PO.BCS01.08 · 生物信息与计算
通过肿瘤数字孪生推进精准肿瘤学：一种用于苏木精-伊红图像中肺腺癌组织病理学亚型分型的通用ViT确定切缘一致模型
Advancing precision oncology through tumor digital twins: A versatile ViT-determined margin-consistent model for lung adenocarcinoma histopathologic subtyping in hematoxylin-eosin images
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：肿瘤学中的数字孪生框架需要一种稳定的、患者特异性的肿瘤结构组织学表征。然而，当前的LUAD亚型分型模型容易受到染色和扫描仪变异、域偏移以及概率校准不良的影响。我们研究一种结合注意力机制和切缘感知训练的状态空间/Transformer混合模型能否生成一个稳健、经校准的“形态孪生”，适合整合到未来的肿瘤数字孪生系统中。
方法：我们实施了一个统一的图块聚合WSI流程，在所有骨干网络中采用相同的分块、Macenko归一化、数据增强和优化。从BMIRDS-LUAD中的143张FFPE H&E WSI中，我们提取了203,226个组织图块（20×下的224×224）。使用ResNet50/101、ViT-L或状态空间/Transformer混合模型（MambaVision）对图块进行编码。通过门控注意力MIL和一个线性分类器（其logit间隙定义决策切缘）生成切片级预测。训练采用交叉熵与监督表征项相结合，无需手工设计的协调处理或测试时自适应。内部开发使用了跨五种LUAD生长模式的WSI分层划分。零样本外部评估使用WSSS4LUAD。终点指标包括准确率、亚型特异性AUC、特征-切缘一致性、内部到外部的性能下降、校准（ECE/Brier）以及10个随机种子间的运行间变异性。
结果：在BMIRDS-LUAD上，MambaVision+注意力取得了96.40±3.32%的准确率，所有亚型的ROC-AUC≥0.99，并具有强的特征-切缘一致性（训练集τ=0.88 / 验证集0.64）。错误主要局限于混合模式或低质量切片。零样本迁移到WSSS4LUAD取得了83.69±7.76%的准确率——在所有骨干网络中性能下降最小（−12.71个百分点）。校准在本站点和站外均有所改善（ECE/Brier：内部0.043/0.098；外部0.087/0.154），相对于ResNet和ViT基线取得了统计学显著的提升。种子间的变异性缩小（3.32%对比ResNet50的6.12%），表明训练稳定性增强。
结论：该框架解决了LUAD亚型分型的关键失败模式——跨站点不稳定、过度自信的边界错误和有限的可重复性。通过整合状态空间建模、Transformer注意力和切缘一致学习，它生成了一个经校准的形态孪生，在域偏移下保持亚型可辨别性，并提供更可信的WSI衍生概率。虽然它并非完整的癌症数字孪生，但它构成了一个稳健的组织学模块，可整合到下一代多模态和时序LUAD数字孪生系统中。
查看英文原文 English abstract
Background: Digital twin frameworks in oncology require a stable, patient-specific histologic representation of tumor architecture. However, current LUAD subtyping models are vulnerable to stain and scanner variability, domain shift, and poor probability calibration. We investigate whether a state-space/Transformer hybrid with attention and margin-aware training can generate a robust, calibrated “morphology twin” suitable for integration into future tumor digital-twin systems.
Methods: A uniform patch-aggregation WSI pipeline was implemented with identical tiling, Macenko normalization, augmentations, and optimization across all backbones. From 143 FFPE H&E WSIs in BMIRDS-LUAD, we extracted 203,226 tissue patches (224×224 at 20×). Patches were encoded using ResNet50/101, ViT-L, or a state-space/Transformer hybrid (MambaVision). Slide-level predictions were produced via gated-attention MIL and a linear classifier whose logit gaps define decision margins. Training employed cross-entropy combined with a supervised representation term, without handcrafted harmonization or test-time adaptation. Internal development used a WSI-stratified split across five LUAD growth patterns. Zero-shot external evaluation used WSSS4LUAD. Endpoints included accuracy, subtype-specific AUC, feature-margin concordance, internal-to-external performance drop, calibration (ECE/Brier), and run-to-run variability across 10 seeds.
Results: On BMIRDS-LUAD, MambaVision+attention achieved 96.40±3.32% accuracy with ROC-AUC ≥0.99 across all subtypes and strong feature-margin alignment (τ=0.88 train / 0.64 validation). Errors were largely confined to mixed-pattern or low-quality slides. Zero-shot transfer to WSSS4LUAD yielded 83.69±7.76% accuracy-the smallest performance drop (−12.71 points) among all backbones. Calibration improved in-site and out-of-site (ECE/Brier: 0.043/0.098 internal; 0.087/0.154 external), with statistically significant gains over ResNet and ViT baselines. Variability across seeds narrowed (3.32% vs 6.12% for ResNet50), indicating enhanced training stability.
Conclusions: This framework addresses key LUAD subtyping failure modes-cross-site instability, overconfident boundary errors, and limited reproducibility. By integrating state-space modeling, Transformer attention, and margin-consistent learning, it produces a calibrated morphology twin that preserves subtype discriminability under domain shift and provides more trustworthy WSI-derived probabilities. While not a complete cancer digital twin, it forms a robust histologic module for integration into next-generation multimodal and temporal LUAD digital-twin systems.
利益披露 Disclosure
M. Khalid, None..
R. Kazemimood, None..
B. Rodd, None.