Development and expert radiologist validation of a custom pipeline for simplification of oncology radiology reports using large language model.

Authors

Abstract

Radiology reports are often written for clinicians and contain complex medical terminology, which patients struggle to understand. This comprehension gap may lead to anxiety and misinformed decisions. This is especially important in cancer care. Large language models (LLMs) provide an opportunity to bridge this gap. Direct use of LLM by patients may potentially be misleading and may not ensure clinical fidelity. We developed an LLM- based tool to automatically translate radiology findings into clear, patient-friendly language which could improve patient-centered care. To develop and validate a bilingual LLM driven tool that simplifies oncology radiology reports into patient-understandable English and Hindi, while maintaining diagnostic fidelity and emotional tone. This study was approved by the Institute's ethics committee. A retrospective corpus of 100 computed tomography (CT) reports (April 2025-July 2025) of patients with colo-rectal cancers were used for the development of the pipeline. Five large language models-GPT-4o, Gemini 2.5 Pro, Claude Opus, LLaMA-3.1-8B and Phi-3.5-mini-were tested. Five iterative prompt versions guided LLMs through successive refinements to ensure medical accuracy, clarity, and inclusion of a standard disclaimer. A custom pipeline; Vernacular Language Coverter(VLC), based on Gemini 2.5 Pro that takes a radiologist's report as input and generates two patient-facing outputs: (1) a simplified English and (2) a Hindi explanation was developed. The tool thus developed was then prospectively validated using 100 de-identified reports. Simplified outputs were reviewed by two radiologists, assessing accuracy, language clarity/Terminology and readability/tone. Completeness was assessed in terms of core diagnostic completeness and minor incompleteness. Flesch Reading Ease (FRE) was calculated for a fraction of reports. Mean rubric scores were high: Accuracy 4.77 ± 0.62, language clarity/Terminology 4.78 ± 0.56 and readability/tone 4.9 ± 0.32. Core diagnostic completeness was attained in all patients. 92% percent of reports were released 'as is. Readability improved markedly (Flesch Reading Ease 49.2→73). Empathy phrases in English and Hindi were appropriate. An AI-driven bilingual framework significantly enhances the clarity, tone, and readability of oncology radiology reports while retaining diagnostic precision. This tool demonstrates the feasibility of safe, patient-centered communication aligned with the goals of personalized medicine.

Tags

Topics