Artificial neural networks for high performance, fully automated particle tracking analysis even at low signal-to-noise regimes

摘要：粒子跟踪是研究分子相互作用、输运过程的一种强有力的生物物理工具。 复杂生物环境中的现象和流变性。不幸的是，PT仍然是一个利基市场 生命科学和物理科学中的工具，用户基础有限，主要是由于大量的时间和技术 从录制的电影中提取准确的时变位置数据的限制。这些约束是 在低信噪比或显著异质性的实验中加剧， 在生物体液中遇到纳米颗粒和病原体。当前可用的软件尝试 要使电影分析过程自动化，几乎完全依赖于基于以下内容分配静态图像滤镜 特定强度、像素大小和信噪比阈值。不幸的是，当应用于实际 实验数据具有很大的时空异质性，目前的软件一般产生 大量的假阳性(即跟踪伪像)或假阴性(即遗漏实际痕迹)，以及 通常两者都有。因此，需要频繁的用户干预，以确保即使在使用 复杂的跟踪软件，显著减少了实验吞吐量，并导致大量用户- 分析数据中的用户差异。精确的粒子跟踪分析所需的时间使PT 与生活中其他常用的实验技术相比，实验成本非常高 科学。这些相同的跟踪分析限制有效地阻止了调查人员进行 更复杂的3D PT，尽管获取此类电影的显微镜功能随时可用 而关键的科学见解可以从3D PT中获得。要通过现有可用的解决方案规避挑战 粒子跟踪软件，我们开发了一种新的粒子识别和跟踪方法，基于 机器学习和卷积神经网络(CNN)。CNN是一种前馈人工神经网络 设计用于在分层连接网络中处理信息的网络，该网络模仿 哺乳动物视网膜和视觉皮质中的真实神经网络。与大多数CNN成像模型不同的是 训练我们对静态图像进行预测，我们已经训练我们的CNN输入相邻的帧，以便每个 预测包括来自过去和未来的信息，从而有效地在这两个空间中执行卷积 以及推断粒子位置的时间。类似的图像分析原理现在正被 自动驾驶汽车技术的开发人员可以区分道路上不同对象的运动。我们 已经将我们的CNN跟踪算法应用于各种2D电影，捕捉 纳米粒子、病毒和高度可移动的细菌，实现了至少30倍的时间节约，几乎不需要 用于人工干预，同时保持稳健的跟踪性能(即，低误报和低误 负利率)。在这份STTR提案中，我们寻求专注于对我们的神经进行进一步的优化和测试 用于2D PT的网络跟踪平台，包括使用云计算(目标1)，以及扩展我们的神经 网络跟踪器，可实现精确的3D PT(目标2)。我们的愿景是将PT作为一种研究工具在 研究人员通过最大限度地减少与PT分析相关的时间和人力成本。

Abstract: Particle tracking (PT) is a powerful biophysical tool for elucidating molecular interactions, transport phenomena and rheological properties in complex biological environments. Unfortunately, PT remains a niche tool in life and physical sciences with a limited user base, in large part due to significant time and technical constraints in extracting accurate time-variant positional data from recorded movies. These constraints are exacerbated in experiments with low signal-to-noise ratios or substantial heterogeneity, as frequently encountered with nanoparticles and pathogens in biological fluids. Currently available software that attempts to automate the movie analysis process rely almost exclusively on assigning static image filters based on specific intensity, pixel size and signal-to-noise ratio thresholds. Unfortunately, when applied to actual experimental data with substantial spatial and temporal heterogeneity, the current software generally produces substantial numbers of false positives (i.e. tracking artifacts) or false negatives (i.e. missing actual traces), and frequently both. Frequent user intervention is thus required to ensure accurate tracking even when using sophisticated tracking software, markedly reducing experimental throughput and resulting in substantial user- to-user variations in analyzed data. The time required for accurate particle tracking analysis makes PT experiments exceedingly expensive compared to other commonly used experimental techniques in life sciences. These same tracking analysis limitations have effectively precluded investigators from undertaking more sophisticated 3D PT, even though the microscopy capability to obtain such movies is readily available and critical scientific insights can be gained from 3D PT. To circumvent the challenges with currently available particle tracking software, we have developed a new approach for particle identification and tracking, based on machine learning and convolutional neural networks (CNN). CNN is a type of feed-forward artificial neural network designed to process information in a layered network of connections that mimics the organization of real neural networks in the mammalian retina and visual cortex. Unlike most CNN imaging models that are trained to make predictions on static images, we have trained our CNN to input adjacent frames so that each prediction includes information from the past and future, thus effectively performing convolutions in both space and time to infer particle locations. Similar principles of image analysis are now being harnessed by developers of autonomous vehicle technologies to distinguish the motions of different objects on the road. We have applied our CNN tracking algorithm to a wide range of 2D movies capturing dynamic motions of nanoparticles, viruses and highly motile bacteria, achieving at least 30-fold time savings with virtually no need for human intervention while maintaining robust tracking performance (i.e. low false positive and low false negative rates). In this STTR proposal, we seek to focus on further optimization and testing of our neural network tracking platform for 2D PT, including the use of cloud computing (Aim 1), and extending our neural network tracker to enable accurate 3D PT (Aim 2). Our vision is to popularize PT as a research tool among researchers by minimizing the time and labor costs associated with PT analysis.