Prompt-Based Vertebral Segmentation Using a Generative AI Approach in OVCF Spinal Radiographs

Background

Osteoporotic vertebral compression fractures (OVCFs) are prevalent among elderly patients, with X-ray imaging serving as the primary diagnostic tool. However, challenges such as organ obstruction and poor contrast after vertebroplasty procedures complicate vertebral segmentation in spinal X-rays.

Most OVCF studies focus on classification tasks rather than segmentation. Precise postoperative vertebrae segmentation remains underexplored despite its importance for treatment monitoring and providing biomechanics and radiomics a firm foundation for further research. X-ray images after vertebroplasty (VP) surgery — with surgical implants including cages, screws, and other hardware — are hard to read, and existing methods relying on handcrafted parameter tuning lead to weak generalizability across datasets.

Method

This research introduces an innovative generative AI framework for vertebral segmentation in spinal X-ray images, combining YOLO-based detection with prompt-driven segmentation inspired by the Segment Anything Model (SAM). The system generates bounding boxes around vertebrae as segmentation prompts and employs an interpolation strategy to address potentially missed compressed vertebrae. By incorporating domain-specific knowledge of vertebral anatomy via the interpolation strategy, the framework enables accurate delineation of vertebral structures in cases of compression fractures.

Results

This generative AI application effectively addresses clinical challenges in vertebral segmentation for OVCF patients, potentially enhancing the accuracy of diagnoses and treatment planning.