Data-driven AI promises support for pathologists to discover sparse tumor patterns in high-resolution histological images. However, three limitations prevent AI from being adopted into clinical practice: (i) a lack of comprehensiveness where most AI algorithms only rely on single criteria/examination; (ii) a lack of explainability where AI models work as 'black-boxes' with little transparency; (iii) a lack of integrability where it is unclear how AI can become part of pathologists' existing workflow. To address these limitations, we propose CrossPath: a brain tumor grading tool that supports top-down, cross data type, multi-criterion histological analysis, where pathologists can shepherd mixed AI models. CrossPath first uses AI to discover multiple histological criteria with H and E and Ki-67 slides based on WHO guidelines. Second, CrossPath demonstrates AI findings with multi-level explainable supportive evidence. Finally, CrossPath provides a top-down shepherding workflow to help pathologists derive an evidence-based, precise grading result. To validate CrossPath, we conducted a user study with pathologists in a local medical center. The result shows that CrossPath achieves a high level of comprehensiveness, explainability, and integrability while reducing about one-third time consumption compared to using a traditional optical microscope.

中文翻译：

---

CrossPath：自上而下、交叉数据类型、通过 Shepherding 混合 AI 模型进行的多标准组织学分析

数据驱动的人工智能有望支持病理学家在高分辨率组织学图像中发现稀疏的肿瘤模式。然而，三个限制阻碍了 AI 被用于临床实践：(i) 缺乏全面性，大多数 AI 算法仅依赖于单一标准/检查；(ii) 缺乏可解释性，其中人工智能模型作为“黑箱”工作，几乎没有透明度；(iii) 缺乏可集成性，目前尚不清楚 AI 如何成为病理学家现有工作流程的一部分。为了解决这些限制，我们提出了 CrossPath：一种支持自上而下、交叉数据类型、多标准组织学分析的脑肿瘤分级工具，病理学家可以在其中引导混合 AI 模型。CrossPath 首先根据 WHO 指南使用 AI 发现 H 和 E 以及 Ki-67 载玻片的多个组织学标准。第二，CrossPath 展示了具有多层次可解释支持性证据的 AI 发现。最后，CrossPath 提供自上而下的引导工作流程，以帮助病理学家得出基于证据的精确分级结果。为了验证 CrossPath，我们与当地医疗中心的病理学家进行了一项用户研究。结果表明，与使用传统光学显微镜相比，CrossPath 实现了高度的综合性、可解释性和可集成性，同时减少了约三分之一的时间消耗。