AI agents in cancer imaging: Concepts, advances, and clinical perspectives.

Abstract

Medical imaging serves as a critical source of evidence for cancer diagnosis, treatment, and follow-up. However, most existing medical imaging artificial intelligence (AI) systems still operate on predefined tasks and inputs, limiting their ability to address the evolving needs of clinical practice and research. AI agents use foundation models as central engines for reasoning and orchestration. They organize analytical processes according to task objectives, invoke external tools and data resources, and adapt subsequent actions based on intermediate results. They may therefore facilitate a shift in cancer imaging AI from isolated model applications toward continuous workflow support. This review outlines the fundamental principles, system components, and evaluation approaches for AI agents from a cancer imaging perspective and summarizes representative advances in image interpretation, clinical decision support, and research workflows. Current evidence suggests that agents can connect previously fragmented imaging, clinical, and knowledge resources, thereby potentially improving the continuity and traceability of complex tasks. Nevertheless, clinical evidence for AI agents remains limited, as most studies have been conducted in controlled settings and have yet to demonstrate consistent value in real-world clinical workflows. Clinical translation is further constrained by the computational demands and potential error propagation inherent in multistep execution, as well as by the dependence of agent performance on specific tools, data environments, and workflow configurations. Future research should move beyond task-level performance toward system-level validation, with particular emphasis on whether agents can reliably organize information, interpret findings, and collaborate effectively with clinicians. Such evidence will be essential to clarify their practical role in cancer imaging care and research.