U2-BENCH
收藏资源简介:
U2-BENCH数据集是首个全面评估大型视觉语言模型在超声理解方面的基准数据集,涵盖15个解剖区域,7241个病例,并定义了8个临床相关任务,包括诊断、视图识别、病变定位、临床价值评估和报告生成等,共涉及50个超声应用场景。数据集旨在解决超声图像解释中存在的挑战,如操作者依赖、噪声和解剖复杂性等问题。该数据集为LVLM研究提供了严格的统一测试平台,有助于推动LVLM在医疗超声影像领域的多模态理解研究。
U2-BENCH is the first comprehensive benchmark dataset for evaluating large vision-language models (LVLMs) on ultrasound comprehension. It encompasses 15 anatomical regions and 7241 clinical cases, and defines 8 clinically relevant tasks covering a total of 50 ultrasound application scenarios, including diagnosis, view recognition, lesion localization, clinical value assessment, report generation and more. This dataset aims to address the challenges existing in ultrasound image interpretation, such as operator dependence, image noise and anatomical complexity. It provides a rigorous and unified testbed for LVLM research, which helps promote multimodal understanding studies of LVLMs in the field of medical ultrasound imaging.

- 1U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding香港浸会大学 · 2025年



