MedPAO-Assist: Edge-Deployable Agentic AI for Structured Radiology Reports on CPU

Forum topic · 小凯 · 2026-10-22

Summary

MedPAO-Assist (arXiv 2610.26913) is an edge-deployable agentic framework that generates structured radiology reports from radiologist dictation and voice commands without cloud services or GPUs. Built around an Extract-Anchore-Generate paradigm, it performs concept extraction, SNOMED-CT ontology anchoring, and structured report generation using only a 0.6B-parameter small language model, followed by voice-driven human-in-the-loop correction. The system runs entirely on a standard laptop CPU with ~3.3 GB peak RAM and ~31 seconds average end-to-end latency via asynchronous multi-threaded execution, supporting Chest X-ray, Chest CT, Abdomen CT, and Brain MRI. It requires no keyboard input, external API calls, or internet connectivity, enabling a fully touch-free workflow suited to privacy-sensitive, offline clinical environments. The paper evaluates concept extraction, ontology anchoring, and concept classification per pipeline stage, plus hardware profiling, demonstrating feasibility of real-time CPU-only clinical deployment.

MedPAO-Assist: Edge-Deployable, Modality-Aware Agentic AI for Radiology Report Structuring

Field: Machine Learning Authors: Shrish Shrinath Vaidya, Minmini Selvam, Ganapathy Krishnamurthi Published: 2026-10-22 arXiv: 2610.26913

Abstract

The adoption of Large Language Models for radiology report structuring has been largely dominated by either proprietary cloud-based models or large open models requiring high-end GPUs for deployment, which is a critical bottleneck for clinical adoption in radiology workflows that demand data privacy and offline operation. To address this, we present MedPAO-Assist, an edge-deployable agentic framework that generates structured radiology reports directly from radiologist dictation and voice commands, requiring only $\approx3.3$ GB peak RAM and achieving an average end-to-end latency of $\approx31$ seconds on a standard laptop CPU through an asynchronous, multi-threaded execution pipeline. The system currently supports four major imaging modalities -- Chest X-ray, Chest CT, Abdomen CT, and Brain MRI.

Key Points

Edge deployment: Runs the full pipeline on a 0.6B-parameter small language model (SLM) — among the first agentic systems to do so — with ~3.3 GB peak memory and ~31 s average end-to-end latency on a standard laptop CPU.
Architecture: A multi-agent design built on an Extract–Anchor–Generate paradigm performing concept extraction, SNOMED-CT ontology anchoring, and structured report generation.
Touch-free workflow: Voice-driven human-in-the-loop correction completes the pipeline with no keyboard input, external API calls, or internet connectivity.
Modality-aware: Supports four major imaging modalities — Chest X-ray, Chest CT, Abdomen CT, and Brain MRI.
Evaluation: Each pipeline stage (concept extraction, ontology anchoring, concept classification) is evaluated individually, with hardware profiling demonstrating feasibility for real-time, CPU-only clinical deployment.