Context-Aware Prompt Engineering and Time-Aware LLM Architecture for Radiology Report Generation

Abstract

Recent advances in large language models (LLMs) have enabled new possibilities for automated radiology reporting, yet key challenges remain, including lack of contextualization, absence of longitudinal reasoning, and risk of clinically inaccurate content. We propose a lightweight, modular architecture that combines a pre-trained LLM with a context-aware Prompt Constructor and a temporal reasoning engine. Prompts are dynamically adapted based on imaging modality, clinical indication, and prior reports, supported by a fuzzy logic-driven rule base and a temporal summarizer to ensure clinical nuance and continuity. Our system is evaluated through realistic case scenarios across multiple imaging types and externally validated on 250 cases from the MIMIC-CXR dataset. Results demonstrate significant gains in BLEU, ROUGE, and BERTScore over baseline prompting, with expert review confirming improved structure, diagnostic alignment, and reduced hallucinations. These findings support the practical integration of explainable and configurable LLMbased systems into real-world radiology workflows.

Ähnliche Arbeiten