AI Agents for Multimodal Oncology Diagnosis: Toward Transparent and Traceable Clinical Decision Support
TL;DR
This Viewpoint defines an AI agent as a feedback-driven system that maintains task state, selects among governed tools, observes results, and revises its plan under explicit safety constraints, and separates agents from multimodal foundation models, retrieval-augmented generation, and fixed workflow automation.
Abstract
Cancer diagnosis depends on data from radiology, digital pathology, molecular profiling, laboratory testing, and longitudinal clinical records. AI performs well in selected tasks, but most systems remain narrow and disconnected from the iterative reasoning required in oncology. This Viewpoint defines an AI agent as a feedback-driven system that maintains task state, selects among governed tools, observes results, and revises its plan under explicit safety constraints. This definition separates agents from multimodal foundation models, retrieval-augmented generation, and fixed workflow automation. We organize the discussion across multimodal data collection, preprocessing, fusion and representation learning, and diagnostic decision support. We distinguished agent-level evidence, component- or infrastructure-level evidence, and prospective propositions throughout. Clinical translation will require resilient failure handling, guideline version control, prospective evaluation, computational and workflow feasibility, and clinician authority over final decisions. The near-term opportunity is therefore transparent and traceable clinical decision support rather than autonomous cancer diagnosis.