Multimodal LLMs Transform Volumetric Radiology AI

This comprehensive review examines the evolving landscape of volumetric radiology AI in the context of advancements in multimodal large language models (MLLMs). It highlights how MLLMs are pushing radiology AI beyond traditional task-specific image analysis towards more holistic multimodal understanding and reasoning. A core challenge identified is the representational mismatch between the full-volume spatial context and quantitative information required for clinical radiological interpretation, and the common MLLM inputs of selected 2D images, compressed visual data, or text derived from reports. The review stresses that reliable volumetric radiology AI necessitates representations that faithfully preserve 3D information and systems capable of accessing, verifying, and integrating this data across clinical workflows. The authors categorize existing literature around volumetric representation, multimodal understanding, and agentic orchestration, linking these to clinical applications and evaluation. They introduce a "Claim-Design-Validation" framework to ensure that technical, workflow, and clinical claims are appropriately supported by design and validation. The review concludes that native volumetric modeling and agentic capabilities are crucial, depending on the specific spatial, quantitative, contextual, and workflow demands of the intended radiological task, and that clinical credibility requires faithful 3D representation, traceable system behavior, and clear human oversight.