Advances in multimodal large language models are pushing radiology AI from single‑image analysis toward multimodal understanding and reasoning. However, volumetric radiology faces a fundamental representation mismatch: clinical interpretation requires full‑volume spatial context and acquisition‑dependent quantitative information, while current MLLMs are typically conditioned on selected 2D images, compressed visual embeddings, or report text.
Reliable volumetric radiology AI therefore must preserve task‑relevant 3D information and enable systems to access, verify, and integrate this information within clinical workflows. This review surveys over 200 papers up to July 2026, organizing the literature around volumetric representation and multimodal understanding at the model level, agentic orchestration at the system level, and their connections to clinical applications and evaluation.
We highlight volumetric foundation models, language alignment and compression strategies, and agentic systems that extend MLLMs through planning, tools, memory, and workflow interaction. Distinctions are drawn between scenarios where selected 2D views or report‑mediated reasoning suffice and those that demand native volumetric modeling.
We also introduce a Claim‑Design‑Validation framework to assess whether technical, workflow, and clinical claims are matched by appropriate design and validation. The literature shows that native volumetric modeling and agentic capabilities depend on the spatial, quantitative, contextual, and workflow requirements of the target task.
Clinical credibility requires faithful volumetric representation, traceable system behavior, claim‑aligned validation, and clearly defined human oversight in realistic workflows. Volume can be expressed as $V = \int_{z} A(z) dz$.
Blogger's Review: This survey provides a comprehensive map of the technical landscape for volumetric radiology AI, offering clear guidance on model and system choices while emphasizing the necessity of aligning claims with rigorous validation to avoid black‑box pitfalls.