Nobrainer: A robust and validated neural network tool suite for imagers

越来越需要高效和强大的软件来处理、集成和提供跨 在BRAIN Initiative和其他项目中正在进行的人口成像工作的多样性。进展 统计学习提供了一套技术，可以解决许多研究应用，使用广泛的 以及项目产生的各种数据。这可以改变我们分析和整合新数据的方式。我们 我建议使用Nobrainer，一个开源Python库，它利用了这些新的学习技术， 该平台大大简化了将深度学习集成到神经成像研究中的过程。使用这个库，我们 构建和分发用于神经成像的用户友好型和支持云的最终用户应用程序 社区在目标1中，我们提供了神经网络模型。我们将创建健壮的、预先训练好的神经网络 使用超过65000人的大脑扫描进行大脑分割和时间序列处理。一旦 经过训练后，这些模型可以用作许多其他应用程序的基础，特别是在减少计算时间方面。 处理.我们随后将使用这些基础网络来执行图像处理、图像校正和 质量控制在目标2中，我们解决了在私有数据集上训练的能力。我们将使用贝叶斯神经网络 网络模型，支持先验信息的原则性使用。我们将利用这些网络来帮助检测 当模型预计在输入上失败时，并提供可视化，以更好地理解 模型正在工作。在目标3中，我们关注维护软件基础设施所需的工程， 提高效率，增加我们训练方法的可扩展性。在这里，我们将扩展，维护， 传播Nobrainer，我们的开源软件框架，连同培训材料和准备使用， 云友好的应用程序。我们还将创建更快的神经网络， 图像处理任务（例如，配准、分割和注释）。Nobrainer开发的工具 通过这些目标将允许用户找到和应用最相关的应用程序和开发人员扩展 该框架支持新的架构并传播新的模型和应用程序。我们预计这些 任何神经影像学研究人员都可以使用的工具， 软件包。这些工具将大大减少数据处理和新模型开发时间， 允许使用公共数据更快地探索假设，并通过更大的 信任模型输出。

There is an increasing need for efficient and robust software to process, integrate, and offer insight across the diversity of population imaging efforts underway across the BRAIN Initiative and other projects. Advances in statistical learning offer a set of technologies that can address many research applications using the extensive and varied data being produced by the projects. This can transform how we analyze and integrate new data. We propose using Nobrainer, an open source Python library that leverages these new learning technologies, as a platform that greatly simplifies integrating deep learning into neuroimaging research. Using this library, we are building and distributing user-friendly and cloud enabled end-user applications for the neuroimaging community. In Aim 1, we provide neural network models. We will create robust, pre-trained neural networks for brain segmentation and time series processing using brain scans from over 65000 individuals. Once trained, these models can then be used as the basis for many other applications, especially in reducing time of processing. We will subsequently use these base networks to perform image processing, image correction, and quality control. In Aim 2, we address the ability to train on private datasets. We will use Bayesian neural network models, which support principled use of prior information. We will use these networks to help detect when the models are expected to fail on an input, and provide visualizations to better understand how the model is working. In Aim 3, we focus on the engineering needed to maintain the software infrastructure, improve efficiency, and increase the scalability of our training methods. Here, we will extend, maintain, and disseminate Nobrainer, our open source software framework, together with training materials and ready to use, cloud-friendly, applications. We will also create much faster, neural network equivalents of time consuming image processing tasks (e.g., registration, segmentation, and annotation). The Nobrainer tools developed through these aims will allow users to find and apply the most pertinent applications and developers to extend the framework to support new architectures and disseminate new models and applications. We expect these tools to be used by any neuroimaging researcher through integration with BRAIN archives and popular software packages. These tools will significantly reduce data processing and new model development time, thus allowing faster exploration of hypotheses using public data and increase reusability of data through greater trust in model outputs.