GPT-4.0 Shows Strong Potential for Real-Time Multilingual Translation of Radiology Reports

Language barriers can be an obstacle for radiologists serving diverse populations, but a new study evaluates whether large language models—specifically Chat GPT-4.0—can provide fast, accurate translations of radiology reports directly within the imaging workflow.

“I grew up across several European countries and encountered language barriers from an early age,” said presenter Robert Terzis, MD, radiologist with the Institute for Diagnostic and Interventional Radiology at University Hospital Cologne in Germany. “In today’s globalized world, many patients present radiology reports clinicians cannot read. That can delay diagnosis and risk suboptimal care.”

In a two-center experimental study, the team selected 100 anonymized radiology reports—X-ray, US, CT or MRI—to reflect a broad spectrum of clinical reporting styles. Reports were translated into French, Spanish, Russian and English using GPT-4.0 with zero-shot prompting, producing 400 translations. Eight bilingual radiologists (two per language) evaluated the results for readability, completeness, factual accuracy, speed and potential harm.

GPT-4.0 delivered translations in under 25 seconds on average, ranging from nine seconds for English X-ray reports to 24 seconds for Russian CT reports.

“Speed is a major benefit,” Dr. Terzis said. “GPT-4.0 can translate across modalities and languages, regardless of complexity.”

High Overall Quality, Notable Language Differences

Translations into English, French and Spanish achieved strong ratings for completeness and factual accuracy. Median overall translation quality was rated a 4.5 out of 5, with English and Spanish receiving the highest scores. Content completeness reached 91%, and factual correctness averaged 79%, with the highest accuracy in English at 84%.

By contrast, Russian translations showed reduced performance, including nine of the study’s 16 potentially harmful errors introduced by AI. Many errors were linked to literal translations or mishandled German medical abbreviations. One example: GPT-4.0 incorrectly translated Verhalt as “behavior,” misinterpreting the clinical phrase abgekapselter Verhalt, which actually indicates an encapsulated abscess.

“Literal renderings that are inappropriate in the target language were a consistent challenge,” Dr. Terzis said.

Still, the model occasionally demonstrated surprising contextual reasoning. In two French cases, GPT-4.0 correctly expanded German cardiac MRI abbreviations into clinically accurate phrases—an unexpected strength.

Translation quality varied little by modality. US reports received slightly higher readability scores than CT, but completeness, accuracy and usefulness for medical translators were high across X-ray, US, CT and MRI. “When we looked across modalities, translation quality was remarkably consistent,” Dr. Terzis said.

Human Oversight Remains Critical

Despite encouraging performance, the authors stress that GPT-4.0 is not ready for autonomous clinical deployment.

“In our dataset, 16 translations (4%) contained potentially harmful errors,” Dr. Terzis said. “Human oversight remains essential. These systems should serve as semi-automated aids that generate draft translations for validation.”

Bilingual radiologists who reviewed the outputs were enthusiastic about the model’s potential, particularly for streamlining workflows and reducing delays for patients who cannot read the hospital’s primary language.

“All text-based medical documentation could benefit from real-time translation,” Dr. Terzis said. API-based integration into the radiology information system could automatically generate draft translations for clinician review, helping hospitals provide same-day translated reports and improving patient understanding.

“With rising medical tourism and increasingly diverse patient populations, the need for scalable solutions will only grow,” he said. “Our next step is to study how GPT-4.0 performs when embedded directly into real clinical workflows.”

Access the presentation, “Integration of GPT-40 in Radiological Workflow for Real-Time Multilingual Translation of Radiology Reports Across Imaging Modalities,” (W3-STCE2-1) on demand at RSNA.org/MeetingCentral.