AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Rethinking Radiology Report Generation: From Narrative Flow to Topic-Guided Findings

Sheng Cheng, Devika Subramanian

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Argues that training radiology VLMs to mimic the narrative flow of expert reports encourages linguistic priors over visual grounding, and proposes topic-guided findings generation as an alternative.

Problem

Vision-language models for radiology report generation are trained to reproduce the narrative flow of human reports, which may weaken grounding in visual evidence and cause factual inaccuracies.

Approach

Reframes report generation from narrative flow to topic-guided findings, hypothesizing that organizing generation around topics rather than inter-sentence narrative reduces reliance on linguistic priors and improves visual grounding.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train on single-topic queries instead of concatenated narrative of all topics (ablation)LLaVA-TA with MLP+LLM tuning · MIMIC-CXR (topic-level)RadGraph F1: 27.8 -> 46.5Improved
replace narrative reports with topic-decomposed data (LLaVA-T) (ablation)LLaVA-Rad · MIMIC-CXR (report-level)CheXpert Macro-F1-14: 40.1 -> 63.0Improved
add topic decomposition (ablation)LLaVA-Rad · MIMIC-CXR (report-level)RadGraph F1: 30.8 -> 34.2Improved
add anatomy masks when full LLM is tuned (ablation)LLaVA-T · MIMIC-CXR (report-level)RadGraph F1: 34.2 -> 34.3No clear effect
add anatomy masks when only MLP is tuned (ablation)LLaVA-T · MIMIC-CXR (topic-level, MLP-only)RadGraph F1: 5.8 -> 41.4Improved
add anatomy masks when only MLP is tuned (ablation)LLaVA-T · IU-Xray (report-level, MLP-only)RadGraph F1: 6.8 -> 34.0Improved
clean reports (remove comparisons and speculation) without topic split (ablation)LLaVA-Rad · MIMIC-CXR (MLP+LLM)RadGraph F1: 30.8 -> 32.4Improved
clean reports without topic split, MLP-only tuning (ablation)LLaVA-Rad · MIMIC-CXR (MLP-only)RadGraph F1: 19.8 -> 10.4Hurt
concatenate all topic sentences as target (allGTTopic) instead of original narrative (ablation)LLaVA-Rad · MIMIC-CXR (MLP+LLM)RadGraph F1: 30.8 -> 27.2Mixed
use topic-guided training instead of narrative training on out-of-distribution dataLLaVA-TA · IU-Xray (MLP+LLM)RadGraph F1: 19.8 -> 31.4Improved
use topic-guided plus anatomy training vs narrative baselineLLaVA-TA · MS-CXRHit score (%): 29.2 -> 59.2Improved
randomly drop 1-2 topics from 5% of training samples (ablation)LLaVA-TA (MLP-only) · MIMIC-CXRRadGraph F1: 41.4 -> 40.9No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy