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

Introduction

Radiology departments are increasingly interested in AI‑driven tools that can turn free‑form dictation into structured reports. Most existing solutions rely on large, cloud‑hosted language models, which raise concerns about data privacy, latency, and the need for high‑end GPUs. MedPAO‑Assist tackles these challenges by delivering a fully offline, edge‑deployable system that runs on a standard laptop CPU with roughly 3.3 GB of RAM. The framework combines a small 0.6 B‑parameter language model with a multi‑agent architecture to handle concept extraction, SNOMED‑CT grounding, and report generation for four major imaging modalities.

Edge‑Ready Architecture

The core of MedPAO‑Assist is an Extract‑Ground‑Generate pipeline orchestrated by three specialized agents:

1. Extraction Agent parses radiologist dictation (or voice input) to identify clinical concepts and findings.
2. Grounding Agent maps those concepts to SNOMED‑CT identifiers, ensuring semantic consistency across reports.
3. Generation Agent assembles the grounded concepts into a modality‑specific structured template (e.g., chest X‑ray, chest CT, abdomen CT, brain MRI).

A fourth, speech‑driven Correction Agent enables a human‑in‑the‑loop workflow: radiologists can issue voice commands to amend or confirm sections without touching a keyboard. All agents run asynchronously on separate threads, keeping the overall latency around 31 seconds from dictation to final structured output.

Modality Awareness and Performance

MedPAO‑Assist is explicitly trained to recognize the nuances of each supported modality. By conditioning the generation step on modality tags, the system produces reports that respect the distinct sections and terminology required for chest X‑ray versus brain MRI. Performance profiling shows:

• Memory footprint: ~3.3 GB peak RAM, well within the limits of typical clinical workstations.
• CPU‑only latency: ~31 seconds end‑to‑end on a mainstream laptop processor, meeting real‑time expectations for most radiology workflows.
• Accuracy benchmarks: Individual stage evaluations (concept extraction, ontology grounding, categorization) indicate competitive precision and recall compared to larger cloud‑based baselines, despite the modest model size.

Practical Implications for Clinical Adoption

The edge‑centric design addresses three persistent barriers to AI uptake in radiology:

1. Data privacy: No patient data leaves the local machine, eliminating regulatory hurdles associated with cloud transmission.
2. Infrastructure cost: Hospitals can deploy the system without investing in expensive GPUs or dedicated servers.
3. Workflow integration: Voice‑only interaction and modality‑aware templates fit naturally into existing reporting pipelines, reducing the learning curve for radiologists.

By demonstrating that a small language model can power a complete, multi‑agent reporting workflow, MedPAO‑Assist opens the door for broader deployment of AI tools in settings where connectivity, budget, or privacy constraints have previously been prohibitive.

Conclusion

MedPAO‑Assist showcases a pragmatic path forward for AI‑augmented radiology: a lightweight, modality‑aware, agentic system that runs entirely on the edge. Its modest hardware requirements, fast latency, and seamless voice‑driven interface make it a compelling candidate for real‑world clinical integration, encouraging further exploration of small‑model, offline AI solutions across other medical domains.