arXivDaily arXiv每日学术速递 周一至周五更新
​通知:联邦节假日调休:哥伦布日 / 原住民日(Columbus Day / Indigenous Peoples' Day),今日(13日)无更新。
arXiv 2606.20477cs.CVcs.CLcs.LG

面向放射学的空间定位2D视觉-语言模型的可扩展训练

Scalable Training of Spatially Grounded 2D Vision-Language Models for Radiology

Computer Vision Group, University of Freiburg, Germany(德国弗莱堡大学计算机视觉组);Department of Radiology, Medical Center -- University of Freiburg, Germany(德国弗莱堡大学医学中心放射科);CRIION-AI Lab, Freiburg, Germany(德国弗莱堡CRIION-AI实验室)

Yusuf Salcan, Simon Ging, Robin Tibor Schirrmeister, Philipp Arnold, Elmar Kotter, Behzad Bozorgtabar, Thomas Brox

AI总结:

提出RefRad2D大规模双语数据集,通过LLM和自动分割生成空间定位数据,训练RadGrounder模型联合完成报告生成、VQA和空间定位,在外部基准上取得竞争性结果。

摘要:

我们研究了如何在没有手动空间标注的情况下,为放射学训练具有视觉定位能力的视觉-语言模型(VLM)。我们引入了RefRad2D,这是一个大规模的双语(德语/英语)数据集,包含来自临床实践的120万对CT和MR图像-文本对,并通过基于LLM的筛选和自动分割自动生成任务特定的VQA和空间定位子集。在此数据上训练的模型RadGrounder联合执行报告生成、视觉问答以及通过边界框检测或分割进行的空间定位。在外部VQA基准(Slake,VQA-RAD)上,RadGrounder取得了与专用医学VLM竞争的结果。将我们的临床数据加入训练混合集,相比于仅在下游数据集上微调,提高了开放式VQA的性能,显示了数据集的迁移性。关键在于,添加定位监督不会降低语言质量,从而在不牺牲VQA性能的情况下实现空间可验证的输出。

英文摘要:

We study how to train visually grounded vision-language models (VLMs) for radiology without manual spatial annotations. We introduce RefRad2D, a large-scale bilingual (German/English) dataset of 1.2M CT and MR image-text pairs derived from clinical practice, with task-specific VQA and spatial grounding subsets generated automatically via LLM-based curation and automated segmentation. Trained on this data, our model RadGrounder jointly performs report generation, visual question answering, and spatial grounding via bounding-box detection or segmentation. On external VQA benchmarks (Slake, VQA-RAD), RadGrounder achieves competitive results with specialized medical VLMs. Adding our clinical data to the training mixture improves open-ended VQA over fine-tuning on the downstream datasets alone, showing the transferability of our dataset. Crucially, adding grounding supervision does not degrade language quality, enabling spatially verifiable outputs at no cost to VQA performance.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑