Radiologist-like Progressive Radiology Report Generation and Benchmarking
Keywords: radiology report generation
TL;DR: new benchmark for radiology report generation
Abstract: Radiology report generation is a critical application at the intersection of radiology and artificial intelligence. It aims to reduce radiologists' workload by automating the interpretation and reporting of medical images. Previous works have employed diverse approaches, with some focusing solely on imaging data while others incorporate the indication but often neglect the interrelationships among different report sections.
Our work identifies and harnesses the intrinsic relationships between the indication, findings, and impression sections of a radiology report.
The indication section provides the clinical context and specifies the reason for the examination, setting the stage for targeted image analysis. The findings section details the radiologist's observations from the image, including identified abnormalities and relevant normal findings. The impression section synthesizes these observations to form a diagnostic conclusion, directly addressing the clinical query presented in the indication.
By mapping these relationships, we propose a Radiologist-Like Progressive Generation (RLPG) framework that mirrors the radiologist's workflow for report generation.
Initially, an image encoder and a large language model process the imaging data alongside the indication to generate detailed findings. Subsequently, the same image, the indication, and the predicted findings are utilized to produce a concise impression. This method improves the alignment between report sections and improves the clinical relevance of the generated reports.
To facilitate research and benchmarking in report generation, we introduce MIMIC-1V3 (i.e., 1 case vs. 3 sections), a curated dataset derived from the MIMIC-CXR by dividing each report into three sections: indication, findings, and impression.
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2287
Loading