Deep learning magnetic resonance imaging algorithm for differentiating metastatic vertebral fractures.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Distinguishing malignant metastatic lesions from benign osteoporotic vertebral compression fractures (VCFs) is a critical diagnostic challenge in spine practice.
• Diagnostic errors can lead to delayed oncologic intervention and suboptimal patient management.

Purpose of the Study:

• To develop and validate a deep learning (DL) object detection algorithm for differentiating malignant spinal metastases from benign VCFs using routine MRI sequences.
• To assess the performance of various DL models in this diagnostic task.

Main Methods:

• A retrospective multicenter study included 2,165 patients with VCFs or spinal metastases across 27,543 vertebral levels.
• Sagittal T1- and T2-weighted MRI series were analyzed using four object detection models (YOLOv5, YOLOv8, YOLOv11, DETR).
• Model performance was evaluated using metrics including mean Average Precision (mAP50-95), Precision, Recall, and F1-score.

Main Results:

• The YOLOv11 model with a ResNet-101 backbone achieved the highest overall performance (mAP50-95: 80.2%, F1-score: 91.9%).
• YOLOv8 demonstrated the highest Recall (93.3%), suggesting utility for screening.
• YOLOv11 provided a balanced Precision/Recall profile, minimizing false negatives and showing robustness in complex cases.

Conclusions:

• Deep learning object detection can accurately differentiate malignant metastatic lesions from benign VCFs on routine MRI.
• The YOLOv11 model demonstrates potential as a decision-support tool to expedite oncologic referrals and enhance diagnostic accuracy.