Agentic Multi-Agent AI for Cancer Pathology Imaging: Autonomous Triage, ROI Navigation, and Adaptive Learning

Digital pathology has unlocked gigapixel whole-slide images (WSIs) for computational analysis in cancer diagnosis.However, conventional deep learning models struggle with the scale and complexity of WSIs, requiring exhaustive patch analysis and extensive annotations.We propose a novel agentic AI approach for pathology imaging that leverages autonomous multi-agent systems to mimic human decision-making in clinical workflows.Our system operates at both patch-level (e.g., PatchCamelyon) and whole-slide levels, integrating triage agents for rapid screening, region-of-interest (ROI) navigation agents for goal-directed slide examination, and adaptive labeling agents for active learning and escalation to experts.Built on foundation models for histopathology, the agents plan and act collaboratively: the triage agent quickly flags suspicious slides or regions, the ROI agent autonomously zooms and pans to gather diagnostic evidence, and the labeling agent requests expert annotation when uncertainty is high, continuously improving the system.We evaluate our approach on public datasets and simulated pathology environments.Empirical results demonstrate that our agentic system achieves high diagnostic accuracy comparable to state-of-the-art MIL methods, while dramatically reducing the number of image patches analyzed per slide.In a metastasis detection task, the multi-agent approach reached an AUC of exhaustive analysis.Through active learning, the system rapidly adapts to new tasks, requiring only a few minutes of expert feedback to attain near-optimal performance.We provide end-to-end implementation in Python and R (PyTorch/torch), and demonstrate that the approach runs on mid-level GPUs (e.g., RTX 3060) and scales to larger compute.Our work illustrates how agentic AI principles autonomous planning, tool integration, and human-in-the-loop learning can transform digital pathology by improving efficiency, adaptability, and clinical integration.