Does Grounding Improve Radiology Report Generation? An Empirical Study on PadChest-GR

Mohamed Aas-Alas, Alberto Albiol, Roberto Paredes

Keywords: Radiology Report Generation, Spatial Grounding, Region-to-Text Generation, Chest X-ray Interpretation

TL;DR: This study shows that injecting spatial grounding into a RRG model substantially boosts linguistic and clinical performance on PadChest-GR.

Abstract: Radiology Report Generation (RRG) aims to automatically produce clinically accurate descriptions of medical images, yet current models often struggle with incomplete findings, generic phrasing, and hallucinations due to the absence of explicit grounding signals. To address these limitations, we propose a grounding-based RRG framework that integrates spatially localized visual evidence into the generation process. Our approach combines a vision encoder ViT with a language decoder LLM GPT-2 through a lightweight transformer-based bridging module inspired by Bridge-Enhanced Vision Encoder–Decoder (VED) architectures. Grounding is introduced using bounding boxes of anatomical regions and pathologies, enabling the model to attend to both global and localized features. We further adopt the region-to-text task, where the model generates findings directly from specific regions of interest. Experiments on the PadChest-GR dataset demonstrate that grounding substantially improves linguistic quality and clinical accuracy, with the full image plus grounding mask configuration achieving the strongest gains across BLEU, ROUGE-L, CIDEr, BERTScore, CheXbert F1, and RadGraph F1. Analyses also show that even partial or noisy grounding yields consistent benefits.

Primary Subject Area: Generative Models

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Reproducibility: https://github.com/LightVED-prhlt/RegionFormer

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 109