With foundation-model features, aggregator complexity does not improve cross-hospital discrimination in gastric whole-slide image classification

📄 arXiv: 2610.41328v1 📥 PDF

作者: Nheng Vanchhay, Ali Hussain, Khan Ziaullah, Moon-Il Joo, Md Ariful Islam Mozumder, Rashedul Islam Sumon, Won-Ae Lee, Min-Sun Cho, Hyeong-Chan Shin, Yujun Park, Baek Hwan Cho, Nam-Hoon Cho, Heejung An, Heung-Kook Choi, Haeyoun Kang, Hyunki Kim, Hee Cheol Kim

分类: eess.IV, cs.AI, cs.CV

发布日期: 2026-10-30


💡 一句话要点

提出CHIME-MIL以解决跨医院胃癌全切片图像分类问题

🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)

关键词: 全切片图像 跨医院分类 多实例学习 特征对齐 病理学 深度学习 医学影像

📋 核心要点

  1. 现有的全切片图像分类器通常在单一医院开发,缺乏跨医院的验证,导致模型泛化能力不足。
  2. 论文提出了一种层次化补丁-区域-切片图聚合器(CHIME-MIL),通过特征对齐和双尺度融合来提升模型的跨医院分类性能。
  3. 实验结果显示,CHIME-MIL在AUC上达到0.837,优于传统聚合器,但在特定条件下未能超越基于注意力的池化方法。

📝 摘要(中文)

全切片图像分类器通常在单一医院开发,且未在未见医院进行测试。本文收集了来自五家韩国医院的5,036张苏木精-伊红染色的胃切片,采用留一医院验证法比较多实例学习聚合器。提出的层次化补丁-区域-切片图聚合器(CHIME-MIL)在特征对齐和双尺度融合下,平均接收者操作特征曲线下面积(AUC)达到0.837,优于六种传统聚合器。尽管在相同条件下未超过基于注意力的池化,全球最大池化的表现更佳(AUC 0.855,P = 0.0002)。对齐方法在没有标签的情况下引入新医院,成本为0.002 AUC,但准确性提升有限且不一致。该模型具有较低的标签校准误差(0.053),但恶性概率不足,移除最关注的补丁显著降低恶性概率。

🔬 方法详解

问题定义:本文旨在解决全切片图像分类器在跨医院应用中的性能不足,现有方法未能有效验证模型在不同医院的泛化能力。

核心思路:提出的CHIME-MIL聚合器通过层次化的特征对齐和双尺度融合,旨在提升模型在不同医院数据上的分类准确性,从而增强模型的适应性。

技术框架:整体架构包括特征提取、特征对齐、双尺度融合和分类四个主要模块。特征提取使用冻结的病理基础模型,特征对齐通过减去每家医院的均值特征向量来实现。

关键创新:最重要的创新在于引入了无标签医院的数据对齐方法,使得模型能够在没有标签的情况下进行新医院的引入,提升了模型的实用性。

关键设计:在模型设计中,采用了层次化的聚合策略,设置了特征对齐的参数,并使用了适合的损失函数以优化分类性能。

📊 实验亮点

实验结果表明,CHIME-MIL在AUC上达到了0.837,优于六种传统聚合器,尽管在特定条件下未能超越基于注意力的池化(AUC 0.855,P = 0.0002)。该模型在标签校准方面表现良好,且对恶性概率的预测具有一定的局限性。

🎯 应用场景

该研究的潜在应用领域包括医学影像分析、病理学自动化诊断和跨医院数据共享。通过提升模型的跨医院适应能力,能够为临床实践提供更为可靠的辅助决策支持,推动个性化医疗的发展。

📄 摘要(原文)

Whole-slide image classifiers are usually developed within one hospital and often go untested at unseen hospitals. We assembled 5,036 haematoxylin-and-eosin-stained gastric slides from five Korean hospitals, scanned at 20x and 40x, and compared multiple-instance learning aggregators under leave-one-hospital-out validation on frozen pathology foundation-model features. A hierarchical patch-region-slide graph aggregator (CHIME-MIL) with site alignment and two-scale fusion reached a mean area under the receiver operating characteristic curve (AUC) of 0.837, above six conventional attention and transformer aggregators trained without these steps. Under a matched recipe it did not exceed attention-based pooling, and global max pooling given the same alignment and fusion was better (AUC 0.855, P = 0.0002). The alignment, which subtracts each hospital's mean feature vector without labels, onboarded a new hospital from ten unlabelled slides and cost 0.002 AUC without held-out-hospital data, but its accuracy gains were small and inconsistent. The model had low top-label calibration error (0.053) but under-confident malignancy probabilities, and removing its most-attended patches lowered the malignancy probability far more than removing random patches. With these features, more complex aggregators did not improve discrimination; what a deployable system adds is label-free onboarding, calibration checks and auditable evidence.