JMIR Publications

ABSTRACT

Cancer diagnosis increasingly depends on heterogeneous data streams, including radiology, digital pathology, molecular profiling, laboratory testing, and longitudinal clinical records. Artificial intelligence (AI) has made substantial progress in isolated tasks such as lesion detection, histopathological classification, and risk prediction. Yet many systems remain narrow, static, and detached from the diagnostic workflow. This limits their value in oncology, where diagnostic reasoning is multimodal, iterative, and context dependent. Large language models, multimodal foundation models, and tool-using autonomous systems now provide the technical basis for AI agents that coordinate diagnostic processes rather than execute isolated predictions. Their clinical value should be framed as supervised orchestration. An agent must select tools, preserve evidence provenance, expose uncertainty, and remain auditable by clinicians. Early oncology-specific evidence supports this direction. An autonomous precision-oncology agent evaluated on 20 realistic multimodal cases achieved 87.5% correct tool use, 91.0% correct clinical conclusions, and 75.5% accurate guideline citation, highlighting both promise and a remaining reliability gap. This perspective is organized around four operational stages: multimodal data collection, multimodal data preprocessing, multimodal fusion and representation learning, and diagnostic decision support. Across these stages, agentic orchestration differs from conventional multimodal AI through task planning, dynamic tool invocation, longitudinal context, uncertainty-aware fusion, traceable reasoning, and physician-in-the-loop governance. Evidence from PubMed-indexed studies of oncology-agent validation, multimodal data-supply-chain deployment, oncology chatbot benchmarks, tumor multidisciplinary team workflows, source-attributed uro-oncology decision support, explainable real-world oncology AI, hallucination safety, and trustworthy-AI governance anchors the discussion.