Abstract
Tuberculosis (TB) screening from chest X-rays is a setting where automated reading could ease severe shortages of radiologists, and vision-language models (VLMs) now attempt it while showing their reasoning step by step through chain-of-thought (CoT) prompting. Recent radiology models report accuracy gains from such reasoning, but mostly after extra training and measured as overall report quality, so whether prompting alone helps a model pick out findings, which reasoning step matters, and whether this holds outside English remain open. We study TB on chest X-rays with radiologist reports in Indonesian. Using one strong general-purpose VLM, we compare direct prompting against a base CoT and six arms that each change one of its reasoning steps, scoring six TB findings for accuracy (F1) and for how often the model reports findings the report does not (false positives), with paired significance tests. No CoT arm improves F1 over direct prompting. Adding a single verification step, in which the model re-checks each finding it reported, is the one clear gain: it significantly lowers false positives at no accuracy cost, and the findings it does report are well grounded in the model’s own earlier observations. This self-grounding explains why some prompts over-report more than others.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SAFER_010.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=PeOuWkIjLr
BibTex
@InProceedings{HanLut_Faithful_MICCAISAT2026,
author = { Hanfan, Luthfi Nur AND Montana, Giovanni AND Setiawan, Cindy Adelia AND Yunus, Reyhan Eddy AND Valindria, Vanya},
title = { { Faithful Reasoning, Not More Reasoning: Chain-of-Thought Ablation for Tuberculosis Findings in Indonesian Chest X-Rays } },
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}
}