Automated Detection of Cervical Spinal Cord Compression on MRI Using YOLO11 Deep Learning Architecture: A Two-Center External Validation Study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Orthopaedic Surgery, The Second Affiliated Hospital of Zunyi Medical University, Zunyi, Guizhou, China.
- Department of Orthopaedic Surgery, Zunyi Hospital of Traditional Chinese Medicine, Zunyi, Guizhou, China.
- School of Medical Information Engineering, Zunyi Medical University, Zunyi, Guizhou, China.
- Department of Orthopaedic Surgery, the Affiliated Hospital of Zunyi Medical University, Zunyi, Guizhou, China.
- Department of Orthopaedic Surgery, Jinsha County People's Hospital, Zunyi, Guizhou, China.
- Innovation Platform of Regeneration and Repair of Spinal Cord and Nerve Injury, Department of Orthopaedic Surgery, The Seventh Affiliated Hospital of Sun Yat-sen University, Shenzhen, China.
- Department of Anatomy, School of Medicine, Shenzhen Campus of Sun Yat-sen University, Sun Yat-sen University, Shenzhen, Guangdong 518106, China.
- Guangdong Provincial Key Laboratory of Orthopedics and Traumatology, The First Affiliated Hospital of Sun Yat-sen University, Guangzhou, China.
Abstract
Retrospective two-center external validation study conducted at two medical centers, collecting cervical spine MRI data from patients suspected of degenerative cervical myelopathy (DCM) between March 2022 and August 2024, forming a consecutive series with external validation. To develop and validate a deep learning model utilizing YOLO11 architecture for automated detection of cervical spinal cord compression on MRI and evaluate its performance against expert annotations. DCM represents the leading cause of non-traumatic spinal cord injury in adults. While MRI facilitates early detection and provides the foundation for timely intervention, image interpretation remains subjective and dependent on physician experience, resulting in diagnostic variability and challenges in clinical consistency. A YOLO11-based deep learning model was implemented with binary classification scheme (Normal vs. Compression). Five physicians annotated 1,431 sagittal T2-weighted cervical MRI images from 735 patients using standardized protocols, achieving excellent inter-observer agreement. Dataset comprised training/validation sets (577 patients, 1,141 images), internal test set (64 patients, 115 images), and external test set (94 patients, 175 images). Five-fold cross-validation assessed model robustness. Standardized preprocessing incorporating contrast enhancement, noise reduction, and normalization was applied. Gradient-weighted Class Activation Mapping enhanced model interpretability. Five-fold cross-validation yielded consistent performance with mAP50 ranging from 0.917 to 0.970, precision from 0.897 to 0.923, and recall from 0.922 to 0.946. External testing demonstrated statistically superior agreement with expert annotations (mAP50=0.944, 95% CI: 0.934-0.953) compared to mid-level physician annotations (mAP50=0.912, 95% CI: 0.908-0.919), with the difference being statistically significant (95% CI of difference: 0.015-0.043, P < 0.05). The YOLO11-based model demonstrated stable two-center performance with close alignment to expert-level clinical standards. The rapid inference, high sensitivity, and integrated visualization system address key challenges related to efficiency and interpretability in clinical AI applications for cervical spinal cord compression assessment.