Automated multi-sequence MRI quantitative assessment and 3D visualization of acute cervical spinal cord injury

Precise evaluation of spinal cord compression and intramedullary lesions is essential for surgical decision-making in spinal cord injury (SCI). However, this process remains intrinsically subjective, frequently challenged by small lesion volumes, elongated morphology, and ambiguous boundaries, which collectively drive high inter-rater variability. We leveraged a heterogeneous multi-center MRI dataset ( n = 711) to develop and validate an uncertainty-aware deep learning architecture. The core multi-modal network captures long-range spatial dependencies, while a Monte Carlo dropout-based refinement module quantifies diagnostic uncertainty to resolve the inherent ambiguity of lesion boundaries. The framework achieved an overall Dice score of 70.44% for lesion segmentation. The integration of uncertainty refinement significantly improved algorithmic alignment with a 7-rater expert consensus (mean Dice: 65.85 vs. 63.59%). The pipeline directly translates raw predictions into an anatomically anchored 3D visualization tool, localizing the maximum compressed level (MCL) with a mean absolute error of 5.72 mm and correctly mapping longitudinal lesion extent to exact vertebral levels in 80.8% of cases. Furthermore, imaging biomarkers exhibited significant inverse correlations with neurological impairment as measured by the American Spinal Injury Association Impairment Scale (AIS) grade and Upper Extremity Motor Score (UEMS) at admission, discharge and 3-month follow-up. This framework advances SCI evaluation from subjective visual assessment to standardized, quantitative outcome monitoring.