PO.BCS01.06 · 生物信息与计算
从H&E染色全切片图像中准确预测微卫星高度不稳定型胃癌
Accurate prediction of microsatellite instability-high gastric cancer from H&E-stained whole slide images
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：尽管胃癌（GC）患者预后较差，但微卫星高度不稳定（MSI-H）疾病患者对检查点抑制反应良好。用于MSI-H检测的下一代测序方法因成本、周转时间和可及性而受到限制。人工智能（AI）驱动的病理学有望改善MSI-H检测。
方法：使用来自TCGA的苏木精-伊红（H&E）染色GC全切片图像（WSI；N=316），并按文献[1]所述确定MSI-H状态的真实标签。构建一个模型，利用可加性多示例学习（aMIL）框架[2]以及来自病理基础模型PLUTO v3.1*[3]（PathAI，波士顿，马萨诸塞州）的嵌入，通过5折交叉验证训练以预测切片级MSI-H状态。使用受试者工作特征曲线下面积（AUROC）分析将模型预测与真实标签进行比较。
结果：模型性能结果汇总于表1。aMIL模型的平均AUROC达到0.86（范围：0.81-0.89）。这些模型预测在各折之间高度准确且一致，提示该模型在预测MSI-H状态方面高度稳健。
结论：在此，我们描述了一个能够从H&E染色WSI中一致且准确地识别MSI-H型GC的AI病理模型。将此类模型应用于常规GC活检有望简化MSI-H及其他分子生物标志物的检测。未来研究将评估与模型注意力相关的组织学特征，并将此工作扩展至其他MSI-H癌症类型，包括结直肠癌和子宫内膜癌。
参考文献：1) JCO Precis Oncol. 2017;1:PO.17.00073。2) arXiv:2206.01794。3) arXiv:2405.07905
*仅供研究使用。不用于诊断程序。
表1. aMIL模型预测胃癌MSI-H的性能。折 aMIL模型AUROC 1（N MSI-H=11；N总=64）0.87 2（N MSI-H=9；N总=63）0.88 3（N MSI-H=11；N总=63）0.89 4（N MSI-H=14；N总=63）0.85 5（N MSI-H=14；N总=63）0.81
查看英文原文 English abstract
Background: While prognosis is poor for patients with gastric cancer (GC), those with microsatellite instability-high disease (MSI-H) respond well to checkpoint inhibition. Next-generation sequencing approaches for MSI-H detection are complicated by cost, turnaround time, and accessibility. Artificial intelligence (AI)-powered pathology has the potential to improve MSI-H detection.
Methods: Hematoxylin and eosin (H&E)-stained GC whole slide images (WSIs; N=316) from TCGA were used, and ground truth MSI-H status was determined as described [1]. A model, utilizing an additive multiple instance learning (aMIL) framework [2] and embeddings from PLUTO v3.1* [3] (PathAI, Boston, MA), a pathology foundation model, was trained to predict slide-level MSI-H status using 5-fold cross-validation. Model predictions were compared to ground truth labels using area under the receiver operating curve (AUROC) analysis.
Results: Model performance results are summarized in Table 1. The aMIL model achieved a mean AUROC of 0.86 (range: 0.81-0.89). These model predictions were highly accurate and consistent across folds, suggesting that the model is highly robust for predicting MSI-H status.
Conclusions: Here, we describe an AI pathology model that consistently and accurately identifies MSI-H GC from H&E-stained WSIs. The application of such models to routine GC biopsies has the potential to streamline the detection of MSI-H and other molecular biomarkers. Future studies will assess histologic features associated with model attention and extend this work to other MSI-H cancer types, including colorectal and endometrial cancer.
References: 1) JCO Precis Oncol. 2017;1:PO.17.00073. 2) arXiv:2206.01794 3) arXiv:2405.07905
*For Research Use Only. Not for use in diagnostic procedures.
Table 1. Performance of aMIL model for prediction of MSI-H in gastric cancer. Fold aMIL Model AUROC 1 (N MSI-H =11; N total =64) 0.87 2 (N MSI-H =9; N total =63) 0.88 3 (N MSI-H =11; N total =63) 0.89 4 (N MSI-H =14; N total =63) 0.85 5 (N MSI-H =14; Nt otal =63) 0.81
利益披露 Disclosure
S. Nofallah,
PathAI Employment, Stock Option.
J. Conway,
PathAI Employment, Stock Option.
J. Brosnan-Cashman,
PathAI Employment, Stock Option.
Keros Therapeutics Independent Contractor.
S. Javed,
PathAI Employment, Stock Option.