Severity Prediction of Traumatic Cervical Spinal Cord Injury With an AI Model Based on MRI Radiomics

Abstract

Background

Traumatic cervical spinal cord injury (TCSCI) often leads to significant patient paralysis. Current clinical diagnosis relies heavily on empirical interpretation of magnetic resonance imaging (MRI) and the American Spinal Injury Association Impairment Scale (AIS) grade, lacking robust quantitative markers to precisely reflect injury severity. This study aimed to build an artificial intelligence (AI) pipeline for AIS grade prediction based on radiomic features extracted from manually defined regions.

Methods

We included 189 patients with TCSCI who underwent MRI within 48 h post-injury. MRI images from 130 patients were used for developing an AI model encompassing image segmentation. Radiomic features were extracted from manually delineated volumes of interest (VOIs). T2-weighted imaging (T2WI) sagittal images were randomly divided into training (n = 104), validation (n = 13), and test (n = 13) sets for segmentation. A total of 183 patients (excluding AIS E) were included in the AIS grade prediction task. Model performance was evaluated using mean dice similarity coefficient (mDICE), mean intersection over union (mIOU), mean specificity, and mean sensitivity.

Results

An optimized UCTransnet network, leveraging a Transformer architecture for formal training, combined with a U-Net++ network for pretraining, achieved promising results in segmenting the spinal cord injury site on T2WI sagittal images (mDICE: 0.777 ± 0.021, mIOU: 0.646 ± 0.025, mean specificity: 0.998 ± 0.001, mean sensitivity: 0.895 ± 0.015). Subsequently, an ensemble model (we named Em-En) constructed using selected radiomic features from the manual VOIs demonstrated superior performance for predicting AIS grades in terms of sensitivity, specificity, accuracy, and clinical decision-making benefit compared to other tested models.

Conclusions

This study presents an AI-assisted pipeline for predicting the severity of TCSCI. The developed resources provide a theoretical foundation for the clinical application of AI-assisted diagnostic methods, potentially lowering the interpretation barrier for MRI and offering clinicians preliminary quantitative indicators of injury severity. The source code is publicly available.