Abstract
Multilingual chest X-ray report generation requires more than fluent translation, because small changes in negation, uncertainty, anatomical location, or severity can alter clinical meaning. We study whether different multilingual generation pathways preserve clinical reliability for the same radiograph across languages. Three pathways are compared under a shared protocol: English report generation followed by translation, direct target-language generation, and LoRA-based adaptation of a medical vision-language model. The experiments use 2,000 chest X-ray studies and cover Arabic, Chinese, English, French, Hindi, Italian, Japanese, Korean, and Spanish. Assessment combines lexical metrics, ontology-normalized Clinical F1, reference-unsupported findings, cross-language consistency, and human clinical review. The adapted medical VLM achieves the highest Clinical F1 (0.901), lowest reference-unsupported rate (16.2%), and highest consistency score (0.886). Direct target-language generation is more reliable than English-first translation, but still shows language-dependent variation. These findings show that multilingual radiology reporting should be evaluated by both clinical correctness within each language and consistency of clinical content across languages.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ELAMI_009.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{DaoBil_Clinical_MICCAISAT2026,
author = { Daoud, Bilel AND ben Ghozlen, Ghazi AND Liu, Yi AND Khalfallah, Ali AND Ding, Shanli},
title = { { Clinical Reliability in Multilingual Chest X-ray Report Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}