Structured representations for advancing radiology report generation

Liao, Yuxiang 2025. Structured representations for advancing radiology report generation. PhD Thesis, Cardiff University. Item availability restricted.

	PDF (Yuxiang Liao, PhD, Thesis) - Accepted Post-Print Version Restricted to Repository staff only until 1 May 2027 due to copyright restrictions. Available under License Creative Commons Attribution No Derivatives. Download (7MB)
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (215kB)

Abstract

Radiology report generation (RRG) lies at the intersection of medical imaging, natural language processing (NLP), and clinical practice. It promises to alleviate radiologist workload and reduce reporting variability but remains constrained by data imbalance, linguistic heterogeneity, and the omission of clinically significant findings. This thesis investigates structured representations as a means of mitigating these limitations and enhancing the clinical reliability of RRG. The research is organised into three phases. First, we developed a NLP pipeline combining coreference resolution (CR) and information extraction (IE) to transform narrative chest X-ray reports into structured graph-based representations. This pipeline normalised linguistic variability and enabled reproducible large-scale structuring of radiology corpora. Second, we proposed a graph-guided RRG approach, which conditions text generation on informative clinical entities derived from structured graphs. This approach enabled tiny models to achieve performance comparable to substantially larger architectures, though challenges remained in balancing generation performance with label precision. Third, we reconstructed reports into observation-centric units to analyse intrinsic data properties affecting RRG. This revealed underestimated failure modes, including long-tail suppression of rare findings and heterogeneous learning dynamics across different clinical observations. Overall, the findings show that structured representations are a viable component of the RRGworkflow, while also revealing key bottlenecks that conventional benchmarks tend to obscure. This thesis contributes a methodology, resources, and analytical framework for incorporating structured representations into RRG. It highlights both the opportunities and challenges of bridging vision-language models with structured clinical knowledge, paving the way toward clinically reliable automated radiology reporting.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Funders:	China Scholarship Council, Cardiff University Scholarship Program for Postgraduate Research
Date of First Compliant Deposit:	1 May 2026
Date of Acceptance:	24 April 2026
Last Modified:	01 May 2026 10:42
URI:	https://orca.cardiff.ac.uk/id/eprint/186729

Actions (repository staff only)

Edit Item