Edge-Deployed Lightweight LLMs for Medical Imaging: Real-Time Dual Report Generation with Progression Aware Summarization

Abstract

Medical imaging is central to modern diagnostics, with more than 3.6 billion procedures conducted annually worldwide and radiology workloads increasing by 30% over the past decade. However, reliance on cloud based large language models (LLMs) for automated report generation raises latency, privacy, and accessibility concerns. To address these challenges, we propose an Edge-Deployed Lightweight LLM framework that integrates Vision Transformer (ViT)-based imaging analysis with compressed LLM reasoning for real time, on-device medical diagnostics. Our framework generates dual outputs: (i) clinician-facing structured diagnostic reports and (ii) patient-friendly summaries, while also incorporating progression aware tracking by comparing previous reports. The system was evaluated on a dataset of 6,736 cases and ∼ 48,500 medical images spanning X-ray, CT, MRI, and ultrasound modalities. Compression techniques; distillation, pruning, and INT8 quantization reduced model size from 1200 MB to 620 MB, achieving a 48.3% reduction without compromising accuracy. The lightweight LLM achieved an AUC of 92.1%, F1-score of 90.3%, and reduced inference time to 0.8 s per report, nearly twice as fast as larger baselines. Bland Altman analysis further confirmed improved agreement (bias –0.01; limits –0.34 to +0.31). These results validate the framework’s potential to deliver explainable, private, and real time AI-driven diagnostics across both advanced hospitals and resource-constrained clinics.

Ähnliche Arbeiten