Cervical spinal cord injury is a severe traumatic disorder of the central nervous system, most commonly caused by high-energy events such as traffic accidents, falls from height, sports injuries, and violent assaults. Because the cervical spinal cord carries motor, sensory, and autonomic signals, such injuries can produce quadriplegia, respiratory dysfunction, and can even be life-threatening. Early and accurate identification matters for guiding treatment decisions, assessing prognosis, and planning rehabilitation, but in emergency settings MRI is often impractical because of long examination times, limited scanner availability, and contraindications in some patients.

The research team, led by Deyan Ou and Shixin Pan with colleagues, collected 631 cervical spine CT image samples retrospectively from two independent clinical centers, comprising 234 cases with cervical spinal cord injury and 397 without. Diagnoses were established from neurological signs, imaging findings, and medical records. All images were exported in DICOM format, converted to NIfTI format, and processed through a unified pipeline that included Z-score normalization and resampling to a uniform 64 by 64 by 32 voxel size using trilinear interpolation. During training only, random flipping along specified axes was applied as a lightweight data augmentation step.

Three-dimensional regions of interest covering the cervical spinal cord were delineated manually using ITK-SNAP software, with interpretation across axial, sagittal, and coronal planes. A clinical surgeon with five years of experience performed the primary segmentation, and 20 percent of samples were independently re-labeled by a senior radiologist with ten years of experience; disagreements were adjudicated by a senior spinal surgeon. Interobserver consistency was quantified with the intraclass correlation coefficient. To prevent data leakage, the second center’s data were kept strictly separate and accessed only after the model architecture and hyperparameters had been finalized using the first center’s data.

The network itself is deliberately simple: four consecutive three-dimensional convolutional blocks, each combining a 3D convolutional layer, batch normalization, ReLU activation, and max-pooling, followed by fully connected layers with dropout for binary classification of injury versus no injury. Training used the PyTorch framework with a cross-entropy loss, the AdamW optimizer at an initial learning rate of 1 x 10^-4, weight decay, and a ReduceLROnPlateau scheduler that adjusted the learning rate based on internal validation performance. The model weights with the highest validation accuracy were retained.

Using a stratified random strategy, the samples were split into a training set of 441, an internal validation set of 95, and an external validation set of 95. The proportions of injured and non-injured cases remained highly consistent across all subsets, and hash-based checks of image file names confirmed an approximately uniform, overlapping distribution with no duplicated or overlapping files between datasets. Training loss fell steadily while validation loss declined and stabilized without rebound, and the gap between training and validation accuracy stayed small, which the authors interpret as evidence against obvious overfitting.

Discriminative performance was notably stable across datasets. The area under the ROC curve reached 0.943 on the training set, 0.953 on internal validation, and 0.948 on external validation, while the area under the precision-recall curve was 0.936, 0.939, and 0.943 respectively. The authors emphasize that this consistency contrasts with a broader literature in which externally validated image-based deep learning algorithms frequently lose performance on external data. Prediction probability distributions showed clear separation, with non-injury cases concentrated at low probabilities and injury cases shifted toward high values across all three datasets.

Beyond a single fixed threshold, the team analyzed how performance changed as the decision threshold varied. An optimal threshold of approximately 0.48 emerged from the F1-score analysis, yielding balanced accuracy, precision, recall, and F1-score across all datasets. Decision curve analysis showed that within a wide threshold probability range of 0.1 to 0.9, the model’s net benefit curve exceeded both the treat-all and treat-none extreme strategies in every dataset, and calibration curves indicated that predicted probabilities matched actual positive rates without obvious overestimation or underestimation. Example cases illustrate the range: a normal CT with intact bone and clear spinal canal received a probability of 0.124, a borderline case scored 0.483, and a case with obvious fracture and dislocation scored 0.978.

The authors are careful to frame the work as a feasibility study rather than a deployable clinical tool. Because the design is retrospective and case-control, and because the model depends on manually delineated regions of interest, the findings cannot be extrapolated directly to emergency triage or resource allocation, and prospective validation is required before any clinical deployment. The manual segmentation step also limits full automation and adds preprocessing workload. The current model performs binary classification only and does not address injury grading, prognosis assessment, or treatment decision support.

Even with those caveats, the study offers a methodological template for medical imaging AI research: standardized preprocessing, transparent architecture, documented training dynamics, and external validation from an independent center. In settings where MRI is unavailable or delayed, a CT-based model of this kind could eventually serve as a supplementary reference for early evaluation of cervical spinal cord injury, provided future prospective studies confirm that the retrospective performance translates into real clinical benefit.

Subject of Research: A deep learning framework for identifying cervical spinal cord injury on cervical spine CT images

Article Title: Construction and validation of a unified 3D convolutional neural network framework for automated identification of cervical spinal cord injury based on CT imaging

Article References: Ou, D., Mo, Y., Huang, Y., Ye, Y., Lin, S., Pan, L., Lin, R., Huang, S., & Pan, S. (2026). Construction and validation of a unified 3D convolutional neural network framework for automated identification of cervical spinal cord injury based on CT imaging. BMC Medical Imaging, 26(1), 556. https://doi.org/10.1186/s12880-026-02626-3

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02626-3

Keywords: cervical spinal cord injury, deep learning, 3D convolutional neural network, computed tomography, BMC Medical Imaging, external validation, trauma imaging, ROC analysis, decision curve analysis, medical imaging AI, multicenter study, MRI limitations