Spine imaging represents a complex diagnostic frontier characterized by anatomical variability, motion artifacts, metallic instrumentation interference, and significant inter-reader diagnostic variability (κ=0.20 across institutions). While conventional discriminative artificial intelligence (AI) models achieve >95% accuracy in detecting degenerative changes, they remain limited by data scarcity, heterogeneous protocols, and poor generalizability. In the spine, these limitations are particularly relevant because clinical decisions can often depend on subtle distinctions (such as differentiating levels of canal or foraminal stenosis, characterizing Modic endplate changes, or assessing pedicle and vertebral morphology), where small inconsistencies can meaningfully alter management or surgical planning. Generative AI (GenAI) systems—including generative adversarial networks (GANs), diffusion models, and vision-language models (VLMs)—offer a paradigm shift by learning underlying data structures to generate high-quality synthetic outputs rather than merely classifying existing data. This narrative review, conducted using SANRA (scale for the assessment of narrative review articles) methodology across PubMed, Scopus, Embase, and Cochrane Library, examined GenAI applications in spine imaging. Eligible studies included observational designs through randomized controlled trials exploring image reconstruction, synthetic computed tomography (CT) generation, segmentation, and surgical planning applications. GAN-generated synthetic magnetic resonance imaging sequences reduce scan times by ~40% while maintaining diagnostic confidence; diffusion models enable radiation-free synthetic CT for preoperative planning; and VLMs generate structured radiology reports with hallucination rates <1.12%. However, critical barriers impede clinical translation: external validation gaps reveal AI performance collapse in real-world cohorts (sensitivity drops to 54.9% in cervical fracture detection); hallucinations and anatomical inaccuracies risk misguiding implant sizing; bias amplification magnifies demographic underrepresentation; and fragmented, small datasets lack standardized benchmarks. Technical fragility, computational demands, clinician trust deficits, and unresolved regulatory frameworks for iteratively-updating systems remain unaddressed. Successful integration requires coordinated development across 5 priorities: (1) multi-institutional datasets with cross-vendor harmonization, (2) federated learning frameworks preserving privacy, (3) uncertainty quantification and explainability tools, (4) outcome-linked clinical validation replacing technical metrics, and (5) workflow-integrated systems with DICOM-native interfaces and provenance tracking.
Ashraf et al. (Wed,) studied this question.