Automated Detection of Cervical Spinal Stenosis and Cord Compression via Vision Transformer and Rules-Based...

Area of Science:

• Artificial Intelligence in Radiology
• Medical Imaging Analysis
• Deep Learning for Diagnostic Support

Background:

• Cervical spinal cord compression can lead to severe neurological deficits and requires timely intervention.
• Current diagnostic methods lack automated tools to alert radiologists to cervical cord compression.
• Accurate and timely detection is crucial to prevent adverse clinical outcomes.

Purpose of the Study:

• To evaluate the efficacy of a vision transformer (ViT) model in detecting cervical spinal cord compression.
• To compare the performance of ViT against conventional convolutional neural network (CNN) models.
• To assess the potential of an automated tool for improving clinical workflow.

Main Methods:

• A cohort of 142 cervical spine MRIs was analyzed, with varying degrees of stenosis and compression.
• A vision transformer (ViT) model was fine-tuned for section-level severity prediction.
• Performance was compared against ResNet50 and DenseNet121 CNN models at section and patient levels.

Main Results:

• The ViT model achieved higher section-level accuracy (82%) compared to ResNet50 (72%) and DenseNet121 (78%).
• ViT demonstrated excellent patient-level classification accuracy (93%), with high sensitivity (0.90) and specificity (0.95).
• The ViT model showed a superior area under the receiver operating characteristic curve compared to both CNNs.

Conclusions:

• A ViT model combined with rules-based classification effectively detects cervical spinal cord compression at the patient level.
• The ViT model significantly outperformed conventional CNNs in this task.
• Clinical implementation of this automated tool could enhance neuroradiology workflow efficiency and consistency.