Development And Applications Of The Open Microscopy Environment (OME)

在过去的几年里，我们重新编写了OME分析系统，使其更加强大和高效（1），增加了一个国际化层，包括将用户界面和底层语义框架完整翻译成西班牙语，并为我们的合作做出了贡献，进一步完善了OME的用户界面及其底层基础设施。 今年该项目的大部分工作是进一步开发和验证生物成像分析中的WND-CHARM模式识别算法。 WND-CHARM是一种通用的模式识别算法，可用于分析任何类型的图像。 在此之前，我们已经使用机器视觉社区用于模式识别的标准基准套件验证了这种方法的准确性和通用性，包括人脸识别，对象识别和其他非生物分类问题。 此外，我们已经使用几种类型的显微镜在生物成像测定中验证了这种方法。 生物学内外的大多数模式识别工作都涉及分类-基本上是将未知图像分类到几个类别中的一个，计算机已被训练使用控制图像进行识别。 虽然这种定性方法是有用的，因为它是客观的，比手动评分更一致，但生物学中的大多数测定依赖于定量测量。 今年，我们重点研究了几种量化模式识别结果的技术，并根据已知的生物学关系对其进行验证。 其中一个验证数据集使用Kelling-Lawrence骨关节炎分级系统来训练分类器，以插入人类膝关节X射线中的形态变化，并研究离散KL等级之间的形态变化程度。 WND-CHARM可用性的一个重要进展来自于模式识别实验规范的开发。 该规范是人类和机器可读的，并提供了如何训练分类器的简明摘要。 该规范将作为今年正在开发的图形用户界面的中间输出。 用于基于图像的测定的模式识别算法消耗巨大的计算机资源，可能比基因组学或蛋白质组学中常用的大几个数量级。 我们已经开始了一个对等分布式计算的实验系统，在本质上类似于分布式文件共享网络。 大多数分布式计算基础设施依赖于大型集中式服务器来执行计算或充当分布式节点的主机。 安装和配置这些中央服务器需要大量的信息技术技能。 相比之下，每个生物实验室都可能包含几个人，他们的项目可以从基于图像的模式识别中受益。 项目与服务器的这种比例更类似于对等拓扑，而不是集中式计算设施。 我们计划使用来自成功的BOINC项目的大部分软件（用于FoldingHome，SETIHome等），同时开发我们自己的迷你服务器，与BOINC客户端一起运行。 这将允许任何生物学家利用本地未使用或空闲的CPU资源进行自己的成像项目，而无需配置集中式分布式计算服务器。

In previous years, we've re-written the OME analysis system to make it more robust and efficient (1), added an internationalization layer that includes a complete translation of both the user interface and the underlying semantic framework into Spanish, and have contributed to our collaborations that have further refined OME's user interface, as well as its underlying infrastructure. The bulk of the work in this project this year has been further development and validation of the WND-CHARM pattern recognition algorithm in biological imaging assays. WND-CHARM is a generalized pattern-recognition algorithm that can be used to analyze any type of image. Previously, we have validated the accuracy and generality of this approach using standard benchmark suites used for pattern recognition by the machine vision community, including face recognition, object recognition, and other non-biological classification problems. Additionally, we have validated this approach in biological imaging assays using several types of microscopy. Most work in pattern recognition both within and outside of biology involves classification - essentially sorting unknown images into one of several classes that the computer has been trained to recognize using control images. While this qualitative approach is useful because it is objective and more consistent than manual scoring, most assays in biology rely on quantitative measurements. This year, we have focused on investigating several techniques for quantifying pattern recognition results, and validating them based on known biological relationships. One of the validation datasets used the Kellgren-Lawrence osteoarthritis grading system to train a classifier to interpolate morphological change in human knee X-rays and investigate the degree of morphological change between the discrete KL grades. A significant advance in the usability of WND-CHARM came from the development of a specification for pattern-recognition experiments. This specification is both human and machine readable and presents a concise summary of how the classifier is trained. This specification will serve as the intermediate output of a graphical user interface being developed this year. Pattern recognition algorithms for image-based assays consume tremendous computer resources, potentially orders of magnitude greater than what is commonly used in genomics or proteomics. We have begun an experimental system for peer-to-peer distributed computing, similar in nature to distributed file sharing networks. Most distributed computing infrastructure relies on a large centralized server to either perform the computations or act as a host for the distributed nodes. The installation and configuration of these centralized servers demands significant information-technology skills. In contrast, each biological laboratory potentially contains several individuals with projects that could benefit from image-based pattern recognition. This ratio of projects to servers is much more similar to a peer-to-peer topology than to a centralized computing facility. We plan to use most of the software from the successful BOINC project (used for FoldingHome, SETIHome, etc.), while developing our own mini-server that will run together with the BOINC client. This will allow any biologist to harness local unused or idle CPU resources for their own imaging projects without having to configure centralized distributed computing servers.