AI training module for Vision Science

项目摘要 用于医疗保健应用的机器学习(ML)模型的开发已经成为一种高度 活跃和快速发展的研究领域，特别是在严重依赖模式的眼科领域 承认。经过训练以解释医学数据的ML模型显示出显著的改进 过去十年的表现，在很大程度上是由深度学习的到来推动的。多款车型现已 已获得FDA批准并正在临床环境中实施，使人工智能(AI)成为 美国眼科学会和整个领域的优先事项。 与传统的最大似然学习算法相比，深度学习利用了大量的训练 用于生成能够在模式识别中实现前所未有的性能的预测模型的数据集 在结构化或非结构化数据中。汇编正确标记的数据集，这些数据集代表 目标患者群体，并且足够大，足以训练深度学习模型，具有挑战性，并且仍然是 在这一领域继续取得进展的主要障碍。由于这些数据稀缺，因此最大限度地利用 通过以一种可用的格式广泛提供它们的效用。研究人员花费了大量的时间和精力 管理用于成功训练其ML模型的数据库，但随后很少出现这些数据集 以公平(可查找、可访问、可互操作和可重复使用)的方式共享。注重结构调整 这些数据以这种方式保护受试者的私人健康信息将增强 开展跨学科合作，促进该领域的进步。 在本补充资料中，我们提出了一个基于Web的、公开可用的数据科学模块，旨在 为不同背景的视觉科学研究人员提供概念性和实践性知识 产生公平的、支持ML的数据所必需的。人工智能模块将通过以下组合来实现这一目标 录制的视频讲座、阅读材料、知识评估和实践作业 即时反馈。该单元将被开发并整合到现有的博士前课程中，并 由俄勒冈健康与科学大学主办，但将向全球观众免费在线观看。 讲师将包括一个跨学科团队，具有在人工智能和 眼科，包括数据科学、医学信息学、机器学习方法、图像方面的专家 加工和公共卫生。

Project Summary The development of machine learning (ML) models for health care applications has become a highly active and rapidly evolving area of research, particularly in ophthalmology, which relies heavily upon pattern recognition. ML models trained to interpret medical data have demonstrated dramatically improved performance in the past decade, driven largely by the advent of deep learning. Multiple models have now received FDA approval and are being implemented in the clinical setting, making artificial intelligence (AI) a priority for the American Academy of Ophthalmology and the field in general. Compared to traditional ML learning algorithms, deep learning leverages massively large training datasets to generate prediction models capable of achieving unprecedented performance in pattern recognition within structured or unstructured data. Assembling correctly labeled datasets, which are representative of the target patient population and are large enough to train a deep learning model, is challenging and remains the primary barrier to continued advancement in this field. Because these data are scarce, it is crucial to maximize their utility by making them broadly available in a useable format. Researchers spend significant time and effort curating the databases used to successfully train their ML models, but rarely are these datasets subsequently shared in a manner that is FAIR (findable, accessible, interoperable, and reusable). Emphasis on structuring these data in such a manner while protecting subjects' private health information would enhance interdisciplinary collaboration and promote advancement of the field. In this supplement, we propose a web-based, publicly available data science module designed to provide vision science researchers from a variety of backgrounds with the conceptual and practical knowledge necessary to produce FAIR, ML-ready data. The AI module will accomplish this goal through a combination of recorded video lectures, reading materials, knowledge assessments, and hands-on assignments with immediate feedback. The module will be developed and integrated into an existing predoctoral curriculum and hosted by Oregon Health & Science University, but will be freely available online to a global audience. Instructors will include an interdisciplinary team with experience operating at the interface of AI and ophthalmology, including experts in data science, medical informatics, machine learning methodology, image processing, and public health.