The adoption of large language models for radiology report structuring has been dominated by proprietary cloud services or open models that require high‑end GPUs, creating a critical bottleneck for clinical settings that demand data privacy and offline operation.
To address this, we introduce MedPAO-Assist, an edge‑deployable agentic framework that generates structured reports directly from radiologist dictation and voice commands. The system uses roughly $3.3\text{ GB}$ peak RAM and achieves an average end‑to‑end latency of about $31\text{ s}$ on a standard laptop CPU, thanks to an asynchronous multi‑threaded execution pipeline.
It currently supports four major imaging modalities: chest X‑ray, chest CT, abdomen CT, and brain MRI. To our knowledge, it is among the earliest agentic systems to run the entire pipeline with a small language model (SLM) in the $0.6\text{B}$‑parameter range.
The architecture follows an Extract‑Ground‑Generate paradigm across multiple agents:
- concept extraction
- SNOMED‑CT grounding
- structured report generation A speech‑driven human‑in‑the‑loop correction phase follows, creating a fully "no‑touch" workflow that requires no keyboard input, external API calls, or internet connectivity.
We evaluate each stage—concept extraction, ontology grounding, concept categorization—and provide hardware profiling results, demonstrating the system’s feasibility for real‑time, CPU‑only clinical deployment.
Review: MedPAO-Assist showcases a practical path toward efficient, privacy‑preserving radiology report automation on resource‑constrained devices, offering valuable insights for edge AI adoption in healthcare.