TRACHOMA SURVEILLANCE AT SCALE: AUTOMATIC DISEASE GRADING OF EYELID PHOTOS

项目摘要 沙眼是全球感染性失明的主要原因。世界卫生组织设定了一个目标， 沙眼已经降到足够低的水平，以至于失明不再是公共卫生问题。控制 定义为上睑板结膜滤泡性沙眼性炎症（TF）的地区患病率 儿童中小于5%，目前通过临床检查确定。虽然目前不需要 定义，强烈的沙眼炎症（TI）与病原体的存在更好地相关， 沙眼衣原体。TF和TI的分级在个体之间差异很大，即使在同一个人中也是如此。 个人随着时间随着病例越来越少，培训新的评分员变得越来越困难。随着地区成为 控制，沙眼预算被削减，分级的机构知识丢失，使检测 剩余病例和潜在复发困难。达到防治沙眼目标的最大障碍之一 是一个不充分的诊断测试。世界卫生组织依赖于TF的现场分级;人的不一致性，分级者偏差， 培训费用正在成为主要障碍，但这并不是必须的。我们建议测试中央 假设一个全自动的深度学习分级器在检测方面可以表现得和训练有素的医生一样好， 和沙眼分级。该假设将在以下具体目标中进行测试：1）自动识别 滤泡和TF分级; 2）自动检测睑板血管和TI分级。我们的方法 包括基于语义的新型图像处理管道的开发、训练和测试。 使用深度学习神经网络和最先进的对象进行分割和疾病分类 侦测本研究中使用的所有数据均来自NEI资助的和其他沙眼临床研究的二级数据。 我们的研究团队进行的试验。我们的目标是促进这些新工具在全球范围内的广泛采用。 沙眼研究和分级社区，通过开源生成代码的可用性和互操作性， 通过使用开放神经网络，生成跨编程语言的机器学习模型 交换格式。我们提出的研究解决的问题，主观性，成本和可靠性的人 沙眼分级成功地完成拟议的具体目标也将是朝着以下目标迈出的关键一步： 未来的研究和发展，为卫生组织和研究团队提供一个新颖，高效， 可扩展的工具，以确保客观的，自动化的，可扩展的沙眼分级，在现场，以提高，或在一些 病例取代，传统的现场分级在沙眼控制的关键收尾阶段，以及监测， 潜在的复苏。

PROJECT SUMMARY Trachoma is the leading cause of infectious blindness worldwide. The WHO has set a goal of controlling trachoma to a low enough level that blindness from the disease is no longer a public health concern. Control is defined as a district-level prevalence of follicular trachomatous inflammation (TF) in the upper tarsal conjunctiva of less than 5% in children, currently determined by clinical examination. While not required for the current definition, intense trachomatous inflammation (TI) correlates better with presence of the causative agent, Chlamydia trachomatis. Grading of both TF and TI vary widely between individuals, and even in the same individual over time. As cases become rarer, training new graders becomes more difficult. As areas become controlled, trachoma budgets are being cut, and the institutional knowledge of grading lost, making detection of remaining cases and potential resurgence difficult. One of the greatest obstacles to reaching our trachoma goals is an inadequate diagnostic test. The WHO relies on field grading of TF; human inconsistency, grader bias, and training costs are becoming major obstacles, but they do not need to be. We propose to test the central hypothesis that a fully automatic, deep learning grader can perform as well as trained physicians in detecting and grading trachoma. The hypothesis will be tested in the following Specific aims: 1) Automatic identification of follicles and grading of TF and 2) Automatic tarsal blood vessels detection and grading of TI. Our approach includes the development, training and testing of novel image processing pipelines based on semantic segmentation and disease classification using deep learning neural networks and state-of-the-art object detection. All of the data to be used in this study is secondary data from NEI-funded and other trachoma clinical trials conducted by our study team. We aim to facilitate widespread adoption of these novel tools across the trachoma research and grading community, by open source availability of generated code and interoperability of generated machine learning models across programming languages through use of the open neural networks exchange format. Our proposed research addresses the problem of subjectivity, cost and reliability of human trachoma grading. Successful completion of the proposed specific aims will also be a key step forward towards future study and development of providing health organizations and research teams with a novel, efficient and extensible tool to ensure objective, automated, scalable trachoma grading in the field to enhance, or in some cases replace, traditional field grading during the critical endgame of trachoma control, as well surveillance for potential resurgence.