Fully automated system predicts osteoporotic vertebral fracture across institutions using lumbar MRI paraspinal muscle signatures.
Authors
Affiliations (11)
Affiliations (11)
- Jiangmen Key Laboratory of Artificial Intelligence in Medical Image Computation and Application, Jiangmen Central Hospital, Jiangmen, Guangdong, China.
- Department of Radiology, Guangzhou First People's Hospital, Guangzhou, Guangdong, China.
- Department of Spinal Surgery, Baishi Orthopedic Hospital, Jiangmen, Guangdong, China.
- Department of Spinal Surgery, The Fifth Affiliated Hospital, Sun Yat-sen University, Zhuhai, Guangdong, China.
- Department of Surgery of Joint and Extremities Osteopathy, Maoming People's Hospital, Maoming, Guangdong, China.
- Department of Spinal Surgery, Jiangmen Central Hospital, Jiangmen, Guangdong, China.
- Laboratory of Intelligent Detection and Information Processing, Guilin University of Aerospace Technology, Guilin, Guangxi, China.
- Jiangmen Central Hospital, Jiangmen, Guangdong, China.
- Department of Radiology, Jiangmen Central Hospital, Jiangmen, Guangdong, China.
- Laboratory of Intelligent Detection and Information Processing, Guilin University of Aerospace Technology, Guilin, Guangxi, China. [email protected].
- Jiangmen Key Laboratory of Artificial Intelligence in Medical Image Computation and Application, Jiangmen Central Hospital, Jiangmen, Guangdong, China. [email protected].
Abstract
Paraspinal muscle (PM) degeneration is a crucial yet frequently overlooked risk factor for osteoporotic vertebral fractures (OVF). We developed PM Segmentation and Classification of OVF (PMSAC-OVF), a fully automated, multi-institutional system that segments lumbar PMs on MRI, extracts federated learning (FL) and radiomics features, and integrates them with clinical variables for OVF prediction. Leveraging a vision foundation model framework, the system enables privacy-preserving, cross-institutional training and lightweight local deployment. Data from 2,884 patients across five institutions (2014-2024) were analyzed. The automated segmentation module demonstrated expert-level accuracy (Dice coefficient: 0.952, Intersection over Union: 0.909) while reducing processing time to seconds. For prediction, FL and radiomics models yielded pooled AUCs of 0.827 (range: 0.819-0.861) and 0.803 (0.793-0.892), respectively. Trimodal models integrating radiomics signatures (RS), FL signatures (FLS), and clinical variables achieved a pooled AUC of 0.840 (0.822-0.916), significantly outperforming clinical-only models (AUC: 0.742, 0.641-0.778). SHapley Additive exPlanations identified RS, FLS, and bone mineral density as the top predictors, highlighting the complementary value of image-derived features. PMSAC-OVF provides a robust, interpretable, and scalable solution for OVF prediction in heterogeneous clinical settings, potentially facilitating early identification and personalized intervention for high-risk individuals.