CervSpineNet: A hybrid deep learning-based approach for the segmentation of cervical spinous processes

Jay Sunil Sawant¹Lama Moukheiber¹Anupama Nair¹Anubha Mahajan¹Jaehui Byun¹Ishwarya Pichaimani¹Sangwook T. Yoon²Christopher T. Martin³Cassie S Mitchell^1*

• ¹Georgia Institute of Technology, Atlanta, United States
• ²Emory University, Atlanta, United States
• ³University of Minnesota Medical School, Minneapolis, United States

Accurate segmentation of cervical spinous processes on lateral X-rays is essential for reliable anatomical landmarking, surgical planning, and longitudinal assessment of spinal deformity. Yet no publicly available dataset provides pixel-level annotations of these structures, and manual delineation remains time-consuming and operator dependent. To address this gap, we curated an expert-labeled dataset of 500 cervical spine radiographs and developed CervSpineNet, a hybrid deep learning framework for automated spinous process segmentation. CervSpineNet integrates a transformer-based encoder that captures global anatomical context with a lightweight convolutional decoder that refines local boundaries. A compound loss function—combining Dice, Focal Tversky, Hausdorff distance transform, and Structural Similarity (SSIM) terms—jointly optimizes region overlap, class balance, structural fidelity, and boundary accuracy. The model was trained and evaluated on original, contrast-enhanced (CLAHE), and augmented dataset variants and benchmarked against four baselines: U-Net, DeepLabV3+, the Segment Anything Model (SAM), and a text-guided SegFormer. Across all experiments, CervSpineNet consistently outperformed competing methods, achieving mean Dice coefficients above 0.93, IoU values above 0.87, and SSIM above 0.98, with substantially lower HD95 distances. The model demonstrated strong agreement with ground truth (global MAE ≈ 0.005) and efficient inference times of 5–10 seconds per image. With a compact footprint (~345 MB), CervSpineNet runs on standard clinical hardware and reduces manual annotation time by approximately 96%, enabling rapid and reproducible segmentation for surgical evaluation, deformity monitoring, and large-scale retrospective studies. Together, the dataset and CervSpineNet provide a scalable, reproducible foundation for advancing AI-assisted cervical spine analysis in both research and clinical practice.