Assuring AI/ML-readiness of digital pathology in diverse existing and emerging multi-omic datasets through quality control workflows

摘要 在一个多组学的时代，组织学仍然是基础、翻译和临床研究的基本方法。 提供有关组织形态的有价值的、低成本的、非破坏性的信息。采用整体式 幻灯片成像(WSI)和数字病理学(DP)已导致大型临床和研究存储库 实例化用于基于图像的生物标记物的计算数据挖掘， 预后和治疗反应。重要是，数据质量在这些WSI的使用中起着关键作用， 尤其是在使用人工智能(AI)和机器学习(ML)方法时。人工产物和批处理 在从活组织检查到数字化的过程中的许多点上可能会出现影响，而几种检测它们的工具 已经开发，但缺乏一致的应用和报告，没有一个是在公共场合常规使用的 储存库。这留下了一个独特的机会，可以立即为现有和未来的NIH提供附加值- 支持的数据集。这项提案是Sage Bionetworks和公平数据共享专家之间的合作 和团队科学，以及WSI自动质量控制(QC)的领导者Andrew Janowczyk博士 带头开发开放源码的DP QC工具，HistoQC。我们建议增强AI/ML 通过提供透明、可重现的检测成像报告，为现有和未来的DP数据做好准备 NIH赞助的数据集中的人工产物和批处理效果通过我们的 现有的QC工作流程。实施DP数据质量透明报告将使研究人员能够排除 以一致的交叉调查者的方式从他们的训练集中提取人工制品。我们的工作将提供更大的信任 数据集重用和实验重复性，同时还简化了AI/ML模型的创建并增强了其 性能。我们将建立在强大的初步数据和原型的基础上，证明两者都有显著的改进 交叉阅读器的质量控制重复性和技术可行性，有三个具体目标。目标1看到了这种丰富 流程将应用于来自NIH支持的公共数据集的WSI，包括TCGA和GTEx，以及NIH/NCI 由多联盟协调(MC2)中心支持的癌症生物学研究计划司 家长助学金。AIM 2采用了从增强原始DP数据中吸取的经验教训，以便在#年准备好AI/ML 目标1部署可扩展的工作流，对来自MC2支持的计划的所有传入DP数据进行质量控制，提供 持续的前瞻性数据丰富，以确保AI/ML准备就绪。最后，Aim 3演示了增强的AI/ML 使用典型的自我监督组织对DP数据进行自动化质量控制过程的准备情况 分类任务。我们的交付成果包括(A)由我们的QC工作流程注释的5000个WSI，并增强为 AI/ML就绪数据集；(B)为AI就绪启用传入数据集处理的工作流；(C)故障率 识别劣质幻灯片的准确率为1%；以及(D)我们的QC比较AI/ML演示取得了改进 由于我们的数据增强，组织分类性能提高了10%。

Abstract In an era of multi-omics, histology remains an essential approach for basic, translational, and clinical research providing valuable, low-cost, and non-destructive information about tissue morphology. The adoption of whole slide imaging (WSI) and digital pathology (DP) has led to large clinical and research repositories being instantiated for computational data mining of image-based biomarkers associated with genotype, diagnosis, prognosis, and therapy response. Importantly, data quality plays a critical role in the usage of these WSI, especially when employing artificial intelligence (AI) and machine learning (ML) methods. Artifacts and batch effects may arise at many points in the process from biopsy to digitization, and while several tools to detect them have been developed, consistent application and reporting are lacking, with none being routinely applied in public repositories. This leaves a unique opportunity to immediately provide added value to existing and future NIH- supported datasets. This proposal sees a collaboration between Sage Bionetworks, experts in FAIR data sharing and Team Science, and Dr. Andrew Janowczyk, a leader in automated quality control (QC) of WSI who has spearheaded the development of an open-source DP QC tool, HistoQC. We propose to enhance the AI/ML readiness of existing and future DP data by providing transparent, reproducible, reporting of detected imaging artifacts and batch effects within NIH-sponsored datasets in an automated fashion via the extension of our existing QC workflows. Implementing transparent reporting of DP data quality will enable researchers to exclude artifacts from their training sets in a consistent cross-investigator manner. Our work will provide greater trust in dataset reuse and experimental reproducibility while also easing AI/ML model creation and enhancing their performance. We will build on strong preliminary data and prototypes, demonstrating both significantly improved cross-reader QC reproducibility and technical feasibility, with three specific aims. Aim 1 sees this enrichment process will be applied to WSI from NIH-supported public datasets, including TCGA and GTEx, and for NIH/NCI Division of Cancer Biology research programs supported by the Multi-Consortia Coordinating (MC2) Center parent grant. Aim 2 employs the lessons learned from the enhancement of raw DP data to be AI/ML ready in Aim 1 to deploy a scalable workflow for QC of all incoming DP data from MC2-supported programs, providing continual prospective data enrichment to assure AI/ML readiness. Lastly, Aim 3 demonstrates enhanced AI/ML readiness of DP data subjected to our automated QC processes using a prototypical self-supervised tissue classification task. Our deliverables include (a) 5000 WSI annotated by our QC workflow and enhanced into AI/ML ready datasets; (b) workflows to enable processing of incoming datasets for AI-readiness, (c) a failure rate of identifying poor quality slides is <1%; and (d) our QC comparative AI/ML demonstration yields an improvement of >10% performance in terms of tissue classification performance as a result of our data enhancements.