Over the last few years, artificial intelligence (AI) has gradually permeated both daily life and clinical practice, reshaping how we approach complex tasks. In the spine field, AI applications have shown incredible potential in different domains, including patient phenotyping, diagnosis, decision-making, intraoperative guidance, and imaging analysis [1-4]. Due to its inherent complexity, heterogeneity, and often semiqualitative interpretation, spine imaging is particularly benefiting from the innovations brought by generative AI (GenAI), which is moving the field beyond conventional anatomical discrimination and classification toward synthesis, reconstruction, and multimodal reasoning [5].

This narrative review [6] carefully dissects the current applications of GenAI in spine imaging, while also providing technical insights that are often overlooked by clinicians yet are essential for understanding and distinguishing between different models. Among these, Generative Adversarial Networks are particularly attractive for tasks where speed and high-resolution output are critical, such as artifact reduction or cross-modality synthesis. Diffusion models, in contrast, tend to offer greater stability and anatomical fidelity, making them more suitable for reconstruction tasks that require anatomical consistency. Vision-language models are less relevant for image generation itself, but hold clear potential for workflow automation, report drafting, and image–text integration. Recognizing these differences is important, as each class of model comes with distinct validation requirements, limitations, and safety considerations.

From a clinical perspective, the potential applications are substantial. The authors highlight how GenAI can be used to enhance training in recognizing underrepresented pathologies (e.g., congenital conditions, infections, or postoperative changes) through synthetic data generation, as well as to improve image quality, assist in report generation, and enable cross-modality translation. The latter is particularly relevant, with recent studies demonstrating that synthetic computed tomography (CT) images can be derived from magnetic resonance imaging and vice versa, with the potential to reduce radiation exposure and avoid repeat imaging, thus decreasing costs and improving patient safety. Beyond image generation, the review also explores the role of generative methods in segmentation, deformity assessment, and surgical planning. The ability to create fully explorable 3-dimensional spinal models is especially notable, with clear implications for preoperative planning, such as evaluating pedicle screw trajectories, planning osteotomies, or improving intraoperative level identification.

At the same time, the authors appropriately caution that visually convincing outputs are not necessarily reliable. This distinction is critical in spine surgery, where even small geometric inaccuracies might lead to disastrous clinical consequences. Issues such as hallucinations and bias, particularly when models are trained on imbalanced datasets (i.e., characterized by overrepresented demographic groups or pathologies), remain significant barriers to safe implementation. Importantly, many of these limitations are not purely technical. Data scarcity, inconsistent annotation standards, and limited generalizability across institutions continue to affect the external validity of published data. Addressing these challenges will require coordinated efforts toward dataset harmonization, shared benchmarks, and prospective multicenter validation, as well as with shared policies to protect patients’ privacy. Nonetheless, the clinical utility of GenAI tools also requires seamless integration into existing workflows through DICOM-compatible outputs, traceable data provenance, and appropriate human oversight. A few additional points warrant further emphasis. First, evaluation should focus on clinically meaningful endpoints rather than relying solely on image similarity metrics, which may fail to capture subtle but important errors. Second, transparent reporting of failure cases, including anatomically implausible outputs or downstream decision errors, will be essential to define safety boundaries and clearly indicate areas for prompt implementation or needed improvement.

Overall, this review offers a thoughtful and forward-looking synthesis of GenAI in spine imaging. Its strength lies not only in summarizing current applications, but in outlining a pragmatic scheme centered on data quality, robust validation, uncertainty quantification, and workflow integration. If these challenges can be addressed, GenAI has the potential to evolve from a promising technology into a meaningful tool for safer, more efficient, and increasingly personalized spine care. This perspective will be essential for promoting multimodal integration of spine imaging with clinical, biomechanical, and biomarker data, promoting the establishment of an effective intelligencebased spine care model [7].

Citations to this article as recorded by