Performance comparison between a deep learning model and spine surgeons in detecting cervical spinal cord compression...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Spinal Surgery

Background:

• Cervical spinal cord compression requires accurate and timely diagnosis.
• Current diagnostic methods can be time-consuming and may require specialized expertise.
• Deep learning offers potential for automated analysis of medical imaging.

Purpose of the Study:

• To develop a deep learning (DL) model for detecting cervical spinal cord compression using cervical radiographs.
• To compare the diagnostic performance of the DL model against experienced spine surgeons.

Main Methods:

• A retrospective study involved 600 patients with cervical spine radiography and MRI data.
• A DL model was trained on radiographs, incorporating segmentation for localization and binary classification for compression detection.
• Performance was evaluated using accuracy, sensitivity, specificity, F1 score, and AUC, with comparison to surgeon diagnoses.

Main Results:

• The DL model achieved high accuracy (94.67% internal, 93.33% external) and AUC (0.9911 internal, 0.9868 external).
• The model significantly outperformed two spine surgeons in diagnostic accuracy (p < 0.05).
• Gradient-weighted class activation mapping (Grad-CAM) highlighted relevant anatomical areas like intervertebral discs and foramina.

Conclusions:

• The developed DL model effectively classifies and localizes cervical spinal cord compression on radiographs.
• The model demonstrates superior diagnostic performance compared to spine surgeons.
• This AI tool has significant potential for clinical application, particularly in resource-limited settings.