Deep Learning Magnetic Resonance Imaging Algorithm for Differentiating Metastatic Vertebral Fractures.
Authors
Affiliations (8)
Affiliations (8)
- Department of Orthopedic Surgery, Bucheon St. Mary's Hospital, The Catholic University of Korea.
- Department of Orthopedic Surgery, Seoul St. Mary's Hospital, The Catholic University of Korea.
- Department of Orthopedic Surgery, Eunpyeong St. Mary's Hospital, The Catholic University of Korea.
- Department of Orthopedic Surgery, Yeouido St. Mary's Hospital, The Catholic University of Korea.
- Department of Orthopedic Surgery, Incheon St. Mary's Hospital, The Catholic University of Korea.
- Department of Orthopedic Surgery, St. Vincent's Hospital, College of Medicine, The Catholic University of Korea.
- Department of Data Science, The Catholic University of Korea.
- Department of Orthopaedic Surgery, Soonchunhyang University Hospital. Electronic address: [email protected].
Abstract
Distinguishing malignant metastatic lesions from benign osteoporotic vertebral compression fractures (VCFs) is a major diagnostic challenge in spine practice; delays or errors can lead to inappropriate management and missed opportunities for timely oncologic intervention. To develop and validate a deep learning-based object detection algorithm using routine MRI sequences to differentiate malignant metastatic lesions from benign VCFs. Retrospective multicenter study conducted across six tertiary hospitals. A total of 2,165 patients with either VCFs or spinal metastases were included, encompassing 27,543 vertebral levels. Sagittal T1- and T2-weighted MRI series were available for all participants and were annotated by spine specialists. Primary outcome: vertebral-level detection/classification performance measured by mean Average Precision across intersection over union thresholds (mAP50-95). Precision, Recall, and F1-score; qualitative error analysis METHODS: : Four object detection models (YOLOv5, YOLOv8, YOLOv11, DETR) were trained and validated on annotated sagittal T1- and T2-weighted images. The dataset was split into training, validation, and test images. Model performance was evaluated using mAP50-95, Precision, Recall, and F1-score on the test set, with qualitative review of misclassifications. Ground-truth was defined by finalized radiology reports and clinical diagnoses; annotators were blinded to model outputs. YOLOv11 with a ResNet-101 backbone achieved the best overall performance (Precision 91.2%, Recall 92.5%, F1-score 91.9%, mAP50-95 80.2%). YOLOv8 showed the highest Recall (93.3%), supporting potential use for screening, whereas YOLOv11 balanced Precision and Recall, minimizing false negatives. Qualitative analysis demonstrated robust detection in difficult scenarios, including multiple concurrent lesions and coexisting benign and malignant fractures. Misclassifications were uncommon and predominantly involved intravertebral vacuum cleft signs that were not explicitly modeled as a class. A deep learning-based object detection approach can accurately distinguish malignant metastatic lesions from benign VCFs on routine MRI. In this large multicenter study, YOLOv11 showed the most balanced performance and may serve as a practical decision-support tool to expedite oncologic referral, improve diagnostic accuracy, and optimize treatment strategies.