Deep learning pipeline detects vertebral fractures with 92.6% sensitivity and 95.8% specificity in older adults during routine CT scans

Vertebral compression fractures are common in older adults but frequently remain undiagnosed, particularly when routine chest or abdominal CT scans are acquired for non-spinal indications. Opportunistic analysis of these scans may help reduce the osteoporosis care gap by enabling automated identification of vertebral abnormalities from imaging already obtained in routine clinical care. This study presents an end-to-end deep learning pipeline for vertebral segmentation, fracture detection, fracture severity grading, and Schmorl's node identification from routine CT scans. A YOLOv11-based segmentation model was trained using CTSpine1K, VerSe2019, and VerSe2020 to localize, segment, and anatomically label vertebrae from C1 to S1. Segmented vertebrae were subsequently processed using a multi-stage YOLOv11 classification framework trained on a retrospectively curated institutional dataset of 5,000 CT scans from the American University of Beirut Medical Center (AUBMC). Vertebra-level annotations were performed by radiologists using a custom DICOM labeling tool and the Genant semi-quantitative grading method. The classification pipeline first categorizes vertebrae as normal, fractured, or unknown; then grades fractured vertebrae as mild, moderate, or severe; and additionally detects the presence of Schmorl's nodes. The complete system was integrated into a web-based platform for visualization, radiologist review, and feedback-driven model improvement. The segmentation model achieved 96% Top-1 anatomical labeling accuracy when at least one spinal extremity was visible and 86% when both extremities were absent, with 100% Top-3 accuracy in both settings and a mean Intersection over Union of 0.94. The primary fracture classification model achieved 92.7% vertebra-level accuracy, with sensitivity of 92.6%, specificity of 95.8%, and an F1-score of 91.6% for fractured vertebrae. Schmorl's node detection achieved 91.9% accuracy and an F1-score of 93.5% for the Schmorl class. Fracture severity grading achieved 85.0% overall accuracy. End-to-end evaluation showed an overall pipeline success rate of 85% and 90% multi-label consistency with radiologist consensus. The proposed pipeline demonstrates the feasibility of automated opportunistic vertebral assessment on heterogeneous, real-world CT imaging. By combining anatomical segmentation, fracture detection, Genant-based severity grading, Schmorl's node detection, and clinical visualization in a single workflow, the system provides a practical framework for integrating AI-assisted vertebral screening into routine radiology practice.