Reproduce “Radiology Report Generation for Low-Quality X-Ray Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hongze Zhu, Chen Hu, Jiaxuan Jiang, Hong Liu, Yawen Huang, Ming Hu, Tianyu Wang, Zhijian Wu and 1 more
This research tackles the performance decline of automated radiology report generation systems when processing low-quality X-ray images. It introduces an Automated Quality Assessment Agent (AQAA) to stratify image quality, establishes a new Low-quality Radiology Report Generation (LRRG) benchmark, and proposes a Dual-loop Training Strategy (DTS) that improves clinical precision from 0.353 to 0.411 and F1 score from 0.291 to 0.321 on mildly suboptimal inputs.
Key insights
- Current RRG systems, despite advanced architectures, exhibit fragility when confronted with real-world low-quality images, revealing a critical gap in clinical applicability.
- A systematic approach combining objective physics-based metrics, perceptual image quality assessment, and VLM-guided inference (AQAA) can reliably stratify image quality, with the AQAA achieving 99% consistency with expert-curated "retake pairs".
- Explicitly encouraging gradient coherence between high-quality and low-quality data during training enables models to learn robust, quality-agnostic diagnostic features rather than spurious quality-specific cues.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.