SAGE: Semantic Explainability of Attention-Based Survival Models in Computational Pathology
SAGE: Semantic Explainability of Attention-Based Survival Models in Computational Pathology
SAGE：计算病理学中基于注意力生存模型的可解释性语义框架
Abstract: Attention-based multiple instance learning (ABMIL) is the predominant approach for slide-level prediction in computational pathology, yet its attention maps provide only local explanations: they indicate where a model focuses but not which histological features drive its predictions or how the model behaves across a patient cohort.
摘要： 基于注意力的多示例学习（ABMIL）是计算病理学中进行切片级预测的主流方法。然而，其注意力图仅能提供局部解释：它们只能指出模型关注的位置，却无法说明哪些组织学特征驱动了预测结果，也无法解释模型在整个患者队列中的行为表现。
We present Semantic Attention Global Explanations (SAGE), a post-hoc framework that extracts global, language-grounded explanations from a frozen ABMIL model. Using a pathology vision-language model, SAGE scores image patches against a dictionary of 25 histological concepts, aggregates these scores according to the model’s learned attention, and quantifies how each concept relates to prediction risk across a cohort.
我们提出了“语义注意力全局解释”（SAGE），这是一个事后（post-hoc）框架，旨在从冻结的 ABMIL 模型中提取基于语言的全局解释。通过利用病理视觉-语言模型，SAGE 针对 25 种组织学概念组成的字典对图像块进行评分，根据模型学习到的注意力权重聚合这些分数，并量化每个概念与队列预测风险之间的关联。
Applied to survival prediction using seven TCGA cancer cohorts and three foundation models, SAGE recovered established prognostic features, such as the adverse association of necrosis, while revealing cancer-specific biology, including a favorable angiogenic signature in renal cell carcinoma consistent with known molecular subtypes.
在应用七个 TCGA 癌症队列和三个基础模型进行生存预测时，SAGE 不仅恢复了已知的预后特征（如坏死与不良预后的关联），还揭示了癌症特有的生物学特性，例如肾细胞癌中与已知分子亚型一致的有利血管生成特征。
Ablation studies demonstrated that these associations depend on the model’s learned attention rather than concept prevalence alone, and that the concept dictionary captures much of the prognostic information encoded by the foundation model features.
消融研究表明，这些关联依赖于模型学习到的注意力，而非仅仅取决于概念的普遍性；同时，该概念字典捕获了基础模型特征中所编码的大部分预后信息。
Through semantically-grounded explanations, SAGE provides a scalable, model-agnostic framework for understanding what ABMIL survival models learn, enabling pathologists to interpret model behavior at the cohort level and offering the potential for biomarker identification.
通过基于语义的解释，SAGE 提供了一个可扩展且与模型无关的框架，用于理解 ABMIL 生存模型学习到的内容。这使病理学家能够在队列层面解读模型行为，并为生物标志物的识别提供了潜力。