📄 中文摘要
基础模型在组织病理学中的应用有望促进高性能和可泛化深度学习系统的发展。然而，当前模型不仅捕捉生物学相关特征，还包括前分析和扫描仪特定的变异，这些因素会对从基础模型特征训练的任务特定模型的预测产生偏倚。研究表明，在下游任务特定模型的训练中引入新颖的鲁棒性损失，可以降低对技术变异的敏感性。通过一个专门设计的综合实验设置，使用来自6155名患者的27,042个全切片图像（WSIs）训练了数千个模型，基于八个流行基础模型的特征进行计算病理学研究。
在组织病理学中实现基础模型的临床应用
出处: Enabling clinical use of foundation models in histopathology
发布: 2026年2月27日
📄 English Summary
Enabling clinical use of foundation models in histopathology
Foundation models in histopathology are anticipated to enhance the development of high-performing and generalizable deep learning systems. However, existing models capture not only biologically relevant features but also pre-analytic and scanner-specific variations that introduce bias into the predictions of task-specific models trained on foundation model features. This study demonstrates that incorporating novel robustness losses during the training of downstream task-specific models can reduce sensitivity to technical variability. A comprehensive experimental setup involving 27,042 whole slide images (WSIs) from 6,155 patients was utilized to train thousands of models based on the features of eight popular foundation models for computational pathology.