Multi-Class Segmentation of Fused Spine Anatomy in X-ray Images Using DeepLabV3+ for Surgical Planning

Abstract:
The accurate segmentation of spinal anatomy is critical for surgical planning in scoliosis correction, yet manual methods remain labor-intensive and subjective. We propose an optimized DeepLabV3+ framework with a ResNet-101 backbone and Atrous Spatial Pyramid Pooling (ASPP) to automate multi-class segmentation of vertebrae and surgical implants (e.g., bone pins) in longitudinal X-ray images. Our model integrates class-weighted training to address rare anatomical variants (T13/L6) and dynamic post-surgical changes, achieving a Dice coefficient of 88.04% and mIoU of 81.49% on a cohort of 36 female patients with idiopathic adolescent scoliosis. Surface-based metrics (HD95: 3.83px, ASSD: 1.37px) demonstrate precise boundary delineation, particularly for post-surgical structures. The framework enables spinal modeling from longitudinal scans (pre-operative to 6-month follow-up), offering surgeons actionable insights for deformity correction and hardware placement. While the dataset’s homogeneity limits generalizability, our approach lays the groundwork for scalable, data-efficient solutions in orthopedic imaging.