3D-CT-GPT++: Enhancing 3D Radiology Report Generation with Direct Preference Optimization and Large Vision-Language Models

ICLR 2025 Conference Submission13584 Authors

28 Sept 2024 (modified: 28 Nov 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Radiology Report Generation, 3D Medical Imaging, Direct Preference Optimization, Multimodal large Language Models
Abstract: Automatically generating radiology reports from three-dimensional medical images, such as 3D CT scans, plays a crucial role in modern diagnostics. Current approaches for generating 3D reports often adopt video processing methods, which struggle to effectively capture the relationships along the Z-axis. Additionally, multimodal large language model-based methods for generating 3D image reports face significant limitations, particularly in terms of the image encoder’s ability to represent 3D structures and the hallucinations that arise in generated content. To address these challenges, we propose the 3D-CT-GPT++ model. This model integrates the optimized 3D image encoder CTViT-V, specifically designed for chest CT scans, and builds upon the LLaVA-1.5 architecture. Furthermore, we introduce \textit{Direct Preference Optimization (DPO)}, where GPT-4 is used to score the outputs of our fully fine-tuned (SFT) model, creating a preference dataset for subsequent DPO training. DPO significantly reduces hallucinations in the report generation process, ensuring the generated reports are more aligned with clinical needs. We fine-tuned the model on both high-quality private and public datasets to ensure clinical relevance. Extensive experiments were conducted using standard natural language generation (NLG) evaluation metrics, including BLEU, METEOR, ROUGE-L, and GREEN, to assess the report generation performance. Experimental results demonstrate that 3D-CT-GPT++ significantly outperforms existing methods in terms of accuracy, fluency, clinical factual consistency, and clinical relevance, advancing the automation of 3D medical report generation.
Primary Area: foundation or frontier models, including LLMs
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 13584
Loading