BRAIN project (Cox)

继续我们的测试开发，我们已经扩展了回归测试，以涵盖更多的Afni代码库。现在，每次将提交推送到GitHub时都会自动执行测试，以确保代码(A)可编译，(B)通过程序测试序列。考虑到Afni包含100多万行源代码，其中一些可以追溯到20世纪80年代，持续测试已经成为使我们的软件在未来许多年保持活力和充满希望的重要部分。这一在集装箱化和云环境中实施的测试基础设施经历了多次迭代，得到了神经成像社区成员和Afni团队的反馈，以便成为一个健壮的解决方案，并便于团队成员使用。 我们继续开发打包Afni的方法，以使其更容易在所有平台上安装，包括容器和云环境。这项工作的一个重要部分是使用Cmake(跨平台Make)构建工具，并使用Conda包管理器指定项目依赖项。去年，通过重构代码和在所有纯计算上下文中消除对非常大的X11图形库的依赖，我们做出了重大努力来缩减Afni容器的大小。此外，Afni源代码和二进制文件中包含的几个库被完全删除，因为它们被用来支持过时的功能(例如，支持过时的Minc-1数据格式)。瘦身也使得其他管道开发人员(例如，fMRIPrep)更容易在他们的产品中使用Afni的组件。 我们直接与斯坦福大学的Fitlins团队(也由Brain Initiative提供部分资金)合作，将Afni的部分计算核心包括在他们实施的BIDS STATS模型中。虽然Fitlins为神经成像中的统计模型拟合创建了一个通用界面，但我们在幕后的工作为Afni的3dREMLfit工具创建了一座桥梁，提供了神经成像统计建模中的所有谨慎决策，并随之而来的算法效率的改进。 我们自己的功能磁共振分析流水线Afni_pro.py创建了一个质量控制HTML页面，使研究能够评估需要考虑的许多细节，并打包在单个交互网页的便捷界面中。

Continuing our testing developments, we have extended the regression testing to cover more of the AFNI codebase. Testing is now carried out automatically with every commit pushed to GitHub, to ensure that the code is (a) compilable, and (b) passes the sequence of program tests. Considering that AFNI comprises over 1 million lines of source code, some of which dates to the 1980s, continual testing has become an important part of keeping our software alive and kicking hopefully for many years to come. This test infrastructure, implemented in a containerized- and cloud-environment has undergone iterations, with feedback from members of the neuroimaging community and from the AFNI team, in order to be a robust solution, and convenient for use by team members. We continue to develop the way in which we package AFNI to make it easier to install across all platforms, including container and cloud contexts. One significant part of this effort has been to use the Cmake (cross-platform make) build tool and to specify the projects dependencies using the conda package manager. Significant efforts to trim down the size of the AFNI containers were made in the last year, by refactoring the code and by eliminating dependence on the very sizable X11 graphics libraries in all purely computational contexts. In addition, several libraries that were included in the AFNI source and binaries were completely removed, as they were used to support features that are obsolete (e.g., support for the outdated MINC-1 data format). Slimming down also makes it simpler for other pipeline developers (e.g., fMRIPrep) to use components from AFNI in their products. We worked directly with the Fitlins team at Stanford (also partly funded by the BRAIN Initiative) to include part of AFNI's computational core in their implementation of the BIDS Statsmodel. While Fitlins creates a general interface for fitting statistical models in neuroimaging, our work under the hood creates a bridge to AFNI's 3dREMLfit tool, providing all the careful decisions in neuroimaging statistical modelling and improvements to algorithmic efficiency that come along with it. Our own fMRI analysis pipeline, afni_proc.py, creates a quality control HTML page, allowing a research to assess many of the details that need to be considered, packaged in the convenient interface of a single interactive webpage.