High accuracy automated tick classification using computer vision

抽象的。在过去的二十年里，美国蜱传疾病的发病率增加了一倍多。今天， 莱姆病是美国最常见的病媒传播疾病，影响超过 50万美国人由于缺乏有效的蜱传疾病疫苗， 预防蜱虫叮咬和早期治疗蜱虫叮咬是减轻疾病的主要重点。蜱媒 监测-监测一个区域，以了解蜱虫种类组成、丰度和空间分布 分发-是向公众提供准确和最新信息的关键， 高风险地区，并在必要时进行精确的病媒控制。尽管向量的重要性 监督，目前的做法是高度资源密集型的，需要大量的劳动力和时间， 收集和鉴定病媒标本。蜱螨学家或实地分类学家的专业知识是有限的资源 蜱虫识别所需的，为国家蜱虫监测造成了重大的能力障碍 实践虽然移动的应用程序，以促进被动监测和报告人蜱 遭遇越来越受欢迎，图像质量可变，参与有限，科学家 对罕见的、入侵的或形态相似的蜱虫物种的错误识别阻碍了这一研究的可扩展性。 approach.迄今为止，还没有建立蜱虫识别能力的自动化解决方案。我们寻求推进 第一阶段的工作，成功地实现了成像和自动识别系统， 即时准确地识别12种成年蜱虫，准确率达98%。这项建议 将首先改进第一阶段光学设计的可扩展性，以适应成像的额外 种内蜱种属变异性，如蜱虫、成年雄性和未进食或饱食的成年雌性。在 同时，我们开发了优化移动的应用程序中蜱虫引导用户成像质量的方法 为大众服务。这将使代表性的图像数据库的发展， 合作伙伴包括TickSpotters、TickCheck、Walter Reed Biosystems Unit（WRBU）等。的 由此产生的数据库将用于训练、验证、测试和部署高精度计算机视觉模型， 两种蜱虫识别产品，供专业公共卫生人员和公众使用。最终 这里开发的方法将使病媒管理组织能够利用图像识别， 一个实用的系统，将提高生物监测的能力和能力，并装备一般 公众可以使用改进的工具在人类遇到蜱虫时识别蜱虫。

Abstract. The incidence of US tick-borne diseases has more than doubled in the last two decades. Today, Lyme disease is the most common vector-borne disease in the United States, impacting over half-a-million Americans each year. Due to lack of effective vaccines for tick-borne diseases, prevention of tick bites and early tick bite treatment is the primary focus of disease mitigation. Tick vector surveillance—monitoring an area to understand tick species composition, abundance, and spatial distribution—is key to providing the public with accurate and up-to-date information when they are in areas of high risk, and enabling precision vector control when necessary. Despite the importance of vector surveillance, current practices are highly resource intensive and require significant labor and time to collect and identify vector specimens. Acarologist or field taxonomist expertise is a limited resource required for tick identification, creating a significant capability barrier for national tick surveillance practice. While mobile applications to facilitate passive surveillance and reporting of human-tick encounters have grown in popularity, variable image quality, limited engagement, and scientist misidentification of rare, invasive, or morphologically similar tick species hinder the scalability of this approach. To date, no automated solutions exist to build tick identification capacity. We seek to advance Phase I work that successfully achieved an imaging and automated identification system capable of instantaneously and accurately identifying twelve adult tick species with 98% accuracy. This proposal will first improve the Phase I optical design for scalability to accommodate imaging of additional intra-specific tick species variability as nymphs, adult males, and unfed or engorged adult females. In parallel, we develop methods to optimize quality of guided user imaging of ticks in a mobile app approach for the general public. This will enable the development of a representative image database with partners including TickSpotters, TickCheck, the Walter Reed Biosystems Unit (WRBU), and others. The resulting database will be used to train, validate, test and deploy high-accuracy computer vision models in two tick identification products for professional public health and the general public. Ultimately the approaches developed here will enable vector management organizations to leverage image recognition in a practical system that will increase capacity and capability for biosurveillance, and equip the general public with improved tools to identify ticks during a human-tick encounter.