Introduction: While central cord syndrome (CCS) is a clinical diagnosis, confirmatory MRI is necessary and directly informs surgical decision-making. Recent evidence suggests a role for early surgery in CCS, reinforcing the need for tools to identify CCS earlier.
Methods: We acquired 405 CT spine series, including 38 (9.3%) series associated with MRI-confirmed CCS. After preprocessing, series were split into 70/30 train/test sets. We evaluated several deep learning architectures (EfficientNet, DenseNet, 3D CNN) for the task of predicting CCS using CT images. A single shot object detector for vertebral fractures was also used in ensemble with the trained deep learning models to evaluate for potential performance gains.
Results: For the primary task of CCS prediction from CT images, EfficientNet (AUC: 0.40 [0.34-0.48]) performed the best in comparison to 3D CNN (AUC: 0.33 [0.28-0.37]) and DenseNet (AUC: 0.36 [0.30-0.39]) architectures. The outputs of the trained EfficientNet and a YoloR system to detect vertebral fractures were passed into a random forest algorithm with achieved appreciable increases in performance with an AUC of 0.57 [0.49-0.61] for the ensemble.
A sensitivity-optimized instance of the ensemble approach demonstrated a sensitivity of 0.79 and a specificity of 0.28. In the test set, model implementation would have resulted in median decrease in time-to-MRI order of 5.4 hours (IQR, 2.5-27.2) for MRI-confirmed CCS patients.
Conclusion : A sensitivity-tuned, ensemble computer vision approach may enable CT-based triage in CCS, reducing delays in diagnosis and surgery. Future work with larger datasets and integration of presentation clinical variables may improve model performance.
.jpg)
