📄 中文摘要
侵袭性肺腺癌的全幻灯片图像分类在真实世界成像扰动下仍然脆弱，影响模型在决策边界的可靠性。提出了一种边际一致性框架，基于BMIRDS-LUAD数据集中143个全幻灯片图像的203,226个图像块进行评估，涵盖五种腺癌亚型。通过结合注意力加权的图像块聚合与边际感知训练，该方法在训练期间实现了0.88的Kendall相关性，在验证期间为0.64，展现了稳健的特征-对数空间对齐。对比正则化在改善类别分离方面有效，但往往导致特征过度聚类并抑制细微的形态变化，为此提出了相应的对策。
基于扰动保真度的侵袭性肺腺癌深层子类型的边际一致性研究
📄 English Summary
Margin-Consistent Deep Subtyping of Invasive Lung Adenocarcinoma via Perturbation Fidelity in Whole-Slide Image Analysis
Whole-slide image classification for invasive lung adenocarcinoma subtyping is vulnerable to real-world imaging perturbations that undermine model reliability at the decision boundary. A margin consistency framework is proposed, evaluated on 203,226 patches from 143 whole-slide images spanning five adenocarcinoma subtypes in the BMIRDS-LUAD dataset. By combining attention-weighted patch aggregation with margin-aware training, the approach achieves robust feature-logit space alignment, measured by Kendall correlations of 0.88 during training and 0.64 during validation. While contrastive regularization effectively improves class separation, it tends to over-cluster features and suppress fine-grained morphological variation; measures to counteract this effect are also introduced.