提出可验证可量化的病理图像解释系统，让AI解释从'能看'迈向'可信'。
From Explainable to Explained AI: Ideas for Falsifying and Quantifying Explanations
- 构建人机协同的视觉语言交互系统，通过滑动窗口实验验证解释主张。
- 利用通用视觉-语言模型量化解释的预测能力，区分不同解释的优劣。
- 适用于数字病理等高风险场景，推动可解释AI向可验证解释演进。
深度学习模型的解释对医学影像分析系统的临床应用至关重要。良好的解释应能揭示模型是否依赖虚假特征，影响泛化性能或伤害特定患者群体，也可能带来新的生物学发现。尽管如GradCAM等技术可识别关键特征，但仅是测量工具，并非完整解释。本文提出一种面向计算病理分类器的真人-机器-视觉语言模型（VLM）交互系统，支持全切片图像的多实例学习。其概念验证包括：（1）集成AI的切片查看器，用于运行滑动窗口实验以检验解释主张；（2）利用通用视觉-语言模型量化解释的预测性。结果表明，该系统可定性验证解释主张，并定量区分竞争性解释。这为数字病理乃至更广泛领域实现了从可解释AI到可解释验证的实践路径。代码与提示语见https://github.com/nki-ai/x2x。
原文摘要 · Abstract (English)
Explaining deep learning models is essential for clinical integration of medical image analysis systems. A good explanation highlights if a model depends on spurious features that undermines generalization and harms a subset of patients or, conversely, may present novel biological insights. Although techniques like GradCAM can identify influential features, they are measurement tools that do not themselves form an explanation. We propose a human-machine-VLM interaction system tailored to explaining classifiers in computational pathology, including multi-instance learning for whole-slide images. Our proof of concept comprises (1) an AI-integrated slide viewer to run sliding-window experiments to test claims of an explanation, and (2) quantification of an explanation's predictiveness using general-purpose vision-language models. The results demonstrate that this allows us to qualitatively test claims of explanations and can quantifiably distinguish competing explanations. This offers a practical path from explainable AI to explained AI in digital pathology and beyond. Code and prompts are available at https://github.com/nki-ai/x2x.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。