Answers are grounded in this paper and cite their sources.
Deep learning has become the dominant paradigm for spine image segmentation, playing a critical role in computer-aided diagnosis, surgical planning, and postoperative evaluation. Despite rapid progress across CT, MRI, X-ray, and ultrasound, the literature remains fragmented across diverse architectures, datasets, and evaluation protocols. This fragmentation hinders objective comparison and obstructs clinical translation. Existing surveys do not fully capture recent methodological shifts, leaving a gap in the collective understanding of the field. To address this, our survey paper provides a systematic review of 172 studies published between 2014 and 2026. We consolidate research across five imaging modalities and three anatomical regions (cervical, thoracic, lumbar). Our primary contribution is a structured taxonomy that organizes deep learning methods by architectural paradigm: encoder-decoder networks, multi-stage frameworks, direct segmentation methods, unsupervised feature extractors, and emerging transformer and Mamba-based methods. We also present the most extensive compilation of publicly available spine datasets to date, accompanied by a summary of quantitative results to establish clear benchmarks and support reproducible research. This survey offers a structured synthesis of the field's evolution and provides concrete guidance for future research aimed at developing robust, generalizable, and clinically deployable spine segmentation systems.
No takes yet. Share an insight, caveat, or question.
Qadri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: