What question did this study set out to answer?

The study aims to develop deep learning models for predicting osteoporosis using lumbar MRI images.

February 2, 2026Open Access

Lumbar MRI-Based Deep Learning for Osteoporosis Prediction

UHUe-Cheung HoNational Taiwan University HLHsueh-Yi LuNational Yunlin University of Science and Technology LKLu-Ting Kuo

Key Points

The study aims to develop deep learning models for predicting osteoporosis using lumbar MRI images.
Retrospective enrollment of 218 patients aged 50 and older who had lumbar MRI and DXA scans.
Segmentation of vertebral bodies from T1- and T2-weighted MRI images to create datasets.
Training of separate convolutional neural network models for T1 and T2 MRI sequences.
Evaluation of model performance using receiver operating characteristic curves and area under the curve (AUC).
EfficientNet b4 classifier showed the best performance with an AUC of 82% for T1-weighted MRI and 83% for T2-weighted MRI.
Sensitivity of 85% and specificity of 79% for the T1-weighted model; sensitivity of 86% and specificity of 80% for the T2-weighted model.
Models outperformed InceptionResNet v2 and ResNet-50 for both MRI sequences.

Abstract

Background: Osteoporosis (OP) is characterized by reduced bone mineral density and increased fracture risk. Many spinal surgery patients have undiagnosed OP due to the lack of preoperative screening, leading to postoperative complications. Magnetic resonance imaging (MRI), a routine, non-invasive tool for spinal assessment, offers potential for opportunistic OP detection. This study aimed to develop deep learning models to identify OP using lumbar MRI. Methods: We retrospectively enrolled 218 patients (≥50 years) who underwent both lumbar MRI and dual-energy X-ray absorptiometry (DXA). After segmentation of vertebral bodies from T1- and T2-weighted MRI images, 738 images per sequence were extracted. Separate convolutional neural network (CNN) models were trained for each sequence. Model performance was evaluated using receiver operating characteristic curves and area under the curve (AUC). Results: Among tested classifiers, EfficientNet b4 showed the best performance. For the T1-weighted model, it achieved an AUC of 82%, with a sensitivity of 85% and specificity of 79%. For the T2-weighted model, the AUC was 83%, with a sensitivity of 86% and specificity of 80%. These results were superior to those of InceptionResNet v2 and ResNet-50 for both sequences. Conclusions: The AI models provided reliable OP classification without additional imaging or radiation. AI-based analysis of standard lumbar MRI sequences can accurately identify OP. These models may assist in early detection of undiagnosed OP in surgical candidates, enabling timely treatment and perioperative strategies to improve outcomes and reduce healthcare burden.

Perguntar à IA

Bookmark

View Full Paper