Improving AI/ML-readiness of FaceBase Research Datasets

项目总结 FaceBase III中心的目标是由国家牙科研究所和 颅面研究(NIDCR)创建一个数据存储库，为整个社区提供服务 牙科和颅面研究人员通过共享与颅面发育相关的不同数据 以及其他可以利用多样化数据的研究社区 在FaceBase存储库中。FaceBase III的一个特别独特和重要的元素是 它有来自11,000多名人类受试者的22,000多张面部图像，其中许多是 根据临床和基因组诊断给症状贴上标签。 面部图像是研究基因和基因相关性的重要资源 表型，并在人工智能(AI)和机器领域引起了强烈的兴趣 学习(ML)研究领域，在自动表型鉴定方面取得了显著进展。而在FaceBase 遵循公平(可查找、可访问、可互操作和可重复使用)原则，有 AI/ML研究特有的问题包括：噪声的存在、标签的不确定性、 和数据集中的偏差。当务之急是，我们必须纠正在使用中的任何限制 FaceBase的面部成像数据，用于AI/ML研究。 在这个项目中，我们建议通过以下方式释放FaceBase面部扫描的巨大潜力 从角度确定数据的特征、格式化和预处理方式方面的差距 它在AI/ML研究和算法开发中的应用。为达致这个目标，我们建议 启动一个试点应用程序，应用由开发的现有深度学习算法 这项建议中的调查人员将现有的FaseBase数据(目标1)。试点的目标是 确定如何改进FaceBase数据的管理、组织和准备，以便 以简化它们在基于ML/AI的调查中的使用。 根据我们从试点中学到的东西，我们将修改当前的FaceBase自我管理 专门围绕面部扫描进行处理(目标2)。这将需要我们精简我们的 与人类主题数据的精选相关的过程，以便我们拥有必要的丰富 描述性元素，同时保持对数据处理的必要限制。 最终，目标是定位FaceBase Hub，以便现有的面部扫描资源 对AI/ML研究人员变得更加有用。更重要的是，我们预计将看到 通过面部扫描数据和其他相关数据类型(如基因分型)提高可用性 和神经功能数据。通过对我们的数据摄取程序进行拟议的改进， 我们预计，这项提议将允许FaceBase扩展到更大的数据集大小， 因此，巩固和扩大其作为一种独特资源的地位 美国国立卫生研究院ML和人工智能研究人员社区。

PROJECT SUMMARY The goal of the FaceBase III Hub was created by the National Institute for Dental and Craniofacial Research (NIDCR) to create a data repository to serve the entire community of dental and craniofacial researchers by sharing diverse data related to craniofacial development and dysmorphia, as well as other research communities that can leverage the diverse data that is in the FaceBase repository. One particularly unique and important element of FaceBase III is that it has over 22,000 facial images from over 11,000 human subjects, many of which are labeled with syndromes based on clinical and genomic diagnoses. Facial images are a critical resource for studying the correlation between genotype and phenotype and have received intense interest within the Artificial Intelligence (AI) and Machine Learning (ML) research field with notable advances in automated phenotyping. While FaceBase embraces the FAIR (Findable, Accessible, Interoperable, and Reusable) principles, there are unique concerns specific to AI/ML research including: presence of noise, uncertainty of labels, and bias within datasets. It is imperative that we remedy any limitations in the utility of FaceBase’s facial imaging data for AI/ML research. In this project, we propose to unlock the tremendous potential of FaceBase facial scans by identifying gaps in how data is characterized, formated, and preprocessed from the perspective of its use in AI/ML research and algorithm development. To accomplish this, we propose to initiate a pilot application that applies existing deep learning algorithms developed by investigators in this proposal to existing FaseBase data (Aim 1). The goal of the pilot is to identify how curation, organization and preparation of FaceBase data might be improved so as to streamline their use in ML/AI based investigations. Based on what we learn from the pilot, we will modify the current FaceBase self curation processes specifically around Facial Scans (Aim 2). This will require us to streamline our process associated with curation of human subject data, so that we have the necessary rich descriptive elements while maintaining required restrictions on data handling. Ultimately, the goal is to position the FaceBase Hub so that the existing facial scan resources become more broadly useful to AI/ML researchers. More significantly, we expect to see an increased availability with facial scan data and other associated data types, such as genotyping and neurofunctional data. By making the proposed improvements to our data ingest procedures, we anticipate that this proposal will allow FaceBase to scale to significantly larger data set sizes, and consequently, cementing and expanding its position as a unique resource to the broader NIH community of ML and AI researchers.