical canal stenosis with obliteration of > 50% of the subarachnoid space without cord deformity; (3) grade 2, cervical canal stenosis with cord deformity but without signal change; and (4) grade 3, signal change in the spinal cord near the level of cord compression on T2-weighted images. Patients with grade 2 or 3 were designated positive for SCC, whereas those with grade 0 or 1 were designated negative, in a binary fashion. In cases of uncertainty or discrepancy between the evaluator’s grading and the radiology reports, the corresponding author finalized the grades.

Development of deep learning algorithm

For each patient, cervical lateral radiographic images were collected in three positions – flexion, neutral, and extension. The images were resized and padded with zeros to a square shape of (224, 224). Each patient provided only one set of dynamic radiographs consisting of three positions, and no modifications such as region-of-interest cropping or color adjustments were applied. In attempts to maximize the generalizability of the model, rigorous data augmentation was applied during the training phase. This included random translation of up to 30 pixels, rotation between ± 15°, scaling of ± 10%, brightness change of ± 10%, contrast of ± 10%, and random horizontal flip. Basic demographic features, including age and sex, were utilized for further model development.

First, binary classification models were trained on radiographs of each position, resulting in three distinct models: flexion, neutral, and extension models, which were collectively termed “single-position models” for subsequent analysis. To ensure a fair and generalizable comparison among methodologies, 10 different ImageNet-pretrained backbone architectures were selected and trained: ResNet-34, VGG-16, EfficientNetV2-S, EfficientNetV2-M, ConvNeXt-Tiny, ConvNeXt-Small, small vision transformer with a patch size of 16 (ViT-S/16), base vision transformer with a patch size of 16 (ViT-B/16), tiny Swin transformer (Swin-Tiny), and small Swin transformer (Swin-Small)^20–25. Binary cross-entropy loss and ADAM optimizer were utilized, and the optimal set of hyperparameters was determined through random search within a commonly accepted search space²⁶. To ensure reproducibility, detailed information regarding the hyperparameter search process and training configurations is provided in Supplementary Table S1.

Afterward, complex models were proposed by combining the results obtained from each of the single-position models. First, the combined model incorporates dynamic series of radiographs by running three single-position models on their respective radiograph sequences and averaging their output values to generate the final prediction. This ensemble approach was selected for practical reasons to reduce computational load and training time, while independently preserving the distinct clinical features extracted from each specific posture. Additionally, the combined-plus-demographics model extends this approach by stacking a logistic regression model that takes the outputs of the single-position models along with age and sex as input. The structures of all models are illustrated in Fig. 2.

Fig. 2.

Structural specification of the implemented models.

The DL models were trained with PyTorch (version 2.6.0), Scikit-Learn (version 1.6.1), and Python (version 3.11) on a computer with the following hardware specifications: Intel(R) Core(TM) i9-14900KF CPU, NVIDIA RTX 4090 24 GB graphics card, and 64 GB of DDR5 RAM.

Evaluation of algorithm performance: internal validation

The area under the receiver operating characteristic curve (AUC) on the test set was considered the primary outcome measure. Additionally, auxiliary metrics, including accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score, were measured at Youden’s J point, a decision threshold at which the sum of sensitivity and specificity is maximized²⁷.

To assess the impact of using a dynamic series of radiographs compared to single-position radiographs, DeLong’s test was utilized to compare the correlated ROC curves of each single-position model with the combined model, comparing the models with the largest validation AUC for each configuration. A comparison between the combined model and the combined-plus-demographics model was also conducted using DeLong’s test to evaluate the effect of demographics on model performance.

Grad-CAM analysis

Grad-CAM analysis generates a heatmap by computing the gradient of the predicted class score with respect to some convolutional layer, emphasizing the most relevant regions in the input image²⁸. This allows for an intuitive interpretation of the model’s decision-making process by revealing which anatomical structures contribute most to the classification. In this study, specifically, this technique was applied to visualize the gradients, highlighting areas the model identified as informative for distinguishing SCC-positive and SCC-negative images. For qualitative assessment, we selected images from the test set to explore the explainability of the best-performing model, the model with the highest test AUC. The Grad-CAM results were compared with T2-weighted sagittal MR images to assess whether the model accurately identified the cord compression site.

External validation of algorithms

To assess the generalizability of our algorithms, external validation was performed using an independent dataset extracted from another institution, a community hospital that represents a different healthcare level compared to the tertiary care referral center. The study protocol was approved by the Institutional Review Board of Bumin Hospital Seoul, and informed consent was waived. All procedures adhered to relevant guidelines and regulations. All consecutive patients who underwent cervical radiographs and MRI between August 2023 and December 2024 with a maximum interval of 180 days were included. After applying the same exclusion criteria, 575 patients were enrolled. The AUC and performance metrics were calculated following the same methodology used for internal validation.

Results

Baseline characteristics

Comparisons between the training, validation, and test sets from the tertiary referral center data set used for training and internal validation showed no significant differences in demographics, including the mean age and sex ratio (Table 2). No significant difference was observed in the ratio of patients with and without cord compression between the groups. The overall prevalence of SCC was approximately 59.5% in the internal dataset (59.6% in the training set, 58.8% in the validation set, and 59.1% in the test set), whereas it was 42.3% in the external validation dataset. Among SCC-positive cases, the distribution of grade 2 and grade 3 was as follows: in the training set, 2,542 (67.7%) were grade 2 and 1,213 (32.3%) were grade 3; in the validation set, 316 (68.3%) were grade 2 and 147 (31.7%) were grade 3; and in the test set, 304 (65.2%) were grade 2 and 162 (34.8%) were grade 3. There was no statistically significant difference in the distribution of grade 2 and grade 3 cases across the datasets (p = 0.529).

Table 2.

Baseline characteristics of internal and external validation datasets.

	Internal validation				External validation (n = 575)
	Training set (n = 6302)	Validation set (n = 787)	Test set (n = 789)	p-value*
Age (years)	56.1 ± 14.3	56.0 ± 14.3	56.3 ± 14.4	0.931	53.3 ± 14.0
Sex				0.137
Male	3254 (51.6)	377 (47.9)	400 (50.7)		285 (49.6)
Female	3048 (48.4)	410 (52.1)	389 (49.3)		290 (50.4)
Cord compression				0.895
Negative (Grade 0,1)	2547 (40.4)	324 (41.2)	323 (40.9)		332 (57.7)
Positive (Grade 2,3)	3755 (59.6)	463 (58.8)	466 (59.1)		243 (42.3)

Values are presented as mean ± standard deviation or number (%). (*) denotes comparisons between groups in the internal validation dataset.

Performance of algorithms: internal validation

The performance metrics of single-position models for each base architecture are presented in Table 3 (top three models by AUC; see Supplementary Table S2 for complete results). Among single-position models, the extension model with ConvNeXt-Small architecture exhibited the highest test AUC of 0.870 (95% confidence interval [CI] 0.846–0.894). At Youden’s J point, it demonstrated an accuracy of 0.783 (95% CI 0.755–0.812), sensitivity of 0.768 (95% CI 0.730–0.807), specificity of 0.805 (95% CI 0.762–0.848), PPV of 0.850 (95% CI 0.816–0.884), NPV of 0.707 (95% CI 0.660–0.753), and F1 score of 0.807 (95% CI 0.771–0.844).

Table 3.

Algorithm performance metrics of single-position models on the test set (top three models by AUC).

	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	F1 score (95% CI)
Flexion
EfficientNetV2-S	0.865 (0.840, 0.890)	0.796 (0.768, 0.824)	0.792 (0.755, 0.829)	0.802 (0.758, 0.845)	0.852 (0.819, 0.886)	0.728 (0.681, 0.774)	0.821 (0.785, 0.856)
EfficientNetV2-M	0.870 (0.846, 0.894)	0.792 (0.764, 0.820)	0.777 (0.739, 0.815)	0.814 (0.772, 0.857)	0.858 (0.824, 0.891)	0.717 (0.671, 0.763)	0.815 (0.779, 0.851)
Swin-Tiny	0.868 (0.844, 0.893)	0.788 (0.760, 0.817)	0.783 (0.746, 0.821)	0.796 (0.752, 0.840)	0.847 (0.813, 0.881)	0.718 (0.671, 0.764)	0.814 (0.778, 0.850)
Neutral
VGG-16	0.875 (0.851, 0.898)	0.800 (0.772, 0.828)	0.811 (0.776, 0.847)	0.783 (0.738, 0.828)	0.844 (0.810, 0.877)	0.742 (0.695, 0.788)	0.827 (0.792, 0.862)
EfficientNetV2-M	0.865 (0.840, 0.889)	0.790 (0.761, 0.818)	0.792 (0.755, 0.829)	0.786 (0.742, 0.831)	0.842 (0.808, 0.877)	0.724 (0.677, 0.770)	0.816 (0.781, 0.852)
ConvNeXt-Tiny	0.862 (0.837, 0.887)	0.797 (0.769, 0.825)	0.839 (0.806, 0.872)	0.737 (0.689, 0.785)	0.821 (0.787, 0.856)	0.760 (0.713, 0.808)	0.830 (0.796, 0.864)
Extension
VGG-16	0.855 (0.829, 0.881)	0.792 (0.764, 0.820)	0.856 (0.824, 0.888)	0.700 (0.650, 0.750)	0.804 (0.770, 0.839)	0.771 (0.723, 0.819)	0.830 (0.796, 0.863)
ConvNeXt-Tiny	0.855 (0.829, 0.880)	0.795 (0.766, 0.823)	0.845 (0.813, 0.878)	0.721 (0.672, 0.770)	0.814 (0.779, 0.849)	0.764 (0.716, 0.812)	0.829 (0.796, 0.863)
ConvNeXt-Small	0.870 (0.846, 0.894)	0.783 (0.755, 0.812)	0.768 (0.730, 0.807)	0.805 (0.762, 0.848)	0.850 (0.816, 0.884)	0.707 (0.660, 0.753)	0.807 (0.771, 0.844)

AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value, NPV, negative predictive value.

The combined model with VGG-16 architecture yielded the highest test AUC of 0.888 (95% CI 0.865–0.910) among all models, representing the best-performing model (Table 4, top three models by AUC; see Supplementary Table S3 for complete results). It achieved an accuracy of 0.814 (95% CI 0.787–0.841), sensitivity of 0.807 (95% CI 0.771–0.843), specificity of 0.824 (95% CI 0.782–0.865), PPV of 0.868 (95% CI 0.837–0.900), NPV of 0.747 (95% CI 0.702–0.792), and F1 score of 0.836 (95% CI 0.802–0.871). The ROC curve and confusion matrix of the model are plotted in Fig. 3.

Table 4.

Algorithm performance metrics of combined and combined-plus-demographics models on the test set (top three models by AUC).

	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	F1 score (95% CI)
Combined
VGG-16	0.888 (0.865, 0.910)	0.814 (0.787, 0.841)	0.807 (0.771, 0.843)	0.824 (0.782, 0.865)	0.868 (0.837, 0.900)	0.747 (0.702, 0.792)	0.836 (0.802, 0.871)
ConvNeXt-Small	0.880 (0.857, 0.903)	0.802 (0.774, 0.830)	0.792 (0.755, 0.829)	0.817 (0.775, 0.859)	0.862 (0.829, 0.895)	0.731 (0.686, 0.777)	0.826 (0.790, 0.861)
Swin-Tiny	0.878 (0.855, 0.902)	0.801 (0.773, 0.829)	0.803 (0.766, 0.839)	0.799 (0.755, 0.842)	0.852 (0.819, 0.885)	0.737 (0.691, 0.783)	0.827 (0.792, 0.861)
Combined-plus-demographics
VGG-16	0.883 (0.860, 0.906)	0.809 (0.781, 0.836)	0.807 (0.771, 0.843)	0.811 (0.768, 0.854)	0.860 (0.828, 0.893)	0.744 (0.699, 0.790)	0.833 (0.798, 0.867)
ConvNeXt-Tiny	0.876 (0.852, 0.899)	0.807 (0.780, 0.835)	0.818 (0.783, 0.853)	0.793 (0.748, 0.837)	0.850 (0.817, 0.883)	0.751 (0.705, 0.797)	0.834 (0.800, 0.868)
ConvNeXt-Small	0.879 (0.856, 0.903)	0.804 (0.776, 0.831)	0.796 (0.760, 0.833)	0.814 (0.772, 0.857)	0.861 (0.828, 0.893)	0.735 (0.689, 0.780)	0.827 (0.792, 0.862)

AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value, NPV, negative predictive value.

Fig. 3.

(a) Receiver operating characteristic curve for the combined model with VGG-16 backbone architecture, the best proposed model with the highest AUC during internal validation. (b) Confusion matrix of the combined model VGG-16 backbone architecture at Youden’s J point. AUC, area under the receiver operating characteristic curve.

The test AUC values of all models are shown in Fig. 4. No significant differences were observed among single-position models (flexion vs. neutral, p = 0.613; neutral vs. extension, p = 0.605; extension vs. flexion, p = 0.980). The combined model demonstrated statistically significant superiority over all single-position models, with test AUC improvements of 0.013 over the neutral model (p = 0.003), 0.018 over the flexion model (p = 0.010), and 0.018 over the extension model (p = 0.021). When comparing the combined model with the combined-plus-demographics model, the inclusion of demographics decreased performance, although the difference was not statistically significant, resulting in a slight AUC decline of 0.004 (p = 0.080).

Fig. 4.

Parallel coordinates plot showing the AUC values of all models with 10 backbone architectures in internal validation. AUC, area under the receiver operating characteristic curve.

Grad-CAM analysis

Grad-CAM analysis was conducted on the final convolutional layer of the best model, the combined model with VGG-16 backbone architecture. Despite the low resolution, we decided to investigate the final layer because it integrates the gradients from earlier layers and best indicates the most decisive features. Hot lesions highlighted by Grad-CAM analysis in the SCC group, which appeared to correspond to potential causal lesions, could be categorized into soft disc herniations, osteophytes around the posterior aspect of the disc space, ossification of the posterior longitudinal ligament (OPLL), and segmental instability (Fig. 5). In the control group, these lesions tended to either be invisible or show decreased activation.

Fig. 5.

Visualization of potential causal lesions of spinal cord compression by Grad-CAM analysis on the combined model with VGG-16 backbone architecture. Identified inspection points can be classified into (a) soft disc herniations, (b) osteophytes around the disc space, (c) ossification of the posterior longitudinal ligament, and (d) segmental instability. Grad-CAM, gradient-weighted class activation mapping.

External validation of algorithms

The performance metrics of single-position models in external validation are presented in Table 5 (top three models by AUC; see Supplementary Table S4 for complete results). Among all models in external validation, the combined-plus-demographics model with ConvNeXt-Small architecture demonstrated the highest AUC of 0.833 (95% CI 0.798–0.869) (Table 6, top three models by AUC; see Supplementary Table S5 for complete results). The best-performing model in internal validation, the combined model with VGG-16 architecture, achieved the AUC of 0.820 (95% CI 0.784–0.857) in external validation. This model yielded an accuracy of 0.757 (95% CI 0.722–0.793), sensitivity of 0.730 (95% CI 0.673–0.786), specificity of 0.777 (95% CI 0.732–0.822), PPV of 0.700 (95% CI 0.643–0.758), NPV of 0.801 (95% CI 0.758–0.845), and F1 score of 0.715 (95% CI 0.658–0.772).

Table 5.

Algorithm performance metrics of single-position models in external validation (top three models by AUC).

	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	F1 score (95% CI)
Flexion
EfficientNetV2-S	0.821 (0.784, 0.857)	0.743 (0.708, 0.779)	0.823 (0.774, 0.871)	0.687 (0.637, 0.737)	0.652 (0.598, 0.706)	0.844 (0.801, 0.888)	0.728 (0.674, 0.781)
EfficientNetV2-M	0.820 (0.784, 0.857)	0.750 (0.715, 0.786)	0.747 (0.691, 0.802)	0.753 (0.707, 0.799)	0.683 (0.627, 0.740)	0.806 (0.762, 0.850)	0.714 (0.657, 0.770)
ConvNeXt-Small	0.810 (0.772, 0.847)	0.747 (0.711, 0.783)	0.751 (0.696, 0.806)	0.744 (0.697, 0.791)	0.677 (0.620, 0.733)	0.807 (0.763, 0.851)	0.712 (0.656, 0.768)
Neutral
EfficientNetV2-M	0.816 (0.779, 0.853)	0.743 (0.708, 0.779)	0.785 (0.732, 0.837)	0.714 (0.665, 0.762)	0.662 (0.607, 0.717)	0.823 (0.779, 0.867)	0.718 (0.663, 0.773)
ConvNeXt-Small	0.812 (0.775, 0.849)	0.750 (0.715, 0.786)	0.747 (0.691, 0.802)	0.753 (0.707, 0.799)	0.683 (0.627, 0.740)	0.806 (0.762, 0.850)	0.714 (0.657, 0.770)
Swin-Small	0.804 (0.766, 0.842)	0.729 (0.693, 0.766)	0.823 (0.774, 0.871)	0.663 (0.612, 0.714)	0.635 (0.581, 0.689)	0.840 (0.795, 0.884)	0.717 (0.663, 0.770)
Extension
EfficientNetV2-S	0.817 (0.780, 0.854)	0.747 (0.711, 0.783)	0.797 (0.746, 0.849)	0.711 (0.662, 0.760)	0.663 (0.608, 0.718)	0.831 (0.787, 0.875)	0.724 (0.670, 0.778)
ConvNeXt-Tiny	0.812 (0.775, 0.850)	0.750 (0.715, 0.786)	0.700 (0.642, 0.759)	0.786 (0.742, 0.830)	0.700 (0.642, 0.759)	0.786 (0.742, 0.830)	0.700 (0.642, 0.759)
ConvNeXt-Small	0.814 (0.777, 0.851)	0.763 (0.728, 0.798)	0.734 (0.678, 0.790)	0.783 (0.739, 0.827)	0.707 (0.650, 0.764)	0.805 (0.762, 0.848)	0.720 (0.664, 0.777)

AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value, NPV, negative predictive value.

Table 6.

Algorithm performance metrics of combined and combined-plus-demographics models in external validation (top three models by AUC).

	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	F1 score (95% CI)
Combined
VGG-16	0.820 (0.784, 0.857)	0.757 (0.722, 0.793)	0.730 (0.673, 0.786)	0.777 (0.732, 0.822)	0.700 (0.643, 0.758)	0.801 (0.758, 0.845)	0.715 (0.658, 0.772)
EfficientNetV2-S	0.824 (0.788, 0.860)	0.764 (0.730, 0.799)	0.768 (0.714, 0.822)	0.762 (0.716, 0.808)	0.697 (0.642, 0.753)	0.821 (0.779, 0.864)	0.731 (0.676, 0.786)
ConvNeXt-Small	0.830 (0.794, 0.866)	0.757 (0.722, 0.793)	0.785 (0.732, 0.837)	0.738 (0.691, 0.785)	0.681 (0.626, 0.737)	0.828 (0.785, 0.871)	0.729 (0.675, 0.784)
Combined-plus-demographics
VGG-16	0.825 (0.789, 0.861)	0.752 (0.717, 0.788)	0.772 (0.719, 0.826)	0.738 (0.691, 0.785)	0.678 (0.622, 0.734)	0.819 (0.776, 0.863)	0.722 (0.667, 0.777)
EfficientNetV2-S	0.832 (0.796, 0.867)	0.773 (0.739, 0.808)	0.738 (0.682, 0.794)	0.798 (0.755, 0.841)	0.723 (0.667, 0.780)	0.810 (0.768, 0.853)	0.731 (0.675, 0.787)
ConvNeXt-Small	0.833 (0.798, 0.869)	0.750 (0.715, 0.786)	0.827 (0.779, 0.875)	0.696 (0.646, 0.745)	0.660 (0.606, 0.714)	0.849 (0.807, 0.892)	0.734 (0.681, 0.787)

AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value, NPV, negative predictive value.

The AUC values of all models in external validation are presented in Fig. 6. No significant differences were observed among single-position models (flexion vs. neutral, p = 0.071; neutral vs. extension, p = 0.206; extension vs. flexion, p = 0.627). Only the neutral model exhibited significantly reduced AUC compared to the combined and combined-plus-demographics models (p < 0.001), while the flexion and extension models did not show a statistically significant difference with the combined models (p > 0.05).

Fig. 6.

Parallel coordinates plot showing the AUC values of all models with 10 backbone architectures in external validation. AUC, area under the receiver operating characteristic curve.

Discussion

We developed a CNN-based DL algorithm to predict cervical SCC using dynamic lateral radiographs. The AUC of our final model was 0.888, indicating a good performance²⁹, and the suggested inspection points appeared to be clinically reasonable. This best-performing model obtained the AUC of 0.820 when tested externally. Therefore, this may be utilized as a screening tool to determine the necessity of cervical MRI and prevent delays in the diagnosis and treatment of DCM. In addition, the potential inspection points suggested by the Grad-CAM may help physicians identify areas that warrant closer examination.

Our algorithm was developed using data from a significantly larger number of consecutive patients compared to previous reports, which included only selected groups of approximately 200 to 250 patients, and provided more favorable performance as evidenced by higher AUC values^9,18. This indicates that our trained model is more credible and has the potential to be generalized after further validation and refinement. In addition, various DL architectures, ranging from ResNet and VGG to up-to-date models such as Vision Transformer and ConvNeXt, were trained and systematically compared based on the type of inputs. Interestingly, while the VGG-16 model achieved the highest AUC in internal validation, the ConvNeXt-Small model performed best in external validation. This discrepancy may be attributed to domain shifts between differing patient populations; earlier architectures like VGG-16, which lack skip connections, might be more susceptible to such variations than modern architectures. Therefore, rather than asserting the absolute superiority of a single architecture, our findings underscore the importance of evaluating architectural stability across diverse populations. Our study also suggests possible inspection points that may serve as diagnostic clues. Although our AUC values were lower than those of previous studies regarding DL algorithms assessing cord compression using MRI¹⁶, the significance of our study lies in its potential as a screening tool based on simple radiographs.

To assess the broader clinical applicability of our model, we performed multi-institutional validation including external validation with patients from an independent healthcare center. The data for development and internal validation of algorithms were obtained from a tertiary referral hospital, which typically treats a high proportion of patients with severe conditions compared to the general population. However, it is well-known that the prevalence of a disease can significantly affect the performance of DL algorithms when applied in medical fields³⁰. Therefore, to evaluate the generalizability, external validation was also performed using a cohort from a community hospital, which is more accessible to the general population and thus offers a more representative sample than tertiary referral centers. However, it is important to note that our retrospective design and sampling framework introduce a selection and spectrum bias. Because the training dataset consists of patients who underwent both radiographs and MRI within 180 days at a tertiary center, indicating sufficient clinical suspicion to warrant advanced imaging, the model was trained on a population with a higher pre-test probability of severe disease. This bias likely explains the performance disparity between the internal (AUC 0.888) and external (AUC 0.820) validations. Therefore, while our model shows promise, its generalizability to a true screening population, such as asymptomatic or minimally symptomatic individuals in primary care, remains limited and requires further prospective validation to assess its real-world utility.

Unlike previous reports, this study included dynamic lateral radiographs, and analyzed the performance of a combined model that considered all three positions together. It is well established that dynamic instability depending on position can be a significant contributor to spinal cord compression^1,18. Notably, the combined model generally achieved superior performance compared to single-position models in both internal and external validations. Moreover, the Grad-CAM analysis may have suggested segments with dynamic instability as one of the inspection points, highlighting the strength of our model in accounting for spatial correlations across different postures by integrating analyses from all three positions. However, single-position models may be more practical and convenient as screening tools in real-world clinical settings compared to multi-position approaches. Therefore, given their acceptable performance metrics, single-position models could serve as viable screening tools in clinical practice as well.

In the evaluation of screening tool performance, both specificity and sensitivity represent critical metrics³¹. For DCM, sensitivity could be considered particularly crucial, as early diagnosis and intervention significantly impact clinical outcomes. Our investigation demonstrated that several DL architectures achieved sensitivity values exceeding 0.8, suggesting the potential clinical utility of our models as effective screening tools for SCC or DCM in real-world clinical settings³². Furthermore, both the combined and combined-plus-demographics models using various backbone architectures achieved F1 scores greater than 0.8, indicating robust and well-balanced performance regarding precision and recall.

Despite not being provided with information on the pathophysiology of DCM, our algorithm may have suggested several possible inspection points, including soft disc herniations, osteophytes from the posterior aspect around the disc space, OPLL, and segmental instability, all of which may represent clinically relevant findings. Among these, activation signals at the area of soft disc herniations could be particularly interesting as they are not directly visualized on radiographs. Our DL model may have identified subtle indirect signs of disc herniation on radiographs that are commonly missed by human observers. Additionally, comparison between MR and Grad-CAM images showed the SCC level prediction within ± 1 level error margin (Fig. 7). Most notably, our study yielded meaningful results by performing level predictions on a larger population of patients with DCM than prior research^9,18.

Fig. 7.

Comparison of activation signals from Grad-CAM images with the levels of spinal cord compression in MRI. Grad-CAM, gradient-weighted class activation mapping; MRI, magnetic resonance imaging.

However, several notable activations were observed outside the cervical spine region in approximately 10% of patients in the test set, particularly in the jaw area. We believe that these activation signals may be associated with dental pathologies such as dental caries or artifacts related to prior dental interventions. These findings may be associated with spurious correlations learned by the model, potentially arising from unintended associations between non-spinal features and SCC-positive labels during the training phase. However, activations outside the cervical spine were not consistently observed across all three radiographic images of the same patients, suggesting that these signal activations are more likely to represent noise or spurious features rather than true pathologic findings. This observation supports the robustness of our combined model in differentiating meaningful findings from noise based on the consistency of activation patterns across multiple images, thereby mitigating the influence of spurious correlations.

In the evaluation of patients with SCC, clinical signs or symptoms play an important role, comparable to MRI findings¹. Treatment planning becomes particularly challenging when discrepancies exist between imaging results and clinical presentations. Since our model included only imaging without clinical findings, it may have limitations in comprehensively assessing the severity of SCC patients. Nevertheless, the significance of this study lies in the fact that the purpose of our model is not to precisely evaluate SCC patients or replace MRI, but rather to assist in determining whether advanced imaging examinations are warranted based on plain radiographs. Moreover, we believe this model may serve as a valuable adjunctive tool, particularly in primary care settings or during initial patient evaluations conducted by physicians with limited experience or without specialized spine training. In practical clinical workflows, this algorithm could be integrated into automated radiological reporting systems or utilized for radiology triage to flag patients requiring prompt MRI referrals. For these screening purposes, the operating decision threshold should be purposefully calibrated to prioritize high sensitivity over specificity, ensuring that early or subtle cases of cord compression are not missed before irreversible neurological deterioration occurs. However, we acknowledge that a high false-positive rate may lead to unnecessary patient anxiety and impose additional burden on clinical workflows, particularly in settings where MRI is readily accessible. Therefore, the decision threshold can and should be flexibly adjusted based on each institution’s resources and MRI accessibility to achieve an optimal balance between sensitivity and specificity.

Since its introduction, several studies have evaluated the effectiveness and validity of the Kang system^33–35. However, grade 2 encompasses a wide range of SCC, complicating the differentiation of severity and treatment planning³⁶. Specifically, this grade includes both patients with minimal cord deformation without symptoms and those with severe cord compression and clear myelopathic symptoms without signal change. In this study, grade 2 and grade 3 were classified as SCC-positive, thereby including patients with mild SCC. Although this labeling strategy may reduce classification specificity and overall discriminative performance, it was based on the objective of our study to develop a model for screening patients who may require further evaluation with MRI, rather than to establish a definitive diagnosis. Nonetheless, future studies incorporating severity-based labeling strategies, including further subdivision of grade 2, may improve clinical applicability by enabling more precise discrimination and facilitating development toward a more definitive diagnostic tool.

The present study has several limitations. First, due to its retrospective nature, selection bias and confounding factors may not have been considered. Second, although patients with a history of cervical spine surgery, tumorous lesions, deformities, and congenital diseases were excluded in this study to enhance the reliability of our model, it would be more meaningful if the DL algorithm performs consistently regardless of the factors excluded. Further studies that train a broader range of images are required in the future. Third, only imaging findings were included without clinical symptoms. Consequently, our purely imaging-based algorithm may overcall or miss cases where radiological findings are discordant with actual clinical presentations, highlighting the need for future multimodal studies to quantify the benefit of integrating clinical data. However, our model is intended to screen patients with the potential for SCC, not to accurately assess the severity of SCC. Fourth, we included patients with SCC regardless of the underlying etiology in this study. Future studies focusing on individual etiologies such as cervical spondylotic myelopathy or OPLL would provide clearer interpretations and greater clinical significance. Fifth, the current study did not evaluate the atlantoaxial segment despite its clinical significance, necessitating further research. Sixth, our combined model simply averaged predictions from single-position models rather than directly learning joint spatial relationships. Future studies exploring alternative architectures, such as multi-input or sequence-based networks, may better exploit temporal and positional correlations to assess dynamic instability. Seventh, ground truth labeling was based on the interpretation of a single spine surgeon, which may raise concerns regarding reliability. To mitigate this limitation, the initial labeling was cross-checked with official radiology reports, which are routinely generated based on the Kang grading system. Cases with ambiguity or discrepancies between the surgeon’s interpretation and the radiology reports were further reviewed by the corresponding author to reach a consensus. Eighth, direct comparison with human interpreters was not performed in this study, which may limit the assessment of the model’s clinical utility. Future studies comparing model performance with human experts would provide more robust validation. Finally, although external validation was performed using data from another institution, our models were trained using data from a single tertiary referral center, potentially introducing bias and limiting generalizability to general clinical settings. Therefore, further studies incorporating multi-institutional training data from various healthcare settings and validation with larger sample sizes are warranted to improve clinical applicability. Despite these limitations, to the best of our knowledge, this study is the first to perform multi-institutional validation using data from various healthcare levels and to use a dynamic series of plain radiographs to predict cervical SCC, potentially helping screen patients who need advanced imaging studies.

In conclusion, we developed a CNN-based DL model for predicting SCC of the cervical spine using dynamic cervical lateral radiographs and performed multi-institutional validation to evaluate its potential as a screening tool, resulting in acceptable performance. Our algorithm may also localize suspected regions, which may correspond to clinically relevant areas. This approach may offer the potential to improve diagnostic efficiency and clinical outcomes in patients with DCM. Further prospective validation studies involving general screening populations in primary care are warranted to evaluate the real-world clinical applicability of this algorithm.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.C. Park: conceptualization, methodology, data curation, formal analysis, writing—original draft. W. Rhee: data curation, formal analysis, visualization, writing—original draft. B.-S. Chang: validation, supervision, writing—review and editing. S.Y. Chang: data curation, validation, writing—review and editing. H. Kim: conceptualization, methodology, supervision, writing—review and editing.

Funding

Not applicable.

Data availability

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

Declarations

Competing interests

The authors declare no competing interests.