arXiv 2605.190272026-05-22cs.CV

MedFM-Robust：医学基础模型的鲁棒性基准测试

MedFM-Robust: Benchmarking Robustness of Medical Foundation Models

Beijing Normal University, China（北京师范大学）；University of Exeter, United Kingdom（埃克塞特大学）；University College London, United Kingdom（伦敦大学学院）；Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates（穆罕默德·本·扎耶德人工智能大学）；University of Surrey, United Kingdom（萨里大学）

Xiangxiang Cui, Tianjin Huang, Yifang Wang, Lijie Hu, Lu Yin

备注：: MICCAI2026

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文提出了一个包含40种扰动类型（12种基础、28种医学特定）的鲁棒性基准，评估了多种医学基础模型在VQA、视觉定位、图像描述和分割任务上的表现，发现微调策略主导鲁棒性、医学特定扰动对分割影响大、零样本VQA鲁棒性依赖模型等关键结论。

摘要：

医学基础模型在临床任务中取得了显著性能，但它们在现实世界扰动下的鲁棒性仍未得到充分探索。我们提出了一个鲁棒性基准，包含8种成像模态下的40种扰动类型（12种基础、28种医学特定），评估了五个视觉语言模型（LLaVA-Med、MedGemma、MedGemma-1.5、Gemini-2.5-flash和GPT-4o-mini）在VQA、视觉定位和图像描述任务上的表现，以及两个分割模型（MedSAM、SAM-Med2D）在五种微调策略下的性能。我们的发现表明：（1）微调策略主导鲁棒性，LoRA的退化程度几乎是全微调的两倍，而SAM-Med2D的Adapter提供了有利的效率-鲁棒性权衡。（2）医学特定扰动对分割造成不成比例的损害，15个最严重扰动中有9个是领域特定的。（3）LoRA微调的视觉定位性能下降超过40个百分点，而零样本图像描述保持稳定（下降<7%）。零样本VQA表现出模型依赖的鲁棒性——医学模型下降不到20%，而Gemini-2.5-flash下降54%。通用视觉语言模型在VQA上准确率更高，但在视觉定位上失败；在医学视觉语言模型中，MedGemma表现出最佳的整体稳定性。这些结果提供了部署指南，并强调了医学AI领域特定鲁棒性评估的必要性。我们的代码可在 https://abnerai.github.io/MedFM-Robust 获取。

英文摘要：

Medical foundation models have achieved remarkable clinical performance, yet their robustness under real-world perturbations remains underexplored. We present a robustness benchmark comprising 40 perturbation types (12 base, 28 medical-specific) across eight imaging modalities, evaluating five VLMs (LLaVA-Med, MedGemma, MedGemma-1.5, Gemini-2.5-flash and GPT-4o-mini) on VQA, visual grounding, and captioning, alongside two segmentation models (MedSAM, SAM-Med2D) with five fine-tuning strategies. Our findings reveal: (1) Fine-tuning strategy dominates robustness, with LoRA exhibiting nearly double the degradation of full fine-tuning, while SAM-Med2D's Adapter offers favorable efficiency-robustness trade-off. (2) Medical-specific perturbations disproportionately damage segmentation, with 9 of 15 top corruptions being domain-specific. (3) LoRA-tuned visual grounding drops over 40 points, whereas zero-shot captioning remains stable (<7% drop). Zero-shot VQA shows model-dependent robustness--medical models drop under 20% while Gemini-2.5-flash drops 54%. General-purpose VLMs achieve higher VQA accuracy but fail on grounding; among medical VLMs, MedGemma demonstrates the best overall stability. These results provide deployment guidelines and underscore the necessity of domain-specific robustness evaluation for medical AI. Our code is available at: https://abnerai.github.io/MedFM-Robust.

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。