Agentic AI in Radiology: Definitions and Applications (Advanced) (2025)-M8-CIN08-2025

Video Transcription

Video Summary

The session, moderated by neuroradiologist Paulo Quiriqui (UT Southwestern), focuses on “agentic AI” in radiology, covering foundations, current use cases, and future directions.<br /><br />Dr. Iman Banerjee (Mayo Clinic) presents Mayo’s in-house agentic framework, “Maya,” which queries consolidated EHR data in a HIPAA-compliant Google Cloud environment and returns answers grounded with source documents. Maya integrates with Epic and internal tools (Teams, SharePoint, ServiceNow). To reduce non-deterministic variation in responses, her team built macro libraries for common clinical queries. She also describes task-specific LLM development: weekly-supervised pipelines to extract critical/incidental findings from radiology reports, end-to-end report summarization producing both technical and patient-friendly “lay” summaries, and a prostate-cancer-specific chatbot trained from scratch to improve domain vocabulary coverage and support Q&A, treatment prediction, and guideline compliance checking. Hallucinations are mitigated through knowledge grounding, clinician verification, and feedback loops.<br /><br />Dr. George Sheep (Cornell) argues many hospitals need self-hosted, open-source models behind the firewall. He highlights no/low-code agent building (e.g., n8n), Model Context Protocol (MCP) for connecting LLMs to systems like PACS, and local hardware (Mac Studio) to run models. Example agents include “High-Res Reports” (resident vs attending report revision analysis) and “RadFlow” (report-based aneurysm workflow alerts).<br /><br />Quiriqui concludes with a roadmap: agents will evolve from embedded tools to orchestrated, semi-autonomous ecosystems, but require human-in-the-loop design, transparency/audit logs, equity, continuous monitoring, and departmental governance to ensure safety and maintain radiology leadership.

Keywords

agentic AI in radiology

Mayo Clinic Maya framework

HIPAA-compliant EHR querying

radiology report summarization

critical and incidental findings extraction

self-hosted open-source LLMs

Model Context Protocol (MCP) for PACS integration

human-in-the-loop governance and audit logs