Automatic Vertebral Body Segmentation Performance in Lateral Thoracic Spine X-Rays: A Comparative Analysis Based on...

Accurate vertebral body segmentation in spine radiographs is essential for quantitative analysis, but metallic implants introduce artifacts that degrade model performance. We evaluated the effects of implant presence and material subtypes on automatic segmentation using real-world thoracic spine lateral X-rays. SOLOv2 and Mask R-CNN were evaluated on radiographs from 264 patients. To analyze the localized impact of surgical hardware, a synchronized cohort of 3637 vertebral bodies correctly delineated by both models was analyzed, comparing non-implanted (n = 3467) and implanted (n = 170) instances. The implanted group was stratified into four subtypes: Bone Cement, Screw, Screw + Cement, and Screw + Etc. Performance was quantified via precision, sensitivity, Dice similarity coefficient (DSC), Hausdorff distance (HD95), and average symmetric surface distance (ASSD), and validated using the Mann-Whitney U test (α = 0.010). Implant presence caused significant performance degradation across all metrics (p < 0.001). The Screw + Etc subgroup showed the most profound degradation versus baseline (DSC: 0.826 vs. 0.930, p < 0.001). The Screw subgroup exhibited significant declines in volumetric overlap and boundary distances (p < 0.010). The Bone Cement subgroup showed a broad multidimensional decline, including a severe drop in sensitivity (p < 0.001), while the Screw + Cement subgroup demonstrated significant distance-based errors (p = 0.005). Implant presence significantly reduces thoracic spine segmentation performance, varying distinctly by subtype. Developers must account for these subtype-specific artifacts to ensure clinical validity.