The need for measurement science in digital pathology.
The need for measurement science in digital pathology.
复制标题
DOI：
10.1016/j.jpi.2022.100157
复制
发表时间：
2022
影响因子：
--
通讯作者：
Turpin, Robert James
中科院分区：
文献类型：
--
作者：
Romanchikova, Marina;Thomas, Spencer Angus;Dexter, Alex;Shaw, Mike;Partarrieau, Ignacio;Smith, Nadia;Venton, Jenny;Adeogun, Michael;Brettle, David;Turpin, Robert James
于COVID-19疫情期间，病理服务的需求激增。病理学工作流程的数字化有助于提高吞吐量，但许多现有的数字化解决方案使用以专有数据格式捕获并由黑盒软件处理的非标准化工作流程，从而产生不同质量的数据。这项研究提出了英国领导的专家组对采用障碍和测量科学所需投入的看法，以改善数字病理学的当前实践。为了支持英国在病理学服务数字化方面的努力，这项研究包括：（1）对现有证据的审查，（2）对领域专家的在线调查，以及（3）与来自医疗保健，监管机构，制药行业，学术界，设备和软件制造商的42名代表的研讨会。讨论主题包括样本处理、数据互操作性、图像分析、设备校准和新型成像方式的使用。80%的与会者表示，数字病理学工作流程中缺乏数据互操作性阻碍了数据查找和导航。所有参与者都强调了整合成像和非成像数据进行诊断的重要性，而80%的人认为数据整合是一项优先挑战。90%的人认为人工智能和机器学习的好处，但认为需要培训和良好的性能指标。校准和提供可追溯性的方法被认为是建立协调的、可重复的样品处理和图像采集管道的关键。供应商中立的数据标准被视为为下游分析提供有意义数据的“必备条件”。用户和供应商需要良好的实践指导，以评估人工智能/机器学习工具的不确定性、适用性和再现性。所有上述需要伴随着病理学劳动力的技能提升。数字病理学需要可互操作的数据格式、可重复和可比较的实验室工作流程以及值得信赖的计算机分析软件。尽管人们对新型成像技术和人工智能工具的使用很感兴趣，但由于缺乏指导和评估工具来评估这些技术对特定临床问题的适用性，它们的采用速度减慢。测量科学在不确定性估计、标准化、参考物质和校准方面的专业知识可以帮助建立实验室程序之间的再现性和可比性，从而产生高质量的数据，并提供更高的诊断信心。对图像和注释的供应商中立开放标准的更改对于改善数据管理、共享和重用至关重要。缺乏标准化数据阻碍了病理学AI/ML工具的开发。仪器校准的频率和范围在实验室之间有很大差异。需要标准化的校准工具来产生一致的可比图像。病理学社区需要自己的指标来评估AI/ML性能。
Pathology services experienced a surge in demand during the COVID-19 pandemic. Digitalisation of pathology workflows can help to increase throughput, yet many existing digitalisation solutions use non-standardised workflows captured in proprietary data formats and processed by black-box software, yielding data of varying quality. This study presents the views of a UK-led expert group on the barriers to adoption and the required input of measurement science to improve current practices in digital pathology. With an aim to support the UK’s efforts in digitalisation of pathology services, this study comprised: (1) a review of existing evidence, (2) an online survey of domain experts, and (3) a workshop with 42 representatives from healthcare, regulatory bodies, pharmaceutical industry, academia, equipment, and software manufacturers. The discussion topics included sample processing, data interoperability, image analysis, equipment calibration, and use of novel imaging modalities. The lack of data interoperability within the digital pathology workflows hinders data lookup and navigation, according to 80% of attendees. All participants stressed the importance of integrating imaging and non-imaging data for diagnosis, while 80% saw data integration as a priority challenge. 90% identified the benefits of artificial intelligence and machine learning, but identified the need for training and sound performance metrics. Methods for calibration and providing traceability were seen as essential to establish harmonised, reproducible sample processing, and image acquisition pipelines. Vendor-neutral data standards were seen as a “must-have” for providing meaningful data for downstream analysis. Users and vendors need good practice guidance on evaluation of uncertainty, fitness-for-purpose, and reproducibility of artificial intelligence/machine learning tools. All of the above needs to be accompanied by an upskilling of the pathology workforce. Digital pathology requires interoperable data formats, reproducible and comparable laboratory workflows, and trustworthy computer analysis software. Despite high interest in the use of novel imaging techniques and artificial intelligence tools, their adoption is slowed down by the lack of guidance and evaluation tools to assess the suitability of these techniques for specific clinical question. Measurement science expertise in uncertainty estimation, standardisation, reference materials, and calibration can help establishing reproducibility and comparability between laboratory procedures, yielding high quality data and providing higher confidence in diagnosis. A change to vendor-neutral open standard for images and annotations is essential to improve data management, sharing and re-use. Lack of standardised data hinders development of AI/ML tools for pathology. Frequency and scope of instrument calibration vary a lot between laboratories. Standardised calibration tools are needed to yield consistent comparable images. Pathology community needs own metrics to assess AI/ML performance.