PathCAM: connecting the digital data pipeline in diagnostic pathology with onboard-camera variable resolution slide imaging

项目摘要：该项目旨在首次实现数字病理图像生成和辅助 通过开发 PathCAM（计算机辅助显微镜）无缝融入标准临床工作流程 病理学家系统。乍一看，数字病理学 (DP) 领域似乎有望取得重大进展。 精准医学协定。然而，作为 DP 核心的数字图像的生成目前还处于外部状态。 用于绝大多数临床玻片诊断的模拟临床工作流程。这是因为， 到目前为止，病理学中的数字图像生成一直只能通过昂贵且缓慢的方式获取 幻灯片成像 (WSI) 系统，不属于标准临床工作流程。因此，当前的模拟微 范围工作流程将成为现在和可预见的未来的流行标准，这意味着任何形式的 目前，通过人工智能（AI）或外部第二意见为病理学家提供数字化援助具有重要意义。 不能脱离实践。此外，无法访问 WSI 系统的医院和患者 常规球员无法从这些助攻中受益。我们提出的低成本解决方案基于一个简单的见解： 使用通常安装在临床显微镜上的“机载摄像机”的视频源、摄像机和关联设备 成熟的软件可以建立临床审查的精确数字记录。通过我们的技术创新，这些 图像将被动创建，无需改变病理学家的行为。我们的方法将提供 病理学家可以实时预览他们在幻灯片上探索的内容，为他们提供有关“整体情况”的反馈 他们的审查和评估的背景。此外，它还允许直接集成人工智能辅助技术 临床审查期间的ogies。我们方法的一个变革性方面是，在审核过程中收集的图像 从跨分辨率尺度的病理学幻灯片中区域选择性地捕获，将被组合成单个 可变分辨率图像 (VRI)，其中包含组织切片的最佳采样，包含精确的 做出临床决策所需的放大倍率和分辨率的平衡。此最佳图像非常适合 立即传输数据以进行远程病理咨询——无需等待单独的数字化系统 扫描幻灯片。 PathCAM 对实时数字协助和远程第二意见的适用性将 在这项工作中学习。除了给临床医生带来直接好处外，PathCAM 方法生成的数据还可以 为未来的 DP 技术带来变革。 PathCAM 可以廉价地对数百万张载玻片数据进行成像 否则将会丢失的审查，并将通过广泛和直接的方式实现更公平的数据收集 通过我们与 Kitware Inc. 的合作进行开源部署。PathCAM 还捕获了丰富的信息 关于病理学家如何做出临床决策（即诊断来源）。这些记录记录了病理学家如何 临床搜索和查看跨分辨率尺度的幻灯片以及时空注意力之间的关系 和诊断可以支持新的自我监督人工智能算法的开发，这些算法可以学习，但不能 受到人机交互数据以及新的人机交互决策支持系统的限制。

Project Summary: This project aims, for the first time, to bring digital pathology image generation and assis- tance seamlessly into the standard clinical workflow by developing PathCAM, a Computer Assisted Microscope system for pathologists. The field of digital pathology (DP) at first glance seems poised to make significant im- pacts in precision medicine. However, the generation of the digital images at the heart of DP is currently outside of the analog clinical workflow used for the overwhelming majority of clinical slide diagnoses. This is because, until now, digital image generation in pathology has been relegated to acquisition in expensive and slow whole slide imaging (WSI) systems, that are outside of the standard clinical workflow. Thus, the current analog micro- scope workflow will be the prevalent standard for now and the foreseeable future, meaning that any form of digital assistance for pathologists, via artificial intelligence (AI) or external second opinions, is currently signifi- cantly disconnected from practice. Moreover, hospitals and patients that do not have access to WSI systems on a routine basis cannot benefit from these assists. Our low-cost proposed solution is based on a simple insight: using video feeds from “onboard cameras” commonly installed on clinical microscopes, the camera and associ- ated software can build an exact digital record of the clinical review. Through our technical innovations, these images will be created passively, requiring no change in pathologist behavior. Our approach will provide a pathologist with a real-time preview of what they explored on a slide, giving them feedback on the “whole picture” of their review and context of their evaluation. Moreover, it allows the direct integration of AI assistance technol- ogies during clinical review. A transformative aspect to our approach is that images collected during review that are captured regioselectively from the pathology slide across resolution scales, will be combined into a single variable resolution image (VRI), which contains an optimal sampling of the tissue section, containing the exact balance of magnification and resolution needed to render the clinical decision. This optimal image is ideal for immediate data transfer for a remote pathology consultation – with no waiting for a separate digitization system to scan the slide. PathCAM’s applicability to both real-time digital assistance and remote second opinions will be studied in this work. Beyond the direct benefits to clinicians, the data produced by the PathCAM approach can be transformative for DP technologies going forward. PathCAM can cheaply image data from millions of slide reviews that would be otherwise be lost, and would enable more equitable data collection via wide and immediate open-source deployment through our collaboration with Kitware Inc. PathCAM also captures rich information about how pathologists reach clinical decisions (i.e., diagnostic provenance). These records of how pathologists clinically search and view slides across resolution scales and relationships between spatiotemporal attention and diagnosis can support the development of new self-supervised AI algorithms that learn from, but are not limited by, human-machine interaction data, as well as new human-in-the-loop decision support systems.