Osteoporosis diagnosis and fracture risk assessment remain challenging. Magnetic Resonance Imaging (MRI) shows promise for evaluating vertebral bone quality, but manual segmentation hinders its widespread use. This study investigates the feasibility of automated lumbar vertebral body segmentation in MRI using lightweight YOLO (You Only Look Once) deep learning models. T1-weighted and T2-weighted MRI scans from 63 postmenopausal women were utilized. Existing semi-automatic segmentations served as ground truth. The data was preprocessed and adapted for YOLO-based instance segmentation. Two models, YOLOv8n-seg and YOLO11n-seg, were trained and evaluated on a test set using Mean Dice Similarity Coefficient (DSC), Mean Intersection over Union (mIoU), and inference time. Both models demonstrated high accuracy. YOLOv8n-seg achieved a mean DSC of 0.9293 and mIoU of 0.8690 (inference: 12.31 ms/image). YOLO11n-seg yielded slightly higher accuracy (DSC: 0.9305, mIoU: 0.8712) but was slower (inference: 15.89 ms/image). These results highlight a trade-off between segmentation speed and accuracy. Lightweight YOLO architectures are effective for automated vertebral body segmentation in MRI, offering a promising approach to streamline analysis workflows. This automation facilitates future research into MRI-based bone fragility assessment and fracture risk prediction.