Radiology Report Generation
A task involving the use of deep learning models to automatically analyze medical images, such as chest X-rays and CT scans, and produce structured textual reports. Its core value lies in longitudinal comparison of multiple examination records to identify changes in lesions over time.
Timeline (last 90 days)
Commonly used RRG evaluation metrics include natural language generation metrics such as BLEU, ROUGE, and CIDEr, as well as clinical accuracy measures such as F1 RadGraph and BERTScore.
General-purpose large language models and multimodal large models, when directly applied to specialized radiology scenarios, struggle to accurately capture the correspondence between medical imaging features and report terminology.
TALON is a temporally aware longitudinal radiology report generation framework that adaptively integrates variable-length patient history without being constrained to a fixed-length history window.
Most mainstream RRG models currently generate reports based on a single examination or, at most, the most recent prior examination, limiting accurate longitudinal comparison.
Recent methods that incorporate multiple prior examinations typically simply aggregate a fixed-length history without explicitly modeling the role relevance of each prior examination before fusion.
All Facts (5)
Commonly used RRG evaluation metrics include natural language generation metrics such as BLEU, ROUGE, and CIDEr, as well as clinical accuracy measures such as F1 RadGraph and BERTScore.
50%UnverifiedGeneral-purpose large language models and multimodal large models, when directly applied to specialized radiology scenarios, struggle to accurately capture the correspondence between medical imaging features and report terminology.
50%UnverifiedTALON is a temporally aware longitudinal radiology report generation framework that adaptively integrates variable-length patient history without being constrained to a fixed-length history window.
50%UnverifiedMost mainstream RRG models currently generate reports based on a single examination or, at most, the most recent prior examination, limiting accurate longitudinal comparison.
50%UnverifiedRecent methods that incorporate multiple prior examinations typically simply aggregate a fixed-length history without explicitly modeling the role relevance of each prior examination before fusion.
50%