X-Gen: Enhancing Radiology Report Generation via LLM-Driven Data Augmentation and Decoupled Training
Date
2025
Authors
Wang, C.
Chen, Q.
To, M.-S.
Kutaiba, N.
Yoo, J.-G.
Xie, Y.
Wu, Q.
Editors
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Conference paper
Citation
Proceedings of the International Conference on Digital Image Computing Techniques and Applications, 2025, pp.450-457
Statement of Responsibility
Chaohan Wang, Qi Chen, Minh-Son To, Numan Kutaiba, Jae-Gon Yoo, Yutong Xie
Conference Name
International Conference on Digital Image Computing Techniques and Applications (DICTA) (3 Dec 2025 - 5 Dec 2025 : Adelaide, Australia)
Abstract
The scarcity and limited accessibility of medical data significantly challenge deep learning applications in medical AI. Radiology report generation (RRG), a key medical AI research area, could greatly improve computer-aided diagnosis through automated X-ray image interpretation. However, obtaining paired X-ray images and reports is labor-intensive and restricted by strict regulations. Large language models (LLMs), such as GPT-4, provide a promising alternative by enabling cost-effective text data augmentation and report rewriting in varied styles. We rigorously assess augmented data's clinical accuracy and stylistic similarity to radiologist-authored reports through expert evaluations. Interestingly, augmented data enhances RRG model performance, yet performance declines when augmented data surpasses original data volume due to style distribution shifts. To mitigate this, we propose integrating a conditional variational autoencoder (cVAE) into the RRG model to separate medical semantics from writing styles during training, enabling better handling of augmented data's distribution shift. Our proposed method, X-Gen, combines data augmentation with decoupled training. Tested on two public Chest X-ray datasets and a private abdomen X-ray dataset, X-Gen significantly improves the performance of baseline models, showcasing its effectiveness and versatility in X-ray report generation.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
© Copyright 2025 IEEE